[{"data":1,"prerenderedAt":5381},["ShallowReactive",2],{"methods-catalog":3},[4,25,47,62,78,92,110,127,145,162,172,186,202,230,244,258,277,295,315,327,340,358,378,391,405,424,438,453,472,488,508,522,537,555,568,581,594,609,625,640,656,671,687,701,718,737,750,764,779,794,803,817,834,850,866,881,898,914,927,944,959,973,989,1003,1016,1030,1043,1060,1075,1089,1103,1118,1133,1150,1163,1181,1194,1212,1226,1240,1258,1272,1290,1307,1323,1339,1348,1367,1383,1398,1411,1425,1438,1456,1467,1483,1499,1514,1528,1546,1563,1577,1594,1608,1618,1634,1650,1663,1680,1694,1707,1721,1734,1746,1765,1780,1793,1809,1819,1828,1840,1855,1867,1880,1899,1911,1922,1937,1949,1962,1979,1993,2009,2021,2039,2051,2065,2081,2094,2113,2127,2138,2152,2165,2179,2191,2207,2220,2237,2250,2263,2276,2289,2304,2318,2334,2348,2359,2373,2388,2403,2419,2438,2456,2473,2489,2502,2515,2528,2544,2559,2576,2590,2604,2614,2631,2647,2661,2678,2698,2713,2727,2745,2770,2789,2803,2815,2832,2845,2860,2874,2888,2903,2915,2928,2941,2955,2965,2981,2994,3007,3026,3042,3055,3069,3084,3095,3115,3131,3146,3163,3175,3187,3200,3217,3236,3257,3273,3291,3312,3326,3337,3351,3363,3379,3394,3407,3420,3436,3448,3464,3479,3497,3514,3527,3540,3555,3568,3581,3591,3603,3616,3630,3646,3658,3672,3686,3700,3715,3731,3747,3763,3780,3796,3809,3823,3837,3849,3866,3880,3894,3911,3924,3939,3952,3970,3983,3998,4014,4032,4045,4058,4071,4086,4102,4119,4134,4144,4157,4173,4189,4202,4217,4230,4243,4255,4272,4287,4301,4314,4328,4341,4358,4371,4385,4398,4414,4431,4444,4458,4477,4492,4508,4523,4538,4553,4569,4584,4598,4610,4627,4641,4653,4666,4678,4692,4706,4719,4733,4746,4761,4775,4792,4807,4820,4838,4853,4865,4879,4896,4915,4935,4952,4967,4980,4993,5007,5021,5038,5053,5069,5084,5101,5116,5131,5149,5161,5176,5186,5201,5217,5234,5247,5262,5275,5290,5305,5324,5338,5354,5368],{"id":5,"label":6,"shortName":7,"title":8,"year":9,"era":10,"cluster":11,"scope":12,"keyIdeaZh":13,"sensors":14,"mapRepresentation":16,"loopClosure":17,"estimator":18,"association":19,"deskew":20,"outputGeometry":21,"fulltextStatus":22,"lidarModels":23,"equipmentCount":24},"acharya2019bimtracker","Acharya et al., 2019","BIM-Tracker","BIM-Tracker: A model-based visual tracking approach for indoor localisation using a 3D building model",2019,"recent","C11b","localization_in_prior_map_or_bim","BIM-Tracker 以建築模型作為地圖，對影像序列做以模型為基礎的視覺追蹤，因此不需要迴圈閉合，誤差也不會累積。每一影格先依前一位姿以 Blender 光線追蹤找出 BIM 中可見的邊，把模型邊依長度取樣成三維點並反投影到影像，再沿垂直方向搜尋 Canny 邊緣建立 3D 對 2D 對應；每個取樣點保留兩側各一個假設，以 MSAC 剔除錯誤對應後用 Gauss-Newton 最小化重投影誤差估計相機位姿，最後以等速度卡爾曼濾波預測下一影格位姿。作者以 Zeb1 掃描建立走廊 BIM，用八組不同解析度、視野、遮擋與動態模糊的照片級合成序列量化誤差，並以智慧型手機真實影片示範擴增實境疊合。",[15],"monocular camera (smartphone Motorola G 1st generation for real data; virtual cameras for synthetic data)","low level-of-detail 3D model derived from an IFC BIM (walls, floor, ceiling, doors), edges subdivided every 50 cm","none needed; each frame is registered to the model, so errors do not accumulate","Model-based visual tracking: per frame, Gauss-Newton minimisation of reprojection errors between sampled 3D model edge points and image edges inside an MSAC framework (two hypotheses per sampled point, jump-out rules, iterative correspondence updates), followed by a constant-velocity Kalman filter that predicts the next pose; measurement covariance from error propagation of the least-squares solution (Sec. 3.7-3.9)","visible BIM edges rendered by Blender ray tracing (BVH), sampled into 3D points, back-projected and matched to Canny edges by searching perpendicular to the projected model edge (Sec. 3.6 states 25 to 40 px as empirically sufficient at 640 x 480 and 30 FPS, while Table 1 lists d_search = 100 px for the 640 x 480 synthetic sets); inter-frame correspondences reused when motion is small (Sec. 3.2-3.6, 3.10)","not_applicable (camera only)","6-DoF camera trajectory in the BIM coordinate system; no point cloud produced","full_text_reviewed",[],5,{"id":26,"label":27,"shortName":28,"title":29,"year":30,"era":10,"cluster":31,"scope":32,"keyIdeaZh":33,"sensors":34,"mapRepresentation":38,"loopClosure":39,"estimator":40,"association":41,"deskew":42,"outputGeometry":43,"fulltextStatus":22,"lidarModels":44,"equipmentCount":46},"affan2026semanticmeshing","Affan et al., 2026","Semantics-aided incremental meshing (LIO + RGB)","Incremental Semantics-Aided Meshing from LiDAR-Inertial Odometry and RGB Direct Label Transfer",2026,"C12","map_representation_or_reconstruction","此方法以 OneFormer 視覺基礎模型對每張 RGB 影像做全景分割，再利用 FAST-LIO2 的 IMU 狀態把標籤投影到已完成運動畸變校正（deskew）的 LiDAR 掃描點，並以遮罩侵蝕、邊界距離與深度不連續檢查剔除不可靠的投影。帶標籤的點進入改良的 TSDF：每個體素保存標籤直方圖，截斷距離依類別與量測距離調整（例如欄杆較窄、牆面較寬），最後以 Marching Cubes 產生帶語意的網格，可匯出為 USD 資產。作者刻意不用迴圈閉合，以免位姿跳動使增量式 TSDF 失效。",[35,36,37],"3D LiDAR","IMU","monocular camera","Voxblox-based TSDF with 0.1 m voxels, per-voxel label histogram with range-dependent log-normal confidence, class- and range-dependent truncation, base weighting either constant or inverse-square (1\u002Fz^2 of the point coordinate in the sensor frame; the mode used in the experiments is not stated), behind-surface weight dropoff, sparsity compensation factor 1.5 and weight cap 255, plus per-voxel semantic and geometric uncertainty scores; mesh via marching cubes","none (authors deliberately use LIO without loop closure to avoid discontinuous pose corrections that would require TSDF re-integration, Sec. 3.1)","FAST-LIO2 tightly coupled LiDAR-inertial odometry (Sec. 3.1)","LiDAR registration inherited from FAST-LIO2 (nearest-neighbour search on ikd-Tree, Sec. 3.1); image-to-LiDAR label association by ego-motion-compensated pinhole-plus-distortion projection filtered by mask erosion, boundary-distance rejection and depth-discontinuity checks (Sec. 3.2); voxel labels by majority vote plus greedy segment-to-map-label overlap assignment (Sec. 3.3)","labels are transferred onto point clouds deskewed by FAST-LIO2 (Sec. 3)","semantically labelled triangle mesh (USD assets mentioned)",[45],"__unnamed",4,{"id":48,"label":49,"shortName":50,"title":50,"year":51,"era":52,"cluster":53,"scope":54,"keyIdeaZh":55,"sensors":56,"mapRepresentation":57,"loopClosure":57,"estimator":58,"association":57,"deskew":57,"outputGeometry":57,"fulltextStatus":59,"lidarModels":60,"equipmentCount":61},"ceres_software","Agarwal et al., 2023","Ceres Solver",2023,"classic","C03","estimation_framework_or_library","Ceres Solver 是 Google 發展的開源 C++ 大規模最佳化程式庫，可求解含邊界限制的非線性最小平方問題與一般無約束最佳化。官方版本歷史說明其於 2010 年開始開發、2012 年 5 月開源釋出，並提供軟體引用格式（無 DOI）。本群集中 Cioffi 等人的連續與離散時間比較即以其 Levenberg-Marquardt 與自動微分求解。",[],"not_applicable","robustified, bounds-constrained nonlinear least squares and general unconstrained optimization: trust-region solvers (Levenberg-Marquardt, traditional and subspace Dogleg) and line-search solvers (nonlinear CG, BFGS, LBFGS); linear solvers DENSE_QR, DENSE_NORMAL_CHOLESKY, SPARSE_NORMAL_CHOLESKY, CGNR, DENSE_SCHUR, SPARSE_SCHUR and ITERATIVE_SCHUR, plus a Schur power series expansion (Power Bundle Adjustment) that is enabled through ITERATIVE_SCHUR options rather than as a separate solver type; automatic, numeric and analytic derivatives; loss functions (Huber, SoftLOne, Cauchy, Arctan, Tolerant, Tukey); Manifold parameterizations (e.g. QuaternionManifold); covariance estimation","partial_sections_reviewed",[],null,{"id":63,"label":64,"shortName":65,"title":66,"year":67,"era":52,"cluster":68,"scope":69,"keyIdeaZh":70,"sensors":71,"mapRepresentation":57,"loopClosure":72,"estimator":73,"association":74,"deskew":57,"outputGeometry":75,"fulltextStatus":22,"lidarModels":76,"equipmentCount":77},"arun1987svd","Arun et al., 1987","SVD closed-form rigid fit (Arun-Huang-Blostein)","Least-Squares Fitting of Two 3-D Point Sets",1987,"C02","registration_component","此文處理已知點對應關係時，兩組三維點之間的最小平方剛體擬合。作者先以兩組點的質心分離平移與旋轉，再對去質心點對構成的 3×3 矩陣做奇異值分解（SVD），以 VUᵀ 作為旋轉，平移由質心差求得，屬於非迭代的封閉解。若所得矩陣的行列式為負，代表得到的是鏡射：點共面時將 V 的第三欄變號即可得到旋轉，其餘情況只會在雜訊極大時出現，作者建議改用類似 RANSAC 的方法。在 VAX 11\u002F780 的模擬中，SVD 法每次執行的 CPU 時間與四元數法相近，明顯短於迭代法。",[],"none","noniterative closed form: centroids decouple translation, H = sum q_i q'_i^t (3x3), SVD H = U L V^t, R = V U^t when det = +1 (sign of third column of V changed in the coplanar case), then T = p' - R p (Sec. II, III-A, IV, VI)","not_applicable (assumes known correspondences)","rigid rotation and translation",[],1,{"id":79,"label":80,"shortName":81,"title":82,"year":9,"era":10,"cluster":11,"scope":12,"keyIdeaZh":83,"sensors":84,"mapRepresentation":86,"loopClosure":87,"estimator":88,"association":89,"deskew":20,"outputGeometry":90,"fulltextStatus":22,"lidarModels":91,"equipmentCount":46},"asadi2019imagebimslam","Asadi et al., 2019","Asadi et al. 2019 SLAM image-to-BIM registration","Real-Time Image Localization and Registration with BIM Using Perspective Alignment for Indoor Monitoring of Construction","作者提出把影片關鍵影格即時對位到設計 BIM 的方法。定位部分以 ORB-SLAM2 為基礎，改用由影片產生的自訂詞袋字典提升缺乏特徵室內場景的追蹤；第一次蒐集時以 MVE 稠密點雲和 BIM 的手動對應角點求相似轉換，建立真實尺度的全域地圖，之後各次蒐集只在該地圖中重新定位。每個關鍵影格先由 SLAM 位姿產生對應的 BIM 視圖，再以梯度下降最小化影像與 BIM 視圖之間消失點距離與消失線夾角，得到精修位姿。作者在走廊與室內施工工地兩段影片驗證，並報告在 Jetson TX1 上的每影格計算時間。",[85],"monocular camera (webcam, 1920 x 1080 at 30 fps, fixed focal length; model not reported)","sparse ORB-SLAM2 map in real-world scale; dense MVE point cloud used once for manual alignment to BIM","ORB-SLAM2 place recognition and relocalization; the perspective step is presented as reducing drift before loop closure","Augmented monocular SLAM built on ORB-SLAM2 with a custom DBoW2 vocabulary generated from the video; first session builds a global map that is scaled by a manual similarity transform to BIM, later sessions relocalize in this map with the stored scale; keyframe poses then refined by gradient-descent alignment of vanishing points and vanishing lines between the keyframe and the rendered BIM view (max 500 iterations, thresholds 1 pixel and 1 deg)","ORB features for tracking and relocalization; Canny edges and Hedau et al. vanishing point voting for keyframe perspective; BIM vanishing points computed directly from model geometry","keyframe camera poses in the BIM coordinate system and the matching BIM views; no point-cloud accuracy product",[],{"id":93,"label":94,"shortName":95,"title":96,"year":97,"era":10,"cluster":98,"scope":99,"keyIdeaZh":100,"sensors":101,"mapRepresentation":103,"loopClosure":72,"estimator":104,"association":105,"deskew":106,"outputGeometry":107,"fulltextStatus":22,"lidarModels":108,"equipmentCount":109},"fasterlio2022","Bai et al., 2022","Faster-LIO","Faster-LIO: Lightweight Tightly Coupled Lidar-Inertial Odometry Using Parallel Sparse Incremental Voxels",2022,"C05","odometry_with_local_mapping","Faster-LIO 以 FAST-LIO2 為基礎，將 ikd-Tree 換成增量式稀疏體素（iVox），以雜湊表與 LRU 快取管理體素，並以近似 k 近鄰查詢取代嚴格 k 近鄰，以換取大幅加速。作者提出線性與偽希爾伯特曲線（PHC）兩種體素內結構，並指出近似鄰點雖會帶來不精確，但在 LIO 配準中影響不大。其貢獻集中在資料結構效率，而非估計架構或全域一致性。",[102,36],"3D LiDAR (solid-state LiDAR in the AVIA dataset from FastLIO2 and 32-line spinning LiDARs; sensor models not named in the paper)","incremental sparse voxels (iVox) in a hash map with LRU cache; linear or pseudo-Hilbert-curve (PHC) in-voxel layout","iterated EKF pipeline inherited from FAST-LIO2 (paper states Faster-LIO is developed under FastLIO2 with code refactoring)","point-to-plane residuals using approximate k-NN from incremental sparse voxels","preprocessing plus undistortion inherited from FAST-LIO2 (Table I footnote)","odometry and point-cloud map; export format not_reported",[45],6,{"id":111,"label":112,"shortName":113,"title":114,"year":115,"era":52,"cluster":53,"scope":54,"keyIdeaZh":116,"sensors":117,"mapRepresentation":120,"loopClosure":121,"estimator":122,"association":123,"deskew":124,"outputGeometry":125,"fulltextStatus":22,"lidarModels":126,"equipmentCount":24},"barfoot2014gp","Barfoot et al., 2014","Exactly sparse GP trajectory (STEAM)","Batch Continuous-Time Trajectory Estimation as Exactly Sparse Gaussian Process Regression",2014,"本文把批次軌跡估計視為以時間為自變數的一維高斯過程（Gaussian process, GP）迴歸，先驗由白雜訊驅動的線性時變隨機微分方程定義（例如等速度模型）。作者證明這類先驗的逆核矩陣為精確稀疏的區塊三對角結構，因此能高效求解量測時間的狀態，並用 GP 內插查詢任意時間的狀態。量測為線性時結果等同傳統離散時間平滑；非線性時則對整條軌跡迭代，並以行動機器人資料示範同時軌跡估計與建圖（STEAM）。",[118,119],"2D laser rangefinder (range\u002Fbearing to tube landmarks)","wheel odometry","landmarks","not_reported","batch GP regression with Markovian priors from linear time-varying SDEs, solved by Gauss-Newton iterations over the whole trajectory and landmarks with a sparse Cholesky decomposition that exploits the block-tridiagonal inverse kernel (O(L^3 + L^2 M) per iteration); GP interpolation queries the state at other times in O(1) each","not_reported (range and bearing measurements to the 17 plastic-tube landmarks of the Tong et al. dataset are used; the paper does not describe how measurements are associated with landmarks)","not implemented; Sec. I motivates continuous-time priors for scanning-while-moving sensors and proposes querying an estimated camera trajectory at every laser acquisition time","trajectory queryable at any time via GP interpolation, plus landmark positions",[45],{"id":128,"label":129,"shortName":130,"title":131,"year":51,"era":10,"cluster":11,"scope":132,"keyIdeaZh":133,"sensors":134,"mapRepresentation":137,"loopClosure":138,"estimator":139,"association":140,"deskew":121,"outputGeometry":141,"fulltextStatus":22,"lidarModels":142,"equipmentCount":144},"sgraphsplus2023","Bavle et al., 2023","S-Graphs+","S-Graphs+: Real-Time Localization and Mapping Leveraging Hierarchical Representations","full_slam_with_global_correction","S-Graphs+ 把關鍵影格位姿圖與三維場景圖放進同一個即時最佳化的因子圖，分成關鍵影格、牆面、房間與樓層四層。前端在每個新關鍵影格以序列 RANSAC 擷取牆面平面，並用以 ESDF 建立的自由空間圖分群，再與牆面組合偵測四面牆與兩面牆房間；樓層中心則由目前所有牆面中距離最寬的一組估計。後端以新的房間對牆面單一代價因子，把房間中心與所屬牆面參數一起最佳化，並沿用 S-Graphs 的掃描匹配迴圈閉合。作者在模擬環境、多處施工中住宅工地的實測資料與公開 TIERS 資料集上，以 VLP-16 資料比較軌跡誤差，以及相對於建築圖所產生三維地圖的點雲 RMSE。",[135,136],"3D LiDAR (VLP-16 data used in all datasets, Sec. VI-A)","Odometry input: robot encoders on the in-house real data (the in-house platform is a legged robot, Fig. 1); LiDAR odometry (VGICP or FLOAM) on simulated and TIERS data","keyframe point clouds with a 3D scene graph of wall planes, four-wall and two-wall rooms and floor centres; local ESDF (clearing radius 10 m) and sparse free-space graph used only for room segmentation (Sec. IV-B, VI-A)","scan-matching loop closure module inherited from S-Graphs (Sec. III)","Four-layer factor graph (keyframes, wall planes, rooms, floors) jointly optimized in real time; keyframes linked by pairwise odometry, walls by pose-plane constraints, rooms by a single room-to-wall cost per room (four-wall and two-wall variants), floors by floor-to-room relative-distance factors, plus a drift node between odometry and map frames (Sec. III, V)","Wall planes extracted from each new keyframe cloud by sequential RANSAC, converted to closest-point form in the map frame and matched to mapped planes by Mahalanobis distance (threshold 0.35 m); rooms matched by L2 distance of centres with wall-id checks (threshold 1 m); loop closures by scan matching as in S-Graphs (Sec. IV, VI-A)","3D point-cloud map plus the four-layer situational graph; map evaluated by point-cloud RMSE against a 3D map generated from architectural plans (Table II)",[143],"Velodyne VLP-16",3,{"id":146,"label":147,"shortName":148,"title":149,"year":150,"era":10,"cluster":151,"scope":132,"keyIdeaZh":152,"sensors":153,"mapRepresentation":155,"loopClosure":156,"estimator":157,"association":158,"deskew":121,"outputGeometry":159,"fulltextStatus":22,"lidarModels":160,"equipmentCount":46},"suma2018","Behley & Stachniss, 2018","SuMa","Efficient Surfel-Based SLAM using 3D Laser Range Data in Urban Environments",2018,"C04","SuMa 以面元（surfel，帶法向量與半徑的小圓盤）地圖表示環境，將掃描投影成球面頂點圖與法向量圖，並從面元地圖繪製（render）同視角的模型圖，以投影式資料關聯（projective data association）執行密集的點到面 frame-to-model ICP，避免最近鄰搜尋。面元以穩定度對數勝算比過濾動態物體與雜訊；迴圈則在非活動地圖中搜尋候選，並以「合成虛擬視圖」檢驗一致性、連續多幀驗證後才加入位姿圖。由於面元綁定建立時的位姿，位姿圖最佳化後可直接更新地圖。",[154],"3D LiDAR (Velodyne HDL-64E S2 via KITTI)","surfel map (position, normal, radius, creation\u002Fupdate timestamps, stability log-odds); active\u002Finactive partition; GPU rolling-grid submaps (Sec. III-B, III-F)","single candidate within a radius searched in the inactive map; ICP with multiple initializations; accepted only if a composed virtual map view is consistent with the scan, then verified over subsequent scans (Sec. III-E)","frame-to-model point-to-plane ICP with projective data association, Gauss-Newton with Huber weights; pose graph optimized with gtsam (Levenberg-Marquardt) in a separate thread (Sec. III-C, III-E, III-F)","dense projective data association between the current vertex\u002Fnormal maps (spherical projection) and vertex\u002Fnormal maps rendered from the surfel map (Sec. III-A, III-C)","globally consistent surfel map \u002F registered point cloud (Fig. 1, Fig. 4)",[161],"Velodyne HDL-64E",{"id":163,"label":164,"shortName":165,"title":166,"year":167,"era":52,"cluster":53,"scope":54,"keyIdeaZh":168,"sensors":169,"mapRepresentation":57,"loopClosure":57,"estimator":170,"association":57,"deskew":57,"outputGeometry":57,"fulltextStatus":22,"lidarModels":171,"equipmentCount":61},"bell1993ikf","Bell & Cathey, 1993","IKF as Gauss-Newton","The iterated Kalman filter update as a Gauss-Newton method",1993,"本文證明迭代卡爾曼濾波（iterated Kalman filter, IKF）的量測更新步驟，就是以 Gauss-Newton 法近似最大概似估計；只迭代一次時即為 EKF 更新，量測函數為仿射時兩者都退化為一般卡爾曼更新。作者以二維雙站測距的解析例子說明，當觀測越來越精確時，IKF 更新與最大概似估計都會正確收斂，而 EKF 更新會收斂到有偏的值，同時其誤差共變異數卻趨近於零。這提供了把迭代濾波理解為單步最佳化的理論基礎。",[],"iterated Kalman filter measurement update shown to generate the same iterates as Gauss-Newton applied to the maximum-likelihood (weighted least-squares) update problem that stacks the predicted state and the observation; a single iterate gives the EKF update",[],{"id":173,"label":174,"shortName":175,"title":176,"year":177,"era":52,"cluster":68,"scope":69,"keyIdeaZh":178,"sensors":179,"mapRepresentation":180,"loopClosure":72,"estimator":181,"association":182,"deskew":57,"outputGeometry":183,"fulltextStatus":22,"lidarModels":184,"equipmentCount":185},"besl1992icp","Besl & McKay, 1992","ICP (point-to-point)","A method for registration of 3-D shapes",1992,"迭代最近點（Iterative Closest Point, ICP）將資料形狀分解為點集後，每次迭代先為每一點找模型形狀上的最近點，再以 Horn 的單位四元數封閉解計算最小平方剛體轉換並更新位姿，直到均方誤差的變化小於門檻。作者證明此演算法對均方距離單調收斂至局部極小值，並提出在更新方向一致時以直線或拋物線外插的加速版本，通常可把 50 次以上的迭代縮減為 15 至 20 次。模型可為點集、折線、參數或隱式曲線與曲面及三角網格；全域配準需從多組初始旋轉出發，局部配準另需多組初始平移。作者也指出，資料中若有大量點不對應模型，此方法並不適用，且易受粗大離群值影響。實驗包含合成點集、雜訊曲線與 Bezier 曲面、NRCC 面具雷射三角量測資料及 Tucson 附近的地形資料。",[],"point sets, curves or surfaces (representation-independent per abstract)","iterative: closest points, then Horn's closed-form unit-quaternion least-squares registration (preferred over SVD in 2-D and 3-D because reflections are not desired), applied to the original data set, until the mean-square error change falls below a threshold; accelerated variant extrapolates the registration vector by a line or parabola when the last update directions agree within about 10 deg, with v_max = 25 ||dq|| (Sec. III-C, IV-A, IV-C)","closest point on the model shape for each data point; point sets, polylines, triangle sets, parametric and implicit curves and surfaces (parametric entities via a simplex approximation followed by Newton iterations; implicit entities via a simplex approximation plus a constrained Lagrange-multiplier solve, although the implemented system handled implicit surfaces through special cases or parametric forms); O(Np Nx) worst case, O(Np log Nx) average; k-d trees suggested as future speed-up (Sec. III, IV-A, VIII)","6-DoF rigid transformation",[],2,{"id":187,"label":188,"shortName":189,"title":190,"year":191,"era":52,"cluster":68,"scope":69,"keyIdeaZh":192,"sensors":193,"mapRepresentation":195,"loopClosure":196,"estimator":197,"association":198,"deskew":121,"outputGeometry":199,"fulltextStatus":22,"lidarModels":200,"equipmentCount":185},"biber2003ndt","Biber & Strasser, 2003","NDT (2D)","The normal distributions transform: a new approach to laser scan matching",2003,"常態分布轉換（Normal Distributions Transform, NDT）將二維平面切成 100 cm 見方的網格，每個至少含三點的網格以點的平均與共變異數建立常態分布，並使用四組錯開半格的重疊網格降低離散化影響，使一次掃描成為分段連續且可微的機率密度。另一次掃描的點經轉換後在此密度上計分，再以牛頓法最佳化位姿，不需建立明確的點對應。作者以此進行相對於關鍵影格的位置追蹤，並以關鍵影格及其全域位姿構成地圖，利用成對匹配的 Hessian 建立二次誤差模型，只在新關鍵影格三條邊以內的子圖上最佳化。實驗以 SICK 雷射掃描儀在未改造的室內走廊、不使用里程計完成建圖，在 1.4 GHz 電腦上離線每秒約可處理 97 次掃描。",[194],"[\"SICK 2D laser scanner, 180 deg field of view, 1 deg angular resolution (Sec. VIII)\"]","per-scan 2D grid of 100 cm cells, each with a normal distribution (mean and covariance, at least three points), using four overlapping grids shifted by half a cell; the map is a collection of keyframes with global poses (Sec. III, VII)","none demonstrated; the authors note that closing a cycle would require optimizing over all keyframes (Sec. VII-B)","Newton's method on the negative NDT score (sum of Gaussian evaluations of transformed points), with analytic gradient and Hessian; Hessian replaced by H + lambda I when not positive definite (Sec. IV, V)","no explicit correspondences; points scored against per-cell normal distributions (abstract)","2D pose (tx, ty, phi) per scan; map of 33 keyframes with poses and the estimated trajectory (Fig. 2, Sec. VIII)",[201],"SICK laser scanner",{"id":203,"label":204,"shortName":205,"title":206,"year":207,"era":10,"cluster":98,"scope":99,"keyIdeaZh":208,"sensors":209,"mapRepresentation":215,"loopClosure":216,"estimator":217,"association":218,"deskew":219,"outputGeometry":220,"fulltextStatus":22,"lidarModels":221,"equipmentCount":229},"molalo2025","Blanco-Claraco, 2025","MOLA-LO","A flexible framework for accurate LiDAR odometry, map manipulation, and localization",2025,"MOLA-LO 主張以「視圖式地圖」（view-based map：帶時間戳的原始感測資料加上位姿）作為基本地圖表示，事後可依任務重新產生各種度量地圖，例如點雲、雜湊體素、佔據體素或類 NDT 地圖。建圖流程可像組合神經網路層一樣以可重用區塊設定，而不需撰寫程式。其 LiDAR 里程計在類 ICP 最佳化中緊耦合估計線速度與角速度，不需 IMU；迴圈閉合則以後處理方式進行，並可加入 GNSS 做地理參考。",[210,211,212,213,214],"3D LiDAR (16 to 128 rings)","2D LiDAR","optional wheel odometry for kinematic prediction","optional consumer-grade GNSS (loop closure and georeferencing only)","IMU not used by LO","View-based map (key-frames with pose, velocities and raw observations) as the stored map; the LO local map is a single hashed-voxel point cloud with at most 20 points per voxel and resolution 1.5% of the estimated maximum sensor range clamped to 0.5 to 1.0 m, updated only when a decider's distance criteria are met; other layers (contiguous point clouds, VDB occupancy voxels, 3D-NDT, 2D grids) can be regenerated from the view-based map","Post-processing only (not concurrent with LO): the view-based map is split into sub-maps with bounding boxes and optional GNSS georeferencing; candidates are sub-map pairs whose expected bounding-box intersection (Monte Carlo over relative poses from Dijkstra on the sub-map graph) exceeds a threshold, with no place-recognition descriptor; each candidate is verified by an ICP pipeline with separate ground and non-ground layers and a voxel-occupancy quality score","ICP-like optimizer with tightly-coupled estimation of linear and angular velocity (no IMU required); self-adaptive parameters via dynamic variables","Default lidar3d-default configuration: point-to-point pairings between the sparser twice-decimated scan layer and the local map, solved by Gauss-Newton on SE(3) with a robust kernel; matching threshold and kernel scale follow an adaptive threshold inspired by KISS-ICP but driven by a proportional feedback controller on ICP quality; the 3D-NDT configuration adds point-to-plane pairings for planar voxels","per-scan de-skewing by trajectory interpolation on SE(3) using the estimated velocities (Eq. 2); ablation on UAL campus data: ATE 7.97 m without vs 6.65 m with de-skewing (Sec. 7.1)","arbitrary metric maps regenerated from view-based maps; georeferenced maps and trajectories exportable to KML (Sec. 5.2)",[222,223,161,224,225,226,227,143,228],"Ouster OS0-128","Ouster OS1-64","Ouster OS0-64","Velodyne HDL-32E","Ouster OS0-32","2 x OS1-16 (Table 1); text names two Velodyne VLP-16","SICK LMS 2D range finder",16,{"id":231,"label":232,"shortName":233,"title":234,"year":235,"era":52,"cluster":236,"scope":99,"keyIdeaZh":237,"sensors":238,"mapRepresentation":239,"loopClosure":72,"estimator":240,"association":241,"deskew":57,"outputGeometry":242,"fulltextStatus":22,"lidarModels":243,"equipmentCount":24},"rovio2015","Bloesch et al., 2015","ROVIO","Robust visual inertial odometry using a direct EKF-based approach",2015,"C08","ROVIO 是單目視覺慣性里程計，把影像塊的像素強度誤差直接當作 EKF 更新的創新項，而非使用特徵點重投影誤差。整個濾波狀態採機器人中心（robocentric）表示，地標以方位向量加上反距離參數化，並以最小維度的流形差分表示不確定度，因此特徵在第二次觀測時即可使用，系統不需額外初始化程序。多層影像塊在影像金字塔上追蹤，強度誤差經 QR 分解降維以維持即時運算，同時線上估計 IMU 偏差與相機外參。",[37,36],"up to 50 robocentric landmarks (bearing vector plus inverse distance) with multilevel patches held in the filter state (Sec. IV-A)","EKF with a fully robocentric state (position, velocity, attitude, IMU biases, camera-IMU extrinsics, and per feature a bearing vector plus inverse-distance parameter) using minimal boxplus differences on SO(3) and S2; linearisation point of uncertain features improved by a patch search; an iterated EKF is named as an alternative (Secs. II-A, II-C)","direct: multilevel 8 x 8 pixel patches on a 4-level image pyramid tracked inside the filter; mean-subtracted intensity errors reduced by QR decomposition to a 2D innovation per feature; affine patch warping from two extra bearing vectors; FAST candidates ranked by a multi-level Shi-Tomasi score with bucketing; Mahalanobis outlier rejection (Secs. II-C, III)","IMU pose, robocentric velocity, IMU biases and camera-IMU extrinsics with covariance; sparse landmark estimates; no dense map",[],{"id":245,"label":246,"shortName":247,"title":248,"year":249,"era":10,"cluster":11,"scope":12,"keyIdeaZh":250,"sensors":251,"mapRepresentation":253,"loopClosure":72,"estimator":254,"association":255,"deskew":121,"outputGeometry":256,"fulltextStatus":22,"lidarModels":257,"equipmentCount":109},"blum2021precisebim","Blum et al., 2021","Localization in architectural 3D plans","Precise Robot Localization in Architectural 3D Plans",2021,"作者主張施工中牆體缺漏、臨時物與實作偏差使 ICP 對整棟 BIM 的對位不可靠，因此提出「局部參考」：先對整個平面圖模型做點到平面 ICP，再只對選定的參考牆面（至少三個互不平行的面）精修，並以影像密度估計網路的分數剔除或加權雜物、人員等離群點後融合到光達點。實驗在真實建築工地以靜止機器人搭配移動工人與雜物進行，並以全測站追蹤機器人上的稜鏡作為參考，且修正參考牆的竣工偏差；表 II 至 IV 的模型偏差是作者把網格上下兩側結構人為拉開 0.3 m 所模擬。",[35,252,36],"3 cameras","3D mesh generated from 2D floor plan (walls same height, planar floor) (Sec. IV-B)","per-scan registration in three steps: ICP of the scan to the full model S initialised from the previous pose, refinement by ICP to a subset R of reference surfaces (at least three mutually non-parallel surfaces; for the approximately rectangular test structures R is often the set of surfaces forming a room corner, and every test location also uses the floor as a reference surface), and rejection of the refined pose when it departs too far from the full-model result; a good initial pose (manual or from global localization) is assumed (Sec. III, III-B, Fig. 5)","point-to-plane ICP of the LiDAR scan against the building model, whose mesh is converted to a sparse point cloud; LiDAR points projected into the rectified camera images receive the per-pixel density score of the density-estimation network of Marchal et al. (trained on the NYU indoor dataset), used as a binary filter (Eq. 1) or a linear weight (Eq. 2); points outside all camera views are rejected (Sec. I, III-A)","robot pose relative to plan (no map output)",[45],{"id":259,"label":260,"shortName":261,"title":262,"year":207,"era":10,"cluster":263,"scope":132,"keyIdeaZh":264,"sensors":265,"mapRepresentation":270,"loopClosure":271,"estimator":272,"association":273,"deskew":274,"outputGeometry":275,"fulltextStatus":22,"lidarModels":276,"equipmentCount":229},"okvis2x2025","Boche et al., 2025","OKVIS2-X","OKVIS2-X: Open Keyframe-Based Visual-Inertial SLAM Configurable With Dense Depth or LiDAR, and GNSS","C07","OKVIS2-X 以關鍵影格式視覺慣性 SLAM（OKVIS2）為核心，可選擇加入深度網路估計的稠密深度、LiDAR 或 GNSS。系統把 supereight2 體素占據子地圖綁定在關鍵影格上，並以影格對地圖、地圖對地圖的占據對齊殘差把子地圖與狀態估計緊耦合，使迴圈閉合或 GNSS 修正後地圖可隨位姿一起更新，形成全域一致、可直接用於導航的稠密地圖。GNSS 以可觀測性準則處理座標對齊與訊號中斷，並支援相機外參線上校正。",[266,36,267,268,269],"one or more cameras","optional learned depth (stereo and multi-view stereo networks)","optional LiDAR","optional GNSS","supereight2 volumetric occupancy submaps anchored to keyframes, meshable into a globally consistent map (Figs. 1, 9, 10)","yes, DBoW2 place recognition with full-graph optimization (Sec. V)","OKVIS2 keyframe-based visual-inertial SLAM (BRISK features, realtime sliding-window estimator plus asynchronous full-graph optimization with posegraph edges from marginalized landmarks, DBoW2 loop closure) extended with volumetric occupancy submaps anchored to keyframes and tightly coupled through frame-to-map and map-to-map occupancy alignment factors (Tukey robustifier), optional GNSS position factors and online camera-IMU extrinsic calibration (Secs. IV-V)","BRISK keypoint matching for landmarks; dense submap alignment by occupancy residuals of depth or LiDAR points against submaps; sky segmentation (Fast-SCNN) to remove sky pixels from depth (Sec. V)","not described as a separate step","trajectory (causal, non-causal and full-BA variants) and dense occupancy submaps or meshes",[45],{"id":278,"label":279,"shortName":280,"title":281,"year":282,"era":52,"cluster":151,"scope":132,"keyIdeaZh":283,"sensors":284,"mapRepresentation":288,"loopClosure":289,"estimator":290,"association":291,"deskew":292,"outputGeometry":293,"fulltextStatus":22,"lidarModels":294,"equipmentCount":24},"borrmann2008_6dlum","Borrmann et al., 2008","6D LUM (6-DoF Lu-Milios GraphSLAM)","Globally consistent 3D mapping with scan matching",2008,"本文把 Lu 與 Milios 的二維全域一致掃描對齊（每幅掃描一個位姿、以相對位姿關係構成網路、以最大概似同時求解）推廣到三維點雲與六自由度位姿，作者稱為 LUM，是直接建立在原始掃描對應點上的 GraphSLAM。每條連結的相對位姿與其共變異數由 ICP 找到的點對經泰勒展開線性化求得，再組成線性方程組以 Cholesky 分解求解，並反覆迭代、每次重算對應點與圖的連結。系統流程是先以 ICP 逐幅配準，當兩個估計位姿距離小於門檻時加入迴圈連結，再啟動 LUM 做全域鬆弛。作者另在網站公開勘誤附錄，修正線性化矩陣在左手座標系下的形式。",[285,286,287],"robot-mounted 3D laser range finder, model not stated anywhere in the article (Hannover data set: 468 scans of 14,000 to 18,000 points, provided by Leibniz Universität Hannover)","high-resolution 3D laser scans provided by RIEGL LMS GmbH (Horn data set, 13 scans of 240,000 to 300,000 points); scanner model not stated; stationary acquisition is inferred from the stop-and-scan formulation and the target-based reference","planar 3-DoF odometry (x, z, theta_y) extrapolated to 6-DoF using the registration matrices of previously registered scans; the odometry source (for example wheel encoders) is not specified","registered 3D point clouds","purely distance-based link creation between estimated poses; no appearance check; authors state that a failed loop closure yields an incorrect map (Sec. 5)","maximum-likelihood pose-graph estimation in the style of Lu and Milios ('LUM'): each link's pose relation and covariance are obtained by linearizing (Taylor expansion) the pose-compounding error of ICP point pairs, the linear system GX = B is solved by Cholesky decomposition, and the process is iterated with correspondences recomputed (Sec. 6.1-6.5)","ICP closest-point pairs within a distance limit (Sec. 4); graph links between scans whose estimated poses are closer than a threshold (5 m, Sec. 5; 7.5 m in the Hannover experiment, Sec. 7.2)","avoided by design (stop-and-scan assumption, Sec. 6.1)","globally consistent registered 3D point cloud with 6-DoF poses and their covariances (abstract, Sec. 6)",[45],{"id":296,"label":297,"shortName":298,"title":299,"year":115,"era":52,"cluster":11,"scope":300,"keyIdeaZh":301,"sensors":302,"mapRepresentation":307,"loopClosure":308,"estimator":309,"association":310,"deskew":311,"outputGeometry":312,"fulltextStatus":22,"lidarModels":313,"equipmentCount":109},"borrmann2014thermalmapping","Borrmann et al., 2014","Irma3D automated thermal 3D mapping","A mobile robot based system for fully automated thermal 3D mapping","downstream_engineering_task","作者提出由機器人 Irma3D 全自動建立建築物熱影像三維模型的系統。平台以 Riegl VZ-400 地面雷射掃描儀為主感測器，上方裝 optris PI160 熱像儀與網路攝影機，以停走方式在各站掃描，每站再以 3DTK 的 6D SLAM 配準成同一座標系。熱像儀以燈泡陣列板做內參與相對掃描儀的外參校正，並以沿光線檢查遮擋的方式把溫度與顏色指派給點雲。站點選擇結合二維 NBV 探索與房間偵測後的三維體素 NBV 規劃，以減少天花板、地板與家具後方的遺漏；最後以行進立方體重建網格、映射溫度場並自動標出熱源。",[303,304,305,306],"terrestrial 3D laser scanner (Riegl VZ-400)","thermal camera (optris PI160)","colour webcam (Logitech QuickCam Pro 9000)","2D laser scanner (SICK LMS100) for obstacle avoidance","registered 3D point cloud with reflectance, thermal and colour values; 0.2 m voxel model for 3D NBV planning; marching-cubes mesh with mapped temperature field (Sec. 4.3, 6)","as provided by 6D SLAM in 3DTK (not described in this paper)","Scan registration with 6D SLAM from 3DTK (The 3D Toolkit) for the scanner poses; GMapping under ROS for robot localization during exploration (Sec. 3.2.5, 4.1)","scan matching of stop-and-go 3D scans in 3DTK (details in cited work); calibration board detected in scans by RANSAC plane fitting plus ICP of a plane model (Algorithm 1)","not_applicable (static scans at each position)","thermal and colour 3D point cloud of a building floor and reconstructed thermal surface model with automatically detected heat sources",[314],"SICK LMS100",{"id":316,"label":317,"shortName":318,"title":319,"year":115,"era":52,"cluster":11,"scope":300,"keyIdeaZh":320,"sensors":321,"mapRepresentation":323,"loopClosure":57,"estimator":57,"association":324,"deskew":57,"outputGeometry":325,"fulltextStatus":22,"lidarModels":326,"equipmentCount":185},"bosche2014flatness","Bosché & Guenet, 2014","TLS + BIM floor flatness control","Automating surface flatness control using terrestrial laser scanning and building information models","作者以 Scan-vs-BIM 原理把工地 TLS 點雲對齊 BIM，並把每個點分派給對應的樓板構件，再自動套用兩種標準平整度檢查法：直尺法（Straightedge，含隨機、方格與作者新提的星形方格三種直尺配置）與依 ASTM E1155 計算 F-number（FF 平整度、FL 水平度）。系統在兩片約 25 年的混凝土實驗室樓板上，與粉筆方格加 2 m 直尺與鋼尺的人工量測比較，作者結論是 TLS 的精度足以執行標準平整度檢查，而且量測更完整、更快。論文第 2 節整理了 BS EN 13670、BS 8204、ACI 117、ASTM E1155 等容許差來源，可作為品質檢查列的需求依據。",[322],"terrestrial laser scanner (FARO Focus3D)","point cloud matched to BIM objects","Scan-vs-BIM: scans registered to the BIM model (Sec. 4 names the plane-based registration of Bosche (2012) [32], while Sec. 7.3 says the approach of Bosche (2010) [31] was used to register the scans and match the points); each TLS point matched to a BIM object and mesh facet by orthogonal-projection proximity and surface-normal similarity, object recognition from the covered surface; principle credited to Bosche and Haas (2008) [1] (Sec. 4, 5.1.1, 7.3)","per-straightedge maximum deviations (Random, Grid-Square, novel Grid-Star generation) and FF\u002FFL F-Numbers with 90% confidence intervals, linked to BIM objects (Sec. 5-6, 8)",[],{"id":328,"label":329,"shortName":330,"title":331,"year":332,"era":52,"cluster":11,"scope":300,"keyIdeaZh":333,"sensors":334,"mapRepresentation":336,"loopClosure":57,"estimator":57,"association":337,"deskew":57,"outputGeometry":338,"fulltextStatus":22,"lidarModels":339,"equipmentCount":77},"bosche2010asbuiltdims","Bosché, 2010","Scan-vs-BIM object recognition and as-built dimensions","Automated recognition of 3D CAD model objects in laser scans and calculation of as-built dimensions for dimensional compliance control in construction",2010,"作者改良先前的方法，先以人工選三組以上對應點把工地雷射掃描粗對齊專案 3D CAD 模型，再以新的 ICP 精對齊整個模型，依與各構件表面相符的點數與覆蓋面積判定構件是否被辨識。接著對每個被辨識的構件個別再做 ICP，求得其竣工位姿，並與設計位姿比較，推算柱垂直度與柱間距等尺寸以檢查是否符合容許差。實驗使用加拿大多倫多一座發電廠鋼構廠房施工期間的五次掃描。作者坦承缺乏真值，且位姿偏差與掃描距離相關，結果尚不足以判斷尺寸合規檢查的精度。",[335],"terrestrial laser scanner (Trimble GX 3D per Sec. 1.1.2 and ref. [44])","point cloud with per-point CAD-object labels","closest orthogonal projection of each scan point onto model facets, accelerated with a bounding-volume hierarchy and frustum and back-face culling; ICP-based model fine registration, then per-object fine registration (Sec. 2.2-3.1); point pairs rejected when their distance exceeds tau_D = max(2 sqrt(MSE of previous iteration), 50 mm) or their normals differ by more than 45 deg; iteration stops when the MSE improvement is below 2 mm2 (Sec. 2.2.2)","recognized objects, as-built object poses, derived dimensions such as column plumb and inter-column distances (Sec. 3, Tables 5-6)",[],{"id":341,"label":342,"shortName":343,"title":344,"year":345,"era":52,"cluster":151,"scope":346,"keyIdeaZh":347,"sensors":348,"mapRepresentation":350,"loopClosure":351,"estimator":352,"association":353,"deskew":354,"outputGeometry":355,"fulltextStatus":22,"lidarModels":356,"equipmentCount":24},"bosse_zlot2009_ctscan","Bosse & Zlot, 2009","Continuous 3D scan-matching (Bosse and Zlot)","Continuous 3D scan-matching with a spinning 2D laser",2009,"odometry","本文處理移動中以旋轉 2D 雷射取得三維點雲時的運動畸變：每半圈（sweep）約需 1 秒，車輛在期間移動會使點雲局部變形。作者不停車，也不依賴里程計或 IMU，而是把 ICP 改成「掃描對掃描」的連續時間配準：先把點分入 0.5 至 8 m 的多解析度體素金字塔，以每個體素的一、二階矩求出橢球及其平面度與圓柱度，再在結合位置與形狀的 9 維描述空間中找對應；第二步不是求單一剛體轉換，而是每 0.2 秒取樣一個軌跡修正量，以匹配、平滑（加速度）與初始條件約束組成線性系統，配合 Cauchy 型穩健權重反覆求解；匹配約束在相鄰樣本間線性內插，最後以三次樣條重建連續軌跡。方法本身沒有迴圈閉合或全域最佳化，屬開迴路里程計。作者在工業園區與輕度林地以滑移轉向裝載機測試，報告局部精確的點雲與六自由度軌跡，但 MATLAB 實作約比即時慢五倍。",[349],"encoder on the spinning mount, described as accurate, giving each 2D scan's pose relative to the vehicle","multi-resolution voxel ellipsoids (moments) for matching; output is the concatenated unwarped point cloud (Sec. II-A, Fig. 2, Fig. 6)","none in the method; the authors mention appearance-based loop-closure investigations with preliminary results (Fig. 10) and place globally consistent mapping outside the paper's scope (Sec. I, III, IV)","ICP variant for sweep-to-sweep matching: small trajectory corrections sampled every 0.2 s (six samples per sweep) solved from a stacked linear system of match, smoothness (acceleration) and initial-condition constraints after first-order linearization; robust M-estimator with Lorentzian\u002FCauchy weights; outer loop (correspondences) usually converges within about five iterations, inner loop (reweighted solve) drops from about seven to one or two iterations (Sec. II-B, Eq. 6-16)","voxel (not point) correspondences: points binned into a pyramid of 3D grids (0.5 m to 8 m cells, several offsets), each voxel timestamped with the mean time of its points; per-voxel first and second moments define an ellipsoid with cylinder-likeness c (Eq. 3) and plane-likeness p (Eq. 4); nearest neighbours searched in a 9D descriptor [alpha*mu; p*v1; c*v3] (Eq. 5) with alpha set to ten times the grid resolution and eigenvector signs fixed (v1 towards the sensor, v3 towards +z); planar matches constrain centroid offset along the normal and cylindrical matches perpendicular to the axis (Eq. 10-12), scaled by inverse square-root eigenvalues; normal-angle constraints derived but omitted in the implementation (Sec. II-A, II-B)","central aim: the recovered sweep trajectory unwarps motion-distorted sweeps; points measured much later than the earliest point in a voxel are excluded to limit distortion (Sec. II-A, Fig. 4)","open-loop 6-DoF sensor trajectory and locally consistent unwarped 3D point cloud; no covariance or map-accuracy metric reported (Sec. III-IV)",[357,45],"SICK LMS291",{"id":359,"label":360,"shortName":361,"title":362,"year":363,"era":52,"cluster":151,"scope":132,"keyIdeaZh":364,"sensors":365,"mapRepresentation":368,"loopClosure":369,"estimator":370,"association":371,"deskew":372,"outputGeometry":373,"fulltextStatus":22,"lidarModels":374,"equipmentCount":377},"zebedee2012","Bosse et al., 2012","Zebedee","Zebedee: Design of a Spring-Mounted 3-D Range Sensor with Application to Mobile Mapping",2012,"Zebedee 把 Hokuyo UTM-30LX 2D 雷射掃描儀與 MicroStrain 3DM-GX2 IMU 裝在同一感測頭，再以彈簧連接手把或載具，利用手持晃動或載具振動讓掃描面不規則擺動而取得三維覆蓋。配套 SLAM 改自作者先前的旋轉 2D 雷射方法：在滑動時間視窗內，以多解析度體素中時空相近的點群建立面元，在位置與法向量構成的 6 維空間做互為最近鄰的配對，並以面元匹配誤差、IMU 角速度與加速度偏差及視窗銜接條件，求解以固定間隔取樣、其間內插的連續時間軌跡修正量，同時估計雷射與 IMU 的時間延遲及 IMU 偏差；另保留少量「固定視角」面元以抑制漂移。開迴路結果再作為初值，對整段軌跡做一次批次全域配準，得到閉迴路軌跡與點雲。摘要所稱「即時」是指處理時間短於資料擷取時間，論文實驗實際是在收集後離線處理。",[366,367],"2D time-of-flight laser Hokuyo UTM-30LX (270 deg FoV, 30 m maximum range, 40 Hz) (Sec. II)","Industrial-grade MEMS IMU MicroStrain 3DM-GX2 at 100 Hz with a rotational rate range of at least 600 deg\u002Fs; the second-generation device uses a MicroStrain 3DM-GX3 (Sec. II, IV-D)","view-based: the trajectory is the full state and raw points are projected when needed; multiresolution surfels for matching plus a small buffer of fixed views (surfels from recent finalized windows) (Sec. III, III-A, III-E)","no explicit loop detection or place recognition; loops are closed implicitly by batch global registration of surfel correspondences over the whole trajectory, which needs a good open-loop initial guess (Sec. III-F)","Sliding-window continuous-time trajectory correction: stacked 6-DoF corrections sampled at regular intervals (linear interpolation assumed for the Jacobians), linearized and solved as Ax=b by iteratively reweighted least squares with a Lorentzian M-estimator and a decreasing outlier threshold; terms are surfel match errors, IMU acceleration and rotational-rate deviations and initial-condition constraints; the state is augmented with laser-IMU latency and IMU bias corrections (Sec. III-B to III-D)","Surfels from spatially and temporally proximal point clusters in a multiresolution voxel grid (resolution doubling per level, two grids offset by half a cell), planarity-filtered; clusters whose rotational velocity normal to the scan plane is below about 15 deg\u002Fs are discarded; approximate kNN in the 6-D position and normal space via a kd-tree, reciprocal matches only, with time separation above half the nominal sweep period; correspondences recomputed every iteration (Sec. III-A)","implicit: laser points are projected with the continuous-time trajectory estimate, and laser-IMU latency is estimated online in each window (Sec. III, III-C)","6-DoF sensor-head trajectory and a 3D point cloud projected with the closed-loop trajectory (Sec. III-F, IV-B; Figs. 7, 8, 16)",[375,376],"Hokuyo UTM-30LX","SICK LMS291 (spinning)",11,{"id":379,"label":380,"shortName":381,"title":382,"year":383,"era":52,"cluster":68,"scope":69,"keyIdeaZh":384,"sensors":385,"mapRepresentation":386,"loopClosure":72,"estimator":387,"association":388,"deskew":57,"outputGeometry":389,"fulltextStatus":22,"lidarModels":390,"equipmentCount":61},"bouaziz2013sparseicp","Bouaziz et al., 2013","Sparse ICP","Sparse Iterative Closest Point",2013,"作者指出一般 ICP 依賴修剪或重新加權對應點的經驗法則來處理離群值與部分重疊，這些法則不穩定且難以調整。論文將配準目標改為對每組對應點的殘差向量施加 p 介於 0 與 1 之間的 ℓp 範數（群組稀疏），保留最近點搜尋步驟，並以交替方向乘子法（ADMM）與收縮運算子求解剛體轉換，點對點與線性化點對平面兩種版本皆可使用，實驗多採 p = 0.4。在虛擬掃描的貓頭鷹模型上，其配準 RMSE 為 4.8e-4，優於 ℓ1-ICP 的 1.6e-2 與使用距離門檻剔除的傳統 ICP。作者也指出 p 越小收斂越慢，而目標幾何含大量離群值或初始位置相距過遠時，最近點步驟仍會使結果落入錯誤的局部極小值。",[],"3D scans \u002F geometric data sets","l_p norm (p in [0, 1], p = 0.4 by default) of per-correspondence residual vectors (group sparsity) minimized with ADMM: a shrinkage step on auxiliary variables, a classical least-squares rigid fit, and a multiplier update; point-to-point and linearized point-to-plane versions (Sec. 4 to 7)","closest point on the target via an l2 kd-tree; unchanged by the l_p metric because |r|^p is monotone (Sec. 5.1)","rigid transformation",[],{"id":392,"label":393,"shortName":394,"title":395,"year":396,"era":10,"cluster":397,"scope":69,"keyIdeaZh":398,"sensors":399,"mapRepresentation":57,"loopClosure":57,"estimator":401,"association":402,"deskew":121,"outputGeometry":403,"fulltextStatus":22,"lidarModels":404,"equipmentCount":77},"brossard2020icpcov","Brossard et al., 2020","3D ICP covariance (unscented)","A New Approach to 3D ICP Covariance Estimation",2020,"C13","作者主張 ICP 結果的不確定性取決於初始值（通常來自里程計）的不確定性，因此以無跡轉換（unscented transform）額外執行 12 次 ICP 配準來傳遞初始化不確定性，並輸出含初始值與 ICP 結果相關項的聯合共變異數；感測器白雜訊與所有點共有的校正偏差（各假設約 5 cm）則以封閉式公式另計。在 Challenging data sets 八個序列、1020 組配準上，所提方法的 NNE 為平移 4.2、旋轉 34，而 Censi 公式與 65 次 Monte Carlo 為 10^2 至 10^3 量級；在軌跡一致性上平均也優於未考慮相關項的組合，但仍略為樂觀。",[400],"3D laser scans from a Hokuyo sensor (Challenging data sets for point cloud registration, Pomerleau et al. 2012; model number not given in the paper)","unscented transform over initialization uncertainty plus closed-form sensor-noise term; point-to-plane ICP","ICP configured as in Pomerleau et al. (2013): 95% random subsampling, kd-tree data association, point-to-plane error metric, 70% closest associations kept for outlier rejection","registration covariance",[45],{"id":406,"label":407,"shortName":408,"title":409,"year":207,"era":10,"cluster":98,"scope":99,"keyIdeaZh":410,"sensors":411,"mapRepresentation":415,"loopClosure":416,"estimator":417,"association":418,"deskew":419,"outputGeometry":420,"fulltextStatus":22,"lidarModels":421,"equipmentCount":423},"steamlio2025","Burnett et al., 2025","STEAM-LIO (GP continuous-time LIO)","Continuous-Time Radar-Inertial and Lidar-Inertial Odometry Using a Gaussian Process Motion Prior","本文以高斯過程（白雜訊加速度，即近似等速）作為連續時間運動先驗，在滑動視窗（約兩個 LiDAR 影格）中批次估計 SE(3) 位姿、機體速度與 IMU 偏差。因角速度屬於狀態，陀螺儀直接作為狀態量測；加速度計則只預積分成相對速度因子，其餘位置積分交給高斯過程。LiDAR 點以連續時間點到平面因子加入，每個點的位姿由相鄰兩個估計時刻的後驗內插取得，因此去畸變與配準在同一最佳化中反覆更新；由於先驗為稀疏的馬可夫形式，預積分與內插的計算量隨估計時刻數線性增加。同一框架也用於二維旋轉雷達，形成雷達慣性里程計。",[412,413,414],"3D spinning LiDAR (Velodyne Alpha-Prime 128-beam on Boreas; 64-beam Ouster in Newer College; 64-beam Velodyne in KITTI-raw, LiDAR-only)","IMU (Applanix raw IMU at 200 Hz on Boreas; Ouster internal IMU at 100 Hz in Newer College)","2D spinning radar (Navtech CIR304-H) for the radar-inertial variant","sliding local voxel point map centred on the robot; on Boreas voxels unobserved for about one second are cleared (Sec. IV, V)","none explicit; on Newer College the incrementally built map allows implicit loop closure when areas are revisited (Sec. V)","sliding-window batch continuous-time estimation (window of two LiDAR frames, about 200 ms) with a white-noise-on-acceleration Gaussian-process prior on SE(3) pose and body-centric velocity, direct gyroscope factors, preintegrated accelerometer relative-velocity factors and bias random-walk priors; Gauss-Newton with outer re-association loops and marginalization (Sec. III, IV, IV-C)","continuous-time point-to-plane factors from a coarsely voxelized scan (default 1.5 m) to a sliding local voxel map (1.0 m voxels, up to 20 points, minimum spacing 0.1 m), weighted by a planarity heuristic; radar uses Doppler-compensated point-to-point factors with a Cauchy loss (Sec. IV, IV-A, IV-D)","each outer iteration undistorts the scan with the posterior continuous-time trajectory of the previous iteration (Sec. IV-A, Alg. 1)","continuous-time trajectory (pose and body-centric velocity with covariance) and lidar or radar point maps (Figs. 1, 9, 18)",[422,45],"Velodyne Alpha-Prime 128-beam",12,{"id":425,"label":426,"shortName":427,"title":428,"year":249,"era":10,"cluster":31,"scope":32,"keyIdeaZh":429,"sensors":430,"mapRepresentation":432,"loopClosure":57,"estimator":433,"association":434,"deskew":57,"outputGeometry":435,"fulltextStatus":22,"lidarModels":436,"equipmentCount":144},"cai2021ikdtree","Cai et al., 2021","ikd-Tree","ikd-Tree: An Incremental K-D Tree for Robotic Applications","ikd-Tree 讓 k-d 樹只以新進點增量更新，支援單點與方盒範圍的插入、重新插入與刪除（刪除採延遲標記），並在樹上同步降採樣：以邊長 L 的立方格劃分空間，每格只保留最接近格心的點。樹以類似 scapegoat 樹的準則監測平衡並局部重建，大型子樹可交由第二執行緒重建，以維持即時性。",[35,431],"IMU (in the FAST-LIO application test)","incremental k-d tree of map points with lazy-delete labels and on-tree voxel downsampling","not_applicable (data structure; tested inside FAST-LIO)","exact (not approximate) k-nearest-neighbour search that uses per-node range bounds and lazy-label pushdown (Sec. III-E)","downsampled point set stored in the tree",[437],"Livox Avia",{"id":439,"label":440,"shortName":441,"title":442,"year":249,"era":10,"cluster":236,"scope":132,"keyIdeaZh":443,"sensors":444,"mapRepresentation":447,"loopClosure":448,"estimator":449,"association":450,"deskew":57,"outputGeometry":451,"fulltextStatus":22,"lidarModels":452,"equipmentCount":109},"orbslam3_2021","Campos et al., 2021","ORB-SLAM3","ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial, and Multimap SLAM","ORB-SLAM3 在 ORB-SLAM2 基礎上加入緊耦合的視覺慣性（visual-inertial）最大後驗估計，包括 IMU 初始化階段，並支援針孔與魚眼相機。其 Atlas 多地圖機制在追蹤失敗時另起新地圖，重訪時再以改良召回率的場所辨識（place recognition）將地圖合併，使 BA 可使用時間上相隔很遠甚至跨作業階段的共視關鍵影格。系統輸出仍是稀疏地圖點與關鍵影格軌跡。",[37,445,446,36],"stereo (pin-hole or fisheye; rectification not required)","RGB-D (supported by the library; no RGB-D experiment reported)","Atlas of disconnected sparse maps (map points plus keyframes), one active (Sec. III)","For each new keyframe, DBoW2 returns the three most similar Atlas keyframes not covisible with it; for each candidate a local window (candidate plus best covisible keyframes) is aligned by RANSAC with Horn's method on 3D-3D matches (Sim(3) for monocular or immature monocular-inertial maps, SE(3) otherwise), refined by guided matching and bidirectional reprojection optimisation, and verified in three covisible keyframes already in the map instead of three consecutive BoW detections; mature visual-inertial maps also require pitch and roll below a threshold. A match in the active map triggers loop correction, a match in another map triggers map merging","Keyframe-based MAP estimation: visual or visual-inertial BA with IMU preintegration on manifold and Huber-robust reprojection terms; tracking optimises only the states of the last two frames with map points fixed; local mapping optimises a sliding window of keyframes and their points with covisible keyframes fixed. IMU initialisation in three MAP steps: 2 s of monocular visual-only BA (10 keyframes at 4 Hz), inertial-only MAP for scale, gravity direction, biases and velocities with a bias prior, then joint visual-inertial MAP; visual-inertial BA again 5 and 15 s after initialisation (scale error about 5% after 2 s and 1% after 15 s)","ORB features and reprojection; DBoW2 keyframe database with a new place-recognition method of improved recall (Sec. III; abstract)","Keyframe trajectories and sparse map points (on EuRoC V202 the four ORB-SLAM3 configurations keep 9,686 to 14,245 map points and 135 to 332 keyframes, Table VI); no dense reconstruction is produced or evaluated in the paper",[],{"id":454,"label":455,"shortName":456,"title":457,"year":396,"era":10,"cluster":263,"scope":346,"keyIdeaZh":458,"sensors":459,"mapRepresentation":464,"loopClosure":465,"estimator":466,"association":467,"deskew":468,"outputGeometry":469,"fulltextStatus":22,"lidarModels":470,"equipmentCount":471},"pronto2020","Camurri et al., 2020","Pronto","Pronto: A Multi-Sensor State Estimator for Legged Robots in Real-World Scenarios","Pronto 是為腿式機器人設計的模組化擴展卡爾曼濾波器：以 IMU 作為高頻過程模型，先融合腿部運動學與接觸偵測得到的速度，再把延遲且低頻的視覺里程計與 LiDAR 點雲配準結果，以鬆耦合的位姿修正方式插入約 10 秒的量測歷史中重新傳播。這樣可在控制迴路中提供 250 至 1000 Hz 的低延遲狀態估測，同時利用外感測器抑制長時間漂移。",[460,461,462,463],"IMU (KVH 1750, KVH 1775, 3DM-GX4-25 or Xsens MTi-100 depending on robot)","joint encoders and force or torque sensing for leg odometry","stereo camera (Carnegie Robotics Multisense SL) or RGB-D camera (Intel RealSense D435) for FOVIS visual odometry","LiDAR (Hokuyo UTM-30LX-EW spinning in the Multisense SL, or Velodyne VLP-16 on ANYmal) for AICP registration","AICP reference point cloud updated after a travelled distance; no global map is maintained by the estimator","no","modular extended Kalman filter with an IMU process model on the real-time control computer; leg-odometry velocity updates; zero-velocity gyro bias update when stationary at least 400 ms; loosely coupled pose or position corrections from FOVIS visual odometry and from AICP LiDAR registration applied through a measurement history of typically 10 s to handle their latency (Secs. 4-5)","leg odometry from kinematics with contact detection (Schmitt trigger on humanoids, probabilistic ground-reaction-force contact on quadrupeds); FOVIS feature-based stereo or RGB-D visual odometry; AICP auto-tuned ICP with planar pre-filtering and overlap-based outlier rejection against a reference cloud (Sec. 4)","not described in the paper; AICP registers accumulated point clouds","base pose and velocity at control rate for closed-loop locomotion",[375,143],14,{"id":473,"label":474,"shortName":475,"title":476,"year":97,"era":10,"cluster":236,"scope":99,"keyIdeaZh":477,"sensors":478,"mapRepresentation":481,"loopClosure":482,"estimator":483,"association":484,"deskew":57,"outputGeometry":485,"fulltextStatus":22,"lidarModels":486,"equipmentCount":487},"gvins2022","Cao et al., 2022","GVINS","GVINS: Tightly Coupled GNSS-Visual-Inertial Fusion for Smooth and Consistent State Estimation","GVINS 在 VINS-Mono 的滑動視窗非線性最佳化中，直接加入 GNSS 原始量測（碼偽距與都卜勒頻移）以及接收器時鐘偏差與漂移因子，與影像及 IMU 緊耦合，提供無漂移的全域六自由度位姿。系統先以單點定位得到粗略錨點，再用都卜勒量測校正區域座標與 ENU 座標間的偏航角，最後以偽距精修錨點，完成線上初始化。對低速、衛星少於 4 顆與完全無 GNSS 的退化情況分別處理，可在室內外轉換時衛星遺失與重新鎖定之間連續運作。",[479,36,480],"monocular camera (left camera of a VI-Sensor)","GNSS receiver raw code pseudorange and Doppler (GPS, GLONASS, Galileo, BeiDou)","sparse features with inverse depth in the sliding window","none (VINS-Mono and VINS-Fusion loop closure also disabled in the comparisons)","tightly coupled sliding-window non-linear optimisation (window size 10) on a factor graph inherited from VINS-Mono: IMU preintegration, visual reprojection, GNSS code pseudorange and Doppler factors, receiver clock bias and drift factors; states include the yaw offset between local world and ENU frames and per-constellation clock biases; two-way marginalisation; robust norm on GNSS factors in urban driving (Secs. V, VI, VIII-B5)","strong corners tracked with iterative Lucas-Kanade optical flow (VINS-Mono front end); GNSS satellites filtered by elevation, health and continuous lock (Secs. V, VI-C)","global 6-DoF pose in ECEF and a local ENU frame, including global yaw; no dense map",[],7,{"id":489,"label":490,"shortName":491,"title":492,"year":207,"era":10,"cluster":98,"scope":99,"keyIdeaZh":493,"sensors":494,"mapRepresentation":497,"loopClosure":498,"estimator":499,"association":500,"deskew":501,"outputGeometry":502,"fulltextStatus":22,"lidarModels":503,"equipmentCount":377},"resple2025","Cao et al., 2025","RESPLE","RESPLE: Recursive Spline Estimation for LiDAR-Based Odometry","RESPLE 把三次 B 樣條（B-spline）直接嵌入狀態空間模型，以遞迴式（濾波）方式估計六自由度連續時間運動，而非以滑動視窗最佳化擬合樣條。狀態向量由位置控制點與姿態控制點增量組成，以修改後的迭代擴展卡爾曼濾波更新，每個 LiDAR 點以其時間戳在樣條上求位姿後計算點到平面殘差。同一骨幹可組成 LiDAR-only、LiDAR-慣性、多 LiDAR 與多 LiDAR-慣性里程計。",[495,496],"one or multiple 3D LiDARs (Ouster OS1-16, Livox Mid70, Mid360, Avia, Hesai XT32)","IMU optional (VN100 or LiDAR built-in IMU)","point map in ikd-Tree","none (backend for global correction listed as future work, Sec. VI)","recursive Bayesian estimator: modified iterated EKF over cubic B-spline control points (position control points and orientation increments), without error-state formulation","point-to-plane residual per point, plane fitted from N=5 neighbours in ikd-Tree","not required: each point evaluated at its own timestamp on the spline","continuous-time trajectory and point map; export format not_reported",[504,505,506,507,437],"Ouster OS1-16","Livox MID70","Hesai XT32 (L1)","Livox MID-360",{"id":509,"label":510,"shortName":511,"title":512,"year":513,"era":52,"cluster":68,"scope":514,"keyIdeaZh":515,"sensors":516,"mapRepresentation":57,"loopClosure":72,"estimator":518,"association":519,"deskew":57,"outputGeometry":520,"fulltextStatus":22,"lidarModels":521,"equipmentCount":77},"censi2007covariance","Censi, 2007","ICP covariance (Censi)","An accurate closed-form estimate of ICP's covariance",2007,"evaluation_method_or_metric","作者以 ICP 最小化的誤差函數為對象，利用隱函數定理推導估計值對量測的一階敏感度，得到封閉形式共變異數，並考慮同一量測被多個對應重複使用與量測彼此相關的情形。論文只處理二維平面（x、y、θ）的定位與掃描匹配，分析對象為點對線段 ICP；以 52 條射線、雜訊標準差 0.03 m 的模擬測距儀，在正方形、走廊與圓形環境各做 300 次蒙地卡羅模擬；在走廊與圓形等約束不足情形，改以 Fisher 資訊矩陣找出不可觀測方向，只比較可觀測子空間上的誤差。",[517],"simulated 2D range finder: 52 rays over 360 deg, zero-mean Gaussian range noise with 0.03 m standard deviation","first-order covariance of the minimizer via the implicit function theorem: cov(x_hat) = (d2J\u002Fdx2)^-1 (d2J\u002Fdz dx) cov(z) (d2J\u002Fdz dx)^T (d2J\u002Fdx2)^-1 evaluated at the estimate; uses only the error function J, not the ICP algorithm internals; requires matrix products and a 3x3 inversion","analysed ICP variant is 2D point-to-segment ('vanilla' ICP): each correspondence uses one point of the current scan and the two reference-scan points that form the closest polyline segment","covariance of the ICP pose estimate",[],{"id":523,"label":524,"shortName":525,"title":526,"year":282,"era":52,"cluster":527,"scope":69,"keyIdeaZh":528,"sensors":529,"mapRepresentation":531,"loopClosure":57,"estimator":532,"association":533,"deskew":57,"outputGeometry":534,"fulltextStatus":22,"lidarModels":535,"equipmentCount":144},"censi2008_plicp","Censi, 2008","PL-ICP and CSM (laser_scan_matcher)","An ICP variant using a point-to-line metric","C01","PL-ICP 是採用點到線（point-to-line）度量的 2D ICP 變體：參考掃描以相鄰點連成折線，目前掃描的每個點對應到最近兩點形成的線段，並以作者推導的精確閉式解最小化點到線距離。作者引用 Pottmann 等的結果說明，點到線度量在零殘差且初值良好時具二次收斂；並證明當參考曲面為折線時，演算法會在有限步內收斂到固定點或循環。附錄另提出利用射線角度排序、提前停止與跳躍表的快速對應搜尋。作者的 CSM 函式庫就是 ROS laser_scan_matcher 增量式 2D 雷射里程計所用的匹配核心。",[530],"2D laser range finder (SICK, 360 rays over 180 deg in the test log)","not_applicable (reference scan as polyline)","iterative point-to-segment matching; each iteration minimizes the point-to-line error in exact closed form via Lagrange multipliers (a fourth-order polynomial in lambda); trimming rejects outliers (Sec. II; App. I)","each transformed point is matched to the segment between its two closest points in the reference scan, which is treated as a polyline; fast search exploits radial ordering, early stopping and precomputed jump tables (Sec. II; App. II)","2D rigid transform (t, theta) between two scans",[536],"Sick range-sensor",{"id":538,"label":539,"shortName":540,"title":541,"year":97,"era":10,"cluster":11,"scope":132,"keyIdeaZh":542,"sensors":543,"mapRepresentation":547,"loopClosure":548,"estimator":549,"association":550,"deskew":551,"outputGeometry":552,"fulltextStatus":22,"lidarModels":553,"equipmentCount":554},"lamp2_2022","Chang et al., 2022","LAMP 2.0","LAMP 2.0: A Robust Multi-Robot SLAM System for Operation in Challenging Large-Scale Underground Environments","LAMP 2.0 是 CoSTAR 團隊為 DARPA 地下挑戰賽開發的集中式多機器人 LiDAR 位姿圖 SLAM。各機器人的前端介面可接不同里程計（LOCUS 或 Hovermap）與不同 LiDAR 配置，先以 HeRO 狀態估計去除掃描畸變、合併多顆 LiDAR，再用自適應體素濾波讓點數一致，並每約 2 m 或 30 度建立關鍵節點與對應掃描送到基地站。基地站的多機器人前端以自適應半徑產生迴圈閉合候選，依可觀測性、圖神經網路預測效益與 RSSI 排序，再以 TEASER++ 或 SAC-IA 初始對齊後用 GICP 精修；後端以 GNC 搭配 Levenberg-Marquardt 在 GTSAM 中做抗離群值的位姿圖最佳化。作者在煤礦、核電廠、SubT 決賽場地與石灰岩礦的四組資料上評估，並釋出含地面真值的資料集。",[544,545,546],"3D LiDARs (three Velodyne lidars on Husky, a single lidar on Spot; models not reported)","Hovermap payload on some robots","odometry input from LOCUS or Hovermap (front-end agnostic)","pose graph with keyed scans (adaptive voxel-filtered point clouds); optimized global point-cloud map formed by transforming keyed scans with optimized poses (Sec. II-D)","intra- and inter-robot loop closures from the multi-robot front-end with two-stage registration (TEASER++ or SAC-IA, then GICP) (Sec. II-C)","Centralized multi-robot pose-graph optimization in GTSAM with Levenberg-Marquardt and Graduated Non-Convexity (GNC) for outlier-robust inlier selection, optionally after Incremental Consistency Maximization (ICM); single-robot front-ends send sparse pose graphs (key nodes every about 2 m or 30 deg) with keyed scans (Sec. II-B, II-D)","Proximity-based loop-closure candidates with adaptive radius, prioritized by observability (ICP information-matrix eigenvalues), a GNN-predicted graph benefit and RSSI beacons; relative pose from TEASER++ or SAC-IA initialization refined by GICP, rejecting poor alignments (Sec. II-C)","motion distortion corrected with the Heterogeneous Robust Odometry (HeRO) local state estimate before merging multi-lidar scans (Sec. II-B)","globally consistent multi-robot trajectories and point-cloud map; map compared with surveyed ground-truth map by cloud-to-cloud error (Fig. 4)",[45],8,{"id":556,"label":557,"shortName":558,"title":559,"year":249,"era":10,"cluster":68,"scope":69,"keyIdeaZh":560,"sensors":561,"mapRepresentation":57,"loopClosure":72,"estimator":564,"association":565,"deskew":57,"outputGeometry":566,"fulltextStatus":22,"lidarModels":567,"equipmentCount":185},"chebrolu2021adaptive","Chebrolu et al., 2021","Adaptive robust kernels","Adaptive Robust Kernels for Non-Linear Least Squares Problems","作者以 Barron 的一般化穩健損失為基礎，把控制核形狀的參數 α 視為未知數，以交替最小化求解：先以一維格點搜尋在 [-10, 2] 內取殘差負對數概似最小的 α，再以迭代重加權最小平方法求解模型參數。為讓 α 可取負值以強力壓低離群值，作者把配分函數的積分截斷在 ±τ（τ = 10c），並預先建立解析度 0.1 的查找表。方法整合進 SuMa 的點對平面投影式 ICP，在 KITTI 里程計序列上不需人工離群值剔除即得到最低的平均平移誤差，並在 CARLA 模擬影像的光束法平差中擴大收斂範圍。",[562,563],"3D LiDAR scans of the KITTI odometry benchmark (scanner model not named in the paper)","monocular camera images simulated in CARLA (car front-looking, UAV nadir, strong shadows, side-looking with motion blur)","alternating minimization: (1) alpha chosen by 1-D grid search over [-10, 2] minimizing the negative log-likelihood of the current residuals under a truncated partition function (tau = 10c, lookup table at 0.1 resolution); (2) model parameters by IRLS with the Barron general kernel at that alpha; scale c fixed a priori","not part of the method; in the ICP experiment it runs inside SuMa frame-to-frame point-to-plane projective ICP without any outlier rejection step; in BA the initial matches come from SIFT with 5-point RANSAC","robust estimate (ICP pose or BA solution)",[45],{"id":569,"label":570,"shortName":571,"title":572,"year":177,"era":52,"cluster":68,"scope":69,"keyIdeaZh":573,"sensors":574,"mapRepresentation":576,"loopClosure":72,"estimator":577,"association":578,"deskew":57,"outputGeometry":579,"fulltextStatus":22,"lidarModels":580,"equipmentCount":144},"chen1992pointtoplane","Chen & Medioni, 1992","Point-to-plane ICP (Chen-Medioni)","Object modelling by registration of multiple range images","此文為點對平面 ICP 的原始期刊版本。作者假設兩個視角已有近似配準，在 P 上以規則格點挑選平滑區域的控制點，沿 P 在該點的法向線與數位曲面 Q 求交（以切平面迭代近似，通常 3 至 5 次），再以 Q 在交點的切平面為目標，最小化控制點到切平面的有號距離平方和，不需要點對點對應。實驗以結構光測距儀取得莫札特半身像、牙齒模型與木塊的距離影像；多視角建模時把新視角對已合併的全部資料配準，以減少逐對配準的誤差累積。",[575],"structured-light range finder after Sato and Inokuchi: projector with a programmable liquid crystal mask and a CCD camera, space coding with projected stripe patterns and triangulation; accuracy about 1 mm; range images at 0.5 mm spatial resolution stored as 32-bit floats","object-centred cylindrical or spherical coordinate map; views are reparameterized by interpolation and averaged in overlaps, with outlier handling","iterative least squares: at each iteration find T minimizing the sum of squared signed distances from transformed control points to the tangent planes of Q at the normal-line intersection points (Eq. 10), compose T^k = T T^(k-1), stop when the change measure of Eq. 11 falls below epsilon_c (0.01 in tests); cost per iteration linear in the number of control points","no point-to-point correspondence: for each control point p_i on P the line along the P-normal is intersected with digital surface Q by a Newton-like tangent-plane iteration (typically 3 to 5 iterations, stop within one sampling unit), and the tangent plane of Q at that intersection is the target; control points (usually 50 to 200) are taken on a regular grid in smooth areas (9x9 plane-fit residual below half a sampling unit)","6-DoF rigid transformation between range views; integrated object model as a spherical coordinate map, rendered views and wireframe",[],{"id":582,"label":583,"shortName":584,"title":585,"year":9,"era":10,"cluster":151,"scope":132,"keyIdeaZh":586,"sensors":587,"mapRepresentation":588,"loopClosure":589,"estimator":590,"association":591,"deskew":121,"outputGeometry":592,"fulltextStatus":22,"lidarModels":593,"equipmentCount":24},"sumapp2019","Chen et al., 2019","SuMa++","SuMa++: Efficient LiDAR-based Semantic SLAM","SuMa++ 在 SuMa 的面元建圖流程中加入 LiDAR 語意分割（RangeNet++，於球面投影影像上推論逐點類別），並以深度一致的洪水填充（flood-fill）修正物體邊界的標籤錯誤。更新地圖時若觀測類別與面元類別不一致，就降低該面元的穩定度，使移動物體逐漸被移除，同時保留停放車輛等靜態物體；ICP 殘差也依語意相容性加權以抑制離群值。",[154],"semantically labeled surfel map; semantic inconsistency penalizes surfel stability to remove moving objects (Sec. III-C, III-E)","SuMa loop closure and pose-graph optimization (Sec. III-B); semantics not used for loop closure (Sec. V)","SuMa frame-to-model ICP (Gauss-Newton) with residual weights combining Huber, semantic compatibility and surfel stability; SuMa pose graph (Sec. III-B, III-F)","projective data association as in SuMa, with per-point labels from RangeNet++ refined by depth-aware flood-fill (Sec. III-C, III-D, III-F)","semantic surfel map with point-wise labels (Sec. III-C, Fig. 4)",[161],{"id":595,"label":596,"shortName":597,"title":598,"year":396,"era":10,"cluster":599,"scope":600,"keyIdeaZh":601,"sensors":602,"mapRepresentation":603,"loopClosure":604,"estimator":605,"association":606,"deskew":121,"outputGeometry":57,"fulltextStatus":22,"lidarModels":607,"equipmentCount":46},"overlapnet2020","Chen et al., 2020","OverlapNet","OverlapNet: Loop Closing for LiDAR-based SLAM","C06","place_recognition_component","OverlapNet 以孿生網路（siamese network）比較兩次光達掃描，輸入由單次掃描產生的距離影像、法向量、強度與語意機率，輸出兩者的重疊率與相對偏航角。系統以位姿共變異數傳播決定迴圈搜尋範圍，取代 SuMa 原本只取最近幀的啟發式迴圈偵測，並可用預測的偏航角作為 ICP 初值。",[35],"not_applicable (host uses surfels)","overlap-based candidate detection within a covariance-propagated search region, replacing SuMa's nearest-frame heuristic; yaw estimate can initialize ICP","not_applicable (integrated into SuMa surfel SLAM with pose graph)","learned siamese network on range, normal, intensity and semantic-probability images predicting scan overlap and relative yaw",[161,608],"a different version of the Velodyne HDL-64E",{"id":610,"label":611,"shortName":612,"title":613,"year":97,"era":10,"cluster":98,"scope":99,"keyIdeaZh":614,"sensors":615,"mapRepresentation":618,"loopClosure":72,"estimator":619,"association":620,"deskew":621,"outputGeometry":622,"fulltextStatus":22,"lidarModels":623,"equipmentCount":487},"dlo2022","Chen et al., 2022a","DLO","Direct LiDAR Odometry: Fast Localization With Dense Point Clouds","DLO 採「速度優先」設計，直接使用輕度降採樣的稠密點雲，以自製 NanoGICP 先做相鄰掃描配準、再對由關鍵影格組成的子地圖配準。子地圖不以半徑搜尋點，而是在關鍵影格空間中選取最近鄰與凸包關鍵影格拼接，使遠處結構也能參與配準。IMU 僅選擇性提供旋轉初值（鬆耦合），且作者明言未做運動畸變校正。",[616,617],"3D LiDAR (Ouster OS1, Velodyne VLP-16)","IMU (optional, gyroscope rotational prior only; VectorNav VN-100 on the field platforms and in the SubT Alpha dataset)","keyframe database; submap built by concatenating k-nearest and convex-hull keyframe clouds (adaptive keyframing via spaciousness metric)","two-stage GICP optimization (scan-to-scan then scan-to-map) with optional loosely-coupled IMU rotational prior","GICP on minimally preprocessed dense clouds (1 m box filter, 0.25 m voxel filter) using custom NanoGICP with data-structure reuse","none (authors state they do not correct motion distortion, Sec. II-B)","pose estimates and keyframe-based map; export format not_reported",[624,143],"Ouster OS1",{"id":626,"label":627,"shortName":628,"title":629,"year":97,"era":10,"cluster":151,"scope":99,"keyIdeaZh":630,"sensors":631,"mapRepresentation":633,"loopClosure":634,"estimator":635,"association":636,"deskew":637,"outputGeometry":638,"fulltextStatus":22,"lidarModels":639,"equipmentCount":554},"ndtloam2022","Chen et al., 2022b","NDT-LOAM","NDT-LOAM: A Real-Time Lidar Odometry and Mapping With Weighted NDT and LFA","NDT-LOAM 把 LOAM 的特徵式前端改成加權的常態分布轉換（NDT）直接配準：每個 NDT 格依量測距離與格內形狀（平面、線狀或立體）給不同權重，並以目前幀對最近關鍵影格配準（Scan2Key）降低逐幀累積誤差。得到的初始位姿再交給沿用 LOAM 建圖模組的局部特徵調整（LFA），以角點與平面點對局部地圖精修。系統只處理前端，沒有迴圈閉合；在 KITTI 上平均平移漂移為 0.899%。",[632],"3D LiDAR only (Velodyne HDL-64E in KITTI; horizontal Velodyne VLP-16 on the Kylin backpack)","LOAM-style corner and surface feature map stored in cubic areas for LFA; keyframe scans (selected by 10 m, 10 deg or 1 s on KITTI; 2 m or 10 deg on the backpack) as NDT targets (Sec. III-C; Sec. III-D; Sec. IV-A; Sec. IV-D)","none (loop closure and graph optimization are named as future work) (Sec. III-C; Sec. V)","two stages: direct odometry by weighted NDT solved with Newton's method and line search against a keyframe (Scan2Key), then local feature adjustment (LFA) that reuses the LOAM mapping module to refine the pose with corner and surface feature residuals against a local map, two iterations (Sec. III)","NDT cells (1 m grid in KITTI tests) weighted by point range and by cell dimensionality from covariance eigenvalues (planar 1.25, volumetric 1.0, linear 0.75); LFA corner and surface correspondences searched in a KD-tree built from map cubes intersecting the scan (Sec. III-B; Sec. III-D; Sec. IV-A)","not described","trajectory and 3D point cloud map (Figs. 5 and 7)",[161,143],{"id":641,"label":642,"shortName":643,"title":644,"year":51,"era":10,"cluster":98,"scope":99,"keyIdeaZh":645,"sensors":646,"mapRepresentation":649,"loopClosure":650,"estimator":651,"association":652,"deskew":653,"outputGeometry":654,"fulltextStatus":22,"lidarModels":655,"equipmentCount":109},"dlio2023","Chen et al., 2023","DLIO","Direct LiDAR-Inertial Odometry: Lightweight LIO with Continuous-Time Motion Correction","DLIO 以由粗到細的方式建構掃描內連續時間軌跡：先以 IMU 數值積分得到離散位姿，再以恆定急動度（jerk）與恆定角加速度的解析式為每個點求得去畸變轉換，可平行計算。去畸變同時產生 GICP 的初值，因此可省去掃描對掃描步驟而直接做掃描對地圖配準。狀態由具全域收斂性質的非線性幾何觀測器（geometric observer）更新，而非卡爾曼濾波或因子圖。",[647,648],"3D mechanical LiDAR (tested: Ouster OS1 with 32 channels at 10 Hz on UCLA data and the Ouster LiDAR of Newer College; Velodyne is named only as an example input)","6-axis IMU (tested: InvenSense MPU-6050 on UCLA data; Ouster internal IMU at 100 Hz on Newer College)","keyframe-based map with submap generation (from DLO)","none (adding loop closures listed as future work, Sec. V)","hierarchical nonlinear geometric observer (contraction-based) updated with GICP scan-to-map pose; IMU propagation between scans","GICP scan-to-map on dense, lightly filtered clouds (no feature extraction; scan-to-scan stage removed)","point-wise continuous-time motion correction that also builds the GICP prior","odometry and keyframe point-cloud map; export format not_reported",[624,45],{"id":657,"label":658,"shortName":659,"title":660,"year":661,"era":10,"cluster":98,"scope":99,"keyIdeaZh":662,"sensors":663,"mapRepresentation":665,"loopClosure":72,"estimator":666,"association":667,"deskew":668,"outputGeometry":669,"fulltextStatus":22,"lidarModels":670,"equipmentCount":24},"iglio2024","Chen et al., 2024","iG-LIO","iG-LIO: An Incremental GICP-Based Tightly-Coupled LiDAR-Inertial Odometry",2024,"iG-LIO 將廣義 ICP（GICP）約束與 IMU 約束緊耦合於最大後驗（MAP）估計，以迭代式誤差狀態更新求解。作者以體素為基礎的表面共變異數估計器降低共變異數計算成本，並以增量式體素地圖儲存環境的機率模型，以減少最近鄰搜尋與地圖管理時間。作者強調所有資料集使用相同參數，效率高於 Faster-LIO 而精度相近。",[664,36],"3D LiDAR (mechanical and solid-state)","incremental voxel map storing probabilistic (point and covariance) models","MAP estimation combining IMU prior and GICP constraints, solved by Gauss-Newton iterations, with error-state covariance propagation analogous to the iterated error-state Kalman filter","GICP with voxel-based surface covariance estimator (VSCE); nearest neighbours via voxel hash indexes","IMU integration (midpoint integration) motion compensation before registration","IMU-rate and LiDAR-rate odometry and voxel map; export format not_reported",[437,225,223,143],{"id":672,"label":673,"shortName":674,"title":675,"year":235,"era":52,"cluster":236,"scope":676,"keyIdeaZh":677,"sensors":678,"mapRepresentation":680,"loopClosure":681,"estimator":682,"association":683,"deskew":684,"outputGeometry":685,"fulltextStatus":22,"lidarModels":686,"equipmentCount":77},"choi2015robustrecon","Choi et al., 2015","Robust Reconstruction of Indoor Scenes (Redwood)","Robust reconstruction of indoor scenes","offline_map_refinement","本文提出離線的 RGB-D 室內場景重建流程。先把影片切成每 50 張影格一段，以 RGB-D 里程計估計段內軌跡並以 TSDF 融合成場景片段（fragment）；再對所有片段兩兩做幾何全域配準（改良的 PCL FPFH 搭配 RANSAC），得到候選迴圈閉合。作者發現即使最佳的配準演算法，精確率也低於 20%，因此在位姿圖最佳化中為每條迴圈閉合邊加入線過程（line process）變數，並以稠密表面對應距離定義對齊誤差，讓錯誤的邊在同一個最小平方問題中自動失效。修剪後再以 ICP 精修與位姿圖求得片段位姿，最後以體積整合輸出全域網格。作者同時擴充 ICL-NUIM 資料集，加入完整掃描軌跡、較真實的深度雜訊模型與辦公室場景表面真值。",[679],"RGB-D video from a consumer depth camera (SUN3D real scenes; augmented ICL-NUIM synthetic sequences with a disparity-based noise and distortion model); fragment-pair registration is geometric (FPFH on fragment point sets), sensor models not named","fragment meshes from volumetric TSDF integration (Curless and Levoy) of 50-frame segments; final global volumetric integration","Geometric: every fragment pair tested by global registration, so loops are found even when images are not similar; false candidates removed by the line-process optimization","Offline pipeline: RGB-D odometry (Kerl et al. 2013) inside 50-frame fragments; pose graph over fragment poses with odometry edges and loop-closure edges weighted by line-process variables l in [0, 1] with prior (sqrt(l) - 1)^2 and balance mu = tau^2 kappa, solved with g2o; edges with l \u003C 0.25 pruned; ICP refinement of the remaining edges and final pose-graph optimization; optional SLAC non-rigid refinement","Pairwise geometric registration of every fragment pair with a modified PCL FPFH plus RANSAC algorithm (four-point samples, normal-angle, edge-length and overlap checks); candidates kept when more than 30% overlap; alignment cost uses dense point correspondences within 5 cm","not_applicable (RGB-D input)","global surface mesh (for example 15.8 million triangles for the 17,391-frame apartment)",[],{"id":688,"label":689,"shortName":690,"title":691,"year":9,"era":10,"cluster":68,"scope":69,"keyIdeaZh":692,"sensors":693,"mapRepresentation":696,"loopClosure":72,"estimator":697,"association":698,"deskew":57,"outputGeometry":699,"fulltextStatus":22,"lidarModels":700,"equipmentCount":46},"choy2019fcgf","Choy et al., 2019","FCGF","Fully Convolutional Geometric Features","FCGF 以 Minkowski Engine 稀疏卷積構成的 ResUNet，一次計算整片點雲每個體素的 32 維幾何特徵，輸入只用座標與常數特徵，不需法向量或局部區塊（patch）前處理。作者提出最難負樣本對比損失與最難三元組損失，並以雜湊方式濾除錨點附近的假負樣本。3DMatch 上特徵匹配召回率 0.952、配準召回率平均 0.82；KITTI 上以 RANSAC 配準的成功率為 97.83% 至 98.92%。屬學習式先驗，跨感測器與場域的泛化需另行驗證。",[694,695],"indoor 3D scan fragments of the 3DMatch benchmark (sensor not named in the paper)","KITTI odometry LiDAR scans (model not named in the paper)","sparse tensor (sparse voxel) representation of point clouds (Sec. 3)","descriptor only; transformations estimated with RANSAC on FCGF correspondences: RANSAC with early termination for the 3DMatch registration recall (Sec. 6.5) and RANSAC for the KITTI RTE and RRE (Sec. 6.6; early termination is not stated for KITTI)","dense 32-D features from a ResUNet of generalized sparse convolutions (Minkowski Engine) on voxel-downsampled points with 1-vectors as input features; trained with hardest-contrastive or hardest-triplet losses using hash-based filtering of false negatives near anchors; matches found by feature similarity","32-dimensional per-point features (abstract)",[45],{"id":702,"label":703,"shortName":704,"title":705,"year":706,"era":52,"cluster":527,"scope":132,"keyIdeaZh":707,"sensors":708,"mapRepresentation":710,"loopClosure":711,"estimator":712,"association":713,"deskew":714,"outputGeometry":715,"fulltextStatus":22,"lidarModels":716,"equipmentCount":77},"cole_newman2006_3dslam","Cole & Newman, 2006","Cole-Newman 3D laser SLAM with an oscillating 2D scanner","Using laser range data for 3D SLAM in outdoor environments",2006,"本文把 2D 延遲狀態（掃描匹配式）SLAM 延伸為戶外起伏地形的 6 自由度 SLAM。一台 SICK 2D 雷射以 0.6 Hz 繞水平軸來回擺動，車輛行進時持續取得 3D 資料，再依里程計行駛距離與姿態變化門檻把資料流切成各自參考一個車輛位姿的點雲，不必停車掃描。狀態向量是一串過去的位姿，里程計負責擴增狀態，相鄰點雲以帶 Geman-McClure 穩健核、並用 Levenberg-Marquardt 最小化的 ICP 式配準提供觀測。作者另以最近鄰距離直方圖訓練高斯分類器，在配準後判斷是否落入錯誤的局部極小值。",[709,119],"custom 3D laser range finder: a standard 2D SICK scanner oscillating at 0.6 Hz about a horizontal axis (model not reported)","state vector of past 6-DoF poses with attached 3D point clouds","prompted when a past pose falls within the current pose uncertainty; the loop registration is re-seeded by perturbation until the integrity check accepts it (Sec. VI; Sec. VII); a companion paper handles robust detection","delayed-state (view-based) EKF whose state is a stack of past 6-DoF vehicle poses (x, y, z, roll, pitch, yaw); odometry augments the state and registration-derived inter-pose transforms are EKF observations (Sec. III)","ICP-like nearest-neighbour point matching with a Geman-McClure robust kernel minimized by Levenberg-Marquardt, approximate k-d tree queries, odometry transform as the initial estimate (Sec. IV; Algorithm 1)","within each segment, scans are transformed into the base pose frame with odometry; a segment ends early when an orientation-change threshold is exceeded and is kept only if it has enough points (Sec. II)","6-DoF trajectory with marginal covariances and the attached point clouds rendered together (Figs. 10-12)",[717],"SICK scanner",{"id":719,"label":720,"shortName":721,"title":722,"year":51,"era":10,"cluster":599,"scope":54,"keyIdeaZh":723,"sensors":724,"mapRepresentation":729,"loopClosure":730,"estimator":731,"association":732,"deskew":733,"outputGeometry":734,"fulltextStatus":22,"lidarModels":735,"equipmentCount":736},"maplab2_2023","Cramariuc et al., 2023","maplab 2.0","maplab 2.0 - A Modular and Multi-Modal Mapping Framework","maplab 2.0 是以因子圖為核心的模組化、多模態建圖框架：一張地圖由多個任務（mission，即單次連續建圖時段）組成，頂點包含位姿、速度、IMU 偏差與地標，可整合視覺、光達與語意地標。新版加入 mapping server，把各機器人的子地圖先各自做局部最佳化與迴圈，再於全域層級做跨機迴圈與聯合最佳化；離線 console 提供批次最佳化、地圖合併、以 ICP\u002FG-ICP 產生並以可切換約束抗離群的光達迴圈，以及 Voxblox 稠密重建。",[35,725,726,727,728],"IMU (optional; framework does not require an IMU)","multi-camera or monocular camera","GNSS (RTK, optional absolute constraints)","wheel encoders and RGB-D landmarks (supported interfaces, not evaluated)","factor-graph map of missions with attached sensor data; dense reconstruction via Voxblox plugin","visual and LiDAR intra- and inter-mission loop closures; loop edges as switchable constraints","factor graph over vertices (6-DoF pose, velocity, IMU biases, landmarks) with batch optimization \u002F bundle adjustment","visual landmarks from ORB detection with BRISK or FREAK binary descriptors (inverted multi-index matching), plus external float descriptors such as SuperPoint with SuperGlue tracking and SIFT with Lucas-Kanade tracking (PCA-compressed 256 to 32, FLANN matching); 2D-3D matches with covisibility filtering and P3P in RANSAC for visual loop closure, or landmark merging; 3D landmarks (RGB-D, LiDAR image keypoints) matched by 3D-3D RANSAC; LiDAR loop closures by ICP or G-ICP registration in the console","not_reported (delegated to odometry source, e.g., FAST-LIO2)","optimized poses, landmarks; dense volumetric reconstruction via Voxblox plugin; map data export (Sec. III-D)",[224,222],9,{"id":738,"label":739,"shortName":740,"title":741,"year":742,"era":52,"cluster":31,"scope":32,"keyIdeaZh":743,"sensors":744,"mapRepresentation":746,"loopClosure":57,"estimator":747,"association":57,"deskew":57,"outputGeometry":748,"fulltextStatus":22,"lidarModels":749,"equipmentCount":144},"curless1996volumetric","Curless & Levoy, 1996","Volumetric range-image integration (TSDF origin; VRIP)","A volumetric method for building complex models from range images",1996,"作者將每張已對齊的距離影像（range image）沿感測器視線轉成有號距離函數與權重，逐一加權累加到體素格網中，最後擷取零等值面成為三角網格；在特定假設下，此等值面在最小平方意義上最佳。體素另外標記為空、未觀測或近表面三種狀態，藉由空間雕刻（space carving）在空與未觀測區域的交界補面，產生無孔洞模型。",[745],"[\"Cyberware 3030 MS laser-stripe optical triangulation scanner (traditional triangulation or spacetime analysis)\"]","cumulative weighted signed distance function on a voxel grid (run-length encoded) with empty\u002Funseen\u002Fnear-surface voxel states","not_applicable (range images are assumed pre-aligned)","watertight triangle mesh extracted as the zero isosurface; hole filling by tessellating between empty and unseen regions",[],{"id":751,"label":752,"shortName":753,"title":754,"year":396,"era":10,"cluster":755,"scope":132,"keyIdeaZh":756,"sensors":757,"mapRepresentation":758,"loopClosure":759,"estimator":760,"association":761,"deskew":57,"outputGeometry":762,"fulltextStatus":22,"lidarModels":763,"equipmentCount":77},"deepfactors2020","Czarnowski et al., 2020","DeepFactors","DeepFactors: Real-Time Probabilistic Dense Monocular SLAM","C09","DeepFactors 把 CodeSLAM 的學習式精簡深度編碼放進標準因子圖：每個關鍵影格的深度由 32 維編碼經以影像為條件的線性解碼器產生，位姿與編碼一起以 GTSAM 的 iSAM2 做批次最大後驗估計。關鍵影格間同時使用稠密光度誤差、BRISK 特徵重投影誤差與稀疏幾何深度一致性誤差三種因子，並有局部與以詞袋檢索的全域迴圈閉合。它在單張 GTX 1080 上可即時運作，但輸出深度解析度只有 256x192，且小編碼難以表示完全平坦的表面。",[37],"key-frame depth maps decoded linearly from a 32-dimensional learned code conditioned on the image (CodeSLAM-style variational auto-encoder) (Sec. III-IV, Sec. VI-A)","local loops by a pose-based criterion within the last 10 key-frames; global loops by bag-of-words (DBoW2) candidates verified by tracking inliers and pose distance, closed with reprojection factors only (Sec. V-D)","batch MAP factor graph over key-frame poses and compact depth codes solved incrementally with iSAM2 in GTSAM; camera tracking by GPU direct whole-image SE(3) Lucas-Kanade against the closest key-frame; tracking and mapping interleaved; one-way frames refine the latest key-frame and are then marginalized (Sec. V)","three pairwise factor types: dense photometric error, BRISK keypoint reprojection error with Cauchy cost, and sparse geometric depth-consistency error with Huber cost (Sec. III)","key-frame poses and dense key-frame depth maps fused for visualization as reconstructions (Figs. 1 and 7)",[],{"id":765,"label":766,"shortName":767,"title":768,"year":769,"era":52,"cluster":236,"scope":132,"keyIdeaZh":770,"sensors":771,"mapRepresentation":773,"loopClosure":774,"estimator":775,"association":776,"deskew":121,"outputGeometry":777,"fulltextStatus":22,"lidarModels":778,"equipmentCount":554},"bundlefusion2017","Dai et al., 2017a","BundleFusion","BundleFusion: Real-Time Globally Consistent 3D Reconstruction Using On-the-Fly Surface Reintegration",2017,"BundleFusion 在每一影格都考慮完整的 RGB-D 歷史資料，以分塊（chunk）的階層式區域到全域最佳化，結合稀疏 SIFT 特徵與稠密幾何、光度對應，即時求得經 BA 的全域位姿。位姿更新後，系統即時將受影響影格從 TSDF 中移除並以新位姿重新融合（re-integration），使稠密模型保持全域一致。由於每張影格都與全部歷史比對，迴圈閉合是隱式處理的。",[772],"RGB-D","TSDF in sparse voxel hashing with 8x8x8 voxel blocks, default 4 mm voxels (1 cm also evaluated); every frame stored with integrated and optimised poses and the 10 frames with the largest pose change re-integrated per new frame","implicit: each frame is globally correlated to all previous frames, so no explicit loop detection; relocalization after gross tracking failure (abstract; introduction)","Two-level hierarchical pose-only optimisation: intra-chunk alignment of 11 consecutive frames (1-frame overlap) and inter-chunk alignment of chunk keyframes with aggregated feature sets; energy = sparse SIFT correspondence distances plus dense photometric (luminance gradient) and point-to-plane geometric terms on 80x60 downsampled frames with linearly increasing dense weight; Gauss-Newton with a GPU data-parallel PCG solver (Jacobi preconditioner), warm-started from the previous frame; about 20 times faster than Ceres on a 101-keyframe sparse problem","GPU SIFT matched against all previous frames (about 150 features per frame, 250 per keyframe) and filtered by a Kabsch-based key point correspondence filter (max residual 0.02 m, condition number limit 100), a surface-area filter (0.032 m2) and two-sided dense geometric and photometric verification (Nmin = 5); after each optimisation all correspondences of a frame pair with residual above 0.05 m are pruned; dense terms use frame pairs within 60 degrees and nonzero overlap","textured surface mesh from the TSDF (authors contrast it with the point-cloud output of ElasticFusion) plus globally optimised per-frame poses",[],{"id":780,"label":781,"shortName":782,"title":783,"year":513,"era":52,"cluster":236,"scope":132,"keyIdeaZh":784,"sensors":785,"mapRepresentation":788,"loopClosure":789,"estimator":790,"association":791,"deskew":57,"outputGeometry":792,"fulltextStatus":22,"lidarModels":793,"equipmentCount":487},"monoslam2007","Davison et al., 2007","MonoSLAM","MonoSLAM: Real-Time Single Camera SLAM","MonoSLAM 以單一延伸卡爾曼濾波器（Extended Kalman Filter, EKF）同時估計相機位姿與稀疏自然地標，並保留兩者之間的完整共變異數（covariance），讓單眼相機即可即時建立持續存在的機率式地圖。作者引入依預測不確定度進行的主動量測（active measurement）與平滑運動模型，以降低影像處理成本。其地圖為稀疏地標而非稠密點雲，作者也將適用範圍描述為室內、房間尺度。",[786,787],"monocular wide-angle camera (field of view near 100 degrees, 30 Hz)","3-axis gyro fused as an internal angular-velocity measurement in the HRP-2 humanoid experiment only","Single state vector and full covariance over the camera and about 100 sparse 3D point features, each stored with an oriented planar patch template; optional surface-normal estimates kept in separate two-parameter EKFs per feature","implicit loop closure inside the single EKF: re-observing early-mapped features after exploration corrects accumulated drift, and active feature selection favours such re-observation; no separate place-recognition module (Sec. 4 AR results; Sec. 5 humanoid circular walk, Fig. 10 'loop closed and drift corrected')","extended Kalman filter over joint camera and landmark state with full covariance (Sec. 3.1)","Shi-Tomasi salient 11x11 pixel patches stored as locally planar templates, warped to the predicted view and matched by normalised cross-correlation only inside 3-sigma innovation-covariance ellipses (typically 15 to 20 pixels across); per frame the 10 to 12 features with the highest innovation covariance are measured; new features start as 3D rays with 100 depth particles between 0.5 and 5 m and become 3D points when depth std over depth falls below 0.3; features failing more than 50% of attempted measurements are deleted","camera trajectory and sparse 3D landmark positions with uncertainty (Sec. 3.1)",[],{"id":795,"label":796,"shortName":797,"title":798,"year":30,"era":52,"cluster":53,"scope":54,"keyIdeaZh":799,"sensors":800,"mapRepresentation":57,"loopClosure":57,"estimator":801,"association":57,"deskew":57,"outputGeometry":57,"fulltextStatus":22,"lidarModels":802,"equipmentCount":61},"gtsam_software","Dellaert & GTSAM Contributors, 2026","GTSAM","GTSAM 4.3.0","GTSAM 是以因子圖與 Bayes network 為運算範式（而非直接操作稀疏矩陣）的 C++ 平滑與建圖程式庫，提供 MATLAB 與 Python 包裝，實作批次最佳化、iSAM2 與固定延遲平滑器。README 說明其預積分 IMU 因子以 Lupton 與 Sukkarieh 的概念為基礎、依 Forster 等人的流形積分改良，而 GTSAM 4 預設改用在 NavState 切空間積分的實作（可用 GTSAM_TANGENT_PREINTEGRATION 切回 RSS 2015 版本）。4.3.0 版發行說明另列出連續時間高斯過程與 Lie 群樣條、GNSS 因子、riSAM 與 GNC 等穩健估計、平行多前沿求解器與實驗性 CUDA 加速；IMU 因子文件並說明可把重力當作最佳化變數，以用於未與重力對齊的 LiDAR 里程計地圖座標系。4.3.0 標籤內 README 的引用區塊仍指向較早的 Zenodo DOI 10.5281\u002Fzenodo.5794541，develop 分支才改為 4.3.0 版記錄。",[],"factor-graph smoothing and mapping library: batch nonlinear optimization (Levenberg-Marquardt named in the 4.3.0 notes), incremental iSAM2, IncrementalFixedLagSmoother (moved to the stable library in 4.3), EKF and invariant-EKF infrastructure, robust GNC and riSAM, constrained (LP, QP, QCQP) and certifiable solvers (README; 4.3.0 release notes)",[],{"id":804,"label":805,"shortName":806,"title":807,"year":706,"era":52,"cluster":53,"scope":54,"keyIdeaZh":808,"sensors":809,"mapRepresentation":811,"loopClosure":812,"estimator":813,"association":814,"deskew":57,"outputGeometry":815,"fulltextStatus":22,"lidarModels":816,"equipmentCount":24},"dellaert2006sqrtsam","Dellaert & Kaess, 2006","Square Root SAM","Square Root SAM: Simultaneous Localization and Mapping via Square Root Information Smoothing","本文把平滑（smoothing）視為 EKF 型 SLAM 的替代方案，將資訊矩陣或量測 Jacobian 分解為平方根形式求解。作者主張此類方法精確且更快，可批次或增量使用，較能處理非線性的運動與量測模型，並能以較低成本得到整條軌跡。欄位排序啟發式能間接利用 SLAM 問題在地理上的局部性。",[810,119],"8-camera rig (visual point features)","landmarks (simulated landmarks observed with bearing and range; 4383 unknown 3D points in the real experiment)","none occurred in the real experiment (Fig. 16 caption); fill-in when closing loops discussed for simulations (Sec. 7.2)","square-root information smoothing by factorizing the information matrix (Cholesky or LDL) or the measurement Jacobian (QR), in batch or incremental mode; best performance with Davis' sparse LDL and colamd or symamd ordering applied to the block (pose and landmark) structure; non-linear problems are relinearized and refactorized at each call","real experiment: features matched between successive frames using RANSAC on a trifocal camera arrangement (Sec. 8); the formulation assumes data association is solved (Sec. 2) and the authors state they ignored data association (Sec. 10)","entire robot trajectory and map",[],{"id":818,"label":819,"shortName":820,"title":821,"year":97,"era":10,"cluster":98,"scope":132,"keyIdeaZh":822,"sensors":823,"mapRepresentation":825,"loopClosure":826,"estimator":827,"association":828,"deskew":829,"outputGeometry":830,"fulltextStatus":22,"lidarModels":831,"equipmentCount":554},"cticp2022","Dellenbach et al., 2022","CT-ICP","CT-ICP: Real-time Elastic LiDAR Odometry with Loop Closure","CT-ICP 以每次掃描的起始與結束兩個位姿參數化掃描內的連續時間軌跡，在點到平面 ICP 中同時估計扭曲，使掃描可「彈性」變形；掃描之間允許不連續，並以位置一致與等速兩項約束抑制過度跳動。地圖為稀疏體素中的稠密點雲。作者再以局部地圖投影成高程影像的迴圈偵測與 g2o 位姿圖構成完整 SLAM，但此迴圈方法假設運動大致在平面上。",[824],"3D LiDAR only (xyz plus per-point timestamps)","dense point cloud in a sparse voxel grid (max 20 points per voxel, 10 cm minimum spacing; voxel 1.0 m driving, 0.8 m high-frequency motion)","elevation images of aggregated local maps with rotation-invariant 2D features, RANSAC and ICP refinement; requires mostly planar ground motion and z-axis aligned with ground normal","scan-to-map continuous-time ICP (robust loss, iterative least squares) over two poses per scan with location-consistency and constant-velocity regularizers; g2o pose-graph back-end","point-to-plane to dense local map, normals and planarity weights from k=20 neighbours in 27 surrounding voxels","elastic: scan distortion estimated jointly with registration","trajectory and aggregated point clouds (Fig. 2); export format not_reported",[161,832,225,833],"simulated 64-channel LiDAR","Ouster 64-channel LiDAR",{"id":835,"label":836,"shortName":837,"title":838,"year":30,"era":10,"cluster":755,"scope":99,"keyIdeaZh":839,"sensors":840,"mapRepresentation":843,"loopClosure":844,"estimator":845,"association":846,"deskew":847,"outputGeometry":848,"fulltextStatus":22,"lidarModels":849,"equipmentCount":46},"deng2026_mcgs_slam","Deng & Gan, 2026","MCGS SLAM","Motion-prior and Confidence-aware Gaussian Splatting (MCGS) SLAM for 3D scene reconstruction of indoor built environments","MCGS-SLAM 以 MonoGS 高斯潑濺 SLAM 為基礎，加入由位姿歷史或加速度計推得的運動先驗（含自適應權重與快速運動偵測）、依位置穩定性、形狀、不透明度、空間與邊緣重要性計算的高斯信心度、依速度、旋轉、位移、可見重疊與視覺豐富度的自適應關鍵影格選擇，以及信心加權的多任務最佳化。評估只用 TUM RGB-D、Replica 與 EuRoC 公開資料集；TUM fr1_desk 的 ATE RMSE 由 2.13 cm 降至 1.52 cm，但 Replica 七個場景有三個變差，EuRoC MH03 至 MH05 誤差仍達 1.79 至 3.78 m，且未在實際建物或工地測試。",[841,842],"RGB-D camera (synchronized RGB-D input)","optional accelerometer and gyroscope for the motion prior; a motion-only mode infers velocity and acceleration from pose history when inertial data are absent","3D Gaussian splatting map (MonoGS baseline) with per-Gaussian confidence driving pruning, densification and reduced learning rates for stable Gaussians","none (authors state the confidence-aware map maintenance avoids descriptor-based loop closure; EuRoC table lists MCGS-SLAM without loop closure)","frame-to-model tracking by minimizing photometric and geometric rendering losses on a 3DGS map (MonoGS baseline) plus an adaptively weighted motion-prior loss (translation, rotation, velocity and acceleration smoothness), with fast-motion detection that raises the prior weight and keyframe rate","dense direct photometric and geometric residuals over visible Gaussians, weighted by per-Gaussian confidence (position stability, shape quality, opacity, spatial importance, edge importance); keyframes chosen by velocity, rotation, displacement, visibility overlap and feature richness","not_applicable (camera frames; no scanning sensor)","3D Gaussian map with rendered RGB and depth and a reconstructed point cloud; geometry assessed qualitatively and by cloud-to-mesh distance ranges on Replica (Fig. 15), no tabulated geometric accuracy",[],{"id":851,"label":852,"shortName":853,"title":854,"year":51,"era":10,"cluster":755,"scope":99,"keyIdeaZh":855,"sensors":856,"mapRepresentation":858,"loopClosure":859,"estimator":860,"association":861,"deskew":862,"outputGeometry":863,"fulltextStatus":22,"lidarModels":864,"equipmentCount":46},"nerfloam2023","Deng et al., 2023","NeRF-LOAM","NeRF-LOAM: Neural Implicit Representation for Large-Scale Incremental LiDAR Odometry and Mapping","NeRF-LOAM 將 LiDAR 里程計與建圖都建立在稀疏八元樹體素嵌入加上共用解碼器的神經 SDF 上，以 SDF 誤差對位姿做梯度下降，並把地面與非地面點分開以抑制 Z 方向漂移。最後以關鍵掃描緩衝區精修地圖與位姿，再用 marching cubes 輸出網格。作者承認目前無法即時且沒有迴圈閉合。",[857],"3D LiDAR (MaiCity: noise-free synthetic 64-beam scans; Newer College: hand-carried LiDAR; KITTI; sensor models not named in the paper)","sparse octree voxels with neural embeddings and a shared 2-layer MLP decoding SDF","none (planned future work)","gradient descent on SE(3) (Lie algebra) minimizing the SDF loss against the current implicit map, initialised by a constant motion model; the shared MLP is frozen after the first K scans to limit catastrophic forgetting; neural mapping jointly optimizes voxel embeddings and fine-tunes poses; final key-scan refinement of map and poses","neural SDF with ground\u002Fnon-ground separation of LiDAR points to limit Z drift","not_reported (Newer College used 'with motion distortion'; no deskewing described in sections read)","dense mesh via marching cubes + refined scan poses",[865,45],"64-beam LiDAR (simulated, noise-free)",{"id":867,"label":868,"shortName":869,"title":870,"year":150,"era":10,"cluster":527,"scope":99,"keyIdeaZh":871,"sensors":872,"mapRepresentation":874,"loopClosure":875,"estimator":876,"association":877,"deskew":878,"outputGeometry":879,"fulltextStatus":22,"lidarModels":880,"equipmentCount":24},"imlsslam2018","Deschaud, 2018","IMLS-SLAM","IMLS-SLAM: Scan-to-Model Matching Based on 3D Data","IMLS-SLAM 只使用 3D 旋轉式 LiDAR，以掃描對模型（scan-to-model）匹配估計位姿。模型是最近 n 個已定位掃描累積而成的點雲，並以隱式移動最小平方（IMLS）曲面表示；每次迭代先把取樣點投影到該曲面，再以線性化的點到平面最小平方求剛體轉換。取樣時依車體座標軸計算九組可觀測性分數，每組各取 s 個點，使旋轉與平移都受到約束；匹配前另以地面偵測與分群刪除尺寸小於門檻的物體，近似處理動態物體。系統沒有迴圈閉合，目前實作也不是即時運算。",[873],"3D spinning LiDAR only (Velodyne HDL32 and HDL64 in the experiments); no IMU, GPS or camera","point cloud of the last n = 100 localized scans with normals, used as an implicit moving least squares (IMLS) surface (h = 0.06 m); the oldest scan is dropped as a new one is added and the k-d tree is rebuilt per scan (Sec. V; Sec. VI.C)","none (drift reported without any loop closure)","iterative scan-to-model registration: each selected sample is projected onto the IMLS surface, then the rigid transform is found by linearized point-to-plane least squares under a small-angle assumption; fixed 20 iterations per scan (Sec. V; Sec. VI)","closest point in the model cloud (FLANN k-d tree) within radius r = 0.20 m; samples chosen from nine lists ranked by their contribution to observability of roll, pitch, yaw and the three translations (planarity-weighted), s = 100 per list (Sec. IV; Sec. VI)","linear interpolation between the previous end pose and the predicted end pose before matching, recomputed with the final pose after matching (Sec. III; Sec. V); not applied on KITTI, whose scans are already de-skewed (Sec. VI.B)","trajectory and accumulated de-skewed point cloud (Figs. 4-6)",[225,161],{"id":882,"label":883,"shortName":884,"title":885,"year":886,"era":52,"cluster":527,"scope":132,"keyIdeaZh":887,"sensors":888,"mapRepresentation":892,"loopClosure":893,"estimator":894,"association":895,"deskew":121,"outputGeometry":896,"fulltextStatus":22,"lidarModels":897,"equipmentCount":24},"dissanayake2001","Dissanayake et al., 2001","EKF-SLAM convergence","A solution to the simultaneous localization and map building (SLAM) problem",2001,"本文以與 Smith 等人相同的估計理論架構，證明線性高斯情形下 EKF-SLAM 的三項性質：相對地圖不確定性單調下降、極限時地標估計完全相關、而絕對誤差下限只由初始車輛不確定性決定。作者強調維持完整地圖共變異數（交互相關項）是收斂與一致性的必要條件，省略它會造成不一致與發散。文中以毫米波雷達與車輛實作，並用測量過的地標位置進行比對，同時指出運算與儲存量隨地標數平方成長。",[889,890,891],"millimetre-wave radar (77 GHz FMCW, beam scanned 360 deg in azimuth)","drive-shaft encoders (vehicle speed)","LVDT on the steering rack (steering for vehicle heading)","point landmark map with full covariance","implicit through full landmark covariance; no separate loop-closure module","extended Kalman filter over vehicle pose and point landmarks (linear-Gaussian analysis for the proofs)","point landmarks from thresholded radar returns; Appendix 2 maintains confirmed and tentative landmark lists and associates an observation to a landmark when its Mahalanobis-type distance is below a threshold (d_min); new landmarks are validated by observation counts and a quality measure","landmark coordinates and vehicle trajectory estimates with covariance",[],{"id":899,"label":900,"shortName":901,"title":902,"year":150,"era":10,"cluster":151,"scope":132,"keyIdeaZh":903,"sensors":904,"mapRepresentation":907,"loopClosure":908,"estimator":909,"association":910,"deskew":911,"outputGeometry":912,"fulltextStatus":22,"lidarModels":913,"equipmentCount":24},"droeschel2018ctslam","Droeschel & Behnke, 2018","MRS continuous-time surfel SLAM (Droeschel and Behnke)","Efficient Continuous-Time SLAM for 3D Lidar-Based Online Mapping","這個方法延續作者的局部多解析度網格地圖：每個 3D 掃描以面元（surfel）配準到以機器人為中心的局部地圖，多個局部地圖再以面元配準連成全域位姿圖。新意在於把每個局部地圖內的掃描位姿建成子圖，形成階層式圖：當地圖累積更多資訊後，可挑選配準不確定性最大的掃描重新對齊並最佳化子圖，再更新上層位姿圖；子圖內以 SE(3) 三次 B 樣條表示連續時間軌跡，內插每條掃描線的位姿以修正掃描期間的運動畸變。作者以平均地圖熵量化點雲清晰度。",[905,906],"3D rotating multi-beam LiDAR (Velodyne VLP-16; two VLP-16 on the Deutsches Museum backpack)","optional IMU or wheel odometry used as registration prior and for motion during acquisition (Sec. III); the MAV carried an IMU measuring attitude (Sec. V-A)","robot-centric local multiresolution grid maps that store measurements, occupancy (ray casting with an approximated 3D Bresenham) and surfels per cell; allocentric graph of local maps with a new map node every 5 m (Sec. III; Sec. V)","after each new local map one candidate map node is drawn with a probability that decays with distance and registered by surfel map-to-map alignment; on revisits, scans from neighbouring map nodes enlarge the local window (Sec. IV-B; Sec. IV-D)","hierarchical graph optimization in g2o: an allocentric pose graph of local multiresolution maps, a sub-graph of 3D scan poses per local map, and a continuous-time cubic B-spline over scan poses; sub-graphs are refined in parallel and global optimization is triggered by loop closures or when a sub-graph reference pose changes by more than 0.01 m or 1 degree (Sec. IV)","surfel-based registration of each 3D scan to a robot-centric local multiresolution grid map (surfel = sample mean and covariance of the points in a cell), and surfel-based map-to-map registration between local maps; scans with the largest entropy of registration covariance are selected for realignment (Sec. III; Sec. IV-A)","motion during acquisition compensated with IMU or wheel odometry when available, then scan-line poses refined by B-spline interpolation within each sub-graph (Sec. III; Sec. IV-C; Fig. 4)","refined 3D point cloud map and trajectory (Figs. 5 and 7)",[143],{"id":915,"label":916,"shortName":917,"title":918,"year":661,"era":10,"cluster":599,"scope":32,"keyIdeaZh":919,"sensors":920,"mapRepresentation":922,"loopClosure":57,"estimator":57,"association":923,"deskew":121,"outputGeometry":924,"fulltextStatus":22,"lidarModels":925,"equipmentCount":554},"dufomap2024","Duberg et al., 2024","DUFOMap","DUFOMap: Efficient Dynamic Awareness Mapping","DUFOMap 不直接偵測動態物，而是辨識「曾被完整觀測為空」的空洞區域（void region）：以射線投射判斷體素是否被完整看空，一旦成立，其他時刻落在其中的點即為動態點。方法以 UFOMap 八元樹實作，並加入考慮量測雜訊與位姿誤差的保守判定，所有情境使用同一組參數，可線上或後處理執行，也能處理非連續的測量級掃描站資料。",[35,921],"terrestrial laser scanner (survey data, qualitative)","UFOMap octree voxels with a void flag","ray casting to classify voxels observed completely empty at least once (void regions); points inside void regions at other times are dynamic","static map and dynamic point labels; online or post-processing",[161,926,143,45,507],"Velodyne VLP-32C",{"id":928,"label":929,"shortName":930,"title":931,"year":396,"era":10,"cluster":151,"scope":132,"keyIdeaZh":932,"sensors":933,"mapRepresentation":936,"loopClosure":937,"estimator":938,"association":939,"deskew":940,"outputGeometry":941,"fulltextStatus":22,"lidarModels":942,"equipmentCount":487},"segmap2020","Dubé et al., 2020","SegMap","SegMap: Segment-based mapping and localization using data-driven descriptors","SegMap 把 LiDAR 點雲切成可重複擷取的片段（segment），每個片段以 CNN 壓縮成 64 維描述子，再以描述子的最近鄰檢索加上片段質心的幾何一致性檢查，得到相對於地圖的六自由度定位。這些定位結果作為迴圈約束，與 ICP 或 LOAM 里程計一起放入 iSAM2 位姿圖，也能在多機器人之間合併地圖。描述子訓練同時兼顧檢索與重建，因此可由描述子重建近似點雲或網格，並分辨車輛、建物與其他類別，以剔除可能移動的物體。",[934,935],"3D LiDAR (KITTI odometry; sensor model not named in the paper)","rotating 2D SICK LMS-151 LiDAR on UGVs that also carried motor encoders and an Xsens MTI-G IMU (search and rescue experiments)","target map of segment centroids with 64-D descriptors (only the last and most complete observation kept); local map radius 50 m; a decoder reconstructs voxel point clouds or marching-cubes meshes from the descriptors (Sec. 3; Sec. 4.3; Sec. 5.5; Sec. 5.8)","segment descriptor retrieval with centroid geometric verification, used for loop closures and for multi-robot global associations; semantic filtering can reject vehicle segments (Sec. 3; Sec. 5.8; Sec. 5.9.1)","incremental pose-graph SLAM (iSAM2) that combines LiDAR odometry (ICP-based, or LOAM loosely coupled) with 6-DoF segment-based localization constraints; multi-robot variant runs one centralized pose graph (Sec. 3; Sec. 5.1; Sec. 5.8; Sec. 5.9)","segments grown incrementally in a dynamic voxel grid (Euclidean clustering after ground removal, or smoothness-based planar growing); 64-D CNN descriptor per segment; k-NN retrieval in descriptor space (64 neighbours in the LOAM experiment) and geometric consistency of segment centroids with at least 7 correspondences (Sec. 3; Sec. 4.1; Sec. 5.8)","in the LOAM plus SegMap system LOAM undistorts the scans (Sec. 5.8); otherwise not described","robot trajectories, a compact segment map, and point clouds or meshes reconstructed from descriptors (Figs. 5, 9, 12, 14)",[943],"SICK LMS-151 (rotating 2D)",{"id":945,"label":946,"shortName":947,"title":948,"year":249,"era":10,"cluster":397,"scope":132,"keyIdeaZh":949,"sensors":950,"mapRepresentation":953,"loopClosure":954,"estimator":955,"association":956,"deskew":121,"outputGeometry":957,"fulltextStatus":22,"lidarModels":958,"equipmentCount":487},"ebadi2021dareslam","Ebadi et al., 2021","DARE-SLAM","DARE-SLAM: Degeneracy-Aware and Resilient Loop Closing in Perceptually-Degraded Environments","DARE-SLAM 先以 ICP 解的特徵分析估計環境的幾何退化程度，並把模糊、不可觀測的區域排除在迴圈閉合搜尋之外，以免錯誤迴圈扭曲整張地圖。再以 LiDAR 點雲的 2D 與 3D 顯著特徵進行對漂移較不敏感的迴圈閉合。作者在礦坑與辦公室資料上評估。",[951,952],"3D LiDAR (Velodyne VLP-16 Puck Lite)","RGB-D camera (RealSense D435, object detection only)","reduced pose graph whose key-nodes carry key-scans; global 3D point-cloud map formed by projecting key-scans into the world frame; per-key-scan 2D occupancy grids used only for place recognition","degeneracy-aware candidate gating plus pose-invariant multi-stage loop closing (SGLC): global pre-matching of 250 x 250-cell (5 m x 5 m) occupancy-grid images over the whole trajectory (maps with 20 or fewer inliers dropped), ICP geometric verification, and PCM pairwise-consistency outlier rejection; same pipeline for inter-robot loops","LOAM-style edge and planar feature extraction (up to 90% point decimation), then two-stage GICP (scan-to-scan, then scan-to-submap) odometry; reduced pose graph with a key-node every 1 m translation or 30 deg rotation, back-end in GTSAM optimized by iterative nonlinear least squares (Levenberg-Marquardt named only as an example); local graphs per robot merged at a base station","GICP correspondences with approximate nearest-neighbour search against a local submap; degeneracy from the condition number of the ICP-derived approximate Hessian (log kappa threshold); loop-closure pre-matching with ORB features on binary occupancy-grid images, FLANN matching and RANSAC homography, scored by correspondence confidence times transformation confidence; ICP geometric verification seeded with the homography yaw","3D point-cloud maps of mines (single-robot and merged multi-robot) and optimized robot trajectories",[143],{"id":960,"label":961,"shortName":962,"title":963,"year":9,"era":10,"cluster":53,"scope":54,"keyIdeaZh":964,"sensors":965,"mapRepresentation":967,"loopClosure":968,"estimator":969,"association":970,"deskew":57,"outputGeometry":971,"fulltextStatus":22,"lidarModels":972,"equipmentCount":46},"eckenhoff2019closedform","Eckenhoff et al., 2019","Closed-form preintegration","Closed-form preintegration methods for graph-based visual-inertial navigation","本文推導 IMU 預積分方程的閉式解，而非以離散取樣近似量測動態，並提出兩種慣性模型：分段常數量測，以及分段常數的局部真實加速度。作者以 Monte Carlo 模擬分析模型選擇對估計的影響，並把此預積分分別用於緊耦合滑動視窗最佳化與鬆耦合直接影像對齊兩種視覺慣性系統。",[36,966],"stereo camera (both real-world systems)","no persistent map: sparse features in inverse-depth parameterization inside the sliding window (features marginalized after 3 s in simulation); the direct VINS keeps keyframes with stereo depth maps for alignment (Sec. V-A; Sec. VI; Sec. VII-B)","direct VINS keeps all states (no marginalization) to allow loop closures and outperforms the indirect VIO on loop-rich EuRoC sequences (Sec. VII-B)","indirect system: tightly coupled sliding-window optimization with marginalization using closed-form preintegration; direct system: loosely coupled direct stereo image alignment whose relative-pose factors and preintegration factors are optimized with iSAM2 (GTSAM) without marginalization","indirect VIO: FAST features on a uniform grid tracked with KLT, stereo correspondences by KLT from left to right image, 8-point RANSAC outlier rejection and Cauchy loss (Sec. VII-A); direct VINS: photometric alignment of high-gradient pixels against keyframes with a Huber cost, keyframe depth from OpenCV StereoSGBM (Sec. V-B; Sec. VII-B)","IMU trajectory (orientation, position, velocity, biases); IMU-camera extrinsics and camera intrinsics estimated online in the UD experiments; no dense map or point cloud (Sec. VII-A2)",[],{"id":974,"label":975,"shortName":976,"title":977,"year":191,"era":52,"cluster":527,"scope":132,"keyIdeaZh":978,"sensors":979,"mapRepresentation":982,"loopClosure":983,"estimator":984,"association":985,"deskew":121,"outputGeometry":986,"fulltextStatus":22,"lidarModels":987,"equipmentCount":144},"dpslam2003","Eliazar & Parr, 2003","DP-SLAM","DP-SLAM: Fast, Robust Simultaneous Localization and Mapping Without Predetermined Landmarks","DP-SLAM 以粒子濾波同時追蹤機器人位姿與地圖假設，不需預先指定地標。為避免每個粒子複製整張地圖，作者提出分散式粒子建圖（DP-mapping）：全系統只保留一張佔據網格，每格以平衡樹記錄曾更新該格的粒子 ID，並維護經修剪與合併的最小粒子祖先樹；粒子查詢某格時，沿祖先找出最近一次的觀測紀錄。如此可同時維持數千張候選地圖，並在沒有明確迴圈閉合步驟下閉合約 60 m 的迴圈。",[980,981],"2D laser range finder (SICK)","wheel odometry (shaft encoders)","single binary occupancy grid (3 cm cells) in which each cell stores a balanced tree keyed by the IDs of particles that updated it, plus a pruned and collapsed minimal particle ancestry tree (Sec. 3.2; Sec. 4)","implicit through maintaining multiple map hypotheses; no explicit loop-closing step or environment assumption (Sec. 4)","particle filter over robot poses and maps with a calibrated odometry motion model; particle culling evaluates the posterior in k passes over disjoint subsets of laser readings and drops poor particles early (Sec. 2, 4)","ray tracing each laser cast through the particle's map to the first obstruction; Gaussian discrepancy with 5 cm standard deviation (Sec. 2.1)","2D occupancy grid (cross-section at the 7 cm laser height)",[988],"SICK laser range finder",{"id":990,"label":991,"shortName":992,"title":993,"year":115,"era":52,"cluster":236,"scope":132,"keyIdeaZh":994,"sensors":995,"mapRepresentation":997,"loopClosure":998,"estimator":999,"association":1000,"deskew":57,"outputGeometry":1001,"fulltextStatus":22,"lidarModels":1002,"equipmentCount":109},"rgbdslamv2_2014","Endres et al., 2014","RGBDSLAMv2","3-D Mapping With an RGB-D Camera","RGBDSLAMv2 只用 RGB-D 相機建立三維地圖。前端從彩色影像擷取 SIFT、SURF 或 ORB 特徵，以深度影像取得三維位置，再用 RANSAC 估計影格間的剛體轉換；候選影格包含前幾張影格、位姿圖上測地鄰域的抽樣以及關鍵影格，用來尋找迴圈閉合。作者提出以光束模型檢查深度影像間的自由空間衝突（EMM），剔除不可信的轉換，後端以 g2o 最佳化位姿圖並刪除誤差過大的邊。最後依軌跡把量測投影成點雲，或以 OctoMap 產生佔據體素地圖。",[996],"RGB-D camera (structured light: Microsoft Kinect, Asus Xtion Pro Live)","globally registered point cloud (optionally surfels) or OctoMap 3-D occupancy grid; paper recommends OctoMap for memory and free-space representation (Sec. III-F)","candidate frames from n immediate predecessors, k frames sampled from the geodesic neighbourhood in the pose graph and l frames sampled from keyframes; validated by RANSAC and the EMM (Sec. III-D)","pose-graph SLAM: pairwise 6-DOF transforms from 3-D feature correspondences via RANSAC with least-squares motion estimation, optional two-frame g2o refinement; global g2o optimisation (CSparse offline, PCG suggested online) with pruning of edges whose error remains high after convergence (Secs. III-B, III-E, IV-C)","sparse visual keypoints (SIFT on GPU, SURF, ORB or Shi-Tomasi plus SURF) matched by nearest to second-nearest ratio (Euclidean, Hellinger or Hamming distance), RANSAC with Mahalanobis inlier test; transforms validated by a beam-based environment measurement model (EMM) on subsampled depth images (Secs. III-B, III-C, IV-D)","optimised camera trajectory plus point cloud created by projecting the original depth measurements, or a textured OctoMap voxel occupancy map (2 cm maps of 4.2 to 25 MB versus 2 to 5 GB for unfiltered point clouds) (Sec. III-F)",[],{"id":1004,"label":1005,"shortName":1006,"title":1007,"year":115,"era":52,"cluster":236,"scope":132,"keyIdeaZh":1008,"sensors":1009,"mapRepresentation":1010,"loopClosure":1011,"estimator":1012,"association":1013,"deskew":57,"outputGeometry":1014,"fulltextStatus":22,"lidarModels":1015,"equipmentCount":185},"lsdslam2014","Engel et al., 2014","LSD-SLAM","LSD-SLAM: Large-Scale Direct Monocular SLAM","LSD-SLAM 為直接法（direct method）單眼 SLAM，不萃取特徵點，而是對影像梯度明顯的像素做光度誤差對齊，並以許多小基線立體比對濾波估計關鍵影格的半稠密（semi-dense）深度圖。新關鍵影格以 sim(3) 直接對齊連接鄰近關鍵影格，以明確偵測尺度漂移，並在位姿圖上做全域最佳化。地圖可輸出為半稠密點雲，但尺度仍非公制。",[37],"Pose graph of keyframes; each keyframe stores the image, a semi-dense inverse depth map and its variance defined only near sufficiently large intensity gradients, scaled to mean inverse depth one; edges hold sim(3) transforms with covariance; the map can be exported as point clouds","candidates from the ten closest keyframes plus an appearance-based candidate (OpenFABMAP, ref. [11]), accepted after a reciprocal sim(3) tracking consistency check (Sec. 3.5 Constraint Acquisition)","weighted Gauss-Newton direct image alignment on se(3)\u002Fsim(3); keyframe pose-graph optimisation with Sim(3) edges (Sec. 2.2, 3.1)","direct photometric alignment on high-gradient (semi-dense) pixels; depth by filtering many small-baseline stereo comparisons (abstract; Sec. 3.1)","semi-dense point cloud from keyframe depth maps and poses (conclusion); monocular, scale not metric",[],{"id":1017,"label":1018,"shortName":1019,"title":1020,"year":150,"era":52,"cluster":236,"scope":99,"keyIdeaZh":1021,"sensors":1022,"mapRepresentation":1023,"loopClosure":1024,"estimator":1025,"association":1026,"deskew":1027,"outputGeometry":1028,"fulltextStatus":22,"lidarModels":1029,"equipmentCount":46},"dso2018","Engel et al., 2018","DSO","Direct Sparse Odometry","DSO 是直接稀疏法的單眼視覺里程計（visual odometry），直接最小化光度誤差，並在滑動視窗內聯合最佳化相機位姿、相機內參、仿射亮度參數與逆深度，舊狀態以邊際化（marginalization）移除。它不使用平滑先驗，而是在影像中均勻取樣具梯度的像素，包括白牆上的弱梯度與邊緣，並整合曝光、暗角與非線性響應的光度校正。DSO 不含迴圈閉合，屬里程計而非完整 SLAM。",[37],"sparse set of points with inverse depth in active keyframes","none (visual odometry; explicit loop closure disabled for ORB-SLAM in comparisons for fairness, Sec. 4)","sliding-window Gauss-Newton (up to 6 iterations per new keyframe, no Levenberg-Marquardt damping) jointly over poses, affine brightness parameters, inverse depths and camera intrinsics, with First Estimate Jacobians and Schur-complement marginalisation; window Nf = 7 keyframes and Np = 2000 active points; keyframes marginalised by a distance score and residuals that would break Hessian sparsity are dropped (about half of all residuals)","direct photometric error of an 8-pixel residual pattern with Huber norm and gradient-dependent weighting, on points sampled with a region-adaptive gradient threshold (32x32 blocks, three passes with lower thresholds); inverse depth in a host frame; candidates tracked by discrete epipolar search before activation; new frames tracked by two-frame direct alignment to the newest keyframe's projected semi-dense depth map with a constant motion model, with up to 27 small-rotation retries on failure","not_applicable (rolling shutter not modelled; simulated as low-frequency geometric noise in Sec. 4.3, where DSO degrades much faster than ORB-SLAM and, according to the authors, optimisation likely fails entirely for noise amplitude above 1.5 (alleviable with a coarser pyramid level); tight rolling-shutter modelling cited as remedy)","point clouds accumulated from the odometry without loop closure, density set by the number of active points (Np = 500 to 10000 shown); monocular scale unobservable (scale is a null space of the energy) and evaluated with Sim(3) alignment and scale drift",[],{"id":1031,"label":1032,"shortName":1033,"title":1034,"year":97,"era":10,"cluster":397,"scope":1035,"keyIdeaZh":1036,"sensors":1037,"mapRepresentation":57,"loopClosure":57,"estimator":1040,"association":57,"deskew":57,"outputGeometry":1041,"fulltextStatus":22,"lidarModels":1042,"equipmentCount":24},"faizullin2022lidarsync","Faizullin et al., 2022","LiDAR sync by GNSS-clock emulation","Open-Source LiDAR Time Synchronization System by Mimicking GNSS-clock","sensing_calibration_sync_preprocessing","本系統以 STM32F4 微控制器模擬 GNSS 時鐘（PPS 與 NMEA GPRMC 訊息）輸入 VLP-16 的硬體同步介面，無需實體 GNSS 接收器，並以中斷為 IMU（MPU-9150）資料打時間戳記。作者以約十分鐘、逾五十萬個 UDP 封包的時間戳週期標準差評估精密度：ROS 抵達時間戳 82.64 µs、LiDAR 內部時鐘 0.31 µs（但無法與其他感測器同步）、本系統 1.35 µs；約 10 µs 的同步準確度則是依感測器文件推估，作者明言實測準確度評估不在本文範圍。",[1038,1039],"3D LiDAR (VLP-16)","IMU (MPU-9150)","not_applicable (hardware timestamping)","synchronized LiDAR and IMU timestamps",[143],{"id":1044,"label":1045,"shortName":1046,"title":1047,"year":30,"era":10,"cluster":11,"scope":132,"keyIdeaZh":1048,"sensors":1049,"mapRepresentation":1053,"loopClosure":1054,"estimator":1055,"association":1056,"deskew":121,"outputGeometry":1057,"fulltextStatus":22,"lidarModels":1058,"equipmentCount":24},"feng2026integratedslam","Feng et al., 2026","Integrated LiDAR SLAM for public-building sites","Integration and evaluation of a 3D LiDAR SLAM system for construction robots in large-scale public building sites","作者不提出新演算法，而是整合並依工地條件調整既有模組：兩階段地面分割（RANSAC 粗分割加法向量一致性精分割）取出樓板地面，快速歐幾里得分群（FEC）處理非地面點以抑制工人與機具等動態物，兩步配準以地面平面特徵估計 z、roll、pitch，再以非地面邊緣特徵估計 x、y、yaw，構成只用光達的里程計，最後以 Scan Context++ 迴圈偵測與位姿圖最佳化修正漂移。平台為搭載 RS-Helios-16P 16 線光達與 NVIDIA Jetson Xavier NX 的阿克曼轉向輪式機器人。以西安某醫院門診大樓的 Gazebo 模擬（1293 m）與同一棟施工中大樓的實地資料（1004 m）比較 F-LOAM 與 LeGO-LOAM，本方法 ATE RMSE 分別為 5.40 m 與 2.87 m，以圖面尺寸評估的實地地圖平均誤差為 0.79%；實地軌跡真值來源未說明。",[1050,1051,1052],"3D LiDAR RS-Helios-16P (16 beams, 10 Hz, +-15 deg vertical FOV, +-2 cm ranging)","nine-axis IMU at 200 Hz (recorded; the evaluated system is LiDAR-only)","camera (recorded; model not reported; not used by the method)","global point cloud map built by scan-to-map feature alignment, saved as PCD (Sec. 3.1, Sec. 5.2)","Scan Context++ descriptor on ground-removed clustered points, run in a parallel thread (Sec. 3.5, Algorithm 1)","LiDAR-only feature-based odometry with two-step registration following LeGO-LOAM (ground planar features estimate z, roll and pitch; edge features from FEC-clustered non-ground points estimate x, y and yaw), scan-to-map mapping, and pose-graph optimization with Scan Context++ loop constraints (Sec. 3.1, Algorithm 1)","plane features from two-stage-segmented ground points (RANSAC fit on two low beams, 6 deg coarse threshold, 3 deg normal-consistency check) and edge features from FEC clusters of non-ground points (dth = 0.20 m, 30 to 50 point minimum cluster); scan-to-map feature alignment (Sec. 3.1, 3.3, 3.4)","point cloud map (PCD) and TUM-format trajectory (Sec. 5.2)",[1059],"RS-Helios-16P",{"id":1061,"label":1062,"shortName":1063,"title":1064,"year":661,"era":10,"cluster":98,"scope":99,"keyIdeaZh":1065,"sensors":1066,"mapRepresentation":1069,"loopClosure":72,"estimator":1070,"association":1071,"deskew":1072,"outputGeometry":1073,"fulltextStatus":22,"lidarModels":1074,"equipmentCount":487},"madicp2024","Ferrari et al., 2024","MAD-ICP","MAD-ICP: It is All About Matching Data - Robust and Informed LiDAR Odometry","MAD-ICP 將每次掃描建成以主成分分析（PCA）切分的 kd 樹，葉節點帶有平均位置與法向量，並以點到平面 ICP 對齊關鍵影格 kd 樹組成的局部地圖。局部地圖只在匹配比例低於門檻時更新，且依位姿共變異數的行列式（D-optimality）挑選不確定性最小的幀加入，以避免把雜訊持續注入地圖。作者主張在不同感測器與運動型態下不失敗，並以同一型號 LiDAR 使用同一組參數。",[1067,1068],"3D LiDAR (all quantitative evaluation)","RGB-D point clouds shown qualitatively in the supplementary material (arXiv Fig. 10: ETH3D SLAM dataset sequences einstein_1 and sofa_1, dataset cited via the BAD SLAM paper [32])","forest of keyframe kd-trees; updated only when fewer than 80% of leaves match, adding the candidate frame with minimum covariance determinant (D-optimality)","Gauss-Newton point-to-plane ICP with Huber kernel; pose covariance from the Gauss-Newton system matrix","PCA-based kd-tree leaves (mean, normal) of the scan matched to a forest of keyframe kd-trees","classic constant motion model for deskewing each scan (Sec. III-D); the velocity smoothed over the last n = 10 poses serves the ICP initial guess, not the deskewing (Sec. III-D)","odometry and keyframe point clouds; export format not_reported",[161,223,222,224],{"id":1076,"label":1077,"shortName":1078,"title":1079,"year":9,"era":10,"cluster":599,"scope":69,"keyIdeaZh":1080,"sensors":1081,"mapRepresentation":1083,"loopClosure":121,"estimator":1084,"association":1085,"deskew":1086,"outputGeometry":1087,"fulltextStatus":22,"lidarModels":1088,"equipmentCount":61},"eigenfactors2019","Ferrer, 2019","Eigen-Factors (EF)","Eigen-Factors: Plane Estimation for Multi-Frame and Time-Continuous Point Cloud Alignment","Eigen-Factors 將每個平面由多個位姿觀測到的點累積為 4×4 齊次點矩陣，平面擬合誤差等於該矩陣的最小特徵值；平面參數不必列為狀態變數，因此複雜度與點數無關，只取決於平面數與位姿數。作者以李代數推導最小特徵值對各位姿的封閉形式梯度，並以簡化的 Nesterov 加速梯度法最佳化軌跡，另提出在 SE(3) 上內插的連續時間軌跡版本。實驗只使用合成平面點雲，且只評估連續時間版本，因為離散多位姿版本會過度擬合。此方法後來成為 BALM2 比較實驗中的平面式多影格配準基準之一。",[1082],"generic 3D point clouds (lidar and RGB-D motivated in Sec. I); evaluated only on synthetic planar point clouds","non-parametric plane landmarks: each plane kept only as per-pose 4x4 matrices S_t of homogeneous points, so raw points and plane parameters need not be stored (Sec. IV-B)","first-order optimization: closed-form gradient of the minimum eigenvalue of each plane's accumulated 4x4 homogeneous point matrix Q with respect to SE(3) poses (Lie algebra, left-hand perturbation), minimized with a simplified Nesterov Accelerated Gradient momentum method (alpha = 0.2\u002F(N_all H), beta = 0.7), compared with plain gradient descent (Sec. IV-C, IV-D, V)","requires a preprocessing segmentation of planes (Sec. I); (inference) the synthetic evaluation samples points per plane, so plane membership appears to be known rather than estimated (Sec. V)","not_reported (synthetic point clouds; no in-scan motion compensation described)","optimized trajectory (time-continuous interpolation) with implicitly estimated planes; no exported map product (Sec. IV-B, V)",[],{"id":1090,"label":1091,"shortName":1092,"title":1093,"year":1094,"era":52,"cluster":68,"scope":69,"keyIdeaZh":1095,"sensors":1096,"mapRepresentation":57,"loopClosure":72,"estimator":1099,"association":1100,"deskew":57,"outputGeometry":1101,"fulltextStatus":22,"lidarModels":1102,"equipmentCount":77},"fischler1981ransac","Fischler & Bolles, 1981","RANSAC","Random sample consensus",1981,"隨機取樣一致（RANSAC）以最少數量的資料點實例化模型，再收集誤差容許範圍內的一致集合；若一致集合大小達門檻 t，就在該集合上以最小平方法重新估計，否則重新抽樣，試驗次數用盡時採用最大一致集合或宣告失敗（Sec. II）。論文推導試驗次數期望值 E(k) = w^(-n) 與達成信心 z 所需的 k。主要應用為由已知位置地標的影像求相機投影中心的定位問題（LDP），給出 P3P 最多四解的封閉形式解，並證明共面 P4P 與一般位置 P6P 有唯一解；原文不涉及點雲配準。點雲配準中的用法依 FGR 論文 Sec. 1 的描述，TEASER 論文 Sec. I 註腳指出其執行時間隨離群比例呈指數成長。",[1097,1098],"aerial photograph from about 4,000 ft with a 6 in. lens, digitized on a 2,000 x 2,000 grid (about 2 ft per pixel)","synthetic landmark-to-image correspondences","hypothesize-and-verify: randomly select a minimal subset of n data points to instantiate the model, collect the consensus set within an error tolerance, and if its size reaches threshold t refit the model (e.g., least squares) on the consensus set; otherwise resample; after k trials use the largest consensus set or fail; expected trials E(k) = w^(-n), and k = log(1-z)\u002Flog(1-w^n) for confidence z","original: landmark-to-image correspondences from error-prone feature detectors (cross correlation in the aerial test); consensus judged with image-plane error ellipses derived from perturbing the three selected points; point cloud registration use with putative 3D correspondences is a later application (secondary)","in the original LDP application: 3-D location of the centre of perspective with an error estimate, and the spatial orientation of the image plane",[],{"id":1104,"label":1105,"shortName":1106,"title":1107,"year":769,"era":52,"cluster":53,"scope":99,"keyIdeaZh":1108,"sensors":1109,"mapRepresentation":1112,"loopClosure":1113,"estimator":1114,"association":1115,"deskew":57,"outputGeometry":1116,"fulltextStatus":22,"lidarModels":1117,"equipmentCount":24},"forster2017preint","Forster et al., 2017a","On-manifold IMU preintegration","On-Manifold Preintegration for Real-Time Visual-Inertial Odometry","本文把兩個關鍵影格之間的大量 IMU 量測預先積分成單一相對運動約束，並提出正確處理旋轉群 SO(3) 流形結構的預積分理論，推導旋轉雜訊的性質、MAP 估計式，以及殘差、雜訊傳播與偏差事後修正的解析 Jacobian。作者將預積分 IMU 因子放入因子圖，以 iSAM2 做增量平滑，並用 structureless 視覺模型避免將三維點納入最佳化。論文亦指出其理論延伸自 Lupton 與 Sukkarieh 的原始預積分概念。",[1110,1111],"monocular camera (left camera of a forward-looking VI-Sensor, 20 Hz)","IMU (ADIS16448 MEMS inside the VI-Sensor, 800 Hz)","3D landmarks eliminated by structureless model (no dense map)","none (VIO); past states are not marginalized, so loop closures could be added (Sec. VIII-B1)","factor-graph MAP estimation solved by iSAM2 incremental smoothing (full smoothing), preintegrated IMU factors on SO(3) with a-posteriori bias correction, structureless vision factors","sparse visual features tracked by the SVO front-end (sparse image alignment)","keyframe trajectory, velocities and IMU biases",[],{"id":1119,"label":1120,"shortName":1121,"title":1122,"year":769,"era":10,"cluster":236,"scope":99,"keyIdeaZh":1123,"sensors":1124,"mapRepresentation":1127,"loopClosure":1128,"estimator":1129,"association":1130,"deskew":57,"outputGeometry":1131,"fulltextStatus":22,"lidarModels":1132,"equipmentCount":736},"svo2017","Forster et al., 2017b","SVO","SVO: Semidirect Visual Odometry for Monocular and Multicamera Systems","SVO 採半直接法（semi-direct）：以直接法追蹤並三角化影像梯度高的像素（含弱角點與邊緣），再以成熟的特徵式方法聯合最佳化結構與運動，並用顯式建模離群值的機率深度濾波器估計深度。期刊版將方法擴充到多相機、邊緣特徵、運動先驗及魚眼等大視角鏡頭。其定位為速度優先的視覺里程計，沒有迴圈閉合。",[37,1125,1126],"multi-camera","fisheye","sparse 3D points and edgelets with depth filters; only a small local map of the last five to ten keyframes is kept (Sec. XI-B-1)","none (visual odometry)","three-step motion estimation: coarse-to-fine sparse image alignment on 4x4 patches with a robust cost for frame-to-frame motion, 2D alignment of 8x8 feature patches against the frame of first observation, then reprojection-error refinement of the latest pose and points, or full keyframe bundle adjustment with iSAM2 (Sec. V, X-B, X-C, XI-B)","semi-direct: direct alignment of high-gradient pixels (FAST corners, or the highest-gradient pixel as an edgelet in cells of 32x32 pixels without corners) plus feature alignment and refinement; depth by epipolar ZMSSD search on 8x8 patches feeding a Gaussian plus uniform depth filter with Beta inlier ratio; at most 180 matched features per frame (Sec. VI, X-C, X-D)","camera trajectory and sparse points",[],{"id":1134,"label":1135,"shortName":1136,"title":1137,"year":97,"era":10,"cluster":151,"scope":132,"keyIdeaZh":1138,"sensors":1139,"mapRepresentation":1143,"loopClosure":1144,"estimator":1145,"association":1146,"deskew":1147,"outputGeometry":1148,"fulltextStatus":22,"lidarModels":1149,"equipmentCount":24},"artslam2022","Frosi & Matteucci, 2022","ART-SLAM","ART-SLAM: Accurate Real-Time 6DoF LiDAR SLAM","ART-SLAM 是模組化的 LiDAR 圖式 SLAM，架構參考 hdl_graph_slam：點雲先降採樣並以八分區平行去除離群點，追蹤模組以完整點雲對最近關鍵影格配準（可選 ICP、GICP、VGICP 或 NDT），並可由多尺度預追蹤或外部里程計提供初值；地面偵測模組估計地面平面，加入高度與姿態約束。迴圈偵測分三步：先依累積距離與位置篩選候選，再以 Scan Context 保留最相似的少數候選，最後逐一配準取最佳結果，所有約束以 g2o 位姿圖最佳化。IMU 與 GPS 為選用輸入，IMU 可用於去除運動畸變。",[1140,1141,1142],"3D LiDAR point clouds (mandatory)","optional IMU for de-skewing in the pre-filterer and orientation constraints in the pose graph","optional GPS constraints and pre-computed odometry (Sec. II-A)","keyframes storing point clouds, poses, timestamps and accumulated distance; 3D map assembled from keyframe clouds (Sec. II-C; Figs. 5-7)","three steps: odometry-based candidate selection (far in accumulated distance, near in estimated position), Scan Context polar grids with a KD-tree to keep k candidates, then scan-to-scan matching and the best match added to the pose graph (Sec. II-F)","keyframe-based scan-to-keyframe registration of full filtered clouds with a user-selected method (ICP, GICP, VGICP or NDT), optionally seeded by a multi-scale pre-tracker or external odometry; g2o pose graph with odometry, floor-plane, loop and optional IMU and GPS constraints (Sec. II-C; Sec. II-D; Sec. II-G)","no feature extraction: downsampled and outlier-filtered full point clouds (octant-parallel filtering) are registered directly; floor plane by RANSAC on near-vertical-normal points or least-squares fitting on rough terrain (Sec. II-B; Sec. II-E)","optional IMU-based de-skewing in the pre-filterer (ART-SLAM IMU variant); not described otherwise (Sec. II-A; Sec. III)","trajectory and 3D point cloud map (Figs. 5-7)",[45],{"id":1151,"label":1152,"shortName":1153,"title":1154,"year":383,"era":52,"cluster":397,"scope":1035,"keyIdeaZh":1155,"sensors":1156,"mapRepresentation":57,"loopClosure":57,"estimator":1159,"association":1160,"deskew":57,"outputGeometry":1161,"fulltextStatus":22,"lidarModels":1162,"equipmentCount":24},"furgale2013unifiedcalib","Furgale et al., 2013","Kalibr (unified temporal-spatial calibration)","Unified temporal and spatial calibration for multi-sensor systems","本文以連續時間 B-spline 表示 IMU 位姿與偏差，把相機與 IMU 之間的固定時間偏移 d 直接寫入影像量測模型，與外參、重力方向及 IMU 偏差一起以 Levenberg-Marquardt 做最大概似批次估計，取代先估時間、再估空間的兩階段作法。以 FPGA 打時間戳的自製視覺慣性感測器在棋盤格前揮動，四種曝光時間各十組資料的時間偏移對曝光時間斜率為 0.498（理論值 0.5），與擬合線的差異都在 ±0.2 ms 內，約為 IMU 5 ms 取樣週期的 4%。只用陀螺儀、只用加速度計或分離估計的 RMS 誤差分別為 0.165、0.572、0.344 ms，本法為 0.054 ms。",[1157,1158],"global-shutter cameras (Aptina MT9V034 image sensors in a custom visual-inertial sensor)","IMU (Analog Devices ADIS16488, tactical grade)","continuous-time batch maximum-likelihood estimation: IMU pose as a sixth-order B-spline and biases as cubic B-splines (50 basis functions per second), jointly estimating gravity direction, camera-IMU transform, time offset, pose and biases with Levenberg-Marquardt and the CHOLMOD sparse solver","checkerboard corner detections with known correspondence to calibration-pattern points (assumed known)","time offset and extrinsic transform",[],{"id":1164,"label":1165,"shortName":1166,"title":1167,"year":235,"era":52,"cluster":53,"scope":54,"keyIdeaZh":1168,"sensors":1169,"mapRepresentation":1174,"loopClosure":1175,"estimator":1176,"association":1177,"deskew":1178,"outputGeometry":1179,"fulltextStatus":22,"lidarModels":1180,"equipmentCount":109},"furgale2015ct","Furgale et al., 2015","Temporal basis functions (continuous-time batch)","Continuous-time batch trajectory estimation using temporal basis functions","作者指出離散時間估計在 IMU、捲簾快門相機或掃描式雷射等高頻感測器下，需為每個量測時間加入位姿變數，使狀態維度過大。本文把完整的 MAP 估計移到連續時間，在高斯假設下推導目標函數，並以少量時間基底函數的係數作為待估狀態，以批次 Gauss-Newton 法求解，再由系統矩陣的逆取得共變異數。論文以三次 B 樣條的矩陣形式實作，旋轉以 Cayley-Gibbs-Rodrigues 參數表示，程式碼已併入 Kalibr 校正工具箱。實驗包含相機與 IMU 外參校正（模擬 1000 次，以及以 Vicon 為參考的實測資料）與捲簾快門相機對圓點圖板的定位，作者並指出節點數量與配置仍是未解問題。",[1170,1171,1172,1173],"stereo camera (Point Grey Research Bumblebee XB3, 24 cm baseline) for camera-IMU calibration","IMU (MicroStrain 3DM-GX2)","rolling-shutter camera (Matrix Vision BlueCougar-X102d)","global-shutter camera (Matrix Vision BlueCougar-X012b) used for comparison","sparse point landmarks estimated in calibration (priors on three landmarks); known dot-grid positions in rolling-shutter localization; no dense map","none (batch calibration and localization experiments)","batch MAP estimation in continuous time under Gaussian assumptions; state represented by coefficients of cubic B-spline basis functions (Cayley-Gibbs-Rodrigues rotation parameters) and solved by Gauss-Newton; covariance recovered from the inverse of the Gauss-Newton system matrix; white-noise Gaussian-process motion prior acts as regularizer (Secs. 3-6)","known correspondences: point landmarks observed by the camera, with initial camera poses from OpenCV solvePnP (Sec. 6.2); known circle-grid positions on the pattern board (Sec. 7.1); no data-association search is described","no LiDAR deskew; rolling-shutter distortion handled by evaluating the spline at the capture time of each image row (Sec. 7.1); sweeping lasers discussed only as related work (Sec. 2)","continuous IMU or camera pose trajectory (spline) with covariance, camera-IMU transform, gravity, IMU bias splines and landmark positions (Secs. 6.1, 6.2, 7.1)",[],{"id":1182,"label":1183,"shortName":1184,"title":1185,"year":150,"era":10,"cluster":236,"scope":132,"keyIdeaZh":1186,"sensors":1187,"mapRepresentation":1188,"loopClosure":1189,"estimator":1190,"association":1191,"deskew":57,"outputGeometry":1192,"fulltextStatus":22,"lidarModels":1193,"equipmentCount":144},"ldso2018","Gao et al., 2018","LDSO","LDSO: Direct Sparse Odometry with Loop Closure","LDSO 把直接稀疏里程計 DSO 擴充為具迴圈閉合的單目視覺 SLAM。它保留 DSO 以梯度選點的直接法追蹤，但讓部分選點偏向可重複的角點，只在關鍵影格上計算 ORB 描述子並建立詞袋資料庫以偵測迴圈。迴圈候選以 ORB 匹配與 RANSAC PnP 初始化，再同時最小化三維點對齊與二維重投影誤差求 Sim(3) 相對位姿；這些約束與滑動視窗的共視相對位姿一起放入 Sim(3) 位姿圖，以 g2o 最佳化修正旋轉、平移與尺度漂移，不做全域光束法平差。",[37],"keyframe pose graph with sparse inverse-depth points from DSO; a point cloud map is shown before and after loop closure (Fig. 7)","DBoW3 query among marginalised keyframes, ORB matching and RANSAC PnP initial guess, then Gauss-Newton Sim(3) estimate minimising 3D point alignment and 2D reprojection terms using depths from the sliding window (Sec. III-C)","DSO sliding-window photometric bundle adjustment (5 to 7 active keyframes, inverse-depth points, affine brightness and exposure) as the odometry front end; back end Sim(3) pose graph built from co-visibility relative poses of the window plus loop constraints, optimised with g2o while the current window poses are kept fixed (Secs. III-A, III-D)","direct photometric alignment for tracking; point selection keeps DSO's gradient-based pixels but favours Shi-Tomasi corners, for which ORB descriptors are computed on keyframes only and stored in a DBoW3 bag-of-words database (Secs. III-B, III-C)","loop-corrected keyframe trajectory up to monocular scale and a sparse point cloud from DSO's points (Figs. 1, 7)",[],{"id":1195,"label":1196,"shortName":1197,"title":1198,"year":9,"era":10,"cluster":11,"scope":12,"keyIdeaZh":1199,"sensors":1200,"mapRepresentation":1205,"loopClosure":1206,"estimator":1207,"association":1208,"deskew":1209,"outputGeometry":1210,"fulltextStatus":22,"lidarModels":1211,"equipmentCount":487},"gawel2019fabricatorloc","Gawel et al., 2019","In situ Fabricator BIM-referenced localization","A Fully-Integrated Sensing and Control System for High-Accuracy Mobile Robotic Building Construction","作者為現地建造用的移動機械臂設計整合感測與控制系統。狀態估計以移動視窗估測器融合三件資訊：運動補償後的 VLP-16 掃描對建築模型網格取樣點雲的點到平面 ICP 位姿、IMU 以及輪式里程計，使機器人直接在建築模型座標中定位。接近施工位置時，系統再以末端執行器上三個互相正交的雷射測距儀，在多個臂姿下量測到附近牆面的距離，與模型網格光線追蹤出的平面比對，只最佳化基座位置以取得局部高精度定位。任務由 COMPAS 建築任務介面自動產生，並由全身 MPC 追蹤末端軌跡；最終點位以 Leica Nova TM50 全測站量測評估。",[1201,1202,1203,1204],"3D LiDAR (Velodyne VLP-16)","IMU (Xsens MTi-100)","wheel encoders","three orthogonal laser distance sensors on the end-effector","point cloud sampled from the 3D CAD triangle mesh of the building model; Octomap occupancy map initialized from the model and updated from LiDAR scans for planning","none (localization in a known building model)","ConFusion moving horizon estimator fusing ICP pose updates against the building model, IMU and wheel odometry, with IMU forward propagation to compensate LiDAR latency; near task locations a separate high-accuracy localization (HAL) optimizes the static base position from laser distance measurements ray-traced against the 3D mesh, keeping orientation fixed (Sec. III-A, III-B)","point-to-plane ICP between motion-compensated LiDAR scans and a point cloud sampled from the 3D CAD triangle mesh; HAL finds the intersected mesh planes by ray tracing with a Cauchy robust cost (Sec. III-A, III-B)","motion-compensated 3D LiDAR scans (method not detailed) (Sec. III-A)","robot base and end-effector poses in the building-model frame; no point-cloud map product evaluated",[143],{"id":1213,"label":1214,"shortName":1215,"title":1216,"year":1217,"era":52,"cluster":236,"scope":99,"keyIdeaZh":1218,"sensors":1219,"mapRepresentation":1221,"loopClosure":72,"estimator":1222,"association":1223,"deskew":57,"outputGeometry":1224,"fulltextStatus":22,"lidarModels":1225,"equipmentCount":144},"stereoscan2011","Geiger et al., 2011","StereoScan (LIBVISO2)","StereoScan: Dense 3d reconstruction in real-time",2011,"StereoScan 以立體影像在單一 CPU 上即時建立三維地圖。前端以斑點與角點遮罩偵測特徵，用 Sobel 響應的稀疏 SAD 比對，並要求左右影像與前後影格四張影像形成環狀匹配；再以 RANSAC 包覆的高斯牛頓法最小化左右影像重投影誤差估計自我運動，並用等加速度卡爾曼濾波平滑速度。另一執行緒以 ELAS 計算稠密視差，依估計位姿把三維點重投影到目前影格，貪婪地合併並平均重複點，產生一致的點雲。",[1220],"stereo camera (calibrated, rectified)","point-based 3D model: ELAS disparity maps converted to 3D and greedily fused by reprojecting previous points into the current image and averaging points that fall on valid disparities (Sec. III-D)","frame-to-frame stereo egomotion: Gauss-Newton minimisation of left and right reprojection errors of triangulated features inside RANSAC (50 iterations of 3-point samples), refinement on all inliers, then a constant-acceleration Kalman filter on the velocity (Sec. III-B)","blob and corner features from 5 x 5 masks with non-maximum and non-minimum suppression; SAD of quantised Sobel responses at 16 sparse locations of an 11 x 11 window; circular matching over left and right images of two frames with 1 pixel epipolar tolerance; Delaunay-neighbourhood support filtering; two-pass search narrowed per 50 x 50 pixel bin; bucketing to 200 to 500 features (Secs. III-A, III-B)","visual odometry trajectory at 25 fps and a fused dense 3D point cloud updated from new depth maps at 3 to 4 fps (abstract; Sec. III)",[],{"id":1227,"label":1228,"shortName":1229,"title":1230,"year":191,"era":52,"cluster":68,"scope":69,"keyIdeaZh":1231,"sensors":1232,"mapRepresentation":1235,"loopClosure":72,"estimator":1236,"association":1237,"deskew":57,"outputGeometry":1238,"fulltextStatus":22,"lidarModels":1239,"equipmentCount":185},"gelfand2003stable","Gelfand et al., 2003","Geometrically stable sampling","Geometrically stable sampling for the ICP algorithm","作者以點對平面 ICP 線性化後的 6x6 共變異數矩陣（力與力矩項）分析幾何穩定性：特徵值偏小的特徵向量對應兩曲面可相互滑動的螺旋運動，並以條件數作為穩定度指標。取樣時先以稀疏隨機樣本估計重疊區的特徵向量，再依各點對每個特徵向量的約束量排序，貪婪補強目前最弱的方向，使條件數接近 1。合成溝槽平面的條件數由 66.1 降至 3.7（選取 30% 點），Forma Urbis Romae 碎片掃描中均勻取樣無法對齊的溝槽得以正確對齊。與長廊、隧道等幾何退化場景的關聯屬推論。",[1233,1234],"range-scanned meshes of Forma Urbis Romae fragments (scanner not named in the paper)","synthetic noisy meshes (incised plane and sphere)","triangle meshes or point sets with normals (normals averaged from adjacent faces or supplied externally)","point-to-plane ICP linearized for small rotations; the 6x6 covariance matrix C of per-pair torque (p x n) and force (n) terms defines the normal equations; stability measured by the condition number of C (ratio of extreme eigenvalues, target close to 1), computed with points and normals of P after centring and scaling the points to unit mean distance","covariance sampling: estimate eigenvectors of C from several hundred random points in the overlap (overlap test by mesh-boundary check of closest points), build six binned lists of candidate points sorted by |x_k . v_i|, and greedily pick the next point from the list of the currently least-constrained eigenvector; closest points on Q then feed point-to-plane minimization; for Forma Urbis Romae meshes overlapping by 25%, several hundred random points sufficed to stabilize the eigenvector estimate","sampling strategy and pose-uncertainty indication",[],{"id":1241,"label":1242,"shortName":1243,"title":1244,"year":150,"era":10,"cluster":98,"scope":132,"keyIdeaZh":1245,"sensors":1246,"mapRepresentation":1249,"loopClosure":1250,"estimator":1251,"association":1252,"deskew":1253,"outputGeometry":1254,"fulltextStatus":22,"lidarModels":1255,"equipmentCount":46},"lips2018","Geneva et al., 2018","LIPS","LIPS: LiDAR-Inertial 3D Plane SLAM","LIPS 以「最近點」（closest point, CP）表示平面：取平面上距參考座標原點最近的三維點，作為最小且可加法更新的平面參數。為避免平面通過原點時的奇異性，每個平面以首次觀測的位姿為錨點表示，並推導錨定平面因子；每個影格點雲先以 RANSAC 找出平面點集，再壓縮成帶共變異數的局部 CP 量測。這些平面因子與連續 IMU 預積分因子一起放入因子圖，以 iSAM2 平滑估計，平面重複觀測即形成迴圈約束。",[1247,1248],"3D LiDAR (8-beam Quanergy M8 in the real test; simulator modelled on it)","IMU (Microstrain 3DM-GX3-25 in the real test; ADIS16448 model in simulation)","sparse landmark map of infinite planes, each stored as a closest-point vector anchored in the frame of its first observation (Sec. V-B)","implicit through re-observation of previously estimated plane landmarks (no separate place recognition) (Sec. VI-B)","graph-based MLE (nonlinear least squares) with continuous IMU preintegration factors and anchored closest-point plane factors, solved incrementally with iSAM2 in GTSAM; Huber loss on plane factors (Sec. III, V-C)","planes extracted from each point cloud with RANSAC plane segmentation (PCL) run offline in the real test; each planar subset compressed to a local closest-point plane with covariance by weighted Gauss-Newton; plane correspondences by a Mahalanobis-distance test (known correspondences in simulation) (Sec. V-C, VI-B, VI-D)","none in the reported experiments; the authors note preintegration could unwarp clouds at high speed but did not use it (Sec. III-B)","IMU trajectory and a set of plane parameters; no dense point cloud output is described",[1256,1257],"Quanergy M8","Quanergy M8 (simulated)",{"id":1259,"label":1260,"shortName":1261,"title":1262,"year":396,"era":10,"cluster":236,"scope":99,"keyIdeaZh":1263,"sensors":1264,"mapRepresentation":1266,"loopClosure":1267,"estimator":1268,"association":1269,"deskew":57,"outputGeometry":1270,"fulltextStatus":22,"lidarModels":1271,"equipmentCount":144},"openvins2020","Geneva et al., 2020","OpenVINS","OpenVINS: A Research Platform for Visual-Inertial Estimation","OpenVINS 是以研究平台定位的開源視覺慣性估測程式庫，核心為流形上的滑動視窗 EKF（MSCKF），採用首次估計 Jacobian（FEJ）維持一致性，並可把部分特徵作為 SLAM 地標保留在狀態中。系統支援相機內參、相機與 IMU 外參及時間偏移的線上校正，以型別化索引系統自動管理狀態與共變異數，另附以 SE(3) B-spline 產生量測的模擬器與軌跡評估工具。作者在模擬與 EuRoC 資料上與多個開源 VIO 比較，顯示其精度具競爭力。",[1265,36],"monocular or stereo camera (arbitrary number of cameras supported)","sliding window of IMU pose clones plus a bounded set of SLAM landmarks (up to 50 in the simulation setup) as 3D points in the state","none (VIO only; authors note a pose-graph optimiser could be appended, Sec. V-B)","modular on-manifold EKF over a sliding window of stochastic IMU pose clones (MSCKF) with First-Estimates Jacobians; optional SLAM landmarks kept in the state and initialised by QR (Givens) splitting of the linearised system; type-based index system manages state and covariance (Secs. II, III-A, III-B, V-A)","sparse visual feature tracking with an OpenCV-based front end; features within the window used in nullspace-projected MSCKF updates; SLAM landmarks in several parameterisations (global 3D, inverse MSCKF, full inverse depth, anchored 3D); errors on raw pixels to allow intrinsic calibration (Secs. II, III-C, V-B)","IMU pose trajectory with covariance, online camera intrinsics, camera-IMU extrinsics and time offset, and sparse landmark positions; no dense map",[],{"id":1273,"label":1274,"shortName":1275,"title":1276,"year":513,"era":52,"cluster":397,"scope":514,"keyIdeaZh":1277,"sensors":1278,"mapRepresentation":57,"loopClosure":57,"estimator":1281,"association":57,"deskew":57,"outputGeometry":1282,"fulltextStatus":22,"lidarModels":1283,"equipmentCount":1289},"glennie2007rigorous","Glennie, 2007","Kinematic LiDAR error budget","Rigorous 3D error analysis of kinematic scanning LIDAR systems","作者把 LiDAR 直接地理定位方程式中的 14 個觀測量（GNSS 位置、IMU 姿態、視準角、掃描角與距離、槓桿臂）做一階展開，以 Jacobian 將典型誤差傳遞為點位的水平與垂直精度，並模擬定翼機、直升機與地面車載三種平台。定翼機的水平誤差主要來自 IMU 與視準角誤差，作者歸納其水平誤差至少約為垂直誤差的 5 倍；直升機（Q-240）以光束發散造成的掃描角誤差為主，水平與垂直誤差比約為 2 至 2.5；地面系統的比值約為 2，誤差預算以雷射掃描儀本身為主，姿態誤差貢獻低於 25%。模型再以定翼機地面控制點、直升機連結點閉合差與既有地面系統高程檢核比對，預測值與實測相近。GNSS 誤差未納入模型，需另行相加。",[1279,1280,36],"LiDAR","GNSS","first-order (Taylor) error propagation of the direct georeferencing equation with 14 observed parameters (GNSS position, IMU roll, pitch and yaw, three boresight angles, scan angle, range, three lever-arm components) through the Jacobians J, K, B and C","expected horizontal and vertical accuracy",[1284,1285,1286,1287,1288],"Terrapoint ALTMS","Riegl Q-140","Optech 3100","Riegl Q-240","Riegl Q-280",10,{"id":1291,"label":1292,"shortName":1293,"title":1294,"year":363,"era":52,"cluster":397,"scope":1035,"keyIdeaZh":1295,"sensors":1296,"mapRepresentation":57,"loopClosure":57,"estimator":1300,"association":1301,"deskew":1302,"outputGeometry":1303,"fulltextStatus":22,"lidarModels":1304,"equipmentCount":109},"glennie2012hdl64","Glennie, 2012","HDL-64E S2 calibration","Calibration and Kinematic Analysis of the Velodyne HDL-64E S2 Lidar Sensor","作者以平面特徵約束的 Gauss-Helmert 最小二乘法，在車載動態資料中同時估計 Velodyne HDL-64E S2 的視準角、槓桿臂與每顆雷射的內部校正參數。資料取自 2010 年在德州 The Woodlands 停車場以原型車載系統（IMAR AirSurv-RQH IMU、Novatel OEM-4 GPS）多次通過所得，萃取 75 個約 4 m² 的平面。可觀測性分析顯示各雷射水平旋轉修正與航向視準角近乎完全相關、z 向槓桿臂幾乎不可觀測、水平與垂直偏移也高度相關，因此移出估計。縮減參數後，平面閉合差 RMSE 由僅做視準校正的 0.047 m 降至 0.034 m（加入靜態校正的水平旋轉修正後為 0.030 m）；限 25 m 內時由 0.037 m 降至 0.025 m，接近 Riegl LMS-Q120i 的 0.020 m，但不及 VZ-400 的 0.013 m。",[1297,1298,1299],"3D LiDAR (Velodyne HDL-64E S2)","IMU (IMAR AirSurv-RQH, navigation grade)","GNSS (Novatel OEM-4 dual-frequency GPS receiver)","planar-feature-conditioned combined (Gauss-Helmert) least-squares adjustment estimating boresight angles, lever arm and per-laser interior parameters together with plane coefficients; after observability analysis the per-laser horizontal rotation corrections, horizontal and vertical offsets and the z lever arm were removed, leaving roll, pitch, yaw, horizontal lever arm and three parameters per laser","semi-automatic extraction of 75 planar surfaces of about 4 m2 each (mostly horizontal or vertical, about 15 near 45 deg), thinned to about 750 returns per plane at 5 to 100 m range","implicit: each return is georeferenced with the GNSS\u002FINS pose at its own measurement time (Eq. 1); no separate deskew step","kinematic interior calibration and boresight parameters with their precision, and planar misclosure statistics of the georeferenced point cloud",[161,1305,1306],"Riegl LMS-Q120i","Riegl VZ-400",{"id":1308,"label":1309,"shortName":1310,"title":1311,"year":150,"era":10,"cluster":263,"scope":346,"keyIdeaZh":1312,"sensors":1313,"mapRepresentation":1316,"loopClosure":1317,"estimator":1318,"association":1319,"deskew":1320,"outputGeometry":1321,"fulltextStatus":22,"lidarModels":1322,"equipmentCount":46},"limo2018","Graeter et al., 2018","LIMO","LIMO: Lidar-Monocular Visual Odometry","LIMO 以單眼相機的特徵追蹤為主，LiDAR 只負責替影像特徵提供深度：先把單次掃描的 LiDAR 點投影到影像，在特徵周圍以深度直方圖切出前景點，再以面積最大的三點平面與視線求交得到特徵深度；地面上的特徵另以 RANSAC 擬合的地面平面處理，超過 30 m 的深度則捨棄。逐影格運動以 PnP 與對極誤差估計後，交給關鍵影格光束法平差，把 LiDAR 深度當成殘差加入，並以關鍵影格與地標篩選、語意剔除動態物件、植生降權及修剪最大殘差維持即時性。系統只做視覺里程計，不含迴圈閉合，也不建立稠密點雲地圖。",[1314,1315],"monocular camera (KITTI grayscale images used for feature tracking; camera model not named in the paper)","3D LiDAR (KITTI; model not named in the paper), used only to give depth to image features","sparse triangulated landmarks inside the bundle-adjustment window, split into near, middle and far bins and thinned by a voxel filter with median filtering; no dense LiDAR map (Sec. V-C)","none; the authors state they aim for visual odometry and perform no loop closure (Fig. 2 caption; Sec. I)","two separate optimizations: (i) frame-to-frame 6-DoF motion from a perspective-n-point cost plus an epipolar cost, each wrapped in a Cauchy loss, used as prior; (ii) windowed keyframe bundle adjustment over reprojection errors, LiDAR depth residuals and a scale regularizer on the oldest motion in the window, with Cauchy losses and a trimmed-least-squares-like removal of the highest residuals after a few iterations (Secs. IV-V, Eq. 7, Algorithm 1)","viso2 feature tracking (about 2000 correspondences in 30-40 ms); feature depth from a local plane through the maximum-area triangle of foreground LiDAR points selected in an image-space neighbourhood by a depth histogram (bin width 0.3 m); ground-plane features use a RANSAC ground fit instead; depth estimates beyond 30 m or at grazing angles rejected; landmarks on dynamic semantic classes rejected and vegetation landmarks weighted (Secs. II, III, V-C, VI)","not_reported (no LiDAR motion compensation step is described; KITTI scans are used as provided)","camera poses and a sparse landmark reconstruction; no registered point-cloud map is produced (Fig. 2 caption)",[45],{"id":1324,"label":1325,"shortName":1326,"title":1327,"year":513,"era":52,"cluster":527,"scope":132,"keyIdeaZh":1328,"sensors":1329,"mapRepresentation":1331,"loopClosure":1332,"estimator":1333,"association":1334,"deskew":121,"outputGeometry":1335,"fulltextStatus":22,"lidarModels":1336,"equipmentCount":487},"gmapping2007","Grisetti et al., 2007","GMapping","Improved Techniques for Grid Mapping With Rao-Blackwellized Particle Filters","GMapping 在 Rao-Blackwellized 粒子濾波（每個粒子攜帶一張佔據網格地圖）上提出兩項改良：以掃描匹配結果與里程計共同計算較準確的提議分布（proposal distribution），以及依有效樣本數選擇性重取樣以減少粒子耗盡。作者報告所需粒子數大約比先前方法少一個數量級。當掃描匹配失敗（如大空曠區域多為最大量程讀值）時，系統退回原始里程計運動模型。",[1330,119],"2D laser range finder","2D occupancy grid per particle","implicit through particle filter (no explicit loop-closure module)","Rao-Blackwellized particle filter; per-particle Gaussian proposal fitted to K samples around the scan-matcher mode, weighted by observation likelihood and the odometry motion model; raw motion model used when scan matching fails; resampling only when Neff drops below N\u002F2 (Sec. III-B to III-E)","per-particle scan matching with the CARMEN 'vasco' matcher: gradient descent on the beam-endpoint likelihood of the current scan against the particle's own grid map, with the search bounded around the odometry-based initial guess (Sec. III-E, IV)","2D occupancy grid map",[1337,1338],"SICK LMS","SICK PLS",{"id":1340,"label":1341,"shortName":1342,"title":1343,"year":769,"era":10,"cluster":1344,"scope":514,"keyIdeaZh":1345,"sensors":1346,"mapRepresentation":57,"loopClosure":57,"estimator":57,"association":57,"deskew":57,"outputGeometry":57,"fulltextStatus":22,"lidarModels":1347,"equipmentCount":61},"grupp2017evo","Grupp, 2017","evo","evo: Python package for the evaluation of odometry and SLAM.","C10","evo 是評估里程計與 SLAM 軌跡的 Python 套件（GPL-3.0 以上授權），提供 evo_ape（絕對位姿誤差）與 evo_rpe（相對位姿誤差）命令列工具，並以 evo_traj 與 evo_res 做軌跡檢查、繪圖與結果比較。它支援 TUM、KITTI、EuRoC 格式與 ROS 1、ROS 2 bag，可選 SE(3) 或 Sim(3) Umeyama 對齊、僅尺度校正或原點對齊，時間關聯預設容許 0.01 s 的時間差，RPE 預設只取相鄰位姿對。作者明言它不是任何特定資料集評估協定的逐一重現；由於選項會改變數值，引用時應同時報告版本與參數（推論）。",[],[],{"id":1349,"label":1350,"shortName":1351,"title":1352,"year":207,"era":10,"cluster":98,"scope":132,"keyIdeaZh":1353,"sensors":1354,"mapRepresentation":1356,"loopClosure":1357,"estimator":1358,"association":1359,"deskew":1360,"outputGeometry":1361,"fulltextStatus":22,"lidarModels":1362,"equipmentCount":1289},"kissslam2025","Guadagnino et al., 2025a","KISS-SLAM","KISS-SLAM: A Simple, Robust, and Accurate 3D LiDAR SLAM System With Enhanced Generalization Capabilities","KISS-SLAM 將 KISS-ICP 延伸為完整 LiDAR-only SLAM：依行進距離切分局部地圖，以關鍵位姿作為位姿圖節點。迴圈偵測將局部地圖地面對齊後投影成鳥瞰密度影像，以 ORB 描述子比對，再以 3D 配準與重疊率（門檻 40%）驗證後加入位姿圖最佳化。處理完成後另做離線細粒度位姿圖最佳化，將殘餘漂移分配到局部軌跡。",[1355],"3D LiDAR only","keypose-anchored local maps (voxel grids) split by travelled distance; output 3D occupancy grid","ground alignment, bird's-eye-view density images, ORB descriptors with database search, RANSAC 2D alignment, then 3D registration and overlap check (accepted above 40%)","KISS-ICP odometry; pose graph over local-map keyposes; offline fine-grained pose graph over scan poses after processing","point-to-point ICP (odometry); loop verification by registration of voxel mean-and-normal clouds","constant-velocity per-point deskew inherited from KISS-ICP","globally corrected trajectory, local point maps and 3D occupancy grid (0.05 m voxels in navigation experiment)",[1363,437,1364,143,1365,1366],"Aeva","Ouster","Hesai XT-32","SICK TiM781S",{"id":1368,"label":1369,"shortName":1370,"title":1371,"year":207,"era":10,"cluster":98,"scope":99,"keyIdeaZh":1372,"sensors":1373,"mapRepresentation":1375,"loopClosure":72,"estimator":1376,"association":1377,"deskew":1378,"outputGeometry":1379,"fulltextStatus":22,"lidarModels":1380,"equipmentCount":487},"kinematicicp2025","Guadagnino et al., 2025b","Kinematic-ICP","Kinematic-ICP: Enhancing LiDAR Odometry with Kinematic Constraints for Wheeled Mobile Robots Moving on Planar Surfaces","Kinematic-ICP 針對在平面上移動、配備 3D LiDAR 的輪式機器人，把單輪車（unicycle）運動學模型放進點到點 ICP 最佳化，並以輪式里程計為初值與正則化項，使估計結果符合平台運動限制。正則化強度依情況自適應調整，以在特徵不足的走廊中更信任輪式里程計。作者回報已部署於 Dexory 倉儲機器人隊伍。",[1374,119],"3D LiDAR (Robosense Bpearl, Hesai XT32)","voxel-grid local map as in KISS-ICP (author-stated, Sec. III)","point-to-point ICP (KISS-ICP based) with unicycle kinematic model and adaptive regularization toward the wheel-odometry initial guess","point-to-point","de-skewing in preprocessing inherited from KISS-ICP (Sec. III); motion source for de-skewing not detailed in text read","planar odometry; export format not_reported",[1381,1382],"Bpearl","Hesai LiDAR XT32",{"id":1384,"label":1385,"shortName":1386,"title":1387,"year":1388,"era":52,"cluster":527,"scope":132,"keyIdeaZh":1389,"sensors":1390,"mapRepresentation":1392,"loopClosure":1393,"estimator":1394,"association":1395,"deskew":121,"outputGeometry":1396,"fulltextStatus":22,"lidarModels":1397,"equipmentCount":46},"gutmann_konolige1999_lrgc","Gutmann & Konolige, 1999","LRGC (Local Registration and Global Correlation)","Incremental mapping of large cyclic environments",1999,"LRGC 以 Lu 與 Milios 的一致位姿估計為核心，分兩種方式使用：每加入一筆新掃描，只與最近 K 個位姿做局部配準，所以每步計算量固定；偵測到迴圈後，才對整個迴圈做一致位姿估計。迴圈偵測不用單一掃描，而是把最新 m 筆掃描組成地圖片段，以相關運算在較舊的地圖中搜尋，並以高匹配分數、低歧異與低變異三個條件拒絕誤判；搜尋範圍與片段大小都隨位置不確定度增加。作者稱這是第一個不需操作者輸入即可在大型循環環境即時產生稠密度量地圖的系統。",[1391,119],"2D laser range finder (180 deg SICK)","undirected graph of robot poses with attached scans; links from dead reckoning, scan matching or correlation (Sec. 2.4)","patch-to-map correlation run every few scans inside a Mahalanobis-gated search area; patch size grows with position uncertainty; an accepted topological link cannot be undone (Sec. 2.3-2.4)","Lu-Milios consistent pose estimation (least-squares pose network) applied to the last K poses for each new scan and, after a loop is found, to the poses along the loop with sparse linear algebra and strong-link reduction (maximum 200 poses) (Sec. 2.2)","pairwise scan matching combining Cox's point-to-line method with Lu-Milios matching (Gutmann-Schlegel method); loop detection by correlating a patch of the newest m scans with the older map on a grid, accepted only if match score is high, ambiguity low and variance low (Sec. 2.1, 2.3)","2D dense metric scan-point maps",[988],{"id":1399,"label":1400,"shortName":1401,"title":1402,"year":661,"era":10,"cluster":755,"scope":99,"keyIdeaZh":1403,"sensors":1404,"mapRepresentation":1406,"loopClosure":72,"estimator":1407,"association":1408,"deskew":57,"outputGeometry":1409,"fulltextStatus":22,"lidarModels":1410,"equipmentCount":185},"gsicpslam2024","Ha et al., 2024","GS-ICP SLAM","RGBD GS-ICP SLAM","GS-ICP SLAM 讓追蹤與建圖共用同一張三維高斯地圖：追蹤端把目前深度影像降採樣反投影後，以 k 近鄰共變異數組成來源高斯，再用廣義 ICP（G-ICP）與地圖中的目標高斯配準求得位姿；建圖端則直接沿用這些共變異數作為新增高斯的初始形狀，並依深度做尺度正規化，因此不需 3DGS 的密化步驟。系統另把追蹤關鍵影格與僅供建圖的關鍵影格分開，以兼顧軌跡精度與渲染品質；整體速度最高達每秒 107 影格，但沒有迴圈閉合，也未評估三維幾何精度。",[1405],"RGB-D camera","single 3D Gaussian map shared by tracking and mapping: new Gaussians inherit G-ICP k-nearest-neighbour covariances with depth-dependent scale normalization (divided by z^p, p = 1.5 best), without densification (Sec. 3.2, Tables 6 and 8)","Generalized-ICP (G-ICP) scan-to-map registration: maximum-likelihood alignment of source Gaussians from the downsampled, reprojected depth image to target Gaussians taken from the 3DGS map, with ellipse scale regularization; 3DGS mapping runs in parallel with L1 and D-SSIM colour and L1 depth losses (Sec. 3, Eq. 1-3); the G-ICP module is built on the VGICP C++ implementation [koide2021vgicp] wrapped with pybind11 (ECCV Supp. Sec. A)","nearest-neighbour correspondences between source and target Gaussians inside G-ICP; keyframes selected when the share of correspondences within a distance threshold falls below a threshold, plus mapping-only keyframes every 10 frames (Sec. 3.1-3.2)","3D Gaussian map rendered to colour and depth; the main text evaluates trajectory (ATE) and rendering (PSNR, SSIM, LPIPS) only (Sec. 4.1); the ECCV supplementary adds average rendered-depth L1 error of 0.030 m on Replica and 0.118 m on TUM, with no mesh or point-cloud accuracy against a reference (Supp. Sec. C.2)",[],{"id":1412,"label":1413,"shortName":1414,"title":1415,"year":150,"era":10,"cluster":236,"scope":132,"keyIdeaZh":1416,"sensors":1417,"mapRepresentation":1419,"loopClosure":1420,"estimator":1421,"association":1422,"deskew":684,"outputGeometry":1423,"fulltextStatus":22,"lidarModels":1424,"equipmentCount":144},"flashfusion2018","Han & Fang, 2018","FlashFusion","FlashFusion: Real-time Globally Consistent Dense 3D Reconstruction using CPU Computing","FlashFusion 是不使用 GPU 運算、可在可攜裝置上即時運作的全域一致稠密 RGB-D 重建系統。定位端以 ORB 特徵對應，把每個新關鍵影格與 MILD 迴圈偵測找出的前 5 個相似關鍵影格做全域配準，並以作者的 FastGO 預先累積對應點的二階統計量，使全域位姿最佳化能在 CPU 上即時求解；另以只修正相對位姿變化最大的 10 對影格來近似 Huber 穩健損失。重建端採空間雜湊 TSDF，以每個區塊 8 個角點的稀疏體素取樣快速挑出含表面的有效區塊（先以 20 mm、再以 5 mm 解析度檢查），且只在關鍵影格做篩選；網格擷取以自適應門檻、查表與鄰近區塊位址表加速。位姿更新後，每個關鍵影格重新整合至多 10 個先前關鍵影格，以 CPU 近似 BundleFusion 的重新整合。",[1418],"RGB-D camera (Asus Xtion for live scanning; TUM RGB-D real sequences; synthetic noisy ICL-NUIM)","spatially hashed TSDF with 8 x 8 x 8-voxel chunks and a second coarse hash of chunk cubes; valid chunks selected by sparse voxel sampling of the 8 chunk corners (first at 20 mm, then 5 mm) only on keyframes; colour stored as colour times weight","MILD multi-index hashing loop detection on ORB features (no training); matched keyframes added to the global optimization","Keyframe global pose optimization (authors' FastGO): Gauss-Newton on SE(3) minimizing distances between corresponding ORB feature points of each keyframe and its top-5 most similar keyframes, with second-order statistics pre-integrated so each frame pair costs O(1); Huber norm approximated by an online correction of the 10 frame pairs whose relative poses change most; local frames fixed relative to their keyframe","About 1000 ORB features per frame; local registration of each frame to its keyframe and global registration of each new keyframe against the top-5 similar previous keyframes found by MILD appearance-based loop detection","coloured triangle mesh with normals from accelerated marching cubes (adaptive per-chunk threshold, one-DoF vertex placement, neighbour-chunk look-up tables); 5 mm voxels",[],{"id":1426,"label":1427,"shortName":1428,"title":1429,"year":30,"era":10,"cluster":11,"scope":300,"keyIdeaZh":1430,"sensors":1431,"mapRepresentation":1433,"loopClosure":57,"estimator":1434,"association":1435,"deskew":57,"outputGeometry":1436,"fulltextStatus":22,"lidarModels":1437,"equipmentCount":24},"han2026nifcyl","Han et al., 2026","NIFCyl tunnel deformation from SLAM LiDAR","Full-field deformation quantification of underground tunnels using SLAM LiDAR point cloud based on unsupervised neural implicit learning","作者提出 NIFCyl：以 8 層 MLP 非監督學習參考點雲的有號距離場，取其梯度作為尺度不變且方向一致的法向，再沿法向以圓柱鄰域平均兩期點雲的投影位置，求得全場變形，不需標註資料或局部 PCA 擬合。資料以手持 Hovermap ST 在西澳 Kalgoorlie 地下硬岩礦取得，兩期點雲以四個噴漆控制點做 SVD 粗配準，再以 ICP 細配準。合成變形情境中 NIFCyl 的 R² 為 0.963、RMSE 0.009 m、總計算 187 秒，三種尺度設定的 M3C2 中最佳者為 0.955、0.010 m、995 秒；現地放置紙箱的試驗中，平面區兩者 MAE 同為 0.022 m，曲面區為 0.032 m 對 0.034 m。實際擴挖案例只以單一點的人工量測（2.513 m）與模型結果（2.505 m）比對。",[1432],"handheld Hovermap ST LiDAR SLAM scanner (Table 1: FoV 360 x 290 deg, range 0.40 to 100 m, LiDAR accuracy +\u002F-30 mm, mapping accuracy +\u002F-15 mm in typical underground and indoor environments, SLAM drift +\u002F-0.03%, up to 300,000 pts\u002Fs single return and 600,000 pts\u002Fs dual return)","neural implicit function of tunnel surface; scale-invariant normal field (per abstract)","not_applicable (deformation analysis of SLAM point clouds)","SDF learned by an 8-layer MLP gives normals via its gradient; deformation measured by averaging reference and compared points inside a cylinder along each normal; epochs aligned by GCP-based SVD then ICP","full-field deformation map",[],{"id":1439,"label":1440,"shortName":1441,"title":1442,"year":661,"era":10,"cluster":397,"scope":69,"keyIdeaZh":1443,"sensors":1444,"mapRepresentation":1450,"loopClosure":1451,"estimator":1452,"association":1453,"deskew":121,"outputGeometry":1454,"fulltextStatus":22,"lidarModels":1455,"equipmentCount":377},"hatleskog2024probdegen","Hatleskog & Alexis, 2024","Probabilistic degeneracy detection (DRPM)","Probabilistic Degeneracy Detection for Point-to-Plane Error Minimization","本法把點與法向量的雜訊傳遞到點對面最佳化的 Hessian，計算每個特徵方向訊號明顯大於雜訊的機率，作為退化判定。更新時不採硬門檻，而以該機率縮放特徵值倒數，平滑地衰減退化方向的更新。參數依 LiDAR 規格書的雜訊特性設定。",[1445,1446,1447,1448,1449],"3D LiDAR (Velodyne VLP-16 in exp. 1-2; Ouster OS0-64 RevD in exp. 3; Ouster OS0-128 RevD in exp. 4)","IMU (Alphasense IMU in exp. 2; PixRacer Pro autopilot IMU in exp. 3; VectorNav VN100 in exp. 4)","monochrome camera for the ROVIO visual-inertial prior (exp. 2)","radar (TI IWR6843AOP-EVM) for a velocity factor (exp. 4 only)","legged odometry (exp. 1)","LOAM-style planar feature map (CompSLAM integration)","none reported","point-to-plane ICP whose update multiplies each inverse eigenvalue of the Hessian by the probability that the signal exceeds the noise tenfold (s = 10); in exp. 3 and 4 the scan-to-map ICP pose enters a GTSAM fixed-lag smoother (3 s lag) as a unary factor with the information matrix of Eq. 28","scan-to-map point-to-plane (planar features within a LOAM-derived mapping module)","registered point cloud map and poses",[143,224,222],{"id":1457,"label":1458,"shortName":1459,"title":1460,"year":1461,"era":10,"cluster":599,"scope":600,"keyIdeaZh":1462,"sensors":1463,"mapRepresentation":57,"loopClosure":1464,"estimator":57,"association":1465,"deskew":121,"outputGeometry":57,"fulltextStatus":22,"lidarModels":1466,"equipmentCount":24},"m2dp2016","He et al., 2016","M2DP","M2DP: A novel 3D point cloud descriptor and its application in loop closure detection",2016,"M2DP 先以質心平移並用 PCA 主軸對齊點雲，再將點雲投影到 4 個方位角乘 16 個仰角共 64 個 2D 平面；每個平面以 8 個同心圓乘 16 個扇區計算點數，組成 64×128 的簽章矩陣，最後以奇異值分解取第一左、右奇異向量，得到 192 維全域描述子，用於光達迴圈偵測。作者在 KITTI00 上報告其描述子計算時間為各 3D 描述子中最短；在 KITTI、Freiburg Campus 與 Ford Campus 的多數序列中，其 100% 精確率下的召回率最高，但在 KITTI00 與 KITTI05 上 SHOT 略高，視覺描述子 GIST 更高。",[35],"descriptor used for loop closure detection","centroid-shifted, PCA-aligned cloud projected onto p x q = 4 x 16 planes (azimuth stride pi\u002Fp, elevation stride pi\u002F(2q)); each plane split into l = 8 concentric rings (radii r, 4r, ..., l^2 r) x t = 16 sectors holding point counts; 64 x 128 signature matrix reduced by SVD to the first left and right singular vectors (192-D); loop decided by thresholding the nearest-neighbour descriptor distance with +\u002F-50 frames excluded",[45],{"id":1468,"label":1469,"shortName":1470,"title":1471,"year":51,"era":10,"cluster":98,"scope":99,"keyIdeaZh":1472,"sensors":1473,"mapRepresentation":1476,"loopClosure":121,"estimator":1477,"association":1478,"deskew":1479,"outputGeometry":1480,"fulltextStatus":22,"lidarModels":1481,"equipmentCount":471},"pointlio2023","He et al., 2023a","Point-LIO","Point‐LIO: Robust High‐Bandwidth Light Detection and Ranging Inertial Odometry","Point-LIO 在每一個 LiDAR 點或 IMU 取樣到達時，就以不迭代的流形擴展卡爾曼濾波（on-manifold EKF）進行傳播與更新，里程計輸出可達 4 至 8 kHz，並從架構上避免掃描內的運動畸變。作者把角速度與線加速度擴增為一階積分隨機過程狀態，將 IMU 量測視為系統輸出，飽和的 IMU 通道直接略過，因此旋轉超出 IMU 量程時仍能估計位姿。每個點以 ikd-Tree 中五個最近鄰擬合平面，計算一維點到平面殘差。",[1474,1475],"Livox Avia solid-state LiDAR with built-in BMI088 IMU (own experiments)","public benchmarks with Livox Horizon, Velodyne HDL-32E and VLP-16 plus their IMUs","ikd-Tree incremental k-d tree point map from FAST-LIO2; local map size 2000 m, spatial downsampling 0.25 m, rebalancing thresholds 0.6 and 0.5, parallel rebuild threshold 1500 points; public benchmarks use FAST-LIO2 default mapping parameters with 1:4 temporal downsampling of raw points","Tightly coupled on-manifold EKF (IKFoM toolbox), deliberately not iterated; 24-dimensional state on SO(3) x R21 with attitude, position, velocity, gyroscope and accelerometer biases, gravity, and angular velocity and linear acceleration modelled as first-order integrator processes; each LiDAR point or IMU sample is propagated and fused at its own timestamp, and saturated IMU channels are skipped","Each point is projected with the propagated pose; its five nearest map points within 5 m in the ikd-Tree are fitted to a plane; if any neighbour lies more than 0.1 m from the plane the point is added to the map without an update, otherwise a one-dimensional point-to-plane residual updates the state","No explicit deskewing step: each point is fused at its own sampling time so frame-level motion distortion does not arise; wall-thickness views show thinner walls than FAST-LIO2 (qualitative, Figs. 5 to 7)","Point map accumulated in the ikd-Tree (points inserted at the updated pose) and odometry at 4 to 8 kHz; no export format or map accuracy evaluation reported",[437,1482,225,143],"Livox Horizon",{"id":1484,"label":1485,"shortName":1486,"title":1487,"year":51,"era":10,"cluster":53,"scope":54,"keyIdeaZh":1488,"sensors":1489,"mapRepresentation":1493,"loopClosure":72,"estimator":1494,"association":1495,"deskew":1496,"outputGeometry":1497,"fulltextStatus":22,"lidarModels":1498,"equipmentCount":24},"he2023ikfom","He et al., 2023b","IKFoM","Symbolic Representation and Toolkit Development of Iterated Error-State Extended Kalman Filters on Manifolds","本文提出在流形上建構迭代誤差狀態擴展卡爾曼濾波（IESEKF）的通用符號化方法：以 ⊞、⊟ 與 ⊕ 運算把機器人系統寫成離散時間的流形標準形式，使濾波各步驟中的流形約束與系統特定部分分離，並證明其最小參數化在整個工作空間內沒有奇異點。作者據此開發 C++ 工具包 IKFoM，支援 R^n、SO(3)、SE_N(3) 與 S^2 等原始流形及其組合，使用者只需提供系統描述即可呼叫預測與更新。期刊版以兩個緊耦合 LiDAR 慣性系統驗證：重新實作 FAST-LIO 並加入線上 LiDAR 與 IMU 外參估計（六組資料的漂移與手推版本相當，執行時間略增），以及以 IKFoM 取代 LINS 的手推濾波器（在 LIO-SAM Campus 資料上執行時間較短）。",[1490,1491,1492],"3D LiDAR (Livox AVIA solid-state LiDAR)","IMU (built into the Livox AVIA)","spinning multiline LiDAR and IMU data of the public LIO-SAM Campus sequences for the LINS re-implementation (sensor models not stated)","global point cloud map (as in FAST-LIO)","iterated error-state extended Kalman filter on compound manifolds in canonical form x_{k+1} = x_k oplus (dt f(x_k,u_k,w_k)); toolkit supports R^n, SO(3), SE_N(3) and S^2(r) primitives; experiment state manifold R3 x R3 x SO(3) x R3 x R3 x S2 x SO(3) x R3 with at most 5 iterations (Sec. III-H; Sec. IV-B; Sec. IV-C)","point-to-plane residuals of LiDAR points to planar features of the map, identical to FAST-LIO [14] (Sec. IV-A, Eq. 30); the LINS re-implementation keeps the original LINS feature extraction and outlier rejection (Sec. IV-C3)","not_reported in the TIE article (system taken from FAST-LIO [14]; motion compensation is not described)","IMU state (position, velocity, rotation, biases, gravity), online LiDAR-IMU extrinsic, and a point cloud map",[437],{"id":1500,"label":1501,"shortName":1502,"title":1503,"year":249,"era":10,"cluster":11,"scope":12,"keyIdeaZh":1504,"sensors":1505,"mapRepresentation":1508,"loopClosure":1509,"estimator":1510,"association":1511,"deskew":121,"outputGeometry":1512,"fulltextStatus":22,"lidarModels":1513,"equipmentCount":144},"hendrikx2021semanticbimloc","Hendrikx et al., 2021","Semantic BIM for 2D LiDAR localization","Connecting Semantic Building Information Models and Robotics: An application to 2D LiDAR-based localization","作者把 IFC 格式 BIM 中的牆與柱轉成機器人可查詢的語意世界模型：先將一層樓的 IFC 匯出為 IFC-JSON 並加上 JSON-LD 語境，再把柱的斷面輪廓與牆的中心線加厚度改寫成二維幾何（牆轉為共用角點的內外兩條折線），連同與感測器的可感知關係存入 PostgreSQL 與 PostGIS。定位時機器人先查詢附近可由 2D LiDAR 感知的 BIM 物件，再依查詢結果設定線、角與矩形偵測器，只接受支持地圖特徵的量測，並把帶有物件編號的關聯加入 GTSAM 的移動視窗因子圖。作者在埃因霍芬理工大學 Atlas 大樓一層以麥克納姆輪平台遙控三條各約 100 m 的路線示範位姿追蹤，其中一條有三位行人干擾。",[1506,1507],"2D LiDAR (Hokuyo UTM30-LX, mounted upside down near the floor, 180 deg FOV, 720 points)","wheel encoder odometry","JSON-LD property graph of IfcWall and IfcColumn entities with 2D Simple Feature geometry (wall inner and outer polylines with shared corner points, column polygons) stored in PostgreSQL with PostGIS (Sec. III)","none (localization against the BIM-derived map)","Moving-horizon factor graph in GTSAM over recent robot poses with range-bearing factors to associated columns and corners and angle-distance factors to walls; horizon truncated to keep N unique semantic map features (three in the experiment) (Sec. IV-V)","map-query-first: features near the current pose (within 6 m) queried from a PostGIS database; line, corner and box features extracted by split-and-merge line fitting and L-shape checks are accepted only if they support a queried BIM feature; associations reference BIM object ids explicitly (Sec. IV)","2D robot trajectory in the BIM coordinate frame with explicit semantic associations; no point cloud produced",[375],{"id":1515,"label":1516,"shortName":1517,"title":1518,"year":363,"era":52,"cluster":236,"scope":132,"keyIdeaZh":1519,"sensors":1520,"mapRepresentation":1522,"loopClosure":1523,"estimator":1524,"association":1525,"deskew":57,"outputGeometry":1526,"fulltextStatus":22,"lidarModels":1527,"equipmentCount":144},"rgbdmapping2012","Henry et al., 2012","RGB-D Mapping (Henry et al.)","RGB-D mapping: Using Kinect-style depth cameras for dense 3D modeling of indoor environments","RGB-D Mapping 以 PrimeSense（等同 Kinect）RGB-D 相機建立室內稠密三維地圖。相鄰影格的配準採用作者提出的 RGB-D ICP：先以 FAST 特徵與 Calonder 描述子加上 RANSAC 求得初始位姿，並以重投影誤差篩選內點，再把固定的特徵對應與稠密點到平面 ICP 誤差合併，以 Levenberg-Marquardt 聯合最佳化；兩階段版本只在 RANSAC 內點不足時才執行 ICP，因此在黑暗或缺乏特徵的區域仍能對齊。系統以關鍵影格、位置預篩選與詞彙樹偵測迴圈，再以 TORO 位姿圖或加入 ICP 點對的稀疏光束法平差（SBA）做全域最佳化，最後把點雲整合為面元（surfel）地圖。",[1521],"RGB-D camera (PrimeSense active-stereo camera equivalent to the Kinect sensor, 640 x 480 registered image and depth at 30 fps; Sec. 1)","dense coloured point cloud of the aligned frames; surfel map (location, orientation, patch size, colour, confidence as a 2D histogram of viewing directions) created as a post-process after global optimisation (Sec. 3.3)","keyframes created when the accumulated rotation or translation since the previous keyframe exceeds a threshold; candidates prefiltered by estimated global pose (keyframes within a few metres) and by a vocabulary tree on Calonder descriptors (ROS implementation), then accepted when RANSAC finds enough geometrically consistent feature matches (Sec. 3.2.1)","frame-to-frame RGB-D ICP: RANSAC over sparse visual feature points (re-projection-error RE-RANSAC refined by two-frame sparse bundle adjustment, or Euclidean-error EE-RANSAC with Horn's method), followed by Levenberg-Marquardt joint minimisation of the fixed feature associations and dense point-to-plane ICP terms (alpha = 0.5 or beta = 1000); a Two-Stage variant runs the ICP stage only when RANSAC inliers are at most phi; with fewer than gamma = 10 inliers the previous relative motion is used as a constant-velocity initial guess (Secs. 3.1.1-3.1.4)","hybrid: FAST keypoints with Calonder descriptors (OpenCV) matched and verified by RANSAC with a 2.0 pixel (re-projection) or 0.03 m (Euclidean) inlier threshold; dense nearest-neighbour point associations via a k-d tree on a source cloud downsampled by a factor of 10, normals from PCA of small neighbourhoods; feature associations are not recomputed during ICP (Secs. 3.1.1-3.1.4)","globally optimised camera trajectory, dense coloured point cloud map and a surfel map (for example 28 million points combined into 1.4 million surfels, Sec. 4.4); export format not_reported",[45],{"id":1529,"label":1530,"shortName":1531,"title":1532,"year":1461,"era":52,"cluster":527,"scope":132,"keyIdeaZh":1533,"sensors":1534,"mapRepresentation":1538,"loopClosure":1539,"estimator":1540,"association":1541,"deskew":1542,"outputGeometry":1543,"fulltextStatus":22,"lidarModels":1544,"equipmentCount":487},"cartographer2016","Hess et al., 2016","Cartographer","Real-time loop closure in 2D LIDAR SLAM","Cartographer 以背包式平台即時產生竣工平面圖：局部端把連續掃描以非線性最佳化對齊到小型子地圖（submap），誤差隨時間累積；全域端把已完成的子地圖與所有掃描做迴圈候選，以分支定界（branch-and-bound）加速的逐像素掃描匹配產生迴圈約束，再以稀疏位姿調整（SPA）定期最佳化。作者的貢獻在於降低迴圈約束計算成本，使數萬平方公尺樓層也能即時完成最佳化。",[1535,1536,1537],"horizontally mounted 2D LIDAR on the backpack (model not reported)","IMU on the backpack (model not reported)","Neato Robotics Revo LDS low-cost laser distance sensor (second experiment)","2D probability-grid submaps (5 cm resolution in the paper)","all finished submaps and scans considered; branch-and-bound scan matching in a search window adds loop constraints in real time","nonlinear least squares (Ceres) for scan-to-submap matching and sparse pose adjustment for global optimization","scan-to-submap matching on probability grids; branch-and-bound pixel-accurate search for loop-closure constraints","not_reported (IMU is used to project scans to the horizontal plane on the unstable backpack)","2D grid floor plan (5 cm resolution) and optimized poses",[1545,1482],"Revo LDS",{"id":1547,"label":1548,"shortName":1549,"title":1550,"year":9,"era":10,"cluster":397,"scope":69,"keyIdeaZh":1551,"sensors":1552,"mapRepresentation":1557,"loopClosure":1558,"estimator":1559,"association":1560,"deskew":121,"outputGeometry":1561,"fulltextStatus":22,"lidarModels":1562,"equipmentCount":24},"hinduja2019degeneracy","Hinduja et al., 2019","Degeneracy-aware factors","Degeneracy-Aware Factors with Applications to Underwater SLAM","本文把退化感知延伸到位姿圖：點對面 ICP 每次迭代以最大與最小特徵值的比值（條件數）作為動態門檻，只沿受約束方向更新（沿用 Zhang 等人的解重映射）；再把結果以部分迴圈閉合因子加入位姿圖，只約束 X、Y 與偏航，深度、俯仰與滾轉則交由深度計與航姿參考系統的先驗處理。作者以 DIDSON 聲納子地圖的水下建圖在模擬與實際資料驗證，顯示可抵抗導航漂移並拒絕退化環境中的不良迴圈閉合（Sec. III, IV）。",[1553,1554,1555,1556],"multibeam imaging sonar (DIDSON, 96 beams, profiling mode with a concentrator lens giving 1 degree vertical FOV)","DVL (Teledyne\u002FRDI Workhorse Navigator, 1.2 MHz)","AHRS with Honeywell HG1700 IMU","depth sensor (Paroscientific Digiquartz)","volumetric sonar submaps in a submap-based pose graph","ICP loop closures inserted as partial factors constraining only well-constrained directions","pose-graph optimization with partially constrained loop-closure factors","degeneracy-aware point-to-plane ICP (PCL-based) with solution remapping between sonar submaps","submap-based 3D reconstruction (point clouds) and poses",[],{"id":1564,"label":1565,"shortName":1566,"title":1567,"year":332,"era":52,"cluster":397,"scope":69,"keyIdeaZh":1568,"sensors":1569,"mapRepresentation":57,"loopClosure":72,"estimator":1571,"association":1572,"deskew":1573,"outputGeometry":1574,"fulltextStatus":22,"lidarModels":1575,"equipmentCount":185},"hong2010vicp","Hong et al., 2010","VICP","VICP: Velocity updating iterative closest point algorithm","一般 ICP 假設同一次掃描的點同時量測，但測距儀是逐點依序量測，快速運動時會產生掃描畸變並累積追蹤誤差。VICP 假設掃描期間速度固定，以前後兩次掃描的相對轉換估計測距儀的本體速度，據以校正每個點後重跑 ICP，再更新速度直到收斂；採用以最後一點為基準的反向補償，使位姿不因掃描時間而延遲，並依感測範圍剔除沒有對應的點。作者以 2D Hokuyo 測距儀在模擬與辦公室推車實驗中驗證（Sec. IV 至 VI）。",[1570],"2D laser rangefinder (Hokuyo URG-04LX, written HOKUYO URG04-LX)","ICP with velocity update","2D point-to-point closest-point ICP solved by SVD (Algorithm 1), with points outside the pie-shaped sensing area of the previous scan rejected as outliers","each point transformed with the constant body velocity estimated from the relative ICP transform (backward difference over the scan interval) and re-estimated in an outer loop until the velocity converges; backward compensation uses the last point as reference so the pose is not delayed by the scan time","2D rangefinder trajectory and motion-compensated scans",[1576],"Hokuyo URG-04LX",{"id":1578,"label":1579,"shortName":1580,"title":1581,"year":661,"era":10,"cluster":755,"scope":32,"keyIdeaZh":1582,"sensors":1583,"mapRepresentation":1586,"loopClosure":1451,"estimator":1587,"association":1588,"deskew":57,"outputGeometry":1589,"fulltextStatus":22,"lidarModels":1590,"equipmentCount":1593},"livgaussmap2024","Hong et al., 2024","LIV-GaussMap","LIV-GaussMap: LiDAR-Inertial-Visual Fusion for Real-Time 3D Radiance Field Map Rendering","LIV-GaussMap 以硬體同步的 LiDAR-慣性系統及尺寸自適應體素取得位姿與平面結構，將體素平面的共變異轉為高斯初始形狀，再用影像光度梯度精修球諧顏色與結構。作者在 FusionPortable 以 Chamfer、EMD 與 F-score 評估結構，並報告光度最佳化會使結構品質略為下降，顯示渲染與幾何之間的取捨。",[1584,36,1585],"3D LiDAR (Livox Avia; Ouster OS1-128; solid-state RealSense L515)","monocular camera (global shutter; rolling shutter on the L515)","surface Gaussians initialized from size-adaptive voxel plane covariances; spherical-harmonic colour","LiDAR-inertial odometry with size-adaptive voxel map provides poses; Gaussians refined by photometric gradients","LiDAR plane covariances for Gaussian initialization; photometric gradients for refinement","Gaussian map with renderings; structure evaluated against ground-truth point clouds (CD, EMD, F-score)",[437,1591,1592],"Ouster OS1-128","Intel RealSense L515",13,{"id":1595,"label":1596,"shortName":1597,"title":1598,"year":207,"era":10,"cluster":263,"scope":99,"keyIdeaZh":1599,"sensors":1600,"mapRepresentation":1602,"loopClosure":1451,"estimator":1603,"association":1604,"deskew":1605,"outputGeometry":1606,"fulltextStatus":22,"lidarModels":1607,"equipmentCount":377},"gslivo2025","Hong et al., 2025","GS-LIVO","GS-LIVO: Real-Time LiDAR, Inertial, and Visual Multisensor Fused Odometry With Gaussian Mapping","GS-LIVO 以三維高斯（3D Gaussians）取代傳統彩色點雲與稀疏區塊地圖：全域高斯地圖以空間雜湊索引的八元樹管理，只將視野內的高斯放入 GPU 上的滑動視窗即時最佳化，以控制顯示記憶體用量。高斯由光達點與影像聯合初始化，里程計沿用 FAST-LIVO2 的序列更新 IESKF，但視覺殘差改為渲染影像與實際影像的光度誤差。作者宣稱這是首個可在 Jetson Orin NX 嵌入式平台即時運作並線上更新地圖的高斯式 SLAM。",[35,36,1601],"camera","Planar 3D Gaussians initialized from LiDAR leaf voxels (normal from LiDAR, color by bilinear sampling) in a global hash-indexed octree in CPU RAM; Gaussians in the current FoV kept in a contiguous CPU buffer mirrored in GPU memory and optimized with Adam; root voxel 0.03 or 0.06 m indoors and 1.0 or 0.5 m outdoors with 2 subdivision levels; window of 100,000 Gaussians (desktop) or 20,000 (Orin NX)","iterated error-state Kalman filter with sequential updates, modified from FAST-LIVO2","LiDAR update with planar features of a size-adaptive voxel map (FAST-LIVO2 and VoxelMap-type LIO); visual update minimizes the photometric loss between the image rendered from Gaussians in the current FoV at the LiDAR-updated pose and the captured image, with Jacobians derived as in MonoGS and chained to the IMU pose inside the IESKF","Not described; the LiDAR-inertial update is taken from FAST-LIVO2 and size-adaptive voxel LIO ([57], [59] in T-RO) without re-describing motion compensation","Gaussian map with photorealistic rendering; 2D occupancy grid derived for navigation (Sec. III-D); point-cloud export not_reported",[45],{"id":1609,"label":1610,"shortName":1611,"title":1612,"year":67,"era":52,"cluster":1344,"scope":514,"keyIdeaZh":1613,"sensors":1614,"mapRepresentation":57,"loopClosure":57,"estimator":1615,"association":1616,"deskew":57,"outputGeometry":57,"fulltextStatus":22,"lidarModels":1617,"equipmentCount":61},"horn1987absolute","Horn, 1987","Horn absolute orientation","Closed-form solution of absolute orientation using unit quaternions","給定兩座標系中三個以上不共線的對應點，作者提出最小平方意義下的閉式解：平移為一組點的形心與另一組點經旋轉、縮放後之形心的差；若採作者建議的對稱誤差式，尺度為兩組點相對形心的均方根偏差之比，且不需先求旋轉；旋轉以單位四元數表示，為一個 4×4 對稱矩陣最大正特徵值對應的特徵向量。作者也指出，若某一組座標的精度遠高於另一組，採非對稱的尺度式可能較合適；附錄說明可加入權重以反映不同點的量測可信度。",[],"closed-form least squares (unit quaternion eigenvector)","known point correspondences",[],{"id":1619,"label":1620,"shortName":1621,"title":1622,"year":383,"era":52,"cluster":31,"scope":32,"keyIdeaZh":1623,"sensors":1624,"mapRepresentation":1629,"loopClosure":57,"estimator":1630,"association":1631,"deskew":121,"outputGeometry":1632,"fulltextStatus":22,"lidarModels":1633,"equipmentCount":109},"hornung2013octomap","Hornung et al., 2013","OctoMap","OctoMap: an efficient probabilistic 3D mapping framework based on octrees","OctoMap 以八元樹（octree）儲存體素的機率佔據值（log-odds），同時表示已佔據、空與未知空間；感測器原點到端點之間的射線更新為空，端點更新為佔據。作者加入機率上下限夾制（clamping）與樹狀剪枝壓縮，使地圖精簡並可做多解析度查詢。",[1625,1626,1627,1628],"[\"2D laser range finder on a pan-tilt unit (SICK LMS, FR-079 corridor)\", \"two fixed laser scanners sweeping to the left and right of the robot (New College Epoch C","models not stated)\", \"dense 3D laser scans (Freiburg campus, 81 scans","sensor model not stated","ranges up to 50 m)\", \"RGB-D camera (Microsoft Kinect, freiburg1_360 sequence)\"]","octree of voxels with log-odds occupancy, clamping thresholds, lossless pruning and multi-resolution queries","not_applicable (mapping with externally supplied poses: FR-079 odometry refined by 3D scan matching, New College trajectory from visual odometry (Sibley et al. 2009), freiburg1_360 aligned by RGB-D SLAM)","ray casting from sensor origin to endpoints with a 3D Bresenham-type voxel traversal; endpoints updated as occupied (l_occ = 0.85, p = 0.7) and traversed voxels as free (l_free = -0.4, p = 0.4); an endpoint voxel is never freed within the same sweep update","occupied\u002Ffree\u002Funknown voxel map at chosen resolution (map files); not a surface model",[1337,45],{"id":1635,"label":1636,"shortName":1637,"title":1638,"year":769,"era":52,"cluster":236,"scope":346,"keyIdeaZh":1639,"sensors":1640,"mapRepresentation":1643,"loopClosure":1644,"estimator":1645,"association":1646,"deskew":1647,"outputGeometry":1648,"fulltextStatus":22,"lidarModels":1649,"equipmentCount":487},"fovis2017","Huang et al., 2017","FOVIS","Visual Odometry and Mapping for Autonomous Flight Using an RGB-D Camera","本章提出供四旋翼無人機自主飛行使用的 RGB-D 視覺里程計，後來以 fovis 函式庫公開。演算法沿用立體視覺里程計的標準流程：灰階影像建立三層高斯金字塔，以自適應門檻的 FAST 角點擷取特徵並分格保留，從深度影像取得特徵深度；先以縮小影像直接估計初始旋轉，藉此限縮搜尋視窗，再以 9×9 影像塊描述子做雙向一致的匹配與次像素修正。內點以「剛體運動保持點間距離」建立一致性圖，並以貪婪法近似最大團挑選；位姿先以 Horn 絕對定向求解，再最小化重投影誤差，並以參考關鍵影格降低懸停時的漂移。里程計與 IMU 以 EKF 融合後在機上即時控制飛行；迴圈閉合與位姿圖最佳化則沿用作者先前的 RGB-D Mapping，在機外筆電執行，並建立 10 公分解析度的佔據體素地圖。",[1641,1642],"RGB-D camera: stripped-down Microsoft Kinect (PrimeSense), 640 x 480 RGB-D at 30 Hz","IMU on the vehicle, fused with the visual odometry in an EKF for control (not used inside the visual odometry)","Offboard 3D log-likelihood occupancy voxel grid at 10 cm resolution from depth downsampled to 128 x 96 (about 1.5 ms per frame); rendered point clouds; offline textured surfaces from sparse bundle adjustment","Not part of the visual odometry. In the paper's full system RGB-D Mapping [14] runs offboard: keyframes every 10 deg or 25 cm, candidates limited to 90 deg and 5 m pose difference and to the 15 best vocabulary-tree matches, RANSAC over FAST keypoints with Calonder descriptors (ratio 0.6, at least 10 inliers), and a two-frame sparse bundle adjustment of the relative pose","Frame-to-reference-keyframe motion from sparse 3D feature matches: Horn absolute orientation on the inliers, refined by nonlinear least-squares minimization of feature reprojection error (bidirectional, ESM), then refined again after discarding matches above a fixed reprojection threshold; the reference frame is replaced only when motion against it fails or has too few inliers. For flight control the VO is fused with IMU data in an EKF, and delayed SLAM corrections are applied retroactively to the state history.","FAST corners on a three-level Gaussian pyramid with an adaptive threshold and 80 x 80 pixel bucketing (25 strongest per bucket), depth read from the depth image; 80-byte descriptors from 9 x 9 intensity patches matched by sum of absolute differences with a mutual-consistency check inside a search window set by an image-based initial rotation estimate; sub-pixel refinement with ESM; inliers from a greedy approximation of the maximal clique of matches whose 3D distances are preserved","not_applicable (RGB-D camera; Kinect rolling shutter named as a limitation at higher speed, Sec. 5)","real-time 6-DoF pose and velocity estimates; occupancy voxel map used for path planning; rendered RGB-D point cloud; offline textured surface model",[],{"id":1651,"label":1652,"shortName":1653,"title":1654,"year":661,"era":10,"cluster":755,"scope":32,"keyIdeaZh":1655,"sensors":1656,"mapRepresentation":1658,"loopClosure":72,"estimator":1659,"association":1660,"deskew":57,"outputGeometry":1661,"fulltextStatus":22,"lidarModels":1662,"equipmentCount":77},"huang2024_2dgs","Huang et al., 2024a","2DGS","2D Gaussian Splatting for Geometrically Accurate Radiance Fields","2DGS 將三維體積壓縮為一組有方向的二維平面高斯圓盤，使基元在多視角下具一致的幾何，並加入深度失真與法向一致性正則化。網格以渲染深度圖經 TSDF 融合取得。作者在 DTU 以 Chamfer 距離、在 Tanks and Temples 以 F1 評估幾何，指出幾何與影像品質之間存在取捨。",[1657],"monocular camera (multi-view images)","2D oriented planar Gaussian disks (surfels) with perspective-correct ray-splat intersection","not_applicable (per-scene optimization; poses from COLMAP)","direct photometric loss with depth-distortion and normal-consistency regularization","mesh by TSDF fusion (Open3D, voxel size 0.004, truncation 0.02) of rendered median depth maps of the training views; median depth outperforms expected depth and screened Poisson reconstruction in the DTU ablation",[],{"id":1664,"label":1665,"shortName":1666,"title":1667,"year":661,"era":10,"cluster":98,"scope":99,"keyIdeaZh":1668,"sensors":1669,"mapRepresentation":1672,"loopClosure":1673,"estimator":1674,"association":1675,"deskew":1676,"outputGeometry":1677,"fulltextStatus":22,"lidarModels":1678,"equipmentCount":554},"loglio2024","Huang et al., 2024b","LOG-LIO","LOG-LIO: A LiDAR-Inertial Odometry With Efficient Local Geometric Information Estimation","LOG-LIO 在 FAST-LIO2 的迭代誤差狀態卡爾曼濾波架構上，加入即時的局部幾何資訊估計。作者提出 Ring FALS：預先依 LiDAR 的環編號與方位角建立方位向量查找表，新掃描到達時只需距離值即可以近似最小平方求得每點法向量，避免鄰域搜尋。地圖以擴充的 ikd-Tree 管理，每個體素節點遞增維護點分布的平均與共變異數，並在收斂後固定。資料關聯先做可見性與法向量一致性檢查，再依序嘗試大尺度面元、小尺度面元，最後才退回點到平面。",[1670,1671],"3D spinning LiDAR with ring index (Velodyne 32-beam in M2DGR; Ouster OS1 16-channel in NTU VIRAL)","9-axis IMU (VectorNav VN100 in NTU VIRAL)","ikd-Tree extended so that each node also stores an incrementally updated point distribution (mean and covariance) of its voxel; distributions are fixed once the Ring FALS normal and the distribution eigenvector agree within 20 deg or after 2 eta = 50 points (Sec. V-D, VI-A)","none (listed as future work, Sec. VII)","error-state iterated EKF adopted from FAST-LIO2, with the MAP update augmented by point-to-surfel residuals besides point-to-plane residuals (Eq. 13, Sec. V-C)","hierarchical: k nearest map points and voxels; visibility check (map normal versus ray) and normal consistency check (mean angle below 60 deg); then large merged surfel, else small fixed surfel of the voxel, else LOAM-style point-to-plane; surfel if planarity above 1.0 and lambda2\u002Flambda1 above 100 (Sec. III-E, V-B)","IMU backward propagation (FAST-LIO2 style) after normal estimation and voxel downsampling (Sec. V-A)","odometry and a voxelized point map with per-voxel normals and surfels",[1679,504],"Velodyne-32",{"id":1681,"label":1682,"shortName":1683,"title":1684,"year":661,"era":10,"cluster":755,"scope":132,"keyIdeaZh":1685,"sensors":1686,"mapRepresentation":1688,"loopClosure":1689,"estimator":1690,"association":1691,"deskew":57,"outputGeometry":1692,"fulltextStatus":22,"lidarModels":1693,"equipmentCount":109},"photoslam2024","Huang et al., 2024c","Photo-SLAM","Photo-SLAM: Real-Time Simultaneous Localization and Photorealistic Mapping for Monocular, Stereo, and RGB-D Cameras","Photo-SLAM 將 ORB-SLAM3 的特徵式定位、局部光束調整與迴圈閉合，與以高斯參數擴充的「超基元」地圖解耦結合，幾何由特徵點與因子圖負責，外觀由高斯潑濺負責。作者明言目標是沉浸式探索的精簡表示而非稠密網格，網格重建評估不在範圍內。",[37,1687,772],"stereo camera (EuRoC MAV dataset; hand-held ZED 2 outdoors)","hyper primitives: ORB map points extended with Gaussian parameters","ORB-SLAM3-style loop closure; keyframes and hyper primitives corrected by a similarity transformation","ORB-SLAM3-based factor graph (Levenberg-Marquardt) localization and local BA; SGD for photorealistic mapping","ORB feature correspondences (geometry); photometric loss (appearance)","sparse ORB points + Gaussians and renderings; mesh reconstruction explicitly out of scope (Sec. 4.1)",[],{"id":1695,"label":1696,"shortName":1697,"title":1698,"year":191,"era":52,"cluster":527,"scope":132,"keyIdeaZh":1699,"sensors":1700,"mapRepresentation":1702,"loopClosure":1703,"estimator":1704,"association":1705,"deskew":121,"outputGeometry":1335,"fulltextStatus":22,"lidarModels":1706,"equipmentCount":46},"hahnel2003_gridfastslam","Hähnel et al., 2003a","Grid-based FastSLAM with scan matching","An efficient FastSLAM algorithm for generating maps of large-scale cyclic environments from raw laser range measurements","本文把 Rao-Blackwellized 粒子濾波與雷射掃描匹配結合：每 k 步先以前 k-1 筆掃描與最近的里程計讀值做掃描匹配，得到修正後的里程量測並用於粒子取樣，再以第 k 筆掃描計算粒子權重，使每筆資料只使用一次。掃描匹配殘差以三參數誤差模型描述，參數由 Intel Research Lab 資料學得，因此取樣分布比原始里程計集中得多，所需粒子數與重取樣次數下降，也減輕粒子耗盡，使機器人能閉合大迴圈。每個粒子各有一張佔據網格地圖，但只用與其可視區域相交的有限掃描更新。",[1701,119],"2D laser range finder (SICK LMS)","2D occupancy grid per particle, updated from a limited number of scans that intersect the particle's visible area (constant-time approximation); 10 cm grid in the Sieg Hall run (Sec. III; Sec. IV.A)","implicit through the particle filter; the scan-matching correction reduces resampling operations and particle depletion so that large loops can be closed (Sec. I, III)","Rao-Blackwellized particle filter over robot paths with one grid map per particle; every k steps a scan-matching-corrected odometry measurement is computed from the k-1 previous scans and the k most recent odometry readings and used for sampling with a learned three-parameter error model, and the k-th scan weights the particles (Sec. III)","grid-based 2D scan matching of a scan against an occupancy grid built from previous measurements, using a beam-endpoint likelihood (for max-range readings the cell 20 cm before the end is assumed free) (Sec. III)",[1337],{"id":1708,"label":1709,"shortName":1710,"title":1711,"year":191,"era":52,"cluster":527,"scope":32,"keyIdeaZh":1712,"sensors":1713,"mapRepresentation":1715,"loopClosure":72,"estimator":1716,"association":1717,"deskew":1718,"outputGeometry":1719,"fulltextStatus":22,"lidarModels":1720,"equipmentCount":144},"hahnel2003_compact3d","Hähnel et al., 2003b","Compact 3D building models from mobile laser scanning","Learning compact 3D models of indoor and outdoor environments with a mobile robot","本文以移動機器人上的雷射測距儀建立室內外建物的精簡 3D 模型。室內機器人以水平雷射做 2D 掃描匹配求位姿，同時以朝上的雷射掃出 3D 結構；室外機器人則以裝在雲台上的單一雷射取得 3D 掃描，並以射線式機率模型對三角網格做 3D 配準。取得的網格全域一致但局部雜訊大，作者以隨機起點的區域成長找出大型平面，把點投影到平面後合併共面多邊形，同時保留門窗等非平面細節；為加速，先由各掃描抽出的線段角度直方圖產生平面候選，再以粗細兩階段的平面掃掠篩選。",[1714,346],"2D laser range finders: indoors a horizontal laser for 2D mapping plus an upward-pointing laser for 3D (SICK PLS used in the Wean Hall run; SICK LMS also named); outdoors one laser on a pan\u002Ftilt unit","triangle mesh from neighbouring scan points, simplified into planar polygons by randomized region-growing plane fitting (delta 30 cm, epsilon 2.8, gamma 10 cm) and merging of coplanar neighbouring polygons; plane candidates found from line-angle histograms and plane sweeps at 5 cm then 1 cm (Sec. 3)","incremental maximum-likelihood pose estimation by hill climbing: 2D scan-to-grid alignment that integrates small Gaussian pose errors; for 3D scans a beam likelihood (Gaussian plus uniform mixture approximated by triangular distributions) against a triangle-mesh model via ray tracing (Sec. 2.1-2.2)","no explicit correspondences; beams are ray-cast into the current grid (2D) or triangle mesh (3D), and max-range beams also contribute (Sec. 2.2)","not_reported (indoor robots moved at 10 cm\u002Fs to obtain adequate 3D point density)","compact 3D polygonal model: large planar polygons and quads plus residual triangles for non-planar regions (Table 2)",[1338,1337],{"id":1722,"label":1723,"shortName":1724,"title":1725,"year":51,"era":10,"cluster":755,"scope":99,"keyIdeaZh":1726,"sensors":1727,"mapRepresentation":1728,"loopClosure":1451,"estimator":1729,"association":1730,"deskew":1731,"outputGeometry":1732,"fulltextStatus":22,"lidarModels":1733,"equipmentCount":487},"loner2023","Isaacson et al., 2023","LONER","LONER: LiDAR Only Neural Representations for Real-Time SLAM","LONER 以點到平面 ICP（以單位矩陣為初始猜測，不使用 IMU）追蹤降採樣至 5 Hz 的 LiDAR 掃描，並在平行執行緒中以關鍵影格視窗（目前關鍵影格加上 7 個隨機選取的過去關鍵影格）聯合最佳化 MLP 與階層特徵格網地圖及關鍵影格位姿。其 JS 散度動態邊界損失依每條射線目前的學習程度調整目標分布寬度：尚未學好的區域用較大邊界以加快收斂，已學好的區域用較小邊界以保留細節。網格只在離線時以虛擬 LiDAR 與 marching cubes 產生；評估時將網格取樣為點雲、降採樣，並把真值裁切至感測器實際觀測範圍。",[35],"MLP with the hierarchical feature-grid encoding of ref. [26] (Instant-NGP multiresolution hash encoding) predicting volume density only, no colour (Sec. III-C)","point-to-plane ICP tracking (identity initial guess, no IMU) + joint optimization of MLP map and keyframe poses in a window","point-to-plane ICP for tracking; depth-supervised neural rendering with a JS-divergence dynamic-margin loss for mapping","constant-velocity motion compensation between scans (Sec. III-B)","mesh generated offline by virtual LiDAR rays + marching cubes (not part of online training)",[45],{"id":1735,"label":1736,"shortName":1737,"title":1738,"year":51,"era":10,"cluster":755,"scope":99,"keyIdeaZh":1739,"sensors":1740,"mapRepresentation":1741,"loopClosure":1451,"estimator":1742,"association":1743,"deskew":57,"outputGeometry":1744,"fulltextStatus":22,"lidarModels":1745,"equipmentCount":77},"eslam2023","Johari et al., 2023","ESLAM","ESLAM: Efficient Dense SLAM System Based on Hybrid Representation of Signed Distance Fields","ESLAM 以多尺度軸對齊特徵平面（tri-plane）取代體素網格，使記憶體隨場景邊長由立方成長降為平方成長，並直接解碼截斷符號距離場（TSDF）以加速收斂。作者承認特徵平面的更新可能影響已重建區域，因此需投入大量計算重播舊關鍵影格以避免遺忘。",[772],"multi-scale axis-aligned feature planes (tri-plane) with shallow decoders to TSDF and RGB","Adam gradient-based tracking of translation + quaternion; periodic joint mapping over keyframes","per-point TSDF (free-space and truncation) losses + depth and colour rendering losses","mesh via marching cubes on a 1 cm TSDF volume, with frustum\u002Focclusion mesh culling before evaluation",[],{"id":1747,"label":1748,"shortName":1749,"title":1750,"year":51,"era":10,"cluster":98,"scope":99,"keyIdeaZh":1751,"sensors":1752,"mapRepresentation":1755,"loopClosure":72,"estimator":1756,"association":1757,"deskew":1758,"outputGeometry":1759,"fulltextStatus":22,"lidarModels":1760,"equipmentCount":1764},"malio2023","Jung et al., 2023","MA-LIO","Asynchronous Multiple LiDAR-Inertial Odometry Using Point-Wise Inter-LiDAR Uncertainty Propagation","MA-LIO 處理多顆非同步、視野與掃描樣式不同的 LiDAR：先以 IMU 離散模型傳播位姿與共變異數，再以 B 樣條內插求得任一點取樣時刻的位姿，把各 LiDAR 的點去畸變並轉換到最後一顆 LiDAR 最新點的座標系，因此不需嚴格硬體同步也不依賴 LiDAR 間重疊。每個點依取樣時刻（狀態共變異數）與距離傳播出點級不確定度，用於加權點到平面殘差與決定是否存入 ikd-Tree 地圖；另以量測法向量奇異值比計算定位權重，在隧道或窄走廊等退化場景中提高 IMU 先驗的比重。狀態估計採迭代誤差狀態卡爾曼濾波。",[1753,1754],"multiple asynchronous 3D LiDARs of different makes and scan patterns (Ouster OS0-64 plus Livox; Velodyne HDL-32E, VLP-16 and LS-16C; Ouster OS2-128 plus Livox Avia and Livox Tele)","IMU (100 to 400 Hz; models not stated)","ikd-Tree point map storing only points whose propagated uncertainty trace is below a threshold, with downsampling that keeps low-uncertainty points near voxel centres (Sec. II-G)","iterated error-state Kalman filter on manifold (FAST-LIO2 style) whose state includes each LiDAR-IMU extrinsic; measurement residuals weighted by point-wise uncertainty rescaled with fixed interval conversion, and a localization weight from the singular values of measurement normals that shifts weight to the IMU prior in degenerate scenes (Sec. II-A, II-E, II-F)","direct point-to-plane: five nearest neighbours in the ikd-Tree define a local plane whose covariance is a weighted sum of neighbour point covariances (Sec. II-E)","B-spline interpolated IMU poses undistort every LiDAR's points to its latest point time and then compensate inter-LiDAR temporal offsets by transforming them to the latest point of the latest LiDAR (Sec. II-C)","odometry and a merged multi-LiDAR point map (Figs. 1, 6)",[224,1482,225,143,1761,1762,437,1763],"LS-16C (written LS-C16 in Table VI)","Ouster OS2-128","Livox Tele",15,{"id":1766,"label":1767,"shortName":1768,"title":1769,"year":282,"era":52,"cluster":53,"scope":54,"keyIdeaZh":1770,"sensors":1771,"mapRepresentation":1774,"loopClosure":1775,"estimator":1776,"association":1777,"deskew":57,"outputGeometry":1778,"fulltextStatus":22,"lidarModels":1779,"equipmentCount":46},"kaess2008isam","Kaess et al., 2008","iSAM","iSAM: Incremental Smoothing and Mapping","iSAM 將 SLAM 表述為平滑（smoothing）問題並保留整條軌跡，使資訊矩陣維持自然稀疏；相對地，濾波在邊際化位姿時會使資訊矩陣變稠密。方法以增量更新平方根資訊矩陣（QR 分解）的方式，只重算受新量測影響的項目。遇到迴圈造成填充（fill-in）時，採週期性變數重排序並重新分解；另提供由分解因子有效取得邊際協方差的演算法，以支援即時資料關聯。",[1772,1773],"laser range data (Victoria Park: tree landmarks from a simple tree detector; Intel: pose constraints from scan matching; MIT Killian Court: data preprocessed into pose constraints)","vehicle odometry (Victoria Park)","landmarks or pose-only graph","handles loops in the trajectory; fill-in controlled by periodic reordering","incremental QR update of the square-root information matrix by Givens rotations (new rows eliminated, new variables appended); periodic block COLAMD variable reordering followed by full refactorization every 100 steps (every 20 steps for the Intel dataset); relinearization performed only at these reordering steps; OCaml implementation with automatic differentiation","maximum likelihood data association: Mahalanobis-distance cost matrix solved as a minimum-cost assignment by the Jonker-Volgenant-Castanon algorithm, using marginal covariances recovered from the square-root factor either exactly (dynamic programming over non-zeros of R) or conservatively (initial landmark uncertainty); nearest neighbour evaluated for comparison; pose-only experiments assume known correspondences","full trajectory and landmark map with access to marginal covariances (Victoria Park map has 140 distinct landmarks); for Intel and Killian Court the figures show the final trajectory with an evidence grid map (Figs. 10b, 11b)",[45],{"id":1781,"label":1782,"shortName":1783,"title":1784,"year":363,"era":52,"cluster":53,"scope":54,"keyIdeaZh":1785,"sensors":1786,"mapRepresentation":1787,"loopClosure":1788,"estimator":1789,"association":1790,"deskew":57,"outputGeometry":1791,"fulltextStatus":22,"lidarModels":1792,"equipmentCount":185},"kaess2012isam2","Kaess et al., 2012","iSAM2","iSAM2: Incremental smoothing and mapping using the Bayes tree","本文提出 Bayes tree 資料結構，把稀疏矩陣分解與圖模型推論連結起來，並據此發展 iSAM2。新量測加入時，iSAM2 只移除並重新消去受影響的樹頂部團（clique），再把未受影響的子樹接回；變數排序以約束式 CCOLAMD 增量進行，把最近存取的變數推向樹根。「流動式重線性化」（fluid relinearization）只在變數增量超過門檻 β 時才更新線性化點，部分狀態更新則在解的變化小於門檻 α 時停止回代。因此 iSAM2 不再需要 iSAM 的週期性批次步驟。作者在模擬與真實的 2D 位姿圖、2D 地標資料集及模擬 3D 位姿圖上，與 iSAM1、HOG-Man、SPA 比較每步計算時間與正規化 χ2。",[],"pose graph and optional landmarks","not_applicable (processes loop-closure factors provided by the front-end)","incremental Gauss-Newton on a factor graph via the Bayes tree: cliques affected by new factors or by relinearization are removed and re-eliminated (incomplete Cholesky within cliques) and orphaned sub-trees re-attached; incremental constrained COLAMD ordering forces recently accessed variables to the root; fluid relinearization when a variable's delta exceeds beta; partial state update stops back-substitution where changes fall below alpha; exponential-map retraction for 3D rotations","not_applicable (back-end; data association supplied externally)","incrementally updated trajectory and landmark estimates",[45],{"id":1794,"label":1795,"shortName":1796,"title":1797,"year":97,"era":10,"cluster":11,"scope":12,"keyIdeaZh":1798,"sensors":1799,"mapRepresentation":1802,"loopClosure":1803,"estimator":1804,"association":1805,"deskew":1806,"outputGeometry":1807,"fulltextStatus":22,"lidarModels":1808,"equipmentCount":109},"kayhani2022tagvio","Kayhani et al., 2022","Tag-based VIO for indoor construction UAVs","Tag-based visual-inertial localization of unmanned aerial vehicles in indoor construction environments using an on-manifold extended Kalman filter","作者為低成本商用無人機提出以平面標籤輔助的視覺慣性定位。AprilTag 的尺寸、編號與在 BIM 座標系中的位姿事先已知，濾波器以機上里程計提供的平移與旋轉速度做預測，並直接把每個偵測到的標籤四個角點的像素座標當作量測來修正，而不是使用偵測器輸出的相機對標籤位姿；狀態以 SE(3) 表示並在流形上以擴展卡爾曼濾波傳遞不確定性。作者也建立可由 BIM 產生施工場景與標籤的 Parrot-Sphinx 與 Gazebo 模擬環境，並在 Vicon 實驗室及模擬中評估位置 RMSE。",[1800,1801],"forward-looking monocular camera of the Parrot Bebop 2 (rectified 856 x 480 at about 30 Hz)","onboard IMU and odometry velocities of the Bebop 2 (about 5 Hz, upsampled to the image rate)","no map is built; tag poses in the BIM reference frame serve as landmarks","none (global tag measurements bound drift)","On-manifold extended Kalman filter on SE(3) with left perturbation; prediction from onboard translational and rotational velocities, correction from pixel coordinates of the four corners of each detected AprilTag whose pose is known in the BIM frame (Sec. 4)","AprilTag detection and ID decoding (AprilRobotics implementation) gives explicit tag-corner correspondences; no natural-feature matching","not_applicable (camera and IMU only)","6-DoF UAV pose and covariance in the BIM frame; no point cloud",[],{"id":1810,"label":1811,"shortName":1812,"title":1813,"year":383,"era":52,"cluster":31,"scope":32,"keyIdeaZh":1814,"sensors":1815,"mapRepresentation":1816,"loopClosure":57,"estimator":57,"association":57,"deskew":57,"outputGeometry":1817,"fulltextStatus":22,"lidarModels":1818,"equipmentCount":77},"kazhdan2013screened","Kazhdan & Hoppe, 2013","Screened Poisson Surface Reconstruction","Screened poisson surface reconstruction","此版本在原泊松重建中加入點位置的軟約束（screening term），使重建等值面更貼近輸入點，以減輕原方法的過度平滑；因約束只定義在稀疏點集上，線性系統的稀疏結構不變，仍可用多重網格（multigrid）求解，並透過演算法改良使時間複雜度對點數呈線性。作者另支援 Neumann 邊界條件，讓缺資料區的曲面可延伸到定義域邊界，而非強制封閉。",[],"implicit function on an adaptive octree with point-value (screening) constraints","triangle mesh isosurface; Dirichlet (closed) or Neumann (open to domain boundary) boundary behavior",[],{"id":1820,"label":1821,"shortName":1822,"title":1822,"year":706,"era":52,"cluster":31,"scope":32,"keyIdeaZh":1823,"sensors":1824,"mapRepresentation":1825,"loopClosure":57,"estimator":57,"association":57,"deskew":57,"outputGeometry":1826,"fulltextStatus":22,"lidarModels":1827,"equipmentCount":61},"kazhdan2006poisson","Kazhdan et al., 2006","Poisson Surface Reconstruction","作者指出定向點（oriented points）的法向量可視為實體指示函數（indicator function）梯度的取樣，於是把表面重建轉為泊松方程式求解，再擷取等值面成為封閉網格。解法一次考慮全部點，不需啟發式分區與混合，因而對雜訊具韌性；並在八元樹上以局部支撐基底形成稀疏且條件良好的線性系統，另處理非均勻取樣。",[],"implicit indicator function on an adaptive octree","watertight triangle mesh (isosurface of the indicator function)",[],{"id":1829,"label":1830,"shortName":1831,"title":1832,"year":661,"era":10,"cluster":755,"scope":99,"keyIdeaZh":1833,"sensors":1834,"mapRepresentation":1835,"loopClosure":72,"estimator":1836,"association":1837,"deskew":57,"outputGeometry":1838,"fulltextStatus":22,"lidarModels":1839,"equipmentCount":144},"splatam2024","Keetha et al., 2024","SplaTAM","SplaTAM: Splat, Track & Map 3D Gaussians for Dense RGB-D SLAM","SplaTAM 以等向性、顏色不隨視角變化的三維高斯為唯一地圖。追蹤時固定高斯，只在剪影值大於 0.99 的已充分觀測像素上，以深度 L1 與權重減半的顏色 L1 最佳化位姿（以等速模型初始化）；建圖時依剪影與深度誤差新增高斯，並以重疊度最高的關鍵影格更新地圖。高斯初始位置直接取自 RGB-D 深度反投影，故幾何來源是量測深度。全文與補充資料都沒有三維表面精度指標，幾何只以渲染深度 L1 評估（ScanNet++ 新視角 2.07 cm、訓練視角 1.28 cm）。Replica、TUM-RGBD 與 ScanNet 的基準數值取自 Point-SLAM 論文而非重跑；在 TUM-RGBD 上平均 ATE 5.48 cm，仍不及特徵式 ORB-SLAM2 的 1.98 cm。在 RTX 3080 Ti 上每影格追蹤約 1.00 s、建圖約 1.44 s，並非即時。",[772],"isotropic 3D Gaussians with view-independent colour","gradient-based camera pose optimization through differentiable splatting (constant-velocity initialization); map update over overlapping keyframes","direct L1 depth + colour rendering losses on silhouette-masked (well-observed) pixels","Gaussian map with rendered RGB and depth; geometry assessed only through rendered depth L1 against ground-truth depth (ScanNet++ 2.07 cm on novel views, 1.28 cm on training views); the full paper and supplement contain no 3D surface metric",[],{"id":1841,"label":1842,"shortName":1843,"title":1844,"year":383,"era":52,"cluster":31,"scope":99,"keyIdeaZh":1845,"sensors":1846,"mapRepresentation":1848,"loopClosure":1849,"estimator":1850,"association":1851,"deskew":1852,"outputGeometry":1853,"fulltextStatus":22,"lidarModels":1854,"equipmentCount":109},"keller2013pointfusion","Keller et al., 2013","Point-based fusion","Real-Time 3D Reconstruction in Dynamic Scenes Using Point-Based Fusion","此系統全程只用一個扁平的點（surfel）清單表示場景，每點存位置、法向量、半徑、信心計數與時間戳，不建立體素或其他空間資料結構。每個影格先以三層階層式稠密 ICP 將深度圖對齊到由模型點渲染出的深度圖來估計 6DoF 位姿，再把模型點渲染成超取樣索引圖做投影式資料關聯；對應點依距影像中心遠近的高斯信心加權平均融合，信心累積到門檻 10 才由不穩定轉為穩定。系統另移除違反自由空間的點，並從 ICP 找不到對應的像素出發做區域成長，把移動物體整體標為動態並排除於位姿估計之外。",[1847],"[\"RGB-D camera (Microsoft Kinect, near mode, 640x480 depth)\", \"time-of-flight camera (PMD CamBoard, 200x200, per-pixel amplitude used for confidence)\"]","flat list of points\u002Fsurfels with position, normal, radius and confidence counter; unstable-to-stable status; no spatial data structure","none (authors state sensor drift is not tackled; loop closure is future work, Sec. 8)","frame-to-model dense hierarchical (pyramid) ICP against the rendered model point map, estimating a 6-DoF pose before fusion","projective data association by rendering the global point model as an index map","not_applicable (each depth map is registered with a single 6DoF camera pose; the paper does not discuss shutter type or intra-frame motion)","fused point (surfel) model with position, normal, radius, confidence and timestamp, optionally the last RGB sample per point; visualised by opaque surface splatting; no mesh extraction",[],{"id":1856,"label":1857,"shortName":1858,"title":1859,"year":51,"era":10,"cluster":755,"scope":32,"keyIdeaZh":1860,"sensors":1861,"mapRepresentation":1862,"loopClosure":72,"estimator":1863,"association":1864,"deskew":57,"outputGeometry":1865,"fulltextStatus":22,"lidarModels":1866,"equipmentCount":185},"kerbl2023_3dgs","Kerbl et al., 2023","3DGS","3D Gaussian Splatting for Real-Time Radiance Field Rendering","3DGS 以具各向異性共變異的三維高斯基元表示場景，並以可微分的分塊光柵化（tile-based rasterization）直接由影像誤差最佳化其位置、形狀、不透明度與球諧顏色。初始化依賴 SfM 相機與稀疏點雲。其目標是即時新視角渲染，而非量測等級的表面幾何。",[1657],"explicit anisotropic 3D Gaussians (position, covariance, opacity, spherical-harmonic colour)","not_applicable (gradient-based optimization of Gaussian parameters with adaptive density control; cameras from SfM)","direct photometric loss via differentiable tile-based rasterization","rendered images and a set of 1 to 5 million anisotropic Gaussians; no surface or mesh extraction; the authors list mesh reconstruction from the Gaussians as future work",[],{"id":1868,"label":1869,"shortName":1870,"title":1871,"year":383,"era":52,"cluster":236,"scope":132,"keyIdeaZh":1872,"sensors":1873,"mapRepresentation":1874,"loopClosure":1875,"estimator":1876,"association":1877,"deskew":57,"outputGeometry":1878,"fulltextStatus":22,"lidarModels":1879,"equipmentCount":144},"dvoslam2013","Kerl et al., 2013","DVO-SLAM","Dense visual SLAM for RGB-D cameras","DVO-SLAM 以稠密方式對齊 RGB-D 影像，同時最小化所有像素的光度誤差與深度誤差，並以雙變量 t 分布自動調整兩項誤差的權重，降低離群值影響。系統採影格對關鍵影格追蹤，以位姿估計共變異數的熵比值決定何時建立新關鍵影格，並用同一熵比值驗證以距離搜尋得到的迴圈閉合候選。所有成功的匹配組成位姿圖，以 g2o 最佳化以修正累積漂移。",[1405],"pose graph of keyframes (RGB-D images); a point cloud can be formed from the optimised trajectory (Fig. 1 text; Sec. IV-C)","metric nearest-neighbour search for keyframes within a sphere around the current keyframe, validated by the entropy ratio at coarse and then higher resolution; extra loop search for every keyframe at the end of a sequence (Sec. IV-B, IV-C)","dense frame-to-keyframe alignment minimising a bivariate t-distributed photometric and depth error by iteratively reweighted Gauss-Newton on se(3), coarse-to-fine over three resolutions up to 320 x 240; covariance from the inverse approximate Hessian; keyframe pose graph optimised with g2o (Secs. III, IV-C)","direct: all pixels warped with the depth map (photometric residual plus depth residual equivalent to point-to-plane ICP with projective lookup); no feature matching (Sec. III-D)","optimised keyframe trajectory; the corrected trajectory allows a consistent point cloud model of the scene, while high-quality 3D model generation is left as future work (Secs. I, IV-C, VI)",[],{"id":1881,"label":1882,"shortName":1883,"title":1884,"year":396,"era":10,"cluster":263,"scope":99,"keyIdeaZh":1885,"sensors":1886,"mapRepresentation":1891,"loopClosure":1892,"estimator":1893,"association":1894,"deskew":1895,"outputGeometry":1896,"fulltextStatus":22,"lidarModels":1897,"equipmentCount":487},"compslam2020","Khattak et al., 2020","CompSLAM","Complementary Multi-Modal Sensor Fusion for Resilient Robot Pose Estimation in Subterranean Environments","CompSLAM 的 ICUAS 版本以鬆耦合方式，把視覺慣性里程計（ROVIO）或熱影像慣性里程計（ROTIO，使用完整輻射溫度影像）接到 LOAM 式 LiDAR 里程計與建圖：相機里程計在新點雲到達時提供掃描對掃描配準的初始值，並以 J^T J 的特徵值判斷掃描對掃描與掃描對地圖配準是否退化；一旦退化，就改用相機里程計的相對位移延續 LiDAR 位姿，並把當前點雲寫入地圖。相機里程計本身以共變異數成長的 D-optimality 與運動界限做健康檢查。熱影像在黑暗與粉塵中仍能提供約束，是本文與僅用可見光相機融合的主要差別。",[1887,1888,1889,1890],"3D LiDAR (Velodyne PuckLITE on the underpass UAV; Ouster OS1-64 in the mine deployment)","IMU (VectorNav VN-100)","visual camera (FLIR Blackfly with shutter-synchronized LEDs) for VIO","LWIR thermal camera (FLIR Tau2, full radiometric imagery) for TIO","LOAM point-cloud map; in ill-conditioned scan-to-map steps the current cloud is inserted with the prior mapping estimate plus the camera-odometry relative transform (Sec. III)","none reported in the ICUAS paper","loosely coupled, degeneracy-aware: ROVIO-based visual-inertial odometry, or its thermal variant on full radiometric thermal images (ROTIO), supplies the relative motion between successive point clouds as the prior for LOAM scan-to-scan matching; degeneracy of scan-to-scan and scan-to-map matching is detected from the eigenvalues of J^T J, and in degenerate steps the previous LiDAR odometry or mapping estimate is propagated with the camera-odometry relative transform; camera odometry is health-checked by the relative growth of its covariance (D-optimality) and by motion bounds (Sec. III, Eqs. 1-4, Fig. 4)","LOAM point-to-line and point-to-plane correspondences for scan-to-scan and scan-to-map matching (Sec. III)","not_reported in the ICUAS paper","robot pose and LiDAR point-cloud map (Figs. 5-7)",[1898,223],"Velodyne PuckLITE",{"id":1900,"label":1901,"shortName":1902,"title":1903,"year":150,"era":10,"cluster":599,"scope":600,"keyIdeaZh":1904,"sensors":1905,"mapRepresentation":1907,"loopClosure":1908,"estimator":57,"association":1909,"deskew":121,"outputGeometry":57,"fulltextStatus":22,"lidarModels":1910,"equipmentCount":24},"scancontext2018","Kim & Kim, 2018","Scan Context","Scan Context: Egocentric Spatial Descriptor for Place Recognition Within 3D Point Cloud Map","Scan Context 以感測器為中心，將單次 3D 光達掃描劃分為 20 個環（ring）乘 60 個扇區（sector）的極座標格網（最大距離 80 m），每格記錄其中點的最大高度，形成 2D 全域描述子，不依賴直方圖或事前訓練。搜尋分兩階段：先以各環佔有率組成的環鍵（ring key）建 kd-tree 取出 10 或 50 個候選，再對候選做所有欄位平移的逐欄餘弦距離比較，最小距離低於門檻即判定為迴圈；平移量同時給出約 6° 解析度的偏航初值，可供 ICP 使用，因此反向重訪與轉角也能偵測迴圈。",[1906],"3D LiDAR (Velodyne HDL-64E on KITTI, HDL-32E on NCLT, two tilted VLP-16 merged on Complex Urban LiDAR)","per-keyframe 2D descriptor database","provides loop candidates with coarse yaw alignment (6 deg resolution); the yaw shift initialises point-to-point ICP, which reduced ICP time and RMSE for KITTI 08 reverse loops (Fig. 7, 8); pose-graph use left to host SLAM","egocentric polar grid Nr = 20 rings x Ns = 60 sectors, Lmax = 80 m, bin value = maximum point height (Eq. 3), empty bins 0; optional root-shift augmentation with Ntrans = 8 translated copies for lane-level offsets; ring key = per-ring occupancy ratio (L0 norm) indexed in a KD tree, 10 or 50 candidates; column-wise cosine distance minimised over all column shifts, accepted below threshold tau; 0.6 m grid downsampling",[161,225,143],{"id":1912,"label":1913,"shortName":1914,"title":1915,"year":396,"era":10,"cluster":599,"scope":32,"keyIdeaZh":1916,"sensors":1917,"mapRepresentation":1918,"loopClosure":57,"estimator":57,"association":1919,"deskew":121,"outputGeometry":1920,"fulltextStatus":22,"lidarModels":1921,"equipmentCount":185},"removert2020","Kim & Kim, 2020","Removert","Remove, then Revert: Static Point cloud Map Construction using Multiresolution Range Images","Removert 以多解析度距離影像（range image）比較查詢掃描與含動態點的累積地圖：先保守地只保留確定的靜態點，再逐步放大查詢與地圖的關聯視窗，把被誤刪的靜態點「回復」（revert），藉此隱式補償位姿估計與配準誤差。方法離線處理，輸入為任一光達里程計或 SLAM 輸出的掃描與位姿。",[35],"static point-cloud map (dynamic points separated)","visibility check in range images: the map (per query frame, using its SE(3) pose) and the query scan are projected to fixed-resolution range images keeping the minimum range per pixel; a map point is marked dynamic when the query-minus-map range difference exceeds a range-adaptive threshold (tau_D times range); batch voting over N randomly ordered scans gives a staticity score (alpha_SM = 0.3, alpha_DM = -0.7, tau_S = -0.1); removal at the finest resolution (vertical FOV \u002F number of rays, 0.4 deg on KITTI) for three steps, then seven revert iterations coarsening by 0.1 deg per iteration","static point-cloud map and parsed dynamic points (README)",[45],{"id":1923,"label":1924,"shortName":1925,"title":1926,"year":97,"era":10,"cluster":599,"scope":132,"keyIdeaZh":1927,"sensors":1928,"mapRepresentation":1930,"loopClosure":1931,"estimator":1932,"association":1933,"deskew":1934,"outputGeometry":1935,"fulltextStatus":22,"lidarModels":1936,"equipmentCount":185},"ltmapper2022","Kim & Kim, 2022","LT-mapper","LT-mapper: A Modular Framework for LiDAR-based Lifelong Mapping","LT-mapper 把長期建圖拆成三個模組：LT-SLAM 以錨節點（anchor node）多時段位姿圖與 Scan Context 跨時段迴圈，對齊原點不同且各自漂移的時段；LT-removert 先移除高動態點，再以集合差分偵測低動態變化，分為新出現（PD）與消失（ND）的點；LT-map 維護最新狀態的即時地圖與持久結構的 meta map，並可串接差異地圖重建任一時間點的地圖。只需單一光達（IMU 可選）。",[35,1929],"IMU (optional, for initial odometry)","keyframe point clouds; live map and meta map; delta maps of positive and negative changes","Scan Context intra- and inter-session loops","multi-session pose-graph optimization with anchor nodes (iSAM2 in GTSAM)","Scan Context inter-session loop candidates verified by ICP between keyframe submaps (loops accepted only with low ICP fitness score, which also sets an adaptive covariance) with a robust back end, followed by radius-search loops by pose proximity; high-dynamic points removed with Removert; low-dynamic changes found by a kd-tree test of whether a point has k target points within r m; weak negative differences reverted by a modified Removert (Sec. IV-A, IV-B)","not_reported (delegated to input odometry, e.g., LIO-SAM)","aligned multi-session point clouds, static maps, positive\u002Fnegative change point sets, maps at any timestamp via change composition",[45],{"id":1938,"label":1939,"shortName":1940,"title":1941,"year":97,"era":10,"cluster":599,"scope":600,"keyIdeaZh":1942,"sensors":1943,"mapRepresentation":1907,"loopClosure":1945,"estimator":1946,"association":1947,"deskew":121,"outputGeometry":57,"fulltextStatus":22,"lidarModels":1948,"equipmentCount":554},"scancontextpp2022","Kim et al., 2022b","Scan Context++","Scan Context++: Structural Place Recognition Robust to Rotation and Lateral Variations in Urban Environments","Scan Context++ 擴充原 Scan Context，提出極座標的 Polar Context（處理航向旋轉）與直角座標的 Cart Context（處理側向平移）兩種描述子。流程分三段：以檢索鍵（retrieval key）建立 kd-tree 做地點檢索，以對齊鍵（aligning key）做 1 自由度半度量定位，最後以完整描述子比對剔除誤判；並以描述子增強同時應對旋轉與側移。方法假設橫滾與俯仰變化不劇烈。",[35,1944],"radar (extension discussed)","provides loop candidates plus 1-DoF initial alignment for ICP in a keyframe pose graph","not_applicable (integration example uses iSAM2 pose graph in SC-LeGO-LOAM, Sec. VII-F)","Polar Context (20 x 60 bins over 0 to 80 m and 360 deg) and Cart Context (40 x 40 bins over -100 to 100 m by -40 to 40 m) with maximum-height bins after 0.5 m voxel downsampling; retrieval key = L1 norm per row (the IROS 2018 version used an L0 occupancy ratio) in a single k-d tree with k = 1 candidate; aligning key (L1 per column) gives the column shift by L2 matching; full-descriptor cosine distance only at that shift for verification; augmentation by +\u002F-2 m lateral root shifts (A-PC) or a double flip (A-CC)",[161,223,225,45],{"id":1950,"label":1951,"shortName":1952,"title":1953,"year":513,"era":52,"cluster":236,"scope":99,"keyIdeaZh":1954,"sensors":1955,"mapRepresentation":1956,"loopClosure":1957,"estimator":1958,"association":1959,"deskew":57,"outputGeometry":1960,"fulltextStatus":22,"lidarModels":1961,"equipmentCount":185},"ptam2007","Klein & Murray, 2007","PTAM","Parallel Tracking and Mapping for Small AR Workspaces","PTAM 將相機追蹤（tracking）與建圖（mapping）拆成兩個平行執行緒：追蹤執行緒以地圖點重投影估計每張影像的位姿，建圖執行緒則對關鍵影格（keyframe）執行計算量較大的光束法平差（bundle adjustment, BA）。此設計讓即時系統可以使用原本多用於離線 SfM 的批次最佳化。作者將其定位為小型擴增實境工作區，並未支援大範圍探索。",[37],"sparse point-feature map with keyframes","no dedicated loop-detection module; authors state the system is not designed to close large loops, although the BA mapping can absorb loops when the camera is placed near the map boundary (Sec. 8)","tracking thread: pose update by ten iterations of reweighted least squares on a Tukey-biweight reprojection objective; mapping thread: Levenberg-Marquardt bundle adjustment with a Tukey M-estimator, run globally or locally over the newest keyframe and its four nearest keyframes (Sec. 5.4, 6.3)","FAST-10 corners on a four-level image pyramid; affine-warped 8x8 patches searched by zero-mean SSD at FAST corners around reprojected map points, 50 coarse-level points then up to 1000 points per frame; new map points triangulated by epipolar search against the nearest keyframe (Sec. 5.1, 5.3, 5.5, 6.2)","keyframe poses and sparse 3D point features",[],{"id":1963,"label":1964,"shortName":1965,"title":1966,"year":235,"era":52,"cluster":236,"scope":99,"keyIdeaZh":1967,"sensors":1968,"mapRepresentation":1972,"loopClosure":1973,"estimator":1974,"association":1975,"deskew":1976,"outputGeometry":1977,"fulltextStatus":22,"lidarModels":1978,"equipmentCount":24},"chisel2015","Klingensmith et al., 2015","CHISEL","Chisel: Real Time Large Scale 3D Reconstruction Onboard a Mobile Device using Spatially Hashed Signed Distance Fields","CHISEL 在 Google Tango 手機與平板上，只用行動裝置的 CPU 即時建立房屋尺度（300 平方公尺以上）的 TSDF 稠密重建，不使用 GPU 通用運算。作者採用 Nießner 等人的空間雜湊兩層結構，把 16×16×16 體素的區塊（chunk）依視錐裁剪結果動態配置，未被更新的區塊即回收，只處理含有表面的空間以節省記憶體與運算。針對 Tango 深度感測器雜訊大、更新率只有 3 至 6 Hz 的問題，加入依雜訊模型調整的動態截斷距離與空間雕刻（space carving）。定位以裝置內建的視覺慣性里程計為輸入，再以掃描對 TSDF 模型的 ICP 修正短距漂移；網格則以增量 marching cubes 依區塊延遲產生，系統本身沒有迴圈閉合。",[1969,1970,1971],"Google Tango 'Peanut' phone: projective depth sensor (6 Hz), 120 deg wide-angle tracking camera (60 Hz), 4 MP colour camera (30 Hz), six-axis gyroscope and accelerometer","Google Tango 'Yellowstone' tablet: projective depth sensor (3 Hz), same tracking camera, 4 MP colour camera (30 Hz)","Kinect RGB-D data of the Freiburg (TUM) benchmark for memory experiments","dynamic spatially hashed TSDF: chunks of 16 x 16 x 16 voxels in a hash map, allocated on frustum intersection and garbage-collected when not updated; each voxel stores a 16-bit fixed-point SDF and 16-bit weight plus 8-bit RGB and colour weight; dynamic truncation from a trained depth-noise model; space carving","none online; Fig. 7 shows a corridor corrected only after offline bundle adjustment","Onboard visual-inertial odometry (EKF fusing wide-angle camera and inertial data with 2D feature tracking at 60 Hz; the paper refers to Kottas et al. and the MSCKF for details) with sparse keypoint mapping used as a black box, incrementally corrected by scan-to-model ICP against the TSDF (residual = TSDF value at each transformed point, gradient by central differences); corrective transforms accumulated per scan","Implicit point-to-TSDF association via the signed distance value and its gradient (first-order projection onto the zero level set)","not_applicable (depth camera)","per-chunk triangle meshes from incremental marching cubes (lazy, asynchronous), coloured by trilinear interpolation; map savable to disk; 2 to 3 cm voxels on the devices",[],{"id":1980,"label":1981,"shortName":1982,"title":1983,"year":1217,"era":52,"cluster":527,"scope":99,"keyIdeaZh":1984,"sensors":1985,"mapRepresentation":1986,"loopClosure":1987,"estimator":1988,"association":1989,"deskew":1990,"outputGeometry":1991,"fulltextStatus":22,"lidarModels":1992,"equipmentCount":109},"hector2011","Kohlbrecher et al., 2011","Hector SLAM","A flexible and scalable SLAM system with full 3D motion estimation","Hector SLAM 結合以 LiDAR 為主的 2D 掃描對地圖（scan-to-map）匹配與以 IMU 為主的 3D 姿態估計：先用估計姿態把掃描轉到穩定座標系，再以 Gauss-Newton 在雙線性內插的佔據網格上求位姿，並用多解析度網格降低陷入局部極小的風險。作者表示在所考慮的小尺度情境中精度足以不需明確的迴圈閉合，並以手持系統在救援競技場與新建築中建圖。",[1330,36],"multi-resolution 2D occupancy grids, used optionally, each coarser level at half the resolution and all levels updated simultaneously from the estimated poses, with matching started at the coarsest level (examples: 20, 10 and 5 cm cells in Fig. 3; 0.25 m grid in the USV map, Fig. 5b) (Sec. IV-C)","none (authors report that explicit loop closing was not required in the considered small-scale scenarios)","Gauss-Newton scan-to-map matching on bilinearly interpolated multi-resolution occupancy grids (coarse to fine), with match covariance from the approximate Hessian; 6-DoF EKF navigation filter at 100 Hz with gyro and accelerometer bias states, fused with the SLAM pose by covariance intersection; the EKF pose projected to the plane serves as the scan-matcher start estimate (Sec. IV-B, IV-C, V)","scan endpoints matched to bilinearly interpolated occupancy grid gradients (no feature extraction)","not_reported (scans are transformed to a stabilized frame using estimated attitude; intra-scan motion compensation not described)","2D occupancy grid map and 2D pose; 6-DoF attitude from EKF",[375],{"id":1994,"label":1995,"shortName":1996,"title":1997,"year":249,"era":10,"cluster":527,"scope":676,"keyIdeaZh":1998,"sensors":1999,"mapRepresentation":2002,"loopClosure":2003,"estimator":2004,"association":2005,"deskew":2006,"outputGeometry":2007,"fulltextStatus":22,"lidarModels":2008,"equipmentCount":144},"interactiveslam2021","Koide et al., 2021a","interactive_slam","Interactive 3D Graph SLAM for Map Correction","interactive_slam 讓使用者透過圖形介面修正自動 3D LiDAR SLAM 產生的地圖。系統把自動 SLAM 的位姿約束與使用者建立的修正約束放在同一個位姿圖中，以 g2o 最佳化並立即顯示結果。修正工具有三種：使用者指定迴圈兩端後以 FPFH 初配準、再以 GICP 精配準的半自動迴圈閉合，以及之後的自動迴圈搜尋；依卡方距離抽樣、重新匹配不一致邊的位姿約束精修；以及點選平面後以區域成長與 RANSAC 擷取平面，並在遠處平面之間加入相同、平行或垂直約束，以修正整體彎曲。",[2000,2001],"3D LiDAR (16-line in the indoor-outdoor sequence; model not reported)","input is any ROS SLAM pose graph or odometry sequence","keyframe point clouds on a pose graph with plane vertices","semi-automatic: the user picks the two ends of a loop, then automatic scan-matching loop search runs on the roughly corrected graph (Sec. III.B)","pose-graph optimization (g2o, Levenberg-Marquardt) over automatic SLAM constraints plus user-created correction constraints, re-optimized after every edit (Sec. III.A)","for user-selected loop pairs, FPFH-based initial alignment the user can fine-tune, then GICP (ICP or NDT selectable); automatic loop search by distance, graph-path and matching-score thresholds with a Huber kernel (Sec. III.B)","not_applicable (operates on keyframes from an upstream SLAM)","corrected, globally consistent 3D point cloud map and pose graph",[],{"id":2010,"label":2011,"shortName":2012,"title":2013,"year":249,"era":10,"cluster":68,"scope":69,"keyIdeaZh":2014,"sensors":2015,"mapRepresentation":2017,"loopClosure":72,"estimator":2018,"association":2019,"deskew":121,"outputGeometry":183,"fulltextStatus":22,"lidarModels":2020,"equipmentCount":144},"koide2021vgicp","Koide et al., 2021b","VGICP","Voxelized GICP for Fast and Accurate 3D Point Cloud Registration","VGICP 延伸 GICP，以體素化取代耗時的最近鄰搜尋：每個體素彙整其內各點的分布（而非像 NDT 直接由點位置計算分布），形成分布對多分布的對應，即使體素內點數少也能得到有效分布。體素化使最佳化容易平行化，作者報告 CPU 約 30 Hz、GPU 約 120 Hz，精度與 GICP 相當且對體素解析度較不敏感。",[2016],"3D LiDAR (Velodyne HDL-32E real and simulated)","voxel map of aggregated point distributions (Gaussian voxelmap)","GICP-style least squares with voxel-based distribution-to-multi-distribution residuals; implementation uses Gauss-Newton-type optimizer (Sec. III, IV-B)","per-point covariances from k nearest neighbours (for example k = 20) found with a KD-tree and regularised to eigenvalues (1, 1, epsilon); the target is voxelized by averaging point means and covariances per voxel; each source point is matched to the voxel it falls in (distribution-to-multi-distribution), so no nearest-neighbour search is needed during optimization",[225],{"id":2022,"label":2023,"shortName":2024,"title":2025,"year":661,"era":10,"cluster":98,"scope":132,"keyIdeaZh":2026,"sensors":2027,"mapRepresentation":2031,"loopClosure":2032,"estimator":2033,"association":2034,"deskew":2035,"outputGeometry":2036,"fulltextStatus":22,"lidarModels":2037,"equipmentCount":2038},"glim2024","Koide et al., 2024","GLIM","GLIM: 3D range-inertial localization and mapping with GPU-accelerated scan matching factors","GLIM 以 GPU 加速的體素化 GICP 配準誤差因子（matching cost factor）取代傳統的掃描對模型配準與以高斯近似的相對位姿約束。里程計以固定延遲平滑（fixed-lag smoothing）在約數秒視窗內持續修正過去狀態，並以關鍵影格作為配準目標，使短暫幾何退化仍可藉由後續觀測回推修正。全域最佳化直接最小化各子地圖之間的配準誤差並緊耦合 IMU，可約束重疊很小的子地圖，但運算量高而需 GPU。",[2028,2029,36,2030],"3D LiDAR (spinning and non-repetitive)","depth cameras (ToF, active stereo, stereo)","optional multi-camera","submaps of points with covariances and GPU voxelmaps","implicit: global matching-cost factors between overlapping submaps rather than explicit place recognition","fixed-lag smoothing (iSAM2 in GTSAM) with GPU voxelized-GICP matching-cost factors, IMU preintegration and keyframes; global factor graph minimizing registration errors between submaps with IMU constraints","voxelized GICP (distribution-to-distribution) with surface-orientation-based correspondence validation and multi-resolution voxelmaps","IMU-based motion prediction transforms points to the IMU frame before covariance estimation (Sec. IV-B)","globally optimized submap point clouds and trajectory; export format not_reported in sections read",[143,437,226,224,1592,222,504],18,{"id":2040,"label":2041,"shortName":2042,"title":2043,"year":661,"era":10,"cluster":68,"scope":69,"keyIdeaZh":2044,"sensors":2045,"mapRepresentation":2047,"loopClosure":72,"estimator":2048,"association":2049,"deskew":121,"outputGeometry":389,"fulltextStatus":22,"lidarModels":2050,"equipmentCount":61},"koide2024smallgicp","Koide, 2024","small_gicp","small_gicp: Efficient and parallel algorithms for point cloud registration","small_gicp 是僅需標頭檔的 C++ 點雲精配準函式庫，平行化下採樣、最近鄰搜尋、局部特徵估計與配準整條流程，以減少 PCL 與 Open3D 僅部分多執行緒所造成的瓶頸。它提供點對點、點對平面與 GICP 誤差、穩健核、Gauss-Newton 與 Levenberg-Marquardt 最佳化器，以及 KdTree、iVox 與高斯體素地圖，並有 Python 綁定。",[35,2046],"range cameras","KdTree, iVox, Gaussian voxelmap","Gauss-Newton or Levenberg-Marquardt least squares with robust kernels (Functionalities)","KdTree, linear iVox and Gaussian voxelmap (incremental insertion, LRU deletion); point-to-point, point-to-plane and GICP error factors (Functionalities)",[],{"id":2052,"label":2053,"shortName":2054,"title":2055,"year":332,"era":52,"cluster":527,"scope":54,"keyIdeaZh":2056,"sensors":2057,"mapRepresentation":2059,"loopClosure":2060,"estimator":2061,"association":2062,"deskew":57,"outputGeometry":2063,"fulltextStatus":22,"lidarModels":2064,"equipmentCount":77},"karto_spa2010","Konolige et al., 2010","Karto SLAM (Sparse Pose Adjustment)","Efficient Sparse Pose Adjustment for 2D mapping","本文提出稀疏位姿調整（Sparse Pose Adjustment, SPA），以 Levenberg-Marquardt 最佳化 2D 位姿圖。作者以有序資料結構一次走訪全部約束即建出稀疏的 H 矩陣，再用 CSparse 的稀疏 Cholesky 分解直接求解線性子問題；每條約束保留完整的精度矩陣，因此能處理非球形的共變異。增量模式採用可接續的 LM，保留上一輪的 lambda，每加入一個節點只做一次迭代。實驗用的位姿圖由 SRI 的 Karto 以相關式掃描匹配產生，包含序列匹配與迴圈閉合約束；SPA 本身只是後端，不處理資料關聯。",[2058],"front-end agnostic optimizer; test graphs were built by the Karto front end from 2D laser logs (laser model not reported)","2D pose graph (in Karto, poses carry laser scans); SPA holds no map of its own","not_applicable to the optimizer; loop-closure constraints are produced by Karto's scan-set matching (Sec. V)","Levenberg-Marquardt nonlinear least squares on a 2D pose graph with full precision matrices per constraint; the linear subproblem is built in one ordered pass (per-column std::map blocks converted to compressed column storage) and solved by sparse direct Cholesky (CSparse with AMD ordering); a 'continuable LM' keeps lambda between incremental iterations (Sec. IV.B-F; Tables I-II)","not_applicable to SPA itself; constraints and covariances come from the Karto front end, which uses the correlation method of Konolige and Chou (extended by Olson) for sequential matching and for loop-closure matching of sets of scans (Sec. V)","optimized 2D poses (x, y, theta); maps are rendered from scans at the optimized poses (Fig. 3)",[],{"id":2066,"label":2067,"shortName":2068,"title":2069,"year":235,"era":52,"cluster":236,"scope":99,"keyIdeaZh":2070,"sensors":2071,"mapRepresentation":2075,"loopClosure":2076,"estimator":2077,"association":2078,"deskew":1976,"outputGeometry":2079,"fulltextStatus":22,"lidarModels":2080,"equipmentCount":554},"infinitam2015","Kähler et al., 2015","InfiniTAM","Very High Frame Rate Volumetric Integration of Depth Images on Mobile Devices","本文把 KinectFusion 式的 TSDF 稠密重建最佳化到能在平板電腦上即時執行。資料結構沿用 Nießner 等人的體素區塊雜湊（每塊 8×8×8 體素），但改成每個桶只有一個表頭、碰撞放入額外鏈結串列的雜湊表，並以免鎖的兩步配置、只檢查上一影格可見區塊的增量可見清單，以及固定大小傳輸緩衝的 GPU 與主記憶體（或磁碟）資料交換，控制每一影格的運算量與延遲。射線投射先以低解析度的區塊投影求出搜尋範圍，並可在相機移動不大時跳過部分影格的射線投射；追蹤採點對面 ICP 或色彩直接對齊，也可直接用平板 IMU 提供旋轉，以減少旋轉漂移。整套程式以 InfiniTAM 框架公開。",[2072,2073,2074],"Depth camera: Microsoft Kinect for XBOX 360 (teddy sequence, 640x480 colour and disparity)","Depth camera: Occipital Structure Sensor (couch sequence, 320x240 depth)","IMU of the tablet (Apple iPad Air 2 orientation for the couch sequence; tablet IMU in the swivel-chair test)","TSDF stored in 8x8x8 voxel blocks addressed by a hash table with one list-head entry per bucket plus an excess linked list; incremental visible-block list; blocks swapped between device memory and host memory or disk through fixed-size transfer buffers","none (stated as not addressed; loop closure is future work)","Frame-to-model point-to-plane ICP (or colour-based direct image alignment) solved by Gauss-Newton on resolution hierarchies, rotation only at coarse levels; optionally the rotation is taken from the tablet IMU via an inertial fusion algorithm and only translation is estimated visually","Projective association between the current depth image and point and normal maps ray-cast from the TSDF (normals computed in image space)","TSDF voxel-block map with raycast point clouds and renderings; optional colour per voxel",[],{"id":2082,"label":2083,"shortName":2084,"title":2085,"year":1217,"era":52,"cluster":53,"scope":54,"keyIdeaZh":2086,"sensors":2087,"mapRepresentation":2088,"loopClosure":2089,"estimator":2090,"association":2091,"deskew":57,"outputGeometry":2092,"fulltextStatus":22,"lidarModels":2093,"equipmentCount":77},"kummerle2011g2o","Kümmerle et al., 2011","g2o","g2o: A general framework for graph optimization","g2o 將 SLAM 與 BA 等可用圖表示的非線性誤差函數，統一寫成以資訊矩陣加權的最小平方問題：節點是待估參數區塊，邊是量測約束。框架以 Gauss-Newton 或 Levenberg-Marquardt 迭代，並以 ⊞ 運算在流形上用最小參數化增量更新旋轉，避開過參數化與奇異性。效率來自利用圖的稀疏結構、Schur 補與可替換的線性求解器（稀疏 Cholesky 或 PCG）；作者在多個 2D\u002F3D 位姿圖與 BA 資料集上報告其效能與特定問題的專用實作相當。",[],"pose graph and optional landmarks (no dense map)","not_applicable (optimizes loop-closure edges if the front-end provides them)","batch nonlinear least squares on a graph (Gauss-Newton or Levenberg-Marquardt) with manifold increments through a box-plus operator; linear solvers: sparse Cholesky via CHOLMOD or CSparse (symbolic decomposition reused across iterations) or block-Jacobi preconditioned conjugate gradient; Schur complement for BA and landmark SLAM; Jacobians numeric or user-supplied analytic","not_applicable (sensor-agnostic back-end; constraints and data association are supplied by a front-end)","optimized poses and landmark or 3D point positions; dense point clouds must be re-projected by the user",[],{"id":2095,"label":2096,"shortName":2097,"title":2098,"year":9,"era":10,"cluster":236,"scope":132,"keyIdeaZh":2099,"sensors":2100,"mapRepresentation":2104,"loopClosure":2105,"estimator":2106,"association":2107,"deskew":2108,"outputGeometry":2109,"fulltextStatus":22,"lidarModels":2110,"equipmentCount":423},"rtabmap2019","Labbé & Michaud, 2019","RTAB-Map","RTAB‐Map as an open‐source lidar and visual simultaneous localization and mapping library for large‐scale and long‐term online operation","RTAB-Map 起源於具記憶體管理的外觀式迴圈偵測，將節點在工作記憶與長期記憶之間轉移，使迴圈偵測在固定時間內完成，以支援大範圍與長期線上運作。擴充版成為以圖為基礎的 SLAM 函式庫，可接收任意來源的里程計，並支援 RGB-D、雙目及 2D\u002F3D LiDAR，後端可選 TORO、g2o 或 GTSAM。論文以同一系統比較視覺與 LiDAR 組態在 KITTI、EuRoC、TUM RGB-D 與 MIT Stata Center 資料上的表現，輸出包括點雲、OctoMap 與佔據格地圖。",[772,2101,211,35,119,2102,2103],"stereo","IMU only through external odometry (wheel and IMU EKF) or integrated visual-inertial odometry such as OKVIS, MSCKF and Google Tango (Sec. 3.1.1, 4.3, 4.4)","the bag-of-words loop closure needs a camera, and without one the authors suggest feeding an empty image and relying on laser proximity detection only (Sec. 6)","pose graph with node sensor data; assembled 2D occupancy grid, OctoMap and point cloud outputs (Fig. 1)","appearance-based loop closure detection with memory management (working, short-term and long-term memory) to bound detection time; proximity detection (abstract; Fig. 1)","graph-based SLAM with selectable back-ends TORO, g2o or GTSAM; odometry from any external or built-in visual\u002FLiDAR source (Fig. 1; back-end paragraph)","visual odometry: GFTT features with BRIEF descriptors matched by NNDR to a local feature map (F2M) or tracked by optical flow to the last keyframe (F2F), constant-velocity prediction, PnP RANSAC and local BA with g2o; lidar odometry: libpointmatcher ICP, point-to-point or point-to-plane, scan-to-scan or scan-to-map; loop closure: incremental bag-of-words with TF-IDF and a Bayes filter, transform from visual PnP optionally refined by ICP; proximity detection by laser scans for nearby graph nodes (Sec. 3.1, 3.4)","no internal deskewing: laser scans are assumed to be motion-distortion corrected before input to RTAB-Map; authors note correction can be ignored when scanner rotation is fast relative to robot speed (Sec. 3.1.2)","assembled point cloud (voxel-filtered, PointCloud2), OctoMap and 2D occupancy grid ROS outputs built from per-node local grids and re-assembled after loop closure (Fig. 1, Fig. 8, Sec. 3.6); file export features of the application not described in the paper",[2111,2112],"Velodyne 64E","UTM30",{"id":2114,"label":2115,"shortName":2116,"title":2117,"year":9,"era":10,"cluster":397,"scope":1035,"keyIdeaZh":2118,"sensors":2119,"mapRepresentation":57,"loopClosure":57,"estimator":2122,"association":57,"deskew":57,"outputGeometry":2123,"fulltextStatus":22,"lidarModels":2124,"equipmentCount":24},"laconte2019lidarbias","Laconte et al., 2019","Incidence-angle LiDAR bias model","Lidar Measurement Bias Estimation via Return Waveform Modelling in a Context of 3D Mapping","作者質疑 LiDAR 量測為零均值高斯雜訊的常見假設，指出與入射角及距離相關的偏差會造成可預期的定位漂移，例如直線隧道的地圖會依靠近哪一側牆而彎曲。本文以回波波形建模解釋此偏差，於實驗裝置量測三款 LiDAR 的偏差，發現高入射角下可達 20 cm，並用模型修正量測以改善地圖與漂移。",[2120,2121],"2D LiDAR (SICK LMS151)","3D LiDAR (Velodyne HDL-32E, Robosense RS-LiDAR-16)","not_applicable (measurement model and correction)","bias-corrected range measurements and maps",[2125,225,2126],"SICK LMS-151","Robosense RS-LiDAR-16",{"id":2128,"label":2129,"shortName":2130,"title":2131,"year":9,"era":10,"cluster":397,"scope":69,"keyIdeaZh":2132,"sensors":2133,"mapRepresentation":57,"loopClosure":57,"estimator":2135,"association":2136,"deskew":121,"outputGeometry":403,"fulltextStatus":22,"lidarModels":2137,"equipmentCount":77},"landry2019cello3d","Landry et al., 2019","CELLO-3D","CELLO-3D: Estimating the Covariance of ICP in the Real World","作者先檢視既有封閉形式共變異數估計在 3D 資料上的限制，再以資料驅動方式學習 ICP 配準的共變異數。訓練與評估使用超過五百萬次配準、1020 組真實點雲對，涵蓋結構化與非結構化、室內與室外環境。",[2134],"3D point clouds from the 'Challenging data sets for point cloud registration algorithms' (Pomerleau et al. 2012); the sensor is not named in this paper","CELLO-style kernel-weighted average of training covariances; an upper-triangular distance metric is learned by SGD with a determinant-plus-trace loss; descriptors per cell of a 4x4x4 grid (25 m x 25 m x 10 m) over the overlap region hold planarity, cylindricality and a 9-bin normal histogram; training covariances come from 5000 ICP samples per pair filtered with DBSCAN","point-to-plane ICP configured in the framework of Pomerleau et al. [24]: maximum-density and random subsampling filters, k-d tree 3-nearest-neighbour matching, trimmed-distance outlier filter keeping the closest 70%, at most 80 iterations",[],{"id":2139,"label":2140,"shortName":2141,"title":2142,"year":51,"era":10,"cluster":263,"scope":99,"keyIdeaZh":2143,"sensors":2144,"mapRepresentation":2146,"loopClosure":1451,"estimator":2147,"association":2148,"deskew":2149,"outputGeometry":2150,"fulltextStatus":22,"lidarModels":2151,"equipmentCount":377},"cocolic2023","Lang et al., 2023","Coco-LIC","Coco-LIC: Continuous-Time Tightly-Coupled LiDAR-Inertial-Camera Odometry Using Non-Uniform B-Spline","Coco-LIC 以非均勻 B 樣條（non-uniform B-spline）表示連續時間軌跡，依 IMU 感知的運動劇烈程度動態配置控制點，在平緩運動時使用較少控制點、劇烈運動時加密，以兼顧精度與計算量。視覺像素的深度直接取自全域光達地圖，建立幀對地圖重投影因子，避免在長滑動視窗中最佳化視覺深度。連續時間表示讓不同頻率、非同步的光達、IMU 與相機量測可在任意時刻查詢位姿後融合。",[2145,36,37],"3D LiDAR (Velodyne VLP-16, HDL-32E, Livox Avia across datasets)","Two LiDAR maps: a local map built from keyscans selected by time and space for scan-to-map point-to-plane matching, and a global LiDAR map stored in 0.1 m voxels whose points are projected into images for frame-to-map visual factors; marginalization acts on control points, not on the map","Factor-graph nonlinear least squares over the new control points and IMU biases of each 0.1 s interval (LiDAR point-to-plane, visual reprojection with Cauchy kernel, raw IMU and bias random-walk factors, marginalization prior), solved with Levenberg-Marquardt in Ceres; separate cubic non-uniform B-splines for rotation and translation; raw IMU used without preintegration","LiDAR planar (surf) points matched to 5 nearest neighbours of the local keyscan map for point-to-plane residuals; global-map LiDAR points tracked across images by KLT optical flow, outliers removed with fundamental-matrix RANSAC and PnP, then used in frame-to-map reprojection factors without visual depth optimization","No separate deskew step: each LiDAR planar point is transformed with the spline pose queried at its own timestamp, so motion distortion removal and trajectory estimation happen simultaneously","not_reported (evaluation is trajectory-based)",[143,437,225],{"id":2153,"label":2154,"shortName":2155,"title":2156,"year":207,"era":10,"cluster":755,"scope":99,"keyIdeaZh":2157,"sensors":2158,"mapRepresentation":2159,"loopClosure":1451,"estimator":2160,"association":2161,"deskew":2162,"outputGeometry":2163,"fulltextStatus":22,"lidarModels":2164,"equipmentCount":109},"gaussianlic2025","Lang et al., 2025","Gaussian-LIC","Gaussian-LIC: Real-Time Photo-Realistic SLAM with Gaussian Splatting and LiDAR-Inertial-Camera Fusion","Gaussian-LIC 以連續時間緊耦合的 LiDAR、慣性與相機里程計（Coco-LIC，每 0.1 秒做一次因子圖最佳化）提供位姿，將著色並降取樣的 LiDAR 點與視覺滑動視窗三角化的 SfM 點一起初始化三維高斯，以補足 LiDAR 未涵蓋的相機視野，並加入天空高斯與曝光仿射模型，以 C++ 與 CUDA 加速達成即時寫實建圖。論文只報告渲染品質（PSNR、SSIM、LPIPS）與執行時間，追蹤比較僅為定性描述，沒有 ATE 或地圖幾何精度數值；作者把提升幾何重建品質列為未來工作。",[35,36,37],"3D Gaussians initialized from colourized LiDAR points plus triangulated visual SfM points; sky and exposure modelling","continuous-time factor-graph sliding-window optimization (Coco-LIC) with point-to-map LiDAR, frame-to-map visual and inertial factors","Coco-LIC odometry with point-to-map LiDAR factors, frame-to-map visual factors and inertial factors (Sec. III-B); a separate VINS-Mono-style visual sliding window tracks Shi-Tomasi corners with KLT only to triangulate SfM points for Gaussian initialization; mapping minimizes an L1 plus D-SSIM re-rendering loss with a per-image exposure affine matrix (Eq. 9)","not described in the full text (v3); poses come from the continuous-time Coco-LIC trajectory optimized every 0.1 s","3D Gaussian map (with sky Gaussians) and rendered images; neither map geometric accuracy nor trajectory accuracy is quantified; authors list improving geometric reconstruction quality as future work",[45],{"id":2166,"label":2167,"shortName":2168,"title":2169,"year":150,"era":10,"cluster":397,"scope":1035,"keyIdeaZh":2170,"sensors":2171,"mapRepresentation":2173,"loopClosure":72,"estimator":2174,"association":2175,"deskew":2176,"outputGeometry":2177,"fulltextStatus":22,"lidarModels":2178,"equipmentCount":144},"legentil2018lidarimucalib","Le Gentil et al., 2018","Lidar-IMU calibration with upsampled preintegration","3D Lidar-IMU Calibration Based on Upsampled Preintegrated Measurements for Motion Distortion Correction","作者指出 LiDAR 點是逐點取樣而非快照，平台快速移動時會產生運動畸變。本法以高斯過程迴歸對 IMU 讀數上取樣，對每個 LiDAR 點計算預積分量以精確去畸變，並以點到平面距離與 IMU 預積分因子聯合估計外參、IMU 狀態與時間偏移。模擬顯示未使用上取樣預積分時，在快速運動下精度明顯下降。",[1038,2172],"IMU (Xsens MTi-3)","set of planes","on-manifold factor-graph optimization of IMU poses, velocities, biases, calibration and time shift","RANSAC (MLESAC) plane detection in every scan; planes tracked between consecutive scans by nearest-neighbour matching of normals after subtracting their centroid; each point assigned to one map plane from the first scan (assumed static) and used in a point-to-plane residual","each LiDAR point reprojected with upsampled preintegrated IMU measurements","extrinsic calibration and time shift",[143],{"id":2180,"label":2181,"shortName":2182,"title":2183,"year":396,"era":10,"cluster":53,"scope":54,"keyIdeaZh":2184,"sensors":2185,"mapRepresentation":57,"loopClosure":57,"estimator":2187,"association":57,"deskew":2188,"outputGeometry":2189,"fulltextStatus":22,"lidarModels":2190,"equipmentCount":24},"legentil2020gpm","Le Gentil et al., 2020","Gaussian Process Preintegration (GPM)","Gaussian Process Preintegration for Inertial-Aided State Estimation","本文以高斯過程（GP）連續表示慣性量測，並對 GP 核函數施加線性運算子，推導出「高斯預積分量測」（GPM）。旋轉僅繞單一軸時，旋轉與速度、位置增量都可解析積分；若含三維旋轉，旋轉增量仍需先以 GP 上取樣再數值積分，速度與位置增量則由重投影到起始 IMU 座標後的加速度計訊號以 GP 推論。此法不需明確運動模型，也沒有離散積分雜訊，適合非同步的慣性輔助估計；作者並推導偏差與感測器間時間偏移的一階修正 Jacobian。模擬比較中 GPM 與 UPM 的誤差約比離散預積分低一個數量級，並整合進 LiDAR 慣性定位與建圖框架 IN2LAAMA 驗證可用性。",[36,2186],"3D LiDAR (validation within IN2LAAMA)","preintegrated measurements from GP models (square exponential kernel) of IMU signals with linear operators on kernels: analytic integral inference for single-axis rotation, numerical integration after GP upsampling for 3D rotation; velocity and position inferred from GP models of accelerometer data reprojected into the IMU frame at t1; first-order post-integration bias and time-shift Jacobians","not described in this paper; the authors state that GPMs can be used with asynchronous platforms and rolling-shutter-like sensors, citing spinning lidars as such sensors (Sec. I); how IN2LAAMA [23] handles lidar motion distortion is not described here","preintegrated measurements (Delta R, Delta v, Delta p with covariance) over any time interval; within IN2LAAMA a lidar point-cloud map and trajectory are produced (Fig. 4)",[143],{"id":2192,"label":2193,"shortName":2194,"title":2195,"year":249,"era":10,"cluster":98,"scope":132,"keyIdeaZh":2196,"sensors":2197,"mapRepresentation":2200,"loopClosure":2201,"estimator":2202,"association":2203,"deskew":2204,"outputGeometry":2205,"fulltextStatus":22,"lidarModels":2206,"equipmentCount":109},"in2laama2021","Le Gentil et al., 2021","IN2LAAMA","IN2LAAMA: Inertial Lidar Localization Autocalibration and Mapping","IN2LAAMA 是以 3D LiDAR 與 6 自由度 IMU 進行離線批次定位、建圖與外參自動校正的框架。它對每個 IMU 軸以高斯過程回歸建立連續慣性訊號，再對每個 LiDAR 點的時間戳記做預積分，得到「上取樣預積分量測」（UPM），因此不需假設等速等運動模型即可精確描述掃描期間的運動並去畸變。後端把點到線、點到平面距離、影格間 IMU 因子、偏差與時間偏移因子放入同一個全批次最佳化，前端則依最新狀態重算特徵，並以大於 360 度的掃描與雙向關聯維持掃描一致性；最後可把 LiDAR 與 IMU 外參加入狀態，不需校正標靶。",[2198,2199],"3D spinning LiDAR with per-point timestamps (Velodyne VLP-16; Velodyne HDL-32 in the MC2SLAM campus drive sequence)","6-DoF IMU (Xsens MTi-3 at 100 Hz; HDL-32 built-in IMU)","dense point cloud reconstructed by projecting all points with the final trajectory and UPMs; no persistent map structure (Sec. VI-E)","yes; proximity-based candidates using radial distance, height offset and z-axis angle between LiDAR frames, optional ICP fitness test, added as lidar factors (Sec. V-D)","offline full-batch MAP optimization in Ceres with analytic Jacobians over all frame poses, velocities, per-frame IMU bias and time-shift corrections (and optionally LiDAR-IMU extrinsics), built incrementally with feature recomputation; bisquare weights on lidar residuals and Cauchy loss on lidar and IMU factors (Sec. III-B, VI)","channel-wise features from a linear-regression curvature score (cosine between left and right fitted lines), planar points and inward or outward edges binned per line; point-to-plane (3 nearest) and point-to-line (2 nearest) associations between frames with spatial-spread and patch-consistency outlier tests; frames longer than 360 deg (520 deg) with back-and-forth association (Sec. V-A, V-C)","every point is projected with its own UPM inside the residuals, so motion distortion is re-corrected whenever the state changes; features are recomputed from the current estimate (Sec. IV-A, V-B, VI-A)","full 6-DoF trajectory, IMU biases, time-shift, LiDAR-IMU extrinsic calibration and dense motion-corrected point cloud map (Figs. 1, 11-13)",[143,225],{"id":2208,"label":2209,"shortName":2210,"title":2211,"year":207,"era":10,"cluster":98,"scope":99,"keyIdeaZh":2212,"sensors":2213,"mapRepresentation":2214,"loopClosure":72,"estimator":2215,"association":2216,"deskew":2217,"outputGeometry":2218,"fulltextStatus":22,"lidarModels":2219,"equipmentCount":61},"genzicp2025","Lee et al., 2025a","GenZ-ICP","GenZ-ICP: Generalizable and Degeneracy-Robust LiDAR Odometry Using an Adaptive Weighting","GenZ-ICP 指出單一誤差度量在不同幾何環境各有弱點：點到平面在長廊等退化場景易病態，點到點在結構化場景精度較低。作者依鄰域平面度把點分為平面與非平面兩類，分別套用點到平面與點到點誤差，並以兩類點數比例自適應調整權重，以避免最佳化在走廊型退化中發散。",[1355],"local map used as ICP target (Fig. 2); its data structure is not described in the paper; metric comparison on Long_Corridor run inside the KISS-ICP framework [5] (Sec. IV-A)","ICP minimizing a weighted sum of point-to-plane and point-to-point residuals with an adaptive weight","per-point planarity classification: point-to-plane for planar neighbourhoods, point-to-point otherwise","not_reported (motion compensation is not described in the paper)","odometry and accumulated map (Fig. 1); export format not_reported",[],{"id":2221,"label":2222,"shortName":2223,"title":2224,"year":207,"era":10,"cluster":263,"scope":99,"keyIdeaZh":2225,"sensors":2226,"mapRepresentation":2229,"loopClosure":465,"estimator":2230,"association":2231,"deskew":2232,"outputGeometry":2233,"fulltextStatus":22,"lidarModels":2234,"equipmentCount":1593},"mins2025","Lee et al., 2025b","MINS","MINS: Efficient and Robust Multisensor-Aided Inertial Navigation System","MINS 以 IMU 為核心，在一個 MSCKF 形式的擴展卡爾曼濾波器中緊耦合相機、輪速計、LiDAR 與 GNSS：每種感測器都有專屬的量測更新，並能線上校正所有感測器的外參、時間偏移與內參。面對非同步量測，系統以高階流形上多項式內插取得任一時刻的位姿並建立內插誤差模型，且依運動動態調整複製狀態的頻率，以兼顧精度與計算量。LiDAR 採用直接點到平面更新並維護 ikd 樹局部地圖。",[36,2227,1203,2228,1280],"cameras (one or more)","LiDAR (one or more)","dense LiDAR local map (ikd-tree) anchored to a clone; no global map","MSCKF-style EKF with the IMU as the backbone; camera updates with nullspace projection and measurement compression; integrated 2D wheel odometry updates with wheel intrinsics; direct LiDAR point-on-plane updates against an ikd-tree local map anchored to a cloned pose; GNSS position updates after 4-DoF alignment, then estimation in the global frame; online spatiotemporal and intrinsic calibration of all sensors (Sec. 3)","visual feature tracks for MSCKF updates; LiDAR points matched to planes in a dense local map (FAST-LIO2-style point-on-plane); wheel and GNSS measurements used directly (Sec. 3)","not described; each point cloud is treated as a measurement at a single time t_k whose pose is obtained through the on-manifold interpolation (Secs. 3.4, 4)","IMU pose, velocity and calibration states at the IMU rate",[2235,2236],"16-channel Velodyne (two units)","64-channel Ouster",{"id":2238,"label":2239,"shortName":2240,"title":2241,"year":661,"era":10,"cluster":755,"scope":69,"keyIdeaZh":2242,"sensors":2243,"mapRepresentation":2245,"loopClosure":57,"estimator":2246,"association":2247,"deskew":57,"outputGeometry":2248,"fulltextStatus":22,"lidarModels":2249,"equipmentCount":46},"mast3r2024","Leroy et al., 2024","MASt3R","Grounding Image Matching in 3D with MASt3R","MASt3R 在 DUSt3R 上增加輸出稠密局部特徵的分支並以匹配損失訓練，同時提出快速互為最近鄰匹配以降低二次複雜度。與 DUSt3R 不同，當訓練真值為公制時不做尺度正規化，使模型可輸出公制尺度點圖。其 DTU 結果是以真值相機對匹配點三角化取得，並非完全無相機的重建。",[2244],"monocular camera (image pairs)","pairwise pointmaps with confidence and dense descriptors","feed-forward pointmap regression (DUSt3R backbone) with an added dense local-feature head","dense learned local features with fast reciprocal nearest-neighbour matching","pointmaps (metric-scale when trained on metric data), dense matches; DTU point clouds by triangulating matches with GT cameras",[],{"id":2251,"label":2252,"shortName":2253,"title":2254,"year":235,"era":52,"cluster":236,"scope":99,"keyIdeaZh":2255,"sensors":2256,"mapRepresentation":2257,"loopClosure":2258,"estimator":2259,"association":2260,"deskew":57,"outputGeometry":2261,"fulltextStatus":22,"lidarModels":2262,"equipmentCount":554},"okvis2015","Leutenegger et al., 2015","OKVIS","Keyframe-based visual-inertial odometry using nonlinear optimization","OKVIS 以非線性最佳化緊耦合（tightly-coupled）融合相機重投影誤差與 IMU 慣性誤差，並只保留有限數量的關鍵影格，透過邊際化維持即時運算。關鍵影格可相隔任意時間，仍以線性化慣性項連結。作者以自製、硬體同步的雙目慣性裝置收集資料，並與 MSCKF 濾波器比較，也示範線上外參校正。",[2101,37,36],"sparse landmarks in a bounded keyframe window","none (odometry)","Nonlinear least squares (Google Ceres) over reprojection errors (keypoint std 0.8 px) and IMU error terms in a window of M keyframes plus the S most recent frames (M = 7, S = 3 in all experiments); frames leaving the window are marginalised by Schur complement with first-estimate Jacobians, dropping non-keyframe landmark observations and marginalising landmarks seen only in the oldest keyframes so that sparsity is kept; optional online camera-IMU extrinsics estimation","Customised multi-scale SSE-optimised Harris corners with BRISK descriptors oriented along the projected gravity direction; brute-force 3D-2D matching against landmarks predicted visible, outliers removed by a Mahalanobis test on the IMU-propagated pose and an OpenGV absolute-pose RANSAC; then brute-force 2D-2D matching with stereo and temporal triangulation (only points with low depth uncertainty initialised) and a relative RANSAC against the newest keyframe. In the comparison all algorithms were fed the same correspondences produced by the stereo pipeline","time series of poses, velocities and IMU biases plus a sparse landmark map (conclusion)",[],{"id":2264,"label":2265,"shortName":2266,"title":2267,"year":383,"era":52,"cluster":236,"scope":346,"keyIdeaZh":2268,"sensors":2269,"mapRepresentation":2270,"loopClosure":2271,"estimator":2272,"association":2273,"deskew":57,"outputGeometry":2274,"fulltextStatus":22,"lidarModels":2275,"equipmentCount":487},"msckf2_2013","Li & Mourikis, 2013","MSCKF 2.0","High-precision, consistent EKF-based visual-inertial odometry","論文比較兩類以 EKF 為基礎的視覺慣性里程計（VIO）：狀態含特徵點的 EKF-SLAM，以及只保留滑動視窗位姿的 MSCKF，並以蒙地卡羅模擬顯示 MSCKF 在精度、一致性與運算量上都較佳。作者證明兩者的線性化模型都讓繞重力軸的偏航角看似可觀測，濾波器因此低估不確定度。MSCKF 2.0 以閉式 IMU 誤差狀態轉移矩陣，並在計算 Jacobian 時固定使用位置與速度的首次估計值，恢復正確的不可觀測子空間；同時把相機與 IMU 外參放入狀態線上估計。",[37,36],"none; features are never kept in the state vector, only a sliding window of poses","none (VIO; loop closure is left to a separate algorithm, Sec. 3.2)","multi-state-constraint EKF: IMU state plus a sliding window of past IMU poses; feature errors removed by left-nullspace projection; Jacobians evaluated at first (propagated) estimates of position and velocity so the linearised model keeps a 4-DOF unobservable subspace; closed-form IMU error-state transition matrix; camera-to-IMU rotation and translation in the state (Secs. 3.3, 5, 7)","Shi-Tomasi corners matched by normalised cross-correlation in the real experiment (about 290 features per image); each feature used once its track ends, triangulated by Gauss-Newton; Mahalanobis gating at the 95th percentile of chi-square with 2N-3 degrees of freedom (Secs. 3.3, 9)","IMU pose, velocity, biases and camera-IMU extrinsics with covariance; no map or point cloud",[],{"id":2277,"label":2278,"shortName":2279,"title":2280,"year":115,"era":52,"cluster":397,"scope":1035,"keyIdeaZh":2281,"sensors":2282,"mapRepresentation":57,"loopClosure":57,"estimator":2285,"association":2286,"deskew":57,"outputGeometry":2287,"fulltextStatus":22,"lidarModels":2288,"equipmentCount":46},"li2014onlinetemporal","Li & Mourikis, 2014","Online temporal calibration (camera-IMU)","Online temporal calibration for camera-IMU systems: Theory and algorithms","本文把相機與 IMU 之間的時間偏移 td 納入 EKF 狀態，與 IMU 位姿、速度、偏差、相機對 IMU 外參及特徵位置一起線上估計，可用於已知地圖定位、EKF-SLAM 與 MSCKF 視覺慣性里程計，只增加一個純量狀態。作者證明除零角速度、等角速度或加速度計讀值固定等少數退化運動外，td 皆為局部可辨識，而這些運動即使已知 td 也會失去可觀性。實驗與模擬顯示線上估計的精度幾乎等同事先已知 td（Sec. 4 至 7）。",[2283,2284],"monocular camera (one camera of a PointGrey Bumblebee2 stereo pair, 20 Hz)","IMU (Xsens MTI-G, 100 Hz)","EKF with td (constant, or random walk when time-varying) and the camera-to-IMU transform in the state; demonstrated as map-based EKF localization, EKF-SLAM (inverse depth then xyz features, modified-Jacobian approach for consistency) and MSCKF 2.0 visual-inertial odometry, where only state augmentation at t + td changes","20 blue LEDs at known positions (map-based and persistent SLAM features); Shi-Tomasi corners tracked in images (about 65 temporary features per image in EKF-SLAM), matched by normalized cross-correlation in VIO","time offset, poses",[],{"id":2290,"label":2291,"shortName":2292,"title":2293,"year":9,"era":10,"cluster":755,"scope":99,"keyIdeaZh":2294,"sensors":2295,"mapRepresentation":2297,"loopClosure":2298,"estimator":2299,"association":2300,"deskew":2301,"outputGeometry":2302,"fulltextStatus":22,"lidarModels":2303,"equipmentCount":24},"lonet2019","Li et al., 2019","LO-Net","LO-Net: Deep Real-Time Lidar Odometry","LO-Net 將相鄰兩幀 LiDAR 點雲以圓柱投影編碼成含距離與強度的資料矩陣，以孿生（Siamese）卷積網路直接迴歸 6 自由度相對位姿；網路內以距離加權的鄰點外積計算逐點法向量（非以可學習權重估計），並同時學習動態物遮罩（mask），以遮罩加權的法向量幾何一致性損失約束訓練。之後利用法向量挑選平滑區點、以遮罩排除移動物，再以點到平面（point-to-plane）的掃描對地圖（scan-to-map）配準精化位姿以降低累積漂移。此方法需要真值位姿做監督訓練，且只在 KITTI 與 Ford 車載資料上評估。",[2296],"Velodyne HDL-64 3D LiDAR (KITTI and Ford data; Sec. 4, 4.5)","sliding point map holding the last n_m = 100 transformed scans (Sec. 3.5, Sec. 4 implementation details)","none (not implemented for any method in the experiments, Sec. 4.2)","supervised Siamese CNN regressing relative 6-DoF pose (translation + quaternion) from two scans, followed by iterative point-to-plane scan-to-map refinement (Sec. 3.3, 3.5)","implicit in the network (cylindrical range\u002Fintensity matrix input) with a mask-weighted normal-consistency loss; normals are computed inside the network by range-weighted cross products of four grid neighbours plus moving-average smoothing (Eq. 3), not by learned weights; mapping uses point-to-plane correspondences to map points, selecting smooth-area points by a convolution over the normal channels and excluding masked points (Sec. 3.2, 3.4, 3.5)","mapping module removes motion distortion by linear interpolation of the LO-Net odometry before scan-to-map matching (Sec. 3.5)","6-DoF trajectory and an accumulated point map shown in figures; map export is not described (Fig. 4)",[161],{"id":2305,"label":2306,"shortName":2307,"title":2308,"year":249,"era":10,"cluster":151,"scope":132,"keyIdeaZh":2309,"sensors":2310,"mapRepresentation":2312,"loopClosure":2313,"estimator":2314,"association":2315,"deskew":637,"outputGeometry":2316,"fulltextStatus":22,"lidarModels":2317,"equipmentCount":144},"saloam2021","Li et al., 2021a","SA-LOAM","SA-LOAM: Semantic-aided LiDAR SLAM with Loop Closure","SA-LOAM 以開源的 F-LOAM 為基礎，先以預訓練的 RangeNet++ 為每個 LiDAR 點加上語意標籤，再把語意用在里程計與迴圈偵測兩處。里程計部分，邊緣與平面特徵只與相同語意的子地圖點配對，依類別分別降採樣以保留小物體，並要求地面平面法向量垂直、建物平面法向量水平，以剔除擬合不良的平面。迴圈部分，把每幀點雲聚成語意圖，以圖匹配網路評分候選，再以語意輔助 ICP 做幾何驗證，最後用 g2o 位姿圖最佳化，得到全域一致的語意地圖。",[2311],"3D LiDAR only (Velodyne HDL-64E in KITTI and Ford Campus); per-point semantics from a pre-trained RangeNet++ network (Sec. IV-A)","edge and planar semantic feature submaps for odometry; global semantic point map plus a lightweight semantic graph map (at most 100 nodes of centre and label per frame) for loop search (Sec. III-A; Sec. III-C)","candidates within an odometry-drift-dependent distance (up to 64 random candidates), similarity scored by the authors' semantic-graph matching network (score above 0.95, top 5 kept), then geometric verification by semantic-assisted ICP to a submap (Sec. III-C; Table I)","F-LOAM based scan-to-submap optimization of point-to-line and point-to-plane distances with semantic-related weights (set equal in the experiments); loop constraints added to a g2o pose graph (Sec. III-B; Sec. III-C; Sec. IV-A)","LOAM-style edge and planar features matched only to submap points of the same semantic label (k-d tree, 5 neighbours), class-specific downsampling, and plane fits kept only when ground normals are vertical and building normals horizontal; submap from the last 20 frames (Sec. III-B; Table I)","trajectory and globally consistent semantic point cloud map (Figs. 1 and 4)",[161],{"id":2319,"label":2320,"shortName":2321,"title":2322,"year":249,"era":10,"cluster":151,"scope":132,"keyIdeaZh":2323,"sensors":2324,"mapRepresentation":2327,"loopClosure":2328,"estimator":2329,"association":2330,"deskew":2331,"outputGeometry":2332,"fulltextStatus":22,"lidarModels":2333,"equipmentCount":1289},"liliom2021","Li et al., 2021b","LiLi-OM","Towards High-Performance Solid-State-LiDAR-Inertial Odometry and Mapping","LiLi-OM 是同時支援固態（Livox Horizon）與機械式 LiDAR 的緊耦合 LiDAR 慣性里程計與建圖系統。前端以輕量的特徵式掃描配準（點到邊、點到面）快速估計運動並自適應挑選關鍵影格；後端以階層式、關鍵影格為基礎的滑動視窗最佳化直接融合 LiDAR 與 IMU 預積分，並以 ICP 驗證迴圈、GTSAM 最佳化全域位姿圖。針對 Horizon 不規則的掃描樣式，論文另提出兩階段的特徵擷取方法。",[2325,2326],"solid-state LiDAR (Livox Horizon, 81.7 x 25.1 deg FoV) or spinning LiDAR (HDL-32E, HDL-64E)","IMU (6-axis sufficient; Xsens MTi-670 in own suite)","keyframe feature maps attached to a global pose graph; frontend local map of about 20 recent frames and backend local map from 30 recent keyframes; features voxel-grid downsampled before building LiDAR constraints","radius search (e.g., 10 m) for spatially close but temporally distant keyframes, ICP fitting score for acceptance, then global pose-graph optimization (Sec. 4)","keyframe-based sliding-window optimization fusing LiDAR features and preintegrated IMU via marginalization (Ceres); in-between frames by local factor graph; global pose graph (GTSAM) at loop closure (Sec. 2, 4, 5.1)","point-to-edge and point-to-plane residuals against local feature maps (frontend about 20 recent frames, backend 30 recent keyframes), each residual weighted by agreement of edge direction or plane normal and by reflectance similarity of the five nearest features; two-stage time-domain extractor for the Livox Horizon (6 x 7-point patches, plane if lambda1\u002Flambda2 \u003C 0.3, else edge test lambda2\u002Flambda3 \u003C 0.25) because LOAM's per-scan-line smoothness cannot be applied to it; LOAM preprocessing only for spinning LiDARs (LiLi-OM*)","rotational de-skew from gyroscope before feature extraction, translational de-skew from frame-to-model odometry estimate (Sec. 2)","keyframe feature map and trajectory (Sec. 4, Fig. 8)",[1482,161,225],{"id":2335,"label":2336,"shortName":2337,"title":2338,"year":513,"era":52,"cluster":397,"scope":1035,"keyIdeaZh":2339,"sensors":2340,"mapRepresentation":57,"loopClosure":57,"estimator":2344,"association":2345,"deskew":57,"outputGeometry":2346,"fulltextStatus":22,"lidarModels":2347,"equipmentCount":24},"lichti2007amcw","Lichti, 2007","AM-CW TLS error model and self-calibration","Error modelling, calibration and analysis of an AM-CW terrestrial laser scanner system","本文以室內標靶控制網對 Faro 880 調幅連續波（AM-CW）地面雷射掃描儀進行自率定：在自由網最小平方平差中同時估計各站外方位、標靶座標與 17 個附加參數，這些參數描述距離、水平方向與高度角的系統誤差，包含可物理解釋的項目（加常數、週期誤差、視準軸誤差、橫軸誤差、指標差）與由殘差分析找出的經驗項。13 個月內 10 組資料的殘差 RMS 明顯下降，獨立檢核點亦改善；但多數參數隨時間顯著變動，入射角大於約 65 度時距離誤差明顯增大。",[2341,2342,2343],"terrestrial laser scanner Faro 880 (AM-CW phase-difference rangefinder; formerly iQsun 880)","integrated dual-axis inclinometers of the Faro 880","total station and 900 mm Leica scale bar (independent check-point survey only)","free-network least-squares self-calibration with inner constraints on object points, estimating exterior orientation, object-point coordinates and 17 additional parameters simultaneously; group weights by variance component estimation; Baarda data snooping for outliers","signalised planar Faro targets (A4; A5 in calibration 1) measured by iQscene contrast centroiding; spherical observations derived from exported Cartesian coordinates","17 calibrated additional parameters (6 range, 7 horizontal direction, 4 elevation angle) with precisions, plus adjusted target coordinates",[],{"id":2349,"label":2350,"shortName":2351,"title":2352,"year":249,"era":10,"cluster":599,"scope":32,"keyIdeaZh":2353,"sensors":2354,"mapRepresentation":2355,"loopClosure":57,"estimator":57,"association":2356,"deskew":121,"outputGeometry":2357,"fulltextStatus":22,"lidarModels":2358,"equipmentCount":77},"erasor2021","Lim et al., 2021","ERASOR","ERASOR: Egocentric Ratio of Pseudo Occupancy-Based Dynamic Object Removal for Static 3D Point Cloud Map Building","ERASOR 假設都市環境中多數動態物體與地面接觸，以自我中心的極座標區塊計算「偽佔據」（pseudo occupancy，區塊內高度差），比較查詢掃描與地圖子集的比值，找出可能含動態點的區塊；再以區域地面平面擬合（R-GPF）保留地面、剔除其上的動態點。方法不依賴光線追蹤或可視性判斷，但假設位姿已事先最佳化。",[35],"static point-cloud map","Egocentric ring-sector bins (R-POD) inside a volume of interest (Lmax 80 m, height from -1.0 m to 3.0 m relative to the ground); pseudo occupancy per bin = max minus min z; the Scan Ratio Test compares query and map pseudo occupancy bin by bin and selects bins whose scan ratio is below 0.2 as potentially dynamic (map bin occupied by an object that is absent in the query); Region-wise Ground Plane Fitting by PCA on the lowest seed points, three iterations, ground threshold tau_g = 0.15 (Sec. II-B to II-E, Sec. IV-A).","static point-cloud map with dynamic points removed",[45],{"id":2360,"label":2361,"shortName":2362,"title":2363,"year":51,"era":10,"cluster":98,"scope":99,"keyIdeaZh":2364,"sensors":2365,"mapRepresentation":2367,"loopClosure":72,"estimator":2368,"association":2369,"deskew":2370,"outputGeometry":2371,"fulltextStatus":22,"lidarModels":2372,"equipmentCount":77},"adalio2023","Lim et al., 2023","AdaLIO","AdaLIO: Robust Adaptive LiDAR-Inertial Odometry in Degenerate Indoor Environments","AdaLIO 以 Faster-LIO 為基礎，針對螺旋樓梯與走廊等狹窄室內空間中固定參數導致對應點驟減而發散的問題，加入自適應參數策略：當體素降取樣後的點數少於一般情況且多數佔用體素靠近感測器原點時，判定為類走廊的退化場景，改用較小的體素（0.2 m 改為 0.1 m）、較小的法向量搜尋半徑（3.0 m 改為 2.0 m）與較嚴格的平面殘差門檻（0.05 m 改為 0.025 m），以保留足夠且可靠的點到平面對應。濾波器、去畸變與體素地圖仍沿用 Faster-LIO。",[35,2366],"IMU (HILTI-Oxford dataset sensors; models not named in the paper)","global voxel map from Faster-LIO (incremental voxels) (Sec. III-B, Fig. 2)","iterated error-state Kalman filter on manifold inherited from Faster-LIO, with IMU forward and backward propagation (Sec. III-B)","Faster-LIO nearest-neighbour search and point-to-plane residuals on a voxelized scan; when degeneracy is detected the voxel size (0.2 to 0.1 m), search radius (3.0 to 2.0 m) and plane residual margin (0.05 to 0.025 m) are reduced (Sec. III-C, Table I)","IMU forward and backward propagation (Sec. III-B)","odometry and accumulated point map (Fig. 4)",[],{"id":2374,"label":2375,"shortName":2376,"title":2377,"year":661,"era":10,"cluster":68,"scope":69,"keyIdeaZh":2378,"sensors":2379,"mapRepresentation":2382,"loopClosure":2383,"estimator":2384,"association":2385,"deskew":121,"outputGeometry":2386,"fulltextStatus":22,"lidarModels":2387,"equipmentCount":554},"lim2024quatropp","Lim et al., 2024","Quatro++","Quatro++: Robust global registration exploiting ground segmentation for loop closing in LiDAR SLAM","Quatro++ 針對 LiDAR SLAM 迴圈閉合中的全域配準，處理機械旋轉式 LiDAR 點雲稀疏、以及離群剔除後剩下不足三個內點造成退化兩個問題。方法先以地面分割移除幾何資訊少的地面點，再做特徵匹配與最大團內點選擇，並假設地面載具以偏航旋轉為主，以 GNC 估計準 SO(3) 旋轉與分量式平移；滾轉與俯仰可由 INS 補償。作者在 KITTI、NAVER LABS、MulRan 與手持式 HiltiOxford 資料上評估，並接入 SLAM 迴圈模組。",[2380,2381],"3D LiDAR (Velodyne HDL-64E, VLP-16, Ouster OS1-64, HESAI XT32 across datasets)","INS optional for roll\u002Fpitch (Sec. 5.5)","LiDAR scans","coarse alignment in the loop-closing module of LeGO-LOAM; QSC-LeGO-LOAM combines ScanContext loop detection, Quatro++ and local registration, with MSE-based false-loop rejection","decoupled estimation on translation-invariant measurements: quasi-SO(3) (yaw-only) rotation by GNC truncated least squares with alternating weight updates (noise bound 0.3, at most 50 iterations, kappa = 1.4), then component-wise translation estimation (COTE); the c2f variant adds G-ICP fine alignment","Patchwork ground segmentation removes ground points; voxel sampling; FPFH descriptors with sensor-specific radii (nu \u003C r_normal \u003C r_FPFH, Table 1); reciprocal-test matching; MCIS-heuristic pruning that keeps the maximal clique found within a time threshold","relative pose (quasi-SE(3)) for loop constraints",[161,143,223,1365],{"id":2389,"label":2390,"shortName":2391,"title":2392,"year":207,"era":10,"cluster":68,"scope":69,"keyIdeaZh":2393,"sensors":2394,"mapRepresentation":2397,"loopClosure":2398,"estimator":2399,"association":2400,"deskew":57,"outputGeometry":2401,"fulltextStatus":22,"lidarModels":2402,"equipmentCount":185},"lim2025kissmatcher","Lim et al., 2025","KISS-Matcher","KISS-Matcher: Fast and Robust Point Cloud Registration Revisited","KISS-Matcher 從整體流程角度重新設計全域點雲配準，組合幾何抑制（如地面分割）、改良自 FPFH 的 Faster-PFH 特徵、以 k-core 為基礎的圖論離群剔除（降低 TEASER++ 最大團搜尋的時間複雜度）與 GNC 求解器，並釋出開源 C++ 函式庫。作者在 KITTI、MulRan 迴圈閉合基準以及以 FAST-LIO2 產生的多機器人地圖層級點雲上測試，報告精度與先進方法相當但速度大幅提升，可由掃描擴展到地圖層級。",[2395,2396],"64-channel 3D LiDAR (KITTI and MulRan, different ray patterns; models not named in the paper)","map clouds produced by FAST-LIO2-based SLAM on the Kimera-Multi dataset (sensors not stated in the paper)","scan-, submap- and map-level point clouds (voxelized) (Sec. IV-A)","evaluated on the loop-closing benchmark of Lim et al. (Sec. IV-A)","maximum k-core pruning of a pairwise-invariant compatibility graph (O(|V|+|E|), CSR storage, beta = 1.5v) followed by a GNC non-minimal solver; the final inlier count is used to reject failed registrations; all parameters scale with voxel size v (r_normal = 3.5v, r_FPFH = 5.0v)","geometric suppression (Patchwork ground segmentation in the experiments); Faster-PFH with one radius search per point, a linearity filter (tau_lin = 0.99) and minimum neighbour count (tau_num = 3); mutual (reciprocity) matching; top N_tau = 3,000 correspondences by descriptor distance ratio","rigid transformation from scan to map level",[45],{"id":2404,"label":2405,"shortName":2406,"title":2407,"year":396,"era":10,"cluster":151,"scope":99,"keyIdeaZh":2408,"sensors":2409,"mapRepresentation":2411,"loopClosure":2412,"estimator":2413,"association":2414,"deskew":2415,"outputGeometry":2416,"fulltextStatus":22,"lidarModels":2417,"equipmentCount":487},"loamlivox2020","Lin & Zhang, 2020","Loam_livox","Loam livox: A fast, robust, high-precision LiDAR odometry and mapping package for LiDARs of small FoV","Loam_livox 把 LOAM 流程改寫給小視野、非重複掃描的固態 LiDAR（Livox Mid-40）。前端依視野邊緣、回波強度、入射角與遮蔽關係剔除不可靠的點，並把反射率突變視為額外的邊緣特徵，以緩解小視野下特徵不足與退化。每一幀直接與全域特徵地圖配準，並以「分段處理」（piecewise processing，將一幀切成三個子幀分別配準）處理手持抖動造成的運動模糊。論文本身不含迴圈閉合。",[2410],"solid-state LiDAR (Livox Mid-40, 38.4 deg circular FoV, non-repetitive scan)","global maps of edge and planar features in memory; raw points saved to disk for possible offline processing (Fig. 5 caption)","none in this paper; a companion preprint (arXiv 1909.11811) and the repository README describe an added loop-closure module","iterative nonlinear least-squares pose optimization with 20% largest-residual trimming (Algorithm 1)","point selection by FoV fringe, intensity, incidence angle and occlusion; LOAM-style smoothness features plus reflectivity-change edges; edge-to-edge and plane-to-plane residuals using 5 nearest map points with eigenvalue checks (Sec. III, IV-A, IV-B)","piecewise processing (three sub-frames matched independently) or linear pose interpolation; piecewise preferred for jerky handheld motion (Sec. IV-C, V-A)","feature maps and 20 Hz odometry; raw points retained on disk (Fig. 5 caption)",[2418],"Livox MID40",{"id":2420,"label":2421,"shortName":2422,"title":2423,"year":97,"era":10,"cluster":263,"scope":99,"keyIdeaZh":2424,"sensors":2425,"mapRepresentation":2431,"loopClosure":2432,"estimator":2433,"association":2434,"deskew":2435,"outputGeometry":2436,"fulltextStatus":22,"lidarModels":2437,"equipmentCount":554},"r3live2022","Lin & Zhang, 2022","R3LIVE","R$^3$LIVE: A Robust, Real-time, RGB-colored, LiDAR-Inertial-Visual tightly-coupled state Estimation and mapping package","R3LIVE 由光達慣性里程計（LIO，沿用 FAST-LIO2）重建幾何結構，視覺慣性里程計（VIO）則為地圖點上色並同時估計狀態。VIO 先以光流追蹤點的 PnP 重投影誤差粗估，再以地圖點 RGB 與當前影像的光度誤差做幀對地圖（frame-to-map）精修，不需擷取顯著視覺特徵。每個地圖點的顏色與共變異數以貝氏更新融合多次觀測，arXiv 預印本另描述離線網格化、貼圖與 pcd、ply、obj 匯出工具，但 ICRA 正式版刪去此節。",[2426,2427,2428,2429,2430],"3D LiDAR (Livox AVIA, FoV 70.4 x 77.2 deg)","IMU (labelled 'Build-in IMU' of the LiDAR in the device figure; model not reported)","global-shutter RGB camera (FLIR Blackfly BFS-u3-13y3c, FoV 82.9 x 66.5 deg)","D-GPS RTK system (reference only)","ArUco marker board (drift reference in GPS-denied tests)","fixed-size voxels (e.g., 0.1 m) holding points with XYZ and RGB; per-point color covariance updated by Bayesian fusion","none (drift reported as trajectories closing without loop detection, Sec. VI-C)","error-state iterated Kalman filter shared by a LiDAR-inertial subsystem (FAST-LIO2) and a visual-inertial subsystem; camera extrinsic, intrinsic and camera-IMU time offset in the state","LiDAR raw points with point-to-plane residuals; VIO first minimizes frame-to-frame PnP reprojection error of optical-flow-tracked map points, then frame-to-map photometric error between map-point RGB and image","motion compensation in the LIO subsystem (Fig. 2; inherited from FAST-LIO2)","dense RGB-colored point cloud in real time (VoR Fig. 1, Fig. 7); offline mesh (Delaunay triangulation plus graph cuts in CGAL) with vertex-color texturing and export to pcd, ply, obj are described only in arXiv v1 Sec. VII-A, not in the ICRA version of record",[437],{"id":2439,"label":2440,"shortName":2441,"title":2442,"year":661,"era":10,"cluster":263,"scope":99,"keyIdeaZh":2443,"sensors":2444,"mapRepresentation":2448,"loopClosure":2449,"estimator":2450,"association":2451,"deskew":2452,"outputGeometry":2453,"fulltextStatus":22,"lidarModels":2454,"equipmentCount":423},"r3livepp2024","Lin & Zhang, 2024","R3LIVE++","R3LIVE++: A Robust, Real-Time, Radiance Reconstruction Package With a Tightly-Coupled LiDAR-Inertial-Visual State Estimator","R3LIVE++ 延伸 R3LIVE，在 VIO 中加入相機光度校正（響應函數與暗角）及曝光時間的線上估計，使地圖點儲存的是與曝光無關的輻射值（radiance）而非原始顏色。作者在 NCLT 公開資料集的 25 個序列上比較定位精度，並以自建資料集評估退化場景的穩健性與輻射地圖誤差。論文亦指出地圖點密度（約 1 cm）與光達原始點密度限制了可重建的影像細節。",[2445,2446,2447],"3D LiDAR (LiVOX AVIA in the R3LIVE-dataset; NCLT 3D LiDAR, model not named in the paper)","IMU (model not reported)","RGB camera (FLIR Blackfly BFS-u3-13y3c global shutter in the R3LIVE-dataset; front-facing camera of the NCLT omnidirectional camera)","radiance map: points with position, RGB radiance, position and radiance covariances and timestamps, stored in 0.1 m voxels","none (Sec. VI-D states the system has no loop detection and correction)","error-state iterated Kalman filter; state includes camera extrinsic, intrinsic, camera-IMU time offset and inverse exposure time","LiDAR point-to-plane (GICP-style) scan-to-map against the five nearest map points in an incremental k-d tree, with a plane fitted only if they lie within about 0.4 m (VoR Sec. IV-A); VIO tracks about 400 map points at least 50 pixels apart by Lucas-Kanade optical flow, first minimizing frame-to-frame PnP error, then frame-to-map radiance error on individual map points using photometrically corrected images (CRF and vignetting); pixels at 0 or 255 are excluded from radiance updates (VoR Sec. V)","in-frame motion of each LiDAR scan compensated by IMU backward propagation (following FAST-LIO) before registration (Sec. IV-A)","radiance (RGB) point map at 1 cm point spacing in the implementation; offline mesh and texture via CGAL and OpenMVS with export to pcd, ply, obj (arXiv v1 Sec. 7.2; the version of record refers to the GitHub utilities, Unreal Engine export and supplementary material); HDR images rendered at chosen exposure times (Sec. VII-A, VIII-B)",[437,45,2455],"planar LiDAR (not used)",{"id":2457,"label":2458,"shortName":2459,"title":2460,"year":249,"era":10,"cluster":263,"scope":99,"keyIdeaZh":2461,"sensors":2462,"mapRepresentation":2466,"loopClosure":2467,"estimator":2468,"association":2469,"deskew":2470,"outputGeometry":2471,"fulltextStatus":22,"lidarModels":2472,"equipmentCount":487},"r2live2021","Lin et al., 2021","R2LIVE","R$^2$LIVE: A Robust, Real-Time, LiDAR-Inertial-Visual Tightly-Coupled State Estimator and Mapping","R2LIVE 在單一誤差狀態迭代卡爾曼濾波器（ESIKF）中，同時以光達平面特徵的點到平面殘差與視覺角點的重投影誤差更新狀態，達成高頻率的緊密耦合里程計。另以滑動視窗因子圖最佳化精修影像關鍵影格位姿與視覺地標，並線上估計相機與 IMU 間的時間偏移。作者以手持裝置在長隧道狀的地鐵站與大型建物內外展示可重建稠密點雲。",[2426,2463,2464,2465],"IMU (model not reported; 200 Hz in the Fig. 3 illustration)","monocular global-shutter camera (FLIR Blackfly BFS-u3-13y3c, FoV 82.9 x 66.5 deg; model named only in the version of record)","D-GPS RTK system (reference only; footnote links the DJI D-RTK product page)","point cloud map (LiDAR frames appended after update) plus sparse visual landmarks","none (Sec. VI-D notes loop returned without loop closure)","on-manifold error-state iterated Kalman filter fusing LiDAR and visual measurements, plus sliding-window factor graph refining image keyframe poses and visual landmarks (LiDAR poses fixed)","LiDAR planar feature points, point-to-plane residual against nearest map points (LOAM\u002FFAST-LIO style); FAST corners tracked by KLT optical flow with PnP reprojection residuals to triangulated landmarks","in-frame motion compensated by IMU backward propagation as in FAST-LIO (Sec. IV-B)","dense 3D point cloud map of building interiors and exteriors (Fig. 1, Fig. 12); RGB coloring not described",[437],{"id":2474,"label":2475,"shortName":2476,"title":2477,"year":51,"era":10,"cluster":31,"scope":99,"keyIdeaZh":2478,"sensors":2479,"mapRepresentation":2482,"loopClosure":2483,"estimator":2484,"association":2485,"deskew":2486,"outputGeometry":2487,"fulltextStatus":22,"lidarModels":2488,"equipmentCount":377},"lin2023immesh","Lin et al., 2023","ImMesh","ImMesh: An Immediate LiDAR Localization and Meshing Framework","ImMesh 以 VoxelMap 的機率平面與迭代卡爾曼濾波估計位姿，並把經空間降採樣、配準後的 LiDAR 點當成網格頂點（以 ikd-Tree 維持頂點最小間距）；每個有新點的體素將其頂點投影到該體素主平面上，以二維 Delaunay 三角化建立三角面，再以類似 git 的 pull、commit、push 步驟增量合併到全域網格。整體在一般桌上型 CPU 上即時執行。",[2480,2481],"3D LiDAR (spinning and solid-state)","IMU (optional)","spatially downsampled, registered LiDAR points kept as mesh vertices with a minimum spacing (0.15 m for mechanical, 0.10 m for solid-state LiDAR) enforced by an ikd-Tree, stored in hashed voxels (0.60 m or 0.40 m) and hashed regions (15 m or 10 m); triangle facets stored per region and indexed in a facet hash table","none (stated limitation, Sec. IX)","iterated Kalman filter with probabilistic planes (built on VoxelMap)","point-to-plane registration against voxel plane features (VoxelMap)","in-frame motion distortion compensated by IMU backward propagation (method of FAST-LIO) before registration (Sec. V-A)","triangle mesh published at scan rate; also point cloud reinforcement via mesh rasterization",[437,161,225,504],{"id":2490,"label":2491,"shortName":2492,"title":2493,"year":661,"era":10,"cluster":755,"scope":132,"keyIdeaZh":2494,"sensors":2495,"mapRepresentation":2496,"loopClosure":2497,"estimator":2498,"association":2499,"deskew":57,"outputGeometry":2500,"fulltextStatus":22,"lidarModels":2501,"equipmentCount":77},"dpvslam2024","Lipson et al., 2024","DPV-SLAM","Deep Patch Visual SLAM","DPV-SLAM 在稀疏影像區塊（patch）視覺里程計 DPVO 上加入兩種迴圈閉合，讓深度學習式單眼 SLAM 可在單張 GPU 上以穩定的影格速率運作。近距迴圈閉合依相機位置偵測重訪，只保留舊影格的區塊特徵並建立指向近期影格的單向邊，再以自製的 CUDA 區塊稀疏光束法平差做全域最佳化；DPV-SLAM++ 另以 DBoW2 影像檢索、特徵匹配與 RANSAC 加 Umeyama 估計 Sim(3) 漂移，於 CPU 執行位姿圖最佳化以修正尺度漂移。輸出僅為相機軌跡與稀疏點。",[37],"patch graph: sparse image patches with inverse depth attached to frames; only sparse 3D reconstruction (Sec. 3.1, Sec. 5)","proximity loop closure: uni-directional long-range edges from stored patches of old frames to recent frames, followed by global BA; optional classical loop closure (DPV-SLAM++) with image retrieval requiring consecutive detections and Sim(3) drift estimation (Sec. 3.2-3.3)","DPVO recurrent update operator predicts sparse patch-flow residuals and confidences; poses and patch inverse depths solved by bundle adjustment on the patch graph; a CUDA block-sparse BA performs global optimization with loop factors; DPV-SLAM++ adds a CPU Sim(3) pose-graph optimization solved by Levenberg-Marquardt (Sec. 3)","sparse, randomly selected p x p patches tracked by learned optical flow from correlation features; loop candidates by camera proximity (DPV-SLAM) and additionally by DBoW2 ORB image retrieval with off-the-shelf keypoint matching, structure-only BA and RANSAC plus Umeyama Sim(3) alignment (DPV-SLAM++) (Sec. 3.1-3.3)","camera trajectory and sparse 3D points of tracked patches (Sec. 5)",[],{"id":2503,"label":2504,"shortName":2505,"title":2506,"year":661,"era":10,"cluster":755,"scope":132,"keyIdeaZh":2507,"sensors":2508,"mapRepresentation":2509,"loopClosure":2510,"estimator":2511,"association":2512,"deskew":57,"outputGeometry":2513,"fulltextStatus":22,"lidarModels":2514,"equipmentCount":24},"loopyslam2024","Liso et al., 2024","Loopy-SLAM","Loopy-SLAM: Dense Neural SLAM with Loop Closures","Loopy-SLAM 在 Point-SLAM 的神經點雲上加入子地圖、詞袋式全域地點辨識與穩健位姿圖最佳化，迴圈閉合後直接剛性平移子地圖中的點以修正地圖，毋須保存全部歷史影格。作者未研究光束調整精修，且實作尚非即時。",[772],"submaps of neural point clouds","Global place recognition with a bag-of-visual-words database (DBoW3) queried when a submap is completed (top K = 4 on Replica, 1 on TUM-RGBD and ScanNet, dynamic similarity threshold); on ScanNet a PGO is triggered on average every 151 frames (Sec. 3.2, Sec. 4, App. E Table 8)","frame-to-model tracking on neural point submaps + robust pose graph optimization (line process) over global keyframes","Frame-to-model tracking by minimizing depth and colour re-rendering losses on the active submap (Point-SLAM style); loop edges from coarse-to-fine dense registration of submap surfaces: all depth frames of a submap are TSDF-fused and points sampled on the marching-cubes surface, FPFH features with RANSAC give the coarse alignment and ICP on full-resolution clouds refines it; loop edges pre-filtered by constraint translation magnitude and a fitness (overlap) score (Sec. 3.1, 3.2)","Mesh from TSDF fusion (1 cm voxels) of depth and colour rendered every fifth frame along the estimated trajectory, following Point-SLAM; rendered RGB-D images (Sec. 4 implementation details)",[],{"id":2516,"label":2517,"shortName":2518,"title":2519,"year":249,"era":10,"cluster":599,"scope":676,"keyIdeaZh":2520,"sensors":2521,"mapRepresentation":2522,"loopClosure":72,"estimator":2523,"association":2524,"deskew":2525,"outputGeometry":2526,"fulltextStatus":22,"lidarModels":2527,"equipmentCount":24},"balm2021","Liu & Zhang, 2021","BALM","BALM: Bundle Adjustment for Lidar Mapping","BALM 將光達束調整（LiDAR bundle adjustment, BA）定義為最小化各特徵點到其所屬邊緣或平面的距離，並證明邊緣與平面參數可用封閉解消去，使最佳化只剩下掃描位姿，因而可以納入大量稠密平面與邊緣特徵。作者推導代價函數對位姿的一階與二階解析導數，並提出自適應體素化（adaptive voxelization），以八元樹遞迴切分空間，直到每個體素只含單一平面或邊緣。此 BA 被整合為 LOAM 架構的後端，在滑動視窗內做局部地圖精修。",[35],"adaptive voxel map of edge and plane features (hash table of octrees); points outside the window summarized by recursive covariance statistics","sliding-window local BA over scan poses with closed-form elimination of edge\u002Fplane parameters and analytical first\u002Fsecond-order derivatives (second-order approximation, LM-type iterations)","edge and plane feature points (LOAM-style extraction) grouped by adaptive voxelization from a default voxel size down to a minimal size (sizes given only as examples: 1 m and 0.125 m) using an eigenvalue test on the point covariance; separate voxel maps for edges and planes stored as hash-indexed octrees; voxels with repeated eigenvalues are skipped and dense voxels may average points per scan; scan-to-map odometry matches each point to the nearest voxel plane or edge instead of five nearest points (Sec. IV Remarks 1-4, Sec. V, Sec. VI-D)","not compensated in the scan-to-map odometry front-end (authors state this in Sec. VII)","registered LiDAR feature point-cloud map and refined scan poses; no other exportable product reported",[1482,2418,143],{"id":2529,"label":2530,"shortName":2531,"title":2532,"year":51,"era":10,"cluster":599,"scope":676,"keyIdeaZh":2533,"sensors":2534,"mapRepresentation":2535,"loopClosure":2536,"estimator":2537,"association":2538,"deskew":2539,"outputGeometry":2540,"fulltextStatus":22,"lidarModels":2541,"equipmentCount":423},"balm2_2023","Liu et al., 2023a","BALM2 (BALM 2.0)","Efficient and Consistent Bundle Adjustment on Lidar Point Clouds","BALM2 延續以點到平面或邊緣之歐氏距離為殘差的光達 BA，並提出「點簇（point cluster）」概念，把同一特徵上的所有原始點壓縮為一組緊湊參數，使代價、導數與不確定度計算都不需逐點列舉。作者推導封閉形式的 Jacobian 與 Hessian、其零空間與稀疏性，據此建立二階求解器，並利用二階資訊估計由量測雜訊造成的位姿不確定度。論文另示範其用於光達慣性里程計、多光達外參校正與全域建圖。",[35],"point clusters per plane\u002Fedge feature inside adaptive voxels","none within the method; relies on the supplied initial trajectory","second-order (Newton\u002FLM-type) batch solver over poses after closed-form elimination of plane\u002Fedge parameters; point-cluster coordinates aggregate raw points; LDLT solve (Eigen); pose covariance from second-order information","raw points to plane or edge features via BALM adaptive voxelization of all points registered with an initial trajectory (root voxel 1 m for Hilti, 2 m for VIRAL and UrbanLoco, at most 3 layers, at least 20 points and threshold 1\u002F25 for the feature test); the incremental ICP trajectory served as the common initialization (Sec. VI-B)","no in-scan motion model inside the BA; real-world scans were deskewed by FAST-LIO2 before BA and its odometry output discarded (Sec. VI-B); the simulation ignored in-frame distortion (Sec. V); continuous-time trajectories are only discussed as an extension (Sec. VIII-C)","optimized scan poses with estimated pose covariance; registered point cloud",[224,504,225,2542,437,2543],"MID-100","16-channel lidar (simulated)",{"id":2545,"label":2546,"shortName":2547,"title":2548,"year":51,"era":10,"cluster":599,"scope":676,"keyIdeaZh":2549,"sensors":2550,"mapRepresentation":2552,"loopClosure":2553,"estimator":2554,"association":2555,"deskew":2556,"outputGeometry":2557,"fulltextStatus":22,"lidarModels":2558,"equipmentCount":144},"hba2023","Liu et al., 2023b","HBA","Large-Scale LiDAR Consistent Mapping Using Hierarchical LiDAR Bundle Adjustment","HBA 針對大場景下原始光達 BA 計算量過大的問題，採「由下而上」分層 BA：在小視窗內做局部 BA 並把視窗內各幀合併為上一層的關鍵影格，逐層向上，最後在頂層做全域 BA；再「由上而下」以位姿圖最佳化把結果平滑回傳到所有原始幀位姿，並以局部 BA 的 Hessian 作為資訊矩陣。作者依計算複雜度推導最佳層數。",[2551],"3D LiDAR (mechanical spinning in the public datasets; solid-state LiDAR of ref. [26], retina-like incommensurable scanning, in the self-collected data)","layered keyframe point clouds; adaptive voxel plane features","no place recognition module; can close gaps when the initial trajectory lacks loop closure if overlapping geometry is associated (Sec. IV-A2)","bottom-up hierarchical local BA in sliding windows (window 10, stride 5, parallel threads) plus global BA on top layer, followed by top-down pose-graph optimization using BA Hessians as information matrices","plane features via adaptive voxelization (BALM) in each layer","input may be raw or deskewed scans (Sec. III-A); deskew not performed by HBA","globally consistent point-cloud map and optimized poses",[45],{"id":2560,"label":2561,"shortName":2562,"title":2563,"year":661,"era":10,"cluster":263,"scope":132,"keyIdeaZh":2564,"sensors":2565,"mapRepresentation":2569,"loopClosure":2570,"estimator":2571,"association":2572,"deskew":2573,"outputGeometry":2574,"fulltextStatus":22,"lidarModels":2575,"equipmentCount":24},"glio2024","Liu et al., 2024","GLIO","GLIO: Tightly-Coupled GNSS\u002FLiDAR\u002FIMU Integration for Continuous and Drift-Free State Estimation of Intelligent Vehicles in Urban Areas","GLIO 在因子圖中緊耦合 GNSS 原始量測、LiDAR 與 IMU：第一階段以滑動視窗融合基準站差分後的雙差虛擬距離、都卜勒、IMU 預積分與 LiDAR 掃描對地圖平面因子；第二階段在獨立執行緒上對關鍵影格做批次最佳化，每個關鍵影格與 12 個相鄰影格建立掃描對多掃描約束，並逐步排除離群量測。GNSS 提供全域約束消除漂移，LiDAR 與 IMU 則在高樓遮蔽、GNSS 品質差時維持連續估測。",[2566,2567,2568],"low-cost GNSS receiver raw pseudorange and Doppler (u-blox F9P) with reference-station corrections","IMU (Xsens Ti-10)","3D LiDAR (Velodyne HDL-32E)","LiDAR keyframe feature map in the global (GNSS) frame","no explicit loop closure; global drift is removed by GNSS factors","two-stage factor graph optimization in Ceres on a LILI-OM base: (1) sliding-window fusion of double-differenced pseudorange, Doppler, IMU preintegration, scan-to-map planar LiDAR factors and marginalization; (2) batch optimization over keyframes with scan-to-multiscan LiDAR factors and relative attitude constraints, run on a separate thread (max 50 iterations or 3 s) with outlier exclusion (Sec. III)","planar LiDAR features (100 per keyframe in the first stage; 25 random planar features per frame pair with 12 adjacent keyframes in the second stage); GNSS raw measurements double-differenced with a reference station; GNSS epochs associated to LiDAR keyframes by interpolation (Sec. III)","not described; the LiDAR front end is adopted from LILI-OM","vehicle trajectory in a global frame and a LiDAR feature map",[225],{"id":2577,"label":2578,"shortName":2579,"title":2580,"year":207,"era":10,"cluster":755,"scope":32,"keyIdeaZh":2581,"sensors":2582,"mapRepresentation":2584,"loopClosure":2585,"estimator":2586,"association":2587,"deskew":57,"outputGeometry":2588,"fulltextStatus":22,"lidarModels":2589,"equipmentCount":77},"slam3r2025","Liu et al., 2025","SLAM3R","SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos","SLAM3R 以前饋式神經網路直接從單眼 RGB 影片產生稠密點雲，而不求解任何相機參數。影片先以滑動視窗切成重疊片段，影像對點雲（I2P）網路以多視角交叉注意力，從 11 張影像回歸視窗中間關鍵影格的點雲；局部對世界（L2W）網路再參考以檢索模組從儲存池挑出的歷史影格，把局部點雲逐步配準到全域座標。兩個網路都以 DUSt3R 權重初始化，在單張 4090D 上約每秒 24 至 25 影格，但因沒有相機參數，無法做全域光束法平差，推導出的位姿也不如專門的 SLAM。",[2583],"monocular RGB camera (video)","global dense point cloud built from registered pointmaps with per-pixel confidence (confidence threshold 3 in evaluation) (Sec. 3, Supp. B)","none as optimization; retrieval of long-term scene frames acts as implicit re-localization during registration (Sec. 4.2)","feed-forward networks without explicit camera parameters: an Image-to-Points (I2P) ViT with multi-view cross-attention regresses keyframe pointmaps from sliding-window clips (L = 11), and a Local-to-World (L2W) network registers each keyframe's pointmap into the global frame using retrieved scene frames; no pose optimization or bundle adjustment (Sec. 3)","implicit through cross-attention between keyframe and supporting or scene-frame tokens; a learned retrieval module selects the top-K scene frames from a reservoir of registered frames (Sec. 3.2, Supp. A)","dense 3D point cloud (colored by input frames); camera poses only as a derived by-product via PnP-RANSAC (Sec. 4.1)",[],{"id":2591,"label":2592,"shortName":2593,"title":2594,"year":30,"era":10,"cluster":98,"scope":132,"keyIdeaZh":2595,"sensors":2596,"mapRepresentation":2597,"loopClosure":2598,"estimator":2599,"association":2600,"deskew":2601,"outputGeometry":2602,"fulltextStatus":22,"lidarModels":2603,"equipmentCount":377},"voxelslam2026","Liu et al., 2026","Voxel-SLAM","Voxel‐SLAM: A Complete, Accurate, and Versatile Light Detection and Ranging‐Inertial Simultaneous Localization and Mapping System","Voxel-SLAM 以同一種自適應體素地圖貫穿初始化、里程計、局部建圖、迴圈與全域建圖五個模組，並依作者所稱的短期、中期、長期與多地圖四類資料關聯設計。局部建圖以滑動視窗 LiDAR-慣性光束法平差（bundle adjustment, BA）同時修正狀態與地圖；迴圈偵測以 BTC 描述子並加上平面約束與漂移比例檢查，觸發位姿圖最佳化與地圖重建。全域建圖以階層式 BA 從關鍵影格視窗到子地圖逐層最佳化，以提升多次作業地圖的一致性；其地圖表示明言沿用 VoxelMap。",[35,36],"adaptive voxel map (planes with point clusters) shared by initialization, odometry, local mapping, loop closure and global mapping","BTC descriptor place recognition with geometric verification, plus checks for plane constraints in three independent directions and a drift-to-travel-distance ratio (e.g., at most 1%); works within and across sessions","Odometry: IMU propagation with motion compensation, then scan-to-map registration against the adaptive voxel map following VoxelMap (point-to-plane distances with per-point noise plus IMU residuals update the current state); the rough LIO inside initialization follows FAST-LIO; local mapping: sliding-window LiDAR-inertial BA (window 10) combining BALM2 point-cluster BA factors with IMU preintegration, solved by Levenberg-Marquardt with analytic Jacobian and Hessian; PGO via GTSAM after loops; hierarchical global BA","point-to-plane to adaptive voxel-map planes; BA over point clusters per voxel","IMU propagation compensates in-scan motion distortion within the odometry module (Sec. VI-A)","globally optimized multi-session point-cloud map and scan poses",[1365,437,225,223],{"id":2605,"label":2606,"shortName":2607,"title":2608,"year":67,"era":52,"cluster":31,"scope":32,"keyIdeaZh":2609,"sensors":2610,"mapRepresentation":2611,"loopClosure":57,"estimator":57,"association":57,"deskew":57,"outputGeometry":2612,"fulltextStatus":22,"lidarModels":2613,"equipmentCount":46},"lorensen1987marchingcubes","Lorensen & Cline, 1987","Marching Cubes","Marching cubes: A high resolution 3D surface construction algorithm","此演算法把三維體積資料（原文為 CT、MR、SPECT 醫學影像）以相鄰兩張切片各四個像素組成邏輯立方體，依八個頂點數值是否達到門檻得到 8 位元索引，查詢由 256 種情形（利用互補與旋轉對稱歸納為 14 種樣式）建立的邊交點表決定三角面拓樸，再以線性內插定出三角形頂點，並以中央差分梯度內插出頂點法向量，藉此擷取等值面（isosurface）成為三角網格。後續 TSDF 管線（如 Voxblox、nvblox、VDBFusion、Mesh-LOAM）與 Poisson 類重建常以它或其八元樹變體把隱式場轉為可輸出的網格；體素雜湊等系統也可改以射線投射（ray casting）直接渲染表面。",[],"scalar volume (3D grid) to triangle mesh","triangle mesh of a constant-value isosurface with per-vertex unit normals from interpolated density gradients (Sec. 4)",[],{"id":2615,"label":2616,"shortName":2617,"title":2618,"year":2619,"era":52,"cluster":527,"scope":132,"keyIdeaZh":2620,"sensors":2621,"mapRepresentation":2622,"loopClosure":2623,"estimator":2624,"association":2625,"deskew":2626,"outputGeometry":2627,"fulltextStatus":22,"lidarModels":2628,"equipmentCount":46},"lu_milios1997","Lu & Milios, 1997","Lu-Milios global scan alignment","Globally Consistent Range Scan Alignment for Environment Mapping",1997,"本文把多幅距離掃描的一致化配準（registration）表述為「位姿網路」上的最佳估計：每幅掃描以機器人位姿為局部座標，掃描對匹配與里程計分別提供強連結與弱連結的相對位姿約束，再以最大概似準則同時求解所有位姿。作者以反覆線性化求解，文中表示約四到五次迭代即收斂。作者明言方法假設機器人停下來取得完整掃描，連續掃描造成的量測時間不一致（即今日的去畸變問題）不在本文範圍內。",[1330,119],"set of registered 2D range scans (point sets) attached to estimated poses","implicit: any sufficiently overlapping scan pair creates a strong link, including revisits","maximum-likelihood (weighted least squares) over a pose network, closed-form linear solution iterated with re-linearization","pairwise scan matching (point-to-point matching or an extension of Cox's point-to-line matching) initialised from odometry, producing corresponding point sets; before matching, points likely not visible from the other pose are discarded, and a strong link is created only when the overlapping spatial extent exceeds a fixed fraction of the extent covered by both scans; with the 220 degree sensor, similar headings are also needed for overlap (Sec. 2.1, 2.2, 5.2)","none; the approach assumes the robot stops to collect each complete scan, and continuous-scan distortion is declared out of scope (Sec. 6)","globally registered 2D scan points and pose estimates with covariance",[2629,2630],"Ladar 2D IBEO Lasertechnik","SICK laser range scanner",{"id":2632,"label":2633,"shortName":2634,"title":2635,"year":363,"era":52,"cluster":53,"scope":99,"keyIdeaZh":2636,"sensors":2637,"mapRepresentation":2641,"loopClosure":2642,"estimator":2643,"association":2644,"deskew":57,"outputGeometry":2645,"fulltextStatus":22,"lidarModels":2646,"equipmentCount":144},"lupton2012preint","Lupton & Sukkarieh, 2012","Lupton-Sukkarieh preintegration","Visual-Inertial-Aided Navigation for High-Dynamic Motion in Built Environments Without Initial Conditions","本文為消防等第一應變人員的人員攜帶定位需求，提出在上一個位姿的機體座標系中積分 IMU 量測，形成不需初始條件的「預積分慣性增量觀測」（位置修正項 Δp+、速度增量與姿態增量，姿態以 Euler 角表示），並以 Jacobian 在積分後修正 IMU 偏差。導航座標系改固定於第一個位姿的機體座標，使初始姿態已知，改以線性方式估計重力向量與初始速度；理論上三個位姿的相對位置即可線性恢復初始條件，不需特殊初始化。系統以資訊空間的圖式 SLAM 實作，並以「滑動視窗強制獨立」（SWFI，30 個影像位姿）移除舊位姿與其觀測以維持固定計算時間；視覺前端為 Harris 角點、金字塔 LK 光流與 RANSAC 基本矩陣剔除。作者以手持 Honeywell HG1900 IMU 與 Bumblebee2 立體相機在雪梨大學辦公建物中驗證。Forster 等人與 GTSAM 文件把 IMU 預積分概念歸於此工作。",[2638,2639,2640],"IMU (Honeywell HG1900, 600 Hz)","stereo camera (Point Grey Research Bumblebee2 with 2.1 mm wide-angle lenses, 6.25 Hz, 12 cm baseline)","monocular camera (left camera of the stereo pair only, Sec. VIII-G)","sparse 3D point landmarks for features currently tracked within the window; removed when no longer observed","none; loop closure detection is listed as future work","information-space graphical SLAM (delayed-state filter that can relinearize past observations and re-assess data association) with preintegrated inertial delta observations; sliding window forced independence removes poses older than the 30-image window together with all their observations; state: positions, velocities and Euler attitudes of window poses, gravity vector, IMU-to-camera rotation and offset, IMU biases and tracked landmarks","Harris corners (20 to 50 per image) tracked with pyramidal Lucas-Kanade optical flow; RANSAC fundamental-matrix test between consecutive images; graph edge-energy test (above 2 sigma) to re-assess data association after insertion; no descriptor-based re-acquisition","platform position, velocity and attitude plus gravity vector, IMU-camera alignment and IMU biases; no dense map",[],{"id":2648,"label":2649,"shortName":2650,"title":2651,"year":396,"era":10,"cluster":397,"scope":1035,"keyIdeaZh":2652,"sensors":2653,"mapRepresentation":2655,"loopClosure":72,"estimator":2656,"association":2657,"deskew":2658,"outputGeometry":2659,"fulltextStatus":22,"lidarModels":2660,"equipmentCount":46},"lv2020licalib","Lv et al., 2020","LI-Calib","Targetless Calibration of LiDAR-IMU System Based on Continuous-time Batch Estimation","LI-Calib 以連續時間 B 樣條表示 IMU 軌跡，使每個 LiDAR 點的取樣時刻都能取得位姿，並直接以原始加速度與角速度殘差和點對面元（surfel）距離聯合最佳化外參。流程先對齊 LiDAR 與 IMU 旋轉初始化外參旋轉，再迭代去除運動畸變、重建面元地圖與更新對應。此法假設兩感測器已硬體同步，只估計空間外參。",[1038,2654],"IMU (three Xsens MTi-100 series)","surfel map (0.5 m cells indoor, 1.0 m outdoor)","continuous-time batch optimization (Levenberg-Marquardt, Kontiki toolkit) with split cubic B-splines in R3 and SO(3), knot spacing 0.02 s; state holds extrinsic rotation and translation, spline control points, gravity alignment and IMU biases; residuals from raw accelerometer and gyroscope readings and point-to-surfel distances","point-to-surfel: map split into cells (0.5 m indoor, 1.0 m outdoor); RANSAC plane fitted in cells whose plane-likeness exceeds 0.6 in the first iteration and 0.7 afterwards; correspondences beyond a distance threshold rejected; raw scans randomly downsampled","points reprojected with spline poses; rotation deskew after initialization, full deskew after each batch iteration","extrinsic LiDAR-IMU transform; undistorted point cloud map",[143],{"id":2662,"label":2663,"shortName":2664,"title":2665,"year":249,"era":10,"cluster":98,"scope":132,"keyIdeaZh":2666,"sensors":2667,"mapRepresentation":2670,"loopClosure":2671,"estimator":2672,"association":2673,"deskew":2674,"outputGeometry":2675,"fulltextStatus":22,"lidarModels":2676,"equipmentCount":109},"clins2021","Lv et al., 2021","CLINS","CLINS: Continuous-Time Trajectory Estimation for LiDAR-Inertial System","CLINS 以兩組累積式均勻三次 B 樣條分別表示位置與旋轉，將 LiDAR 慣性系統的軌跡建模為連續時間函數。每個新掃描到達時先以 IMU 積分初始化新增控制點，再在局部視窗內把 LOAM 邊緣與平面特徵以各點自身時間戳記的位姿投影到關鍵掃描子地圖，與原始加速度計、陀螺儀殘差一起做非剛性配準，同時估計控制點與 IMU 偏差，因此去畸變與位姿估計在同一最佳化中完成。迴圈閉合時採兩階段修正：先對關鍵掃描的離散位姿做位姿圖最佳化，再以修正後的位姿為錨點、以原軌跡的局部線速度與角速度維持局部形狀，重新擬合樣條控制點。",[2668,2669],"3D spinning LiDAR (Velodyne VLP-16 in the YQ and KAIST Urban tests)","IMU (Xsens MTi-300)","local submap of key-scans selected by spatio-temporal distance (feature points); global point map assembled from undistorted scans along the continuous trajectory","yes; loop closures between key-scans trigger a two-stage correction (loop detection method not described in the paper) (Sec. IV-C)","sliding-window batch nonlinear least squares (MAP) over the active B-spline control points and IMU biases, tightly coupling LiDAR feature residuals with raw accelerometer and gyroscope residuals; Levenberg-Marquardt in Ceres with automatic differentiation; static control points enter the problem but stay fixed (Sec. IV-B)","LOAM-style edge and planar features selected by local curvature; nearest-neighbour correspondences in a local submap of key-scans; point-to-line and point-to-plane residuals evaluated with the pose at each point's own timestamp (Sec. IV-B)","implicit in the non-rigid registration: every point is mapped with the spline pose at its timestamp; key-scans added to the submap are undistorted to scan start with the non-active trajectory (Sec. IV-B)","continuous-time trajectory queryable at any timestamp (evaluated at 100 Hz) and an assembled 3D point cloud map; the trajectory was also used to assemble 2D SICK LMS-511 scans into a dense 3D reconstruction (Fig. 1, Sec. V-C)",[143,2677],"SICK LMS-511",{"id":2679,"label":2680,"shortName":2681,"title":2682,"year":51,"era":10,"cluster":263,"scope":132,"keyIdeaZh":2683,"sensors":2684,"mapRepresentation":2687,"loopClosure":2688,"estimator":2689,"association":2690,"deskew":2691,"outputGeometry":2692,"fulltextStatus":22,"lidarModels":2693,"equipmentCount":736},"clic2023","Lv et al., 2023","CLIC","Continuous-Time Fixed-Lag Smoothing for LiDAR-Inertial-Camera SLAM","CLIC 以分段三次 B 樣條表示連續時間軌跡，在固定時間長度的滑動視窗內做平滑：LiDAR 點到平面、原始 IMU、偏差與視覺重投影因子都在各自量測時刻取軌跡位姿，並推導解析雅可比矩陣、以邊緣化保留舊狀態的資訊，使連續時間方法可即時運行。框架可接入一至兩顆 LiDAR 與相機，並線上估計相機與 IMU 相對 LiDAR 的時間偏移。",[2685,36,2686],"one or two 3D LiDARs","monocular camera (optional)","LiDAR feature map plus visual landmarks (Fig. 10)","yes, Euclidean distance based detection with the two-stage continuous-time trajectory correction of CLINS (Sec. V)","continuous-time fixed-lag smoothing over a split cubic B-spline trajectory (knot spacing 0.03 s) with a 0.12 s LiDAR-inertial temporal window and a 10-keyframe visual window; Levenberg-Marquardt in Ceres with analytic Jacobians; marginalization keeps prior information when the window slides (Secs. III-V)","LiDAR point-to-plane factors from planar features, raw IMU factors, bias factors and visual reprojection factors with inverse depth; LiDAR points are evaluated on the continuous-time trajectory at their own timestamps (Sec. V)","implicit: LiDAR points are associated with trajectory poses at their own timestamps on the continuous-time trajectory","continuous-time trajectory and LiDAR map",[2694,2695,2696,2697],"16-beam Ouster (two units)","64-beam Ouster with internal IMU","16-beam LiDAR","16-beam LiDAR (YQ and Vicon Room rig)",{"id":2699,"label":2700,"shortName":2701,"title":2702,"year":249,"era":10,"cluster":527,"scope":132,"keyIdeaZh":2703,"sensors":2704,"mapRepresentation":2707,"loopClosure":2708,"estimator":2709,"association":2710,"deskew":121,"outputGeometry":2711,"fulltextStatus":22,"lidarModels":2712,"equipmentCount":185},"slamtoolbox2021","Macenski & Jambrecic, 2021","SLAM Toolbox","SLAM Toolbox: SLAM for the dynamic world","SLAM Toolbox 是以 SRI 的 Open Karto 為基礎的 ROS 2D 雷射位姿圖 SLAM 套件，提供同步建圖、非同步建圖與純定位三種模式，也支援多次作業（multi-session）建圖。它把完整的原始掃描與位姿圖一起序列化，因此可以日後載入並繼續精修或擴充地圖，也能手動調整位姿圖節點、協助困難的迴圈閉合，或以運動學方式合併多張地圖。純定位模式把目前作業的量測以滾動緩衝加入原位姿圖，舊量測過期後即移除，作者稱為彈性位姿圖變形。作者並把原本的 SBA 最佳化介面改為 Ceres 外掛，重寫量測匹配以取得約十倍加速。",[2705,2706],"laser scanner (paper); planar 2D scans per the repository README, which the paper itself does not state","wheel or other odometry supplied as the odom-to-base transform (repository README; not stated in the paper)","pose graph with the complete raw scan data serialized (not submaps as in Cartographer); 2D maps rendered for navigation (Figs. 1, 3)","loop closures inside the pose graph, including between separate mapping sessions (Fig. 3); detection method not described in the paper","graph-based SLAM derived from Open Karto; the provided Sparse Bundle Adjustment optimization interface was replaced by Google Ceres behind a run-time dynamically loaded optimizer plugin interface (Features)","Open Karto measurement matching, restructured for about a 10x speed-up and multi-threading; new processing modes and K-D tree search for localization and multi-session mapping (Features); the matching algorithm itself is not described","2D map for navigation and localization; serialized pose graph and raw scans",[45],{"id":2714,"label":2715,"shortName":2716,"title":2717,"year":207,"era":10,"cluster":755,"scope":132,"keyIdeaZh":2718,"sensors":2719,"mapRepresentation":2721,"loopClosure":2722,"estimator":2723,"association":2724,"deskew":57,"outputGeometry":2725,"fulltextStatus":22,"lidarModels":2726,"equipmentCount":185},"vggtslam2025","Maggio et al., 2025","VGGT-SLAM","VGGT-SLAM: Dense RGB SLAM Optimized on the SL(4) Manifold","VGGT-SLAM 將 VGGT 產生的子地圖逐步對齊，指出在未校正相機下重建只確定到 15 自由度的射影變換，因此以 SL(4) 流形上的單應矩陣取代相似變換對齊子地圖，並加入以 SALAD 檢索的迴圈約束。作者明言重建不具公制尺度，影像須先去除鏡頭畸變，且當多張影像只看到單一平面（TUM 僅拍地板的片段）時單應估計會退化並使重建發散。NeurIPS 正式版補充的焦距統計顯示，VGGT 在同一場景內估計的焦距會明顯波動（桌面與路樁場景標準差 37.1 與 51.8 像素），這正是 SL(4) 明顯優於 Sim(3) 的情境；但在 7-Scenes 與 TUM 等一般場景，Sim(3) 版本表現相近，附錄中 w = 8 的 Sim(3) 版本在 TUM 平均 ATE 甚至更低（0.040 m 對 0.053 m）。每個子地圖的 VGGT 推論約 662 ms，SL(4) 對齊只多約 17 ms。建築室內若出現只拍到大面樓板或牆面的連續影格，可能觸發同類退化，需以實測驗證（推論）。",[2720],"monocular camera (uncalibrated)","VGGT dense point-cloud submaps","SALAD image-descriptor retrieval + relative homography loop constraints","nonlinear factor-graph optimization on the SL(4) manifold estimating 15-DoF homographies between VGGT submaps","shared frames between submaps; homography estimation with 5-point RANSAC on VGGT points","dense point cloud and trajectory; reconstruction not in metric scale (Sec. 1)",[],{"id":2728,"label":2729,"shortName":2730,"title":2731,"year":513,"era":52,"cluster":68,"scope":69,"keyIdeaZh":2732,"sensors":2733,"mapRepresentation":2738,"loopClosure":72,"estimator":2739,"association":2740,"deskew":2741,"outputGeometry":183,"fulltextStatus":22,"lidarModels":2742,"equipmentCount":109},"magnusson2007ndt3d","Magnusson et al., 2007","3D-NDT","Scan registration for autonomous mining vehicles using 3D-NDT","作者把 Biber 與 Strasser 的二維 NDT 推廣為三維：把模型掃描切成固定格網，每格以點的平均與共變異數表示常態分布，再以牛頓法最佳化資料點落在分布上的分數，不需最近鄰搜尋。論文比較取樣方式、格子大小，以及八叉樹、加成式、迭代式細分、連結格與無限外界等變體，並以 Kvarntorp 礦坑的原型雷射與 SICK LMS 200 資料對照 ICP：迭代細分加無限外界的 3D-NDT 在 50 對機器人掃描中高精度配準 45 對；在相同取樣比例下，3D-NDT 通常比 ICP 快將近三倍，但起始誤差較大時比 ICP 更早失敗。",[2734,2735,2736,2737],"Optab Optronikinnovation AB prototype 3D laser range finder (modulated infrared laser on a rotating mirror, phase-shift ranging; pitching scans for TUNNEL, yawing scans for JUNCTION)","SICK LMS 200 2D laser scanner on a pan-tilt unit giving pitching 3D scans of about 95,000 points (KVARNTORP-LOOP)","2D wheel odometry of the robot for initial pose estimates (KVARNTORP-LOOP)","total station measuring three marked points on the scanner (TUNNEL); not accurate enough as ground truth, used only as initial estimate","regular grid of cells (1 m baseline, 0.5 m to 3 m tested) storing mean and covariance of the points in each cell that holds more than a minimum number of points (Sec. 2.2 calls five points per cell a reasonable limit); octree-forest, additive and iterative subdivision variants","Newton's method with line search on the negated NDT score (maximum step 0.05, convergence when the change of p is below 0.0001); 7-parameter axis-angle transform (Eq. 13), noted as redundant; ICP baseline: point-to-point least squares with a 1 m fixed outlier threshold and approximate kd-tree search (ANN)","each data-scan point is scored against the normal distribution of the cell it falls in (point-to-distribution); variants: octree, additive and iterative subdivision, linked cells and infinite outer bounds","not_applicable (robot kept stationary during each scan)",[2743,2744],"Optab prototype 3D laser range finder","SICK LMS 200",{"id":2746,"label":2747,"shortName":2748,"title":2749,"year":345,"era":52,"cluster":68,"scope":69,"keyIdeaZh":2750,"sensors":2751,"mapRepresentation":2760,"loopClosure":2761,"estimator":2762,"association":2763,"deskew":2764,"outputGeometry":2765,"fulltextStatus":22,"lidarModels":2766,"equipmentCount":1593},"magnusson2009thesis","Magnusson, 2009","3D-NDT thesis","The three-dimensional normal-distributions transform: an efficient representation for registration, surface analysis, and loop detection","此博士論文以常態分布轉換（NDT）作為三維掃描的通用表面表示，並用於掃描配準、迴圈偵測與表面結構分析。配準部分將 Biber 與 Straßer 的二維 NDT 擴展到三維，以 z-y-x Euler 角參數化、解析梯度與 Hessian 搭配牛頓法與 Moré-Thuente 線搜尋求解，並提出由粗到細的迭代離散化（2 m、1 m、0.5 m 格子）、連結格子與三線性內插等延伸。作者在 Kvarntorp 礦坑隧道、模擬場景與飛行時間相機資料上，以 100 組預設初始偏移進行受控測試，並以里程計為初值處理兩段礦坑掃描序列，與 ICP 比較後結論為 NDT 對初始旋轉誤差較穩健且較快；Hessian 的反矩陣可估計位姿變異，作為配準成功與否的信心指標。論文另提出以色彩核函數擴充的 Colour-NDT、不需位姿資訊而以 NDT 表面形狀直方圖進行的外觀式迴圈偵測，以及用於礦堆巨石偵測的局部表面粗糙度分類。",[2752,2753,2754,2755,2756,2757,2758,2759],"SICK 2D lidar on a pan\u002Ftilt unit producing pitching 3D scans on Tjorven (180 deg horizontal, about 100 deg vertical field of view; SICK model not named for Tjorven)","SICK lidar on a continuously rotating slip-ring mount (yawing omnidirectional scans) and a Hokuyo 2D lidar for 2D localisation on Alfred","tiltable SICK laser scanner on Kurt3D (pitching scans)","PMD[vision] 19k time-of-flight camera combined with a Matrix-Vision Blue Fox colour camera (Colour-NDT data)","SwissRanger time-of-flight camera (3D-Cam scan pair, collected by Jacobs University Bremen)","SICK lidar on a Schunk PowerCube via slip-ring contacts with a digital camera (Kemi mine muck-pile scans)","simulated yawing lidar (Sci-Fi and Sim-Mine scan pairs)","wheel-encoder odometry for initial pose estimates (Kvarntorp-Loop, Mission-4)","NDT cell grids: iterative multi-resolution cells (2, 1, 0.5 m for lidar scans; 0.5, 0.25, 0.125 m for time-of-flight data), with octree and k-means variants evaluated; Colour-NDT stores three colour-weighted Gaussians per cell; loop detection summarises overlapping 0.5 m cells as 55-bin surface-shape histograms","appearance-based loop detection with NDT surface-shape histograms (1 spherical, 9 planar and 1 linear class in 5 range intervals; orientation normalised by dominant plane directions); threshold chosen manually or from an EM-fitted Gamma mixture; detects loop candidates only","Newton's method with Moré-Thuente line search on the NDT score (Gaussian approximation of a normal-plus-uniform mixture) with analytic gradient and Hessian and z-y-x Euler parametrisation; baseline uses iterative discretisation (2, 1, 0.5 m cells) with linked cells, 20% spatially distributed subsampling of the current scan and a step-size convergence limit of 1e-6; a BFGS quasi-Newton variant was less robust","each current-scan point is scored against the Gaussian of the cell it falls in; linked cells use the nearest occupied cell (kD tree of occupied cells); trilinear interpolation weights the eight nearest cells; Colour-NDT weights per-cell colour-kernel Gaussians; the ICP baseline uses point-to-point closest points with a fixed 0.5 m outlier threshold (0.1 m for 3D-Cam)","none applied: the mobile-robot registration data sets were acquired stop-and-scan (robot stopped every few metres); Sec. 3.2 only reviews motion-compensation methods for scanning while moving","6-DoF relative pose; NDT surface representation; loop detection output",[45,2767,2768,2769],"SICK lidar on continuously rotating motor with slip-ring contacts (Alfred)","tiltable SICK laser scanner (Kurt3D)","SICK lidar on a Schunk PowerCube via slip-ring contacts (with a digital camera)",{"id":2771,"label":2772,"shortName":2773,"title":2774,"year":30,"era":10,"cluster":98,"scope":99,"keyIdeaZh":2775,"sensors":2776,"mapRepresentation":2778,"loopClosure":72,"estimator":2779,"association":2780,"deskew":2781,"outputGeometry":2782,"fulltextStatus":22,"lidarModels":2783,"equipmentCount":2788},"rkolio2026","Malladi et al., 2026","RKO-LIO","Robust Approach for LiDAR-Inertial Odometry Without Sensor-Specific Modeling","RKO-LIO 不採用卡爾曼濾波或預積分因子圖，而是假設相鄰 LiDAR 幀間線加速度與角速度固定，以簡化模型積分 IMU 取得 ICP 初值與逐點去畸變，再以掃描對地圖 ICP 精修。作者在 ICP 中加入依 IMU 加速度資訊自適應調整的姿態正則化，並主張不需感測器特定的雜訊模型與校正，即可用同一組參數跨車載、背包、四足與無人機平台。",[35,2777],"IMU (consumer to industrial grade)","VDB voxel grid (voxel size 1.0 m) storing a fixed number of points per voxel; double downsampling","No filter or factor graph for the pose: IMU samples between two scans are bias- and gravity-compensated, averaged, and integrated with a constant linear acceleration and angular velocity model to give the ICP initial guess and per-point deskew; scan-to-map point-to-point ICP adds an accelerometer-based orientation cost weighted by 1\u002Fbeta, with beta = beta0 (1 + sigma_a^2) and beta0 = 200, where body acceleration comes from a small Kalman filter with maximum expected jerk 3 m\u002Fs3; biases assumed constant and estimated from the first inter-scan interval","Scan-to-map ICP building on the KISS-ICP scan-alignment module: point-to-point residuals with a fixed association threshold of 0.5 m in a VDB voxel map (1.0 m voxels), double downsampling at 0.5 v for map update and 1.5 v for registration (optional off switch for sparse LiDARs), points within 1 m of the sensor clipped","per-point transform from IMU-based motion between LiDAR frames (Sec. III-B)","odometry and local voxel map; export format not_reported",[2784,143,2785,437,2786,2787,1591,1365],"Hesai QT64","Aeva Aeries II","Hesai XT32, QT32 and QT64 (different sessions)","Ouster OS0",19,{"id":2790,"label":2791,"shortName":2792,"title":2793,"year":661,"era":10,"cluster":755,"scope":99,"keyIdeaZh":2794,"sensors":2795,"mapRepresentation":2797,"loopClosure":2798,"estimator":2799,"association":2800,"deskew":57,"outputGeometry":2801,"fulltextStatus":22,"lidarModels":2802,"equipmentCount":46},"monogs2024","Matsuki et al., 2024","MonoGS (Gaussian Splatting SLAM)","Gaussian Splatting SLAM","MonoGS 是首個以三維高斯為唯一表示的單目 SLAM，以解析的李群雅可比直接最佳化相機位姿，並提出等向性正則化避免高斯沿視線拉長。有深度時加入幾何殘差。單目結果沒有公制尺度，評估時需做尺度對齊；地圖品質只以渲染指標評估，作者並指出高斯不顯式表示表面。",[37,772,2796],"stereo (depth from stereo, tested only on EuRoC Machine Hall in Supp. 9.5)","anisotropic 3D Gaussians with isotropic shape regularization","none (authors compare mainly against methods without loop closure)","direct photometric (plus geometric with depth) pose optimization with analytic Lie-group Jacobians; windowed keyframe mapping","direct photometric residual; depth residual when available","Gaussian map and renderings; authors note Gaussians do not explicitly represent the surface (Sec. 5)",[],{"id":2804,"label":2805,"shortName":2806,"title":2807,"year":396,"era":10,"cluster":755,"scope":32,"keyIdeaZh":2808,"sensors":2809,"mapRepresentation":2810,"loopClosure":72,"estimator":2811,"association":2812,"deskew":57,"outputGeometry":2813,"fulltextStatus":22,"lidarModels":2814,"equipmentCount":144},"nerf2020","Mildenhall et al., 2020","NeRF","NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis","NeRF 以多層感知器（MLP）將三維位置與觀看方向映射為體密度與顏色，並透過可微分體積渲染（volume rendering）以多視角影像的光度誤差最佳化網路。方法本身不估計相機位姿，實景資料需先以 COLMAP 等 SfM 取得位姿與內參。幾何只隱含在密度場中，是後續神經隱式 SLAM 共用的表示與渲染原理。",[1657],"neural implicit radiance field (MLP: volume density + view-dependent colour)","not_applicable (per-scene gradient-based optimization of an MLP; camera poses are inputs)","direct photometric loss through differentiable volume rendering","Novel-view images; geometry only implicit in the density field; the paper reports no surface extraction and no geometric accuracy evaluation; real forward-facing scenes are optimized in normalized device coordinates that use disparity rather than metric depth (Sec. 3 to 6, App. A, App. C)",[],{"id":2816,"label":2817,"shortName":2818,"title":2819,"year":150,"era":10,"cluster":236,"scope":132,"keyIdeaZh":2820,"sensors":2821,"mapRepresentation":2825,"loopClosure":2826,"estimator":2827,"association":2828,"deskew":2829,"outputGeometry":2830,"fulltextStatus":22,"lidarModels":2831,"equipmentCount":24},"cblox2018","Millane et al., 2018","C-blox","C-blox: A Scalable and Consistent TSDF-based Dense Mapping Approach","C-blox 把場景表示為一組相互重疊的 TSDF 子體積（subvolume），每個子體積固定附著在 ORB-SLAM2 的一個關鍵影格上。迴圈閉合後，只要以最佳化後的關鍵影格位姿更新子體積座標系，就能修正稠密地圖，不必重新整合深度影像。為避免子體積數量隨軌跡長度無限增加，作者先以地標共視圖找出可能重複觀測同一區域的子體積對，再從光束法平差資訊矩陣計算兩者相對定位的條件共變異，只有定位品質足夠時才把兩個子體積融合，藉此限制地圖成長。整個系統只用 CPU，並在 voxblox 中加入多執行緒快速整合器，可在無人機上的 Intel NUC 即時執行。",[2822,2823,2824],"Stereo camera pair of a VI-sensor (global shutter, tightly synchronized) for ORB-SLAM2 tracking on the MAV","RGB-D camera Intel RealSense D415 (coloured pointclouds) for dense integration on the MAV","Synthetic RGB-D input (ICL-NUIM) and simulated RGB plus noiseless depth (CARLA) in the evaluations","collection of overlapping TSDF subvolumes (voxblox) with no spatial partitioning; a new subvolume starts after a maximum number of keyframes or a large sparse-map change; redundant well-localized subvolumes are fused by trilinear interpolation","ORB-SLAM2 loop detection and bundle adjustment; the dense map is corrected by updating subvolume base frames with the optimized keyframe poses","Modified ORB-SLAM2 (keyframe bundle adjustment of feature re-projection errors with Huber cost) supplies camera poses; each subvolume is rigidly attached to a keyframe and moved when keyframe poses are re-optimized; no dense tracking against the TSDF","Sparse ORB features for tracking; subvolume-fusion candidates from a landmark covisibility graph (edge weight = number of shared landmarks), accepted only if the relative localization quality q = 1\u002F||Sigma_i|j|| from the bundle-adjustment information matrix exceeds a threshold (Schur complement, constrained AMD reordering and Cholesky-based covariance recovery)","not_applicable (depth camera input)","TSDF subvolume collection and fused mesh; voxel size 0.02 m on ICL-NUIM and 0.5 m on CARLA",[],{"id":2833,"label":2834,"shortName":2835,"title":2836,"year":661,"era":10,"cluster":31,"scope":32,"keyIdeaZh":2837,"sensors":2838,"mapRepresentation":2839,"loopClosure":57,"estimator":2840,"association":2841,"deskew":121,"outputGeometry":2842,"fulltextStatus":22,"lidarModels":2843,"equipmentCount":24},"millane2024nvblox","Millane et al., 2024","nvblox","nvblox: GPU-Accelerated Incremental Signed Distance Field Mapping","nvblox 將 Voxblox 的分層體素地圖移到 GPU：以雜湊表索引 8x8x8 體素區塊，並行更新 TSDF 或佔據層，定期以平行 marching cubes 產生網格；另提出以區塊內掃掠與跨區塊傳遞交替進行的增量式 GPU ESDF 演算法，採完整歐氏距離而非近似距離。支援 RGB-D 與 LiDAR，並可在嵌入式 GPU 上執行。",[772,35],"block-hashed voxel layers (8x8x8 voxels per block) on GPU: TSDF, ESDF, occupancy, color, mesh","not_applicable (poses supplied; e.g., FAST-LIO for the drone LiDAR example)","projective TSDF update of voxels in view (LiDAR depth images with linear interpolation)","TSDF, ESDF, occupancy and marching-cubes mesh",[2844],"Ouster OS1 (64-beam)",{"id":2846,"label":2847,"shortName":2848,"title":2849,"year":2850,"era":52,"cluster":527,"scope":132,"keyIdeaZh":2851,"sensors":2852,"mapRepresentation":2854,"loopClosure":2855,"estimator":2856,"association":2857,"deskew":121,"outputGeometry":2858,"fulltextStatus":22,"lidarModels":2859,"equipmentCount":185},"fastslam2002","Montemerlo et al., 2002","FastSLAM","FastSLAM: A Factored Solution to the Simultaneous Localization and Mapping Problem",2002,"FastSLAM 利用「給定機器人路徑時各地標條件獨立」的性質，把 SLAM 後驗分解為路徑分布與各地標的條件分布：以粒子濾波器（particle filter）取樣路徑，每個粒子再為每個地標維持一個小型 EKF。作者以樹狀資料結構使每次更新的時間複雜度降為 O(M log K)，並讓每個粒子各自做資料關聯（data association），因此可同時追蹤多種關聯假設。模擬中地標數擴充到 50,000 個，實體機器人實驗則以人工測得的地標位置比對。",[2853],"[\"2D laser range finder (SICK)\",\"robot controls u_t (odometry sensor not named in the paper)\"]","point landmarks stored in a balanced binary tree per particle","implicit through particle weighting and resampling; no explicit loop-closure module","Rao-Blackwellized particle filter (particles over robot path, one small EKF per landmark per particle)","per-particle maximum-likelihood landmark association with new-landmark threshold","landmark positions and robot path per particle",[988],{"id":2861,"label":2862,"shortName":2863,"title":2864,"year":191,"era":52,"cluster":527,"scope":132,"keyIdeaZh":2865,"sensors":2866,"mapRepresentation":2868,"loopClosure":2869,"estimator":2870,"association":2871,"deskew":121,"outputGeometry":2872,"fulltextStatus":22,"lidarModels":2873,"equipmentCount":24},"fastslam2_2003","Montemerlo et al., 2003","FastSLAM 2.0","FastSLAM 2.0: An Improved Particle Filtering Algorithm for Simultaneous Localization and Mapping that Provably Converges","FastSLAM 2.0 修改原 FastSLAM 的取樣方式，在抽樣機器人位姿時同時考慮最新量測，而不只依賴運動模型。作者證明對線性高斯 SLAM，在所有特徵被無限次觀測且已知一個特徵位置的條件下，單一粒子即可在期望值意義上收斂到正確地圖（Sec. 5）。在 Victoria Park 公開資料上，作者報告其精度明顯優於原版 FastSLAM。",[2867],"[\"range finder (type not stated in the paper)\",\"vehicle odometry (described as relatively inaccurate)\",\"GNSS (DGPS used for evaluation only)\"]","point landmarks (per-particle EKFs)","implicit; no explicit module","Rao-Blackwellized particle filter with pose proposal conditioned on the current measurement","per-particle maximum-likelihood association that accounts for the sampled pose (Sec. 4.4); new features created when the measurement probability falls below a threshold, and spurious features removed by a log-odds existence filter (Sec. 4.5)","landmark map and vehicle path",[],{"id":2875,"label":2876,"shortName":2877,"title":2877,"year":1217,"era":52,"cluster":527,"scope":99,"keyIdeaZh":2878,"sensors":2879,"mapRepresentation":2881,"loopClosure":2882,"estimator":2883,"association":2884,"deskew":2885,"outputGeometry":2886,"fulltextStatus":22,"lidarModels":2887,"equipmentCount":144},"velodyneslam2011","Moosmann & Stiller, 2011","Velodyne SLAM","Velodyne SLAM 專為 Velodyne HDL-64E 的連續旋轉取樣與較高量測雜訊設計，只使用 LiDAR 資料。每轉一圈的資料排成 870 乘 64 的距離影像，先估計各點的法向量與平面信心，再以位置加法向量的 6D 最近鄰 ICP 將掃描對整張地圖配準，並在配準前後各做一次以線性內插的去畸變。地圖是每格最多保存一個曲面元素的 3D 網格；在平面信心高的區域，新量測會先沿法向量調整再加入，以降低雜訊造成的厚度。作者另提出離線精修步驟，以儲存的原始量測重建更細緻的地圖。",[2880],"3D spinning LiDAR only (Velodyne HDL-64E S2); no wheel-speed, inertial or other information","3D grid of resolution g (5 cm default) holding at most one surface (point, normal, normal confidence) per cell; in flat regions new measurements are moved along their normals to fit neighbouring surfaces before insertion; cells are replaced by closer or more confident measurements and never erased (Sec. II.E)","none (named as future work)","scan-to-map ICP minimizing point-to-plane distances (Chen-Medioni) with nearest neighbours searched in 6D (position and normal), initialized by constant-motion prediction; 1000 surfaces sampled from the upper and 500 from the lower half of the range image (Sec. II.C)","6D nearest neighbour (px, py, pz, nx, ny, nz) between scan surfaces and map surfaces (Sec. II.C)","linear-interpolation de-skewing applied twice, before ICP with the predicted pose and after ICP with the final pose (Sec. II.D)","vehicle trajectory and a detailed 3D point cloud of surfaces; an offline refinement builds a new map from the stored original measurements (Sec. II.F)",[161],{"id":2889,"label":2890,"shortName":2891,"title":2892,"year":249,"era":10,"cluster":11,"scope":12,"keyIdeaZh":2893,"sensors":2894,"mapRepresentation":2897,"loopClosure":2898,"estimator":2899,"association":2900,"deskew":121,"outputGeometry":2901,"fulltextStatus":22,"lidarModels":2902,"equipmentCount":144},"moura2021bimslam","Moura et al., 2021","BIM-based localization and mapping (COBOLLEAGUE)","BIM-based Localization and Mapping for Mobile Robots in Construction","作者在歐盟 COBOLLEAGUE 專案中提出把 BIM 轉成 SLAM 位姿圖的介面。IFC 模型先依樓層（IfcStorey 高程）拆分，排除門、窗與空間後轉成網格、體素化並存成八元樹；再由樓層八元樹投影出可通行的二維佔據格，以 Voronoi 路網與覆蓋路徑規劃產生虛擬機器人的軌跡，在每個路點以光線投射產生虛擬掃描、里程計與 IMU 資料，組成不需最佳化的 Cartographer 狀態檔。實際機器人載入此凍結的 BIM 位姿圖後，可在沒有先行探勘的情況下重新定位，並把新掃描加入新軌跡以記錄與模型不同的現況。作者在 Gazebo 模擬與 Eurecat 工業實驗室的實測中示範重新定位與地圖更新。",[2895,2896,36],"3D LiDAR (model not reported)","odometry encoders","BIM-derived serialized Cartographer state (.pbstream): per-floor trajectories of virtual scans generated from an octree of the voxelized IFC model; new sessions append trajectories that can be exported as point clouds (Sec. III)","Cartographer global constraints between the new trajectory and the frozen BIM trajectory (inter-submap constraints) (Sec. V-B1)","Google Cartographer 3D graph SLAM with its global solver matching new trajectory data to frozen BIM-derived submaps (default configuration, lowered global localization minimum score) (Sec. III-C)","Cartographer scan-to-submap matching and global constraint search against the BIM-derived submaps (Sec. III-C)","robot pose in the BIM frame and an updated 3D point-cloud map; deviations from the model highlighted by cloud-to-mesh distance (Sec. V-B2)",[45],{"id":2904,"label":2905,"shortName":2906,"title":2907,"year":513,"era":52,"cluster":53,"scope":346,"keyIdeaZh":2908,"sensors":2909,"mapRepresentation":2910,"loopClosure":72,"estimator":2911,"association":2912,"deskew":57,"outputGeometry":2913,"fulltextStatus":22,"lidarModels":2914,"equipmentCount":46},"mourikis2007msckf","Mourikis & Roumeliotis, 2007","MSCKF","A Multi-State Constraint Kalman Filter for Vision-aided Inertial Navigation","MSCKF 是以擴展卡爾曼濾波（EKF）為基礎的視覺輔助慣性導航演算法。其核心是推導一種量測模型：當靜態特徵被多個相機位姿觀測時，直接以這些位姿間的幾何約束更新濾波器，而不必把三維特徵座標放進狀態向量。因此計算量只與特徵數呈線性關係，狀態中只保留有限數量的過去相機位姿副本。作者以車載相機與 IMU 在都市住宅區 3.2 km 的行駛資料驗證。",[37,36],"no persistent map (features not kept in state)","EKF whose state holds the evolving IMU state (quaternion, gyro and accelerometer biases, velocity, position; 15-dimensional error state) plus up to Nmax cloned camera poses; each feature's stacked residual is projected onto the left nullspace of its feature Jacobian with Givens rotations, residuals of all features are compressed by QR decomposition before the update; when Nmax is reached, Nmax\u002F3 evenly spaced poses starting from the second oldest are removed after processing their features, and the oldest pose is always kept; Nmax = 30 in the experiment","SIFT feature extraction and matching; each feature triangulated by Gauss-Newton with an inverse-depth parameterization while camera poses are treated as known; multi-view geometric constraints of static features; simple Mahalanobis distance test to discard features on moving objects","IMU pose and velocity trajectory with covariance",[],{"id":2916,"label":2917,"shortName":2918,"title":2919,"year":769,"era":10,"cluster":236,"scope":132,"keyIdeaZh":2920,"sensors":2921,"mapRepresentation":2922,"loopClosure":2923,"estimator":2924,"association":2925,"deskew":57,"outputGeometry":2926,"fulltextStatus":22,"lidarModels":2927,"equipmentCount":554},"orbslam2_2017","Mur-Artal & Tardos, 2017","ORB-SLAM2","ORB-SLAM2: An Open-Source SLAM System for Monocular, Stereo, and RGB-D Cameras","ORB-SLAM2 將 ORB-SLAM 擴充到雙目（stereo）與 RGB-D 相機，把近距與遠距雙目特徵納入 BA，使尺度可觀測，迴圈閉合改以剛體 SE(3) 位姿圖最佳化並在另一執行緒進行全域 BA。系統另提供只做定位的地圖重用模式。作者明言目標是長期且全域一致的定位，而非最精細的稠密重建；論文中的稠密點雲是以估計的關鍵影格位姿反投影感測器深度圖所得。",[37,2101,772],"sparse map points plus keyframes (covisibility graph, spanning tree)","DBoW2 detection with geometric validation; rigid-body SE(3) pose-graph when stereo\u002Fdepth makes scale observable, followed by full BA (Sec. III-D)","BA with monocular and stereo constraints (motion-only, local, and full BA in a separate thread after pose-graph optimisation) (Sec. III, III-D)","ORB extracted on both rectified stereo images (or on the RGB image); a stereo keypoint (uL, vL, uR) comes from matching each left ORB along the same row with subpixel patch-correlation refinement; for RGB-D the depth d is converted to a virtual right coordinate uR = uL - fx*b\u002Fd with b approximated to 8 cm for Kinect and Asus Xtion; keypoints with depth below 40 times the baseline are close (triangulated from one frame, carry scale), others far (triangulated only from multiple views), unmatched ones stay monocular; DBoW2 place recognition as in ORB-SLAM","keyframe trajectory and sparse map points; the dense point clouds shown are obtained by back-projecting sensor depth maps from estimated keyframe poses (Sec. IV-C, Fig. 7)",[],{"id":2929,"label":2930,"shortName":2931,"title":2932,"year":235,"era":52,"cluster":236,"scope":132,"keyIdeaZh":2933,"sensors":2934,"mapRepresentation":2935,"loopClosure":2936,"estimator":2937,"association":2938,"deskew":57,"outputGeometry":2939,"fulltextStatus":22,"lidarModels":2940,"equipmentCount":736},"orbslam2015","Mur-Artal et al., 2015","ORB-SLAM","ORB-SLAM: A Versatile and Accurate Monocular SLAM System","ORB-SLAM 以同一組 ORB 特徵同時支援追蹤、局部建圖、重定位（relocalization）與迴圈閉合（loop closure），分成三個平行執行緒。系統以共視圖（covisibility graph）限定局部 BA 範圍，偵測到迴圈後估計相似變換 Sim(3) 以校正單眼尺度漂移，再於稀疏的 Essential Graph 上做位姿圖最佳化。寬鬆建立、嚴格剔除關鍵影格與地圖點的策略使地圖只在場景內容改變時成長。",[37],"sparse map points plus keyframes with covisibility graph","Per keyframe, BoW candidates scoring above the lowest score among covisible neighbours (theta_min 30) are kept and a loop is accepted only after three consecutive consistent candidates; a Sim(3) is estimated from 3D-3D ORB matches with Horn's method inside RANSAC and refined with guided matching, serving as geometric validation; duplicated points are fused and covisibility edges added; then Sim(3) pose-graph optimisation on the Essential Graph (spanning tree, covisibility edges with theta_min 100, loop edges; 10 LM iterations in the experiments)","motion-only BA in tracking, local BA in mapping, pose-graph optimisation over Sim(3) constraints on the Essential Graph; Levenberg-Marquardt in g2o (Sec. III-B)","ORB (oriented multi-scale FAST with 256-bit descriptor): FAST corners on 8 scale levels (factor 1.2), 1000 corners for 512x384 to 752x480 images and 2000 for KITTI 1241x376, spread by a per-level grid; constant-velocity prediction with guided search of last-frame points, then projection of a covisibility-based local map (keyframes sharing points plus their neighbours) with viewing-angle and scale-range checks; DBoW2 bag-of-words place recognition with an offline ORB vocabulary, covisibility-grouped scores and all matches above 75% of the best score","keyframe trajectory and sparse 3D map points (up to scale in monocular mode)",[45],{"id":2942,"label":2943,"shortName":2944,"title":2945,"year":207,"era":10,"cluster":755,"scope":132,"keyIdeaZh":2946,"sensors":2947,"mapRepresentation":2949,"loopClosure":2950,"estimator":2951,"association":2952,"deskew":57,"outputGeometry":2953,"fulltextStatus":22,"lidarModels":2954,"equipmentCount":24},"mast3rslam2025","Murai et al., 2025","MASt3R-SLAM","MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors","MASt3R-SLAM 以 MASt3R 雙視角重建先驗為核心建構即時單目稠密 SLAM，只假設單一相機中心，用迭代投影做點圖匹配、以 Sim(3) 位姿處理預測間不一致的尺度，並以影像檢索做迴圈閉合與重定位，後端為二階全域最佳化。幾何評估在 EuRoC Vicon 房間以軌跡對齊結構掃描真值，在 7-Scenes 以 ICP 對齊深度反投影參考，並報告 RMSE。作者承認全域最佳化不精修幾何，且畸變大的相機會降低預測品質。",[2948],"monocular camera (uncalibrated, generic central camera)","per-keyframe canonical pointmaps with local fusion","incremental ASMK image retrieval + MASt3R decoder matching; relocalisation via retrieval","Gauss-Newton second-order optimization of Sim(3) keyframe poses minimizing ray error; sparse Cholesky backend in CUDA","MASt3R pointmap matching via iterative projective ray search","dense point cloud from fused pointmaps + trajectory (scale via Sim(3), not guaranteed metric)",[],{"id":2956,"label":2957,"shortName":2958,"title":2959,"year":383,"era":52,"cluster":31,"scope":32,"keyIdeaZh":2960,"sensors":2961,"mapRepresentation":2962,"loopClosure":57,"estimator":57,"association":57,"deskew":57,"outputGeometry":2963,"fulltextStatus":22,"lidarModels":2964,"equipmentCount":77},"museth2013vdb","Museth, 2013","VDB \u002F OpenVDB","VDB: High-resolution sparse volumes with dynamic topology","VDB 是一種淺而寬、高度平衡的階層式稀疏體積資料結構，概念近似 B+ 樹：常見設定為以雜湊表或 std::map 實作的可動態擴充根節點，接兩層固定分支 32³ 與 16³ 的內部節點，最底層為 8³ 體素的葉節點，並以位元遮罩分開編碼拓樸與數值。由於深度與分支在編譯期固定，可直接由全域座標以位元運算求得各層偏移，隨機存取平均為 O(1)，再以快取最近走訪節點的反向（由下而上）走訪加速空間連貫的存取。原為電影特效的動態稀疏體積與位準集設計，後被 VDBFusion 用來儲存 TSDF。",[],"shallow height-balanced B+tree-like sparse grid: dynamic root node (hash map or std::map) over internal nodes with 32^3 and 16^3 branching and 8^3-voxel leaf nodes; bit masks encode active topology and child pointers, values stored at any level as voxels or tiles; optional out-of-core leaf buffers","sparse volumetric grids such as narrow-band level sets and density volumes; meshes and rendered images are downstream products (Fig. 4, Sec. 4.5-4.7)",[],{"id":2966,"label":2967,"shortName":2968,"title":2969,"year":9,"era":10,"cluster":98,"scope":99,"keyIdeaZh":2970,"sensors":2971,"mapRepresentation":2974,"loopClosure":2975,"estimator":2976,"association":2977,"deskew":2978,"outputGeometry":2979,"fulltextStatus":22,"lidarModels":2980,"equipmentCount":487},"mc2slam2019","Neuhaus et al., 2019","MC2SLAM","MC2SLAM: Real-Time Inertial Lidar Odometry Using Two-Scan Motion Compensation","MC2SLAM 以兩個連續 LiDAR 掃描一起估計第一個掃描期間的運動：先以 IMU 積分（無 IMU 時以線性外推）預測兩掃描的軌跡，再在其上加一個隨時間線性增長的六自由度偏差，用點到平面殘差把稀疏取樣的查詢點配準到最近約 100 個掃描構成的局部地圖，完成配準與去畸變；殘差尺度以中位數絕對偏差穩健估計。每次只補償前一個掃描並插入近似 Poisson 圓盤取樣的局部地圖，得到的相對位姿再與 IMU 預積分因子放入因子圖，估計速度與偏差並使軌跡與重力對齊。",[2972,2973],"3D multi-beam LiDAR (Velodyne HDL-32 with built-in IMU in the authors' data; Velodyne HDL-64 in KITTI without IMU)","IMU (built-in HDL-32 IMU)","local map of all compensated points from the last about 100 sweeps, inserted only if farther than about 5 cm from stored points (approximate Poisson disk), stored in a uniform grid ordered by time (Sec. 3.2)","not in the reported system; authors state their implementation can close loops but the chapter focuses on odometry (footnote 1)","two parts: per-sweep robust nonlinear least squares for a 6-DoF bias on an IMU-predicted trajectory over two sweeps (Tukey loss, MAD-based residual scale, Ceres), and an online factor graph with pose prior, laser odometry and IMU preintegration factors solved in Ceres every five sweeps (Sec. 3.2, 4)","Poisson-disk-like query points (minimum spacing about 20 cm, then about 500 random points) matched point-to-plane to a local map within radius epsilon; plane normal by eigen analysis of neighbours (Sec. 3.1-3.2)","two-scan motion compensation: the trajectory over sweeps k and k+1 is estimated jointly with registration, then only sweep k is compensated and inserted into the local map (Sec. 3.2)","odometry trajectory and motion-compensated sweeps accumulated into a point cloud map (Fig. 1)",[225,161],{"id":2982,"label":2983,"shortName":2984,"title":2985,"year":1217,"era":52,"cluster":236,"scope":99,"keyIdeaZh":2986,"sensors":2987,"mapRepresentation":2988,"loopClosure":2989,"estimator":2990,"association":2991,"deskew":57,"outputGeometry":2992,"fulltextStatus":22,"lidarModels":2993,"equipmentCount":144},"dtam2011","Newcombe et al., 2011a","DTAM","DTAM: Dense tracking and mapping in real-time","DTAM 不擷取特徵點，而是以每個像素的光度資料在關鍵影格上估計稠密深度圖，並以空間正則化能量函數求解，形成大量頂點的表面拼貼。相機位姿則以整張影像對稠密模型進行直接對齊（direct alignment）追蹤。此方法依賴 GPU 平行運算，並假設靜態場景。",[37],"overlapping keyframes, each with an RGB reference image, pose, inverse depth map and an M x N x S photometric cost volume; the Fig. 3 example keyframe has nearly 300 x 10^3 estimated points versus about 1000 PTAM point features in the same frame","none described; the method has no loop detection or map correction, and a relocaliser is mentioned only as disabled during the PTAM comparison","Mapping: per-keyframe inverse depth map minimising a photometric cost volume (average L1 error over tens to hundreds of overlapping frames at S inverse-depth samples) plus an edge-weighted Huber regulariser; the energy is decoupled with an auxiliary variable, solved by primal-dual updates for the convex part and a point-wise exhaustive search over the cost volume whose feasible range shrinks each iteration, with one embedded Newton step for sub-sample accuracy (theta from 0.2 to 1e-4). Tracking: Lucas-Kanade style iterative least squares, first inter-frame rotation on coarse pyramid levels, then 6DOF forward-compositional alignment of the live image to a view synthesised from the dense model, coarse to fine.","direct photometric every-pixel association: cost volume built by projecting reference pixels into overlapping frames for each inverse-depth sample; tracking compares every pixel of the live image with the model-predicted image, rejecting pixels whose photometric error exceeds a threshold that decreases during coarse-to-fine iterations","textured dense inverse depth maps; a triangle mesh is computed from each keyframe depth map (oblique edges culled) and used for tracking, forming a surface patchwork with millions of vertices",[],{"id":2995,"label":2996,"shortName":2997,"title":2998,"year":1217,"era":52,"cluster":236,"scope":99,"keyIdeaZh":2999,"sensors":3000,"mapRepresentation":3001,"loopClosure":3002,"estimator":3003,"association":3004,"deskew":121,"outputGeometry":3005,"fulltextStatus":22,"lidarModels":3006,"equipmentCount":144},"kinectfusion2011","Newcombe et al., 2011b","KinectFusion","KinectFusion: Real-time dense surface mapping and tracking","KinectFusion 將 Kinect 深度串流即時融合到單一全域截斷符號距離函數（Truncated Signed Distance Function, TSDF）體素模型中，並以光線投射（raycasting）產生的模型表面預測，用由粗到細的 ICP（point-to-plane、投影式資料關聯）追蹤感測器位姿。追蹤對象是累積模型而非前一影格，因而在房間尺度內漂移有限。所有步驟皆可在 GPU 上平行化。",[772],"single fixed-extent dense TSDF volume in GPU memory storing truncated distance and weight (16 bits per component); projective TSDF with nearest-neighbour depth lookup and weighted running average, optional weight cap for moving-average reconstruction of dynamic scenes; 256^3 voxels in the turntable experiments, 64^3 to 512^3 evaluated","none explicit; loops are closed only implicitly by frame-to-model tracking (turntable loop-closing frames nearly overlap after one pass and more tightly after four passes); on tracking failure an interactive relocalisation asks the user to align the live depth frame with the prediction from the last known pose","coarse-to-fine point-to-plane ICP against the raycast model prediction over the bottom 3 pyramid levels with at most 4, 5 and 10 iterations (coarse to fine); small-angle linearisation gives per-correspondence 6x6 systems summed on the GPU by tree reduction and solved by Cholesky on the CPU; a null-space check and an increment-magnitude check switch the system into relocalisation mode","projective data association on bilateral-filtered depth (vertex and normal map pyramid, L = 3), rejecting pairs by vertex distance and normal-angle thresholds; raw depth, not the filtered depth, is fused into the TSDF","dense TSDF surface rendered by raycasting; mesh export not reported in sections read",[],{"id":3008,"label":3009,"shortName":3010,"title":3011,"year":97,"era":10,"cluster":263,"scope":99,"keyIdeaZh":3012,"sensors":3013,"mapRepresentation":3018,"loopClosure":3019,"estimator":3020,"association":3021,"deskew":3022,"outputGeometry":3023,"fulltextStatus":22,"lidarModels":3024,"equipmentCount":736},"viralfusion2022","Nguyen et al., 2022b","VIRAL-Fusion","VIRAL-Fusion: A Visual-Inertial-Ranging-Lidar Sensor Fusion Approach","VIRAL-Fusion 以滑動視窗最佳化融合三類觀測：IMU 預積分、UWB 測距，以及機上既有自我定位系統（如 VINS-Fusion 與 A-LOAM）輸出的相鄰位姿變化量。UWB 測距模型同時考慮天線相對機體中心的偏移與量測時刻落在兩個狀態之間的時間差，因此能約束姿態與速度。由於 UWB 錨點在現場先行自我定位並定義場址座標系，估測結果不隨時間累積漂移，也不需要迴圈閉合。",[3014,3015,3016,3017],"[\"IMU (VectorNav VN-100, per footnote link)\", \"body-offset UWB ranging (Humatics P440, per footnote link","two UAV nodes with two antennae each and three or four anchors)\", \"camera system (footnote links the ueye_cam driver","used through VINS-Fusion odometry)\", \"two LiDARs (Ouster OS1, per footnote link","horizontal and vertical, used through A-LOAM odometry)\"]","UWB anchor positions in a site frame from anchor self-localization; no dense map is built by the fusion","no; drift is bounded by ranging to fixed UWB anchors instead of loop closure","sliding-window nonlinear least squares in Ceres over orientation, position, velocity and IMU biases; cost combines UWB body-offset range factors, IMU preintegration factors and pose-displacement factors from one or more onboard self-localization (OSL) systems such as VIO or LiDAR odometry (Sec. IV-A, Eq. 31; Sec. V)","loosely coupled at the OSL level: the relative pose change reported by each OSL system between two time steps is used as an observation (Sec. V-A); UWB ranges are checked by SNR, leading-edge quality, rate of change and an IMU-predicted range test before use (Sec. IV-B)","not described in the fusion method; LiDAR data enter only through the A-LOAM odometry estimates","UAV pose and velocity in the anchor-defined world frame at the optimization rate plus IMU-propagated high-rate pose",[3025],"OS1 (two units)",{"id":3027,"label":3028,"shortName":3029,"title":3030,"year":51,"era":10,"cluster":98,"scope":132,"keyIdeaZh":3031,"sensors":3032,"mapRepresentation":3035,"loopClosure":3036,"estimator":3037,"association":3038,"deskew":3039,"outputGeometry":3040,"fulltextStatus":22,"lidarModels":3041,"equipmentCount":1289},"slict2023","Nguyen et al., 2023","SLICT","SLICT: Multi-Input Multi-Scale Surfel-Based Lidar-Inertial Continuous-Time Odometry and Mapping","SLICT 以 UFOMap 八元樹維護全域多尺度面元（surfel）地圖，每個節點只存點數、座標和與散佈矩陣，因此子節點新增或刪除時可遞增更新父節點面元，不必反覆重建整張地圖的 k-d 樹。前端把一顆或多顆 LiDAR 的點合併成單一串流，每個原始點依時間戳記在滑動視窗兩個相鄰狀態間內插位姿，直接與多個尺度中平面度足夠的面元建立點到面元因子，再與 IMU 預積分因子一起以 Ceres 最佳化。後端在 K 個最近關鍵影格中找出時間差足夠大的候選作為迴圈，以 ICP 計算相對位姿，做位姿圖最佳化後重算全域地圖。",[3033,3034],"one or several 3D LiDARs merged into one stream (Ouster OS1-128 plus prism-based Livox Mid-70 in-house; horizontal and vertical LiDARs in NTU VIRAL; Ouster 64-channel in Newer College)","IMU (VectorNav VN100 in-house; built-in 100 Hz Ouster IMU in Newer College)","global multi-scale surfel map in an octree built on UFOMap; each node stores point count, sum and scatter so surfels can be added or removed incrementally with Welford-type updates (Sec. II-D)","yes; proximity-based candidate among K nearest keyframes with sufficient time difference, relative pose by ICP (Sec. III-I)","sliding-window MAP optimization in Ceres over several states per LiDAR sweep (2 to 8 new states per cloud; 400 ms window with 16 intervals in NTU VIRAL), with IMU preintegration factors and continuous-time point-to-surfel factors; the deskew, association and optimization steps can be iterated (Sec. II-C, III-A, III-F)","multi-scale point-to-surfel: each point is matched to every surfel at octree depths 1 to Dmax whose voxel intersects a sphere around the point, with enough points and planarity above a threshold, and whose plane distance is below dmax; five scales from 2 to 32 times the 0.1 m leaf size were used (Sec. III-D, IV-A)","IMU-propagated poses interpolated with slerp per point for association; the optimization itself uses raw points with time-interpolated states (Sec. III-C, III-D)","trajectory, keyframe poses with deskewed keyframe clouds, and a global point and surfel map (Figs. 8 and 12)",[1591,505,45],{"id":3043,"label":3044,"shortName":3045,"title":3046,"year":383,"era":52,"cluster":236,"scope":32,"keyIdeaZh":3047,"sensors":3048,"mapRepresentation":3049,"loopClosure":3050,"estimator":3051,"association":3052,"deskew":121,"outputGeometry":3053,"fulltextStatus":22,"lidarModels":3054,"equipmentCount":46},"voxelhashing2013","Nießner et al., 2013","Voxel Hashing","Real-time 3D reconstruction at scale using voxel hashing","體素雜湊（voxel hashing）以簡單的空間雜湊表只在有量測的表面附近配置 TSDF 體素區塊，避免規則網格或階層式資料結構的記憶體負擔。資料可在 GPU 與主機之間串流進出雜湊表，讓感測器移動時重建範圍可擴大。它主要是地圖表示與融合的資料結構，而非完整 SLAM。",[772],"TSDF in 8x8x8 voxel blocks (8 bytes per voxel: SDF, RGB, weight) indexed by a spatial hash table of 2^21 entries with bucket size 2; blocks outside an active sphere of 8 m radius centred 4 m in front of the camera streamed to host memory in 1 m^3 chunks and streamed back when revisited (Sec. 4, 8, 9.1)","none (authors state that no drift correction is explicitly handled; results section)","frame-to-model point-to-plane ICP against the raycast surface with projective data association, linearised on the GPU and solved by SVD on the CPU (camera tracking paragraph)","projective data association for point-plane ICP, with an optional colour weighting term (camera tracking paragraph)","TSDF with per-voxel colour and weight; isosurface extracted by raycasting for tracking and display, and the authors state isosurfaces can be extracted by raycasting or polygonisation; output meshes are shown in Fig. 10 (Sec. 3, 7, Fig. 10)",[],{"id":3056,"label":3057,"shortName":3058,"title":3059,"year":3060,"era":52,"cluster":236,"scope":346,"keyIdeaZh":3061,"sensors":3062,"mapRepresentation":3064,"loopClosure":72,"estimator":3065,"association":3066,"deskew":57,"outputGeometry":3067,"fulltextStatus":22,"lidarModels":3068,"equipmentCount":24},"nister2004vo","Nistér et al., 2004","Visual Odometry (Nistér et al.)","Visual odometry",2004,"本文提出並命名「視覺里程計」，只用影像即時估計單一相機或立體相機的運動。前端在每張影像偵測 Harris 角點，以正規化互相關在視差限制內比對並做雙向一致性檢查，再把匹配串成軌跡。單目版以五點法估計三視角相對方向、三角化後以三點法求位姿並估計尺度；立體版則直接三角化再以三點法求位姿並以左右影像共同評分，全部用搶先式 RANSAC 與迭代精修。定期重新三角化形成「防火牆」，阻止錯誤傳播。",[3063,37],"stereo camera (calibrated)","none persistent; locally triangulated sparse 3D points only","frame-to-frame visual odometry: monocular scheme with three-view 5-point relative orientation, triangulation, 3-point resection and scale estimation, all in preemptive RANSAC with iterative refinement; stereo scheme with triangulation followed by 3-point pose from the left image, scored and refined on reprojection errors in both images; Cauchy robust cost; periodic re-triangulation as a firewall against error propagation (Sec. 4)","Harris corners without absolute thresholds (up to 5000 features in 10 x 10 buckets); 11 x 11 normalised cross-correlation within a disparity limit (typically 10% of image size); mutual-consistency check; matches linked into tracks (Secs. 2, 3)","6-DoF camera trajectory in metric scale (stereo); obstacle maps when combined with a separate stereo obstacle-detection module (Fig. 8)",[],{"id":3070,"label":3071,"shortName":3072,"title":3073,"year":249,"era":10,"cluster":755,"scope":346,"keyIdeaZh":3074,"sensors":3075,"mapRepresentation":3079,"loopClosure":72,"estimator":3080,"association":3081,"deskew":121,"outputGeometry":3082,"fulltextStatus":22,"lidarModels":3083,"equipmentCount":109},"nubert2021delora","Nubert et al., 2021","DeLORA","Self-supervised Learning of LiDAR Odometry for Robotic Applications","DeLORA 以自監督（self-supervised）方式訓練 LiDAR 里程計網路：推論時只輸入由原始掃描投影成的球面距離影像，網路直接輸出相鄰兩幀的相對位姿；訓練時以 KD-tree 在三維空間尋找對應點，計算點到平面與平面到平面的幾何損失，因此不需要真值位姿或標註資料。法向量在訓練前以主成分分析（PCA）離線計算，只隨預測旋轉而轉動，使梯度不必穿過法向量計算。作者在足式機器人 ANYmal（建物地下室長廊）、履帶機器人（DARPA SubT Urban Circuit 資料）與 KITTI 上測試，並把位姿接到 LOAM 建圖模組產生地圖。",[3076,3077,3078],"Velodyne VLP-16 Puck Lite (ANYmal, Sec. IV-A)","Ouster OS1-64 (DARPA SubT Urban, Sec. IV-B)","KITTI odometry LiDAR (sensor model not named in the paper, Sec. IV-C)","none inside the method; maps in the experiments are built by the LOAM mapping module fed with DeLORA poses (Sec. IV-A, IV-B)","self-supervised CNN (ResNet-like blocks) regressing relative 6-DoF pose (translation + quaternion) from two spherical range images; poses optionally passed to the LOAM mapping module (Sec. III-B, IV-A)","training only: 3D nearest-neighbour correspondences via KD-tree, with point-to-plane and plane-to-plane losses on PCA normals precomputed offline; inference uses the range image (x, y, z, range) only (Sec. III-C, III-D)","6-DoF relative poses; point-cloud maps only when combined with LOAM mapping (Figs. 3-4)",[143,223],{"id":3085,"label":3086,"shortName":3087,"title":3088,"year":97,"era":10,"cluster":397,"scope":69,"keyIdeaZh":3089,"sensors":3090,"mapRepresentation":57,"loopClosure":57,"estimator":3092,"association":3093,"deskew":121,"outputGeometry":57,"fulltextStatus":22,"lidarModels":3094,"equipmentCount":554},"nubert2022learninglocalizability","Nubert et al., 2022b","Learning-based localizability","Learning-based Localizability Estimation for Robust LiDAR Localization","本文以神經網路直接由單一 LiDAR 掃描預測掃描對掃描配準在六個自由度上是否可定位，不需先建立對應或求解配準最佳化即可提早偵測失效。網路只用模擬資料訓練，並取代 CompSLAM 中以特徵值門檻判斷退化的模組；在礦坑隧道、開闊混凝土場地與辦公室的現地測試中，以地圖與曲線定性展示同一網路不需重新調參，也能轉用到 Ouster OS0-128。",[3091],"3D LiDAR (Velodyne VLP-16; Ouster OS0-128 for sensor-transfer test)","sparse 3D convolutional ResUNet feature extractor (MinkowskiEngine; 4000 sampled points, 0.2 m voxels) with global max pooling and a 5-layer MLP giving six sigmoid outputs; multi-label binary classification of localizability along x, y, z, roll, pitch and yaw trained with binary cross-entropy; per-direction probability thresholds chosen from a validation precision-recall curve","not_applicable (prediction before registration)",[143,222],{"id":3096,"label":3097,"shortName":3098,"title":3099,"year":30,"era":10,"cluster":263,"scope":54,"keyIdeaZh":3100,"sensors":3101,"mapRepresentation":3108,"loopClosure":3109,"estimator":3110,"association":3111,"deskew":3112,"outputGeometry":3113,"fulltextStatus":22,"lidarModels":3114,"equipmentCount":487},"holisticfusion2026","Nubert et al., 2026","Holistic Fusion","Holistic Fusion: Task- and Setup-Agnostic Robot Localization and State Estimation With Factor Graphs","Holistic Fusion 是以 GTSAM 因子圖為核心的通用狀態估測框架：IMU 為骨幹，外部模組提供的位姿、位置、速度與地標量測都可作為因子接入；各個參考座標系（例如會漂移的 LiDAR 地圖座標系、里程計座標系與 GNSS 世界座標系）之間的對齊關係被當成隨機漫步的狀態一起最佳化，並沿路徑以局部關鍵影格對齊。線上以固定延遲平滑器提供 IMU 頻率的輸出，並另外提供平滑、不跳動的里程計座標系；離線則做整段批次最佳化，可作為事後處理的參考軌跡。",[3102,3103,3104,3105,3106,3107],"IMU (core)","LiDAR registration poses (e.g., Open3D SLAM, CompSLAM, Coin-LIO outputs)","GNSS (single or dual antenna)","leg kinematics or wheel encoder","mm-wave radar velocity","cabin rotary encoder (HEAP)","none internally; aligns external map frames (e.g., LiDAR map, odometry frame) to the world frame","not part of the framework; handled by upstream modules","GTSAM factor graph fusing IMU pre-integration with absolute pose, absolute position (GNSS), 3D landmark, local velocity and relative measurements; dynamic reference-frame alignment states modelled as an SE(3) random walk, local keyframe alignment along the path, landmark and global calibration states; online iSAM2 fixed-lag smoother with IMU-rate prediction and asynchronous updates, and offline batch optimization for post-processed ground truth (Sec. IV)","front-end agnostic: consumes poses, positions, velocities and landmarks from external modules (e.g., LiDAR scan-to-map registration, leg odometry, GNSS); outlier handling by robust norms (Sec. IV)","not applicable (no raw point clouds are processed by the framework)","robot state in world, map and smooth odometry frames at IMU rate; offline trajectory used as post-processed ground truth",[143,222],{"id":3116,"label":3117,"shortName":3118,"title":3119,"year":513,"era":52,"cluster":151,"scope":132,"keyIdeaZh":3120,"sensors":3121,"mapRepresentation":3124,"loopClosure":3125,"estimator":3126,"association":3127,"deskew":3128,"outputGeometry":3129,"fulltextStatus":22,"lidarModels":3130,"equipmentCount":487},"nuchter2007_6dslam","Nüchter et al., 2007","6D SLAM (Kurt3D, stop-scan-go ICP SLAM)","6D SLAM, 3D mapping outdoor environments","本文提出以三維雷射掃描為基礎的 6D SLAM（六自由度同時定位與建圖）：機器人以停下、掃描、再前進（stop-scan-go）的方式取得每一幅三維點雲，先把輪式里程計外推為六自由度初值，再以八元樹（octree）由粗到細搜尋初始對齊，之後用 ICP（Iterative Closest Point）逐幅配準（registration）。偵測到迴圈時，把閉合誤差依行經路徑長度比例分攤給迴圈內各幅掃描；資料收集完成後，再以同步配準（simultaneous matching）式的全域鬆弛（global relaxation）反覆將每幅掃描對其重疊鄰居重新配準。為壓低計算量，作者使用點數縮減、近似 k-d tree 與快取 k-d tree（cached k-d tree）搜尋。整套系統只維持單一位姿假設，未使用機率式不確定性表示。",[3122,3123],"3D laser range finder built from a SICK 2D scanner on a servo-driven pitch mount (Sec. 5.1)","wheel odometry used only for initial pose extrapolation (Sec. 3.1)","set of registered 3D point clouds (scans); octree used only for initial alignment; voxel view shown for a RoboCup arena map (Fig. 19)","loop hypothesis from maximum laser range and current pose, revised by octree matching; accepted when the number of closest-point pairs exceeds a threshold; closing error distributed over the loop's scans in proportion to travelled path length (translation linearly, rotation by quaternion interpolation) (Sec. 3.4)","single-hypothesis deterministic pose estimation: odometry extrapolated to 6 DoF, octree-based coarse-to-fine initial alignment, then ICP with closed-form SVD solution per scan; explicitly no probabilistic filter or covariance (Sec. 3, Sec. 6)","point-to-point closest-point correspondences (ICP) using k-d tree, approximate k-d tree and cached k-d tree search; octree cube-overlap counting for the initial guess (Sec. 3.2-3.3, Sec. 4)","avoided by design: the robot stands still while each 3D scan is taken (stop-scan-go, Sec. 3 and Sec. 5.1); no in-motion distortion correction","globally registered 3D point cloud and 6-DoF scan poses; a continuous trajectory is reconstructed afterwards by distributing the gaps between trajectory patches (Sec. 5.4, Fig. 18); no covariance output",[45],{"id":3132,"label":3133,"shortName":3134,"title":3135,"year":249,"era":10,"cluster":151,"scope":12,"keyIdeaZh":3136,"sensors":3137,"mapRepresentation":3139,"loopClosure":3140,"estimator":3141,"association":3142,"deskew":3143,"outputGeometry":3144,"fulltextStatus":22,"lidarModels":3145,"equipmentCount":24},"rloam2021","Oelsch et al., 2021","R-LOAM","R-LOAM: Improving LiDAR Odometry and Mapping With Point-to-Mesh Features of a Known 3D Reference Object","R-LOAM 延伸 LOAM（A-LOAM 實作）的建圖模組：假設環境中有一個幾何與全域位姿皆已知的參考物件，先以物件包圍盒裁切掃描點，再透過 AABB 樹找出每個掃描點在三角網格上的最近虛擬點，形成點到網格（point-to-mesh）殘差，與 LOAM 的角點、面點殘差經正規化後共同最佳化，網格權重隨迭代次數以對數方式增加。驗證在 Gazebo 模擬的機庫 B737、廂型車與艾菲爾鐵塔三種情境，分別模擬 Velodyne VLP-16 與 Ouster OS1-128；相對 LOAM，三種情境的中位數 APE 平均降幅分別超過 92%、69% 與 94%，但需要 15 至 35 次迭代，且完全依賴網格與物件位姿的正確性。",[3138],"Simulated Velodyne VLP-16 (16 lines, up to 30,000 points per scan, 10 Hz, Gaussian noise sigma 0.03) and simulated Ouster OS1-128 (128 lines, up to 262,144 points per scan, 10 Hz, sigma 0.05), one at a time, on a 1-DoF gimbal tilting between -0.6 and +0.6 rad on a quadcopter UAV in Gazebo (Sec. IV-A, Fig. 4)","LOAM feature map of corner and surface points stored in a cube structure; map building unchanged (Sec. III-A, III-B3)","none; the paper notes LOAM performs no loop closure and R-LOAM does not add one (Sec. II-A, III)","LOAM (A-LOAM) mapping optimization extended to a joint cost of normalized corner, surface and point-to-mesh residuals with Huber loss; mesh weight lambda raised logarithmically from 0.1 to 40 over up to 35 correspondence and optimization iterations; Ceres Levenberg-Marquardt trust region; A-LOAM's real-time abort was disabled so every scan is map-optimized (Sec. III-B3, IV-B)","LOAM corner and surface point correspondences plus point-to-mesh correspondences: scans are cropped to the object's bounding box plus a buffer (2 m, 200 m for the Eiffel Tower, ground removed below 0.5 or 1 m), then for each remaining point the closest virtual point on the triangular mesh is found via an AABB tree and the IGL library (Sec. III-B1, III-B2, IV-B)","not_reported (the paper does not describe motion or gimbal-actuation compensation beyond LOAM defaults)","per-scan 6-DoF poses and the LOAM feature map; improved 3D reconstruction is claimed but no map accuracy metric is reported (Sec. V, VII)",[143,1591],{"id":3147,"label":3148,"shortName":3149,"title":3150,"year":97,"era":10,"cluster":151,"scope":12,"keyIdeaZh":3151,"sensors":3152,"mapRepresentation":3156,"loopClosure":3157,"estimator":3158,"association":3159,"deskew":3160,"outputGeometry":3161,"fulltextStatus":22,"lidarModels":3162,"equipmentCount":554},"roloam2022","Oelsch et al., 2022","RO-LOAM","RO-LOAM: 3D Reference Object-based Trajectory and Map Optimization in LiDAR Odometry and Mapping","RO-LOAM 是可外掛在 LiDAR SLAM 上的「參考物件式軌跡與地圖最佳化」：LOAM 本身不修改，每隔 L 幅掃描便把最近 M+1 幅裁切後的掃描以 ICP 對齊到已知參考物件的稠密點雲模型，再以 EKF 運動先驗檢查最後一個對齊位姿是否與前序一致（0.05 m 與 0.5 度內），通過者以高權重加入兩次修正之間的位姿圖，最佳化後重新插入過去掃描以修正 LOAM 地圖。實驗以八旋翼無人機搭載由 Dynamixel 致動器旋轉的 Velodyne VLP-16，在機庫內沿 B737 單側飛行三次，以 Leica Nova MS60 追蹤稜鏡作為地面真值；啟用後 LOAM 的中位數 APE 由約 67 至 81 cm 降到約 6.4 至 6.8 cm，且可在邊緣雲端伺服器上線上執行。",[3153,3154,3155],"Velodyne VLP-16 (10 Hz) on a Dynamixel actuator rotating the LiDAR about the roll axis within +\u002F-40 deg","actuator readings transform scans to the robot frame","Intel NUC logs data to SSD (Sec. III-A, IV-A)","LOAM voxelized feature map in a cube structure; corrected by reinserting previous scans with pose-graph-optimized poses (Sec. III-A, III-F)","none; global correction comes from model-aligned trajectory and map optimization (TMO); authors state it may complement relocalization or loop closure methods (Sec. VI)","Unmodified LOAM (A-LOAM) map-optimized poses; every L scans the latest M+1 isolated scans (each with more than 50 points, downsampled to 500) are aligned to the reference point-cloud model by ICP with Huber loss and Levenberg-Marquardt (1 m max correspondence distance, 100 iterations, 2 s cap); the last pose is a candidate if its MSE is below 0.001 and it lies within 0.05 m and 0.5 deg of an EKF motion prior (ROS robot_localization, R = 0.01 I) built from the previous aligned poses; accepted poses become high-confidence constraints in a Ceres pose graph; final setting M = 9, L = 15 (Sec. III-C to III-E, IV-B, V)","raw scans transformed to the world frame with actuator readings and LOAM poses, cropped to the reference object by a bounding box, then point-to-point nearest-neighbour correspondences to the dense model point cloud within 1 m (Sec. III-B, III-C)","actuator motion compensated with actuator readings before LOAM; no ego-motion deskew beyond LOAM defaults is described (Sec. III-A)","corrected trajectory and corrected LOAM feature map (Fig. 5); no map accuracy metric is reported",[143],{"id":3164,"label":3165,"shortName":3166,"title":3167,"year":769,"era":10,"cluster":31,"scope":32,"keyIdeaZh":3168,"sensors":3169,"mapRepresentation":3170,"loopClosure":72,"estimator":3171,"association":3172,"deskew":121,"outputGeometry":3173,"fulltextStatus":22,"lidarModels":3174,"equipmentCount":377},"oleynikova2017voxblox","Oleynikova et al., 2017","Voxblox","Voxblox: Incremental 3D Euclidean Signed Distance Fields for on-board MAV planning","Voxblox 以體素雜湊（voxel hashing）儲存 TSDF，並提出兩項整合策略：同一體素內的點先分組取加權平均再只射線投射一次（grouped raycasting），以及考量深度平方雜訊與表面後方線性衰減的權重函數。接著以波前傳播（raise\u002Flower wavefront）由 TSDF 增量建立歐氏有號距離場（ESDF）供無人機路徑規劃，並可隨時以 marching cubes 輸出網格。",[772,2101],"voxel-hashed TSDF layer, incremental ESDF layer, mesh layer","not_applicable (poses supplied externally, e.g., Vicon or visual-inertial pose)","grouped (merged) ray casting of points into TSDF voxels","TSDF, ESDF and incremental marching-cubes mesh",[],{"id":3176,"label":3177,"shortName":3178,"title":3179,"year":332,"era":52,"cluster":397,"scope":1035,"keyIdeaZh":3180,"sensors":3181,"mapRepresentation":57,"loopClosure":57,"estimator":3184,"association":57,"deskew":57,"outputGeometry":3185,"fulltextStatus":22,"lidarModels":3186,"equipmentCount":24},"olson2010passivesync","Olson, 2010","Passive synchronization","A passive solution to the sensor synchronization problem","許多商用感測器不支援同步，只能在資料抵達主機時打時間戳記，而緩衝與非即時作業系統造成的抖動在高負載時可達數百毫秒。本文利用延遲不可能為負的因果關係，搭配感測器時鐘速率漂移的上界模型，以取最大值的規則估計感測器與主機時鐘的偏移，再還原每筆資料的主機時間。演算法可單向即時執行或前後兩次處理，作者證明結果不會劣於直接以抵達時間打戳記，但實驗僅為合成資料。",[3182,3183],"generic sensors without synchronization support (examples in text: SICK and Hokuyo LIDARs and Xsens IMUs behind USB-to-serial converters)","deployment context only: 12 SICK LIDARs, a Velodyne HDL-64E, 15 Delphi ACC radars and an Applanix IMU\u002FGPS on MIT's DARPA Urban Challenge vehicle (not evaluated in this paper)","max-rule lower-bound estimate of the sensor-to-host clock offset using causality (latency is non-negative) and a bounded clock-rate drift model with parameters alpha1 and alpha2; two-pass O(N) algorithm (causal forward pass plus optional backward pass)","corrected timestamps",[161,45],{"id":3188,"label":3189,"shortName":3190,"title":3191,"year":9,"era":10,"cluster":236,"scope":99,"keyIdeaZh":3192,"sensors":3193,"mapRepresentation":3195,"loopClosure":72,"estimator":3196,"association":3197,"deskew":684,"outputGeometry":3198,"fulltextStatus":22,"lidarModels":3199,"equipmentCount":46},"refusion2019","Palazzolo et al., 2019","ReFusion","ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals","ReFusion 是以 TSDF 為模型的 RGB-D 稠密 SLAM，目標是在有多個移動物體的室內場景中只重建靜態部分。位姿估計不渲染合成視圖，而是把目前影格的點直接帶入 TSDF，以內插得到的符號距離作為殘差，並加入體素色彩的光度誤差。第一次配準後，殘差超過門檻的像素視為動態區域，再以考慮深度的區域成長（flood fill）擴展成遮罩，排除後重新配準並整合。另把相機視錐內確定為空的體素標記為自由空間，之後落在自由空間的量測即視為動態物體而拒絕。方法純幾何、不依賴語意偵測，體素以雜湊配置並在 GPU 平行處理；作者同時發布以動作捕捉提供軌跡、以 Leica BLK360 地面雷射掃描提供靜態場景真值的 Bonn RGB-D 動態資料集。",[3194],"RGB-D camera (ASUS Xtion Pro LIVE in the Bonn RGB-D Dynamic Dataset; TUM RGB-D dynamic sequences)","TSDF with weight and colour per voxel in dynamically allocated voxel-hashed blocks (1 cm voxels, 0.1 m truncation); voxels seen empty in the camera frustum are marked as free space (SDF set to the truncation distance)","Frame-to-model direct alignment: points of the current frame are transformed into the TSDF and the interpolated SDF value is the geometric residual, plus a photometric residual against voxel colours (weight 0.025); Levenberg-Marquardt on three coarse-to-fine levels, GPU-parallel; a second registration is run after masking dynamic pixels","Correspondence-free point-to-implicit residuals; dynamic pixels are those whose residual exceeds t = gamma times tau squared (gamma 0.5, tau 0.1 m), grown by depth-aware flood fill (threshold 0.007) and dilation","mesh of the static part of the scene, trajectory",[],{"id":3201,"label":3202,"shortName":3203,"title":3204,"year":249,"era":10,"cluster":263,"scope":99,"keyIdeaZh":3205,"sensors":3206,"mapRepresentation":3210,"loopClosure":3211,"estimator":3212,"association":3213,"deskew":3214,"outputGeometry":3215,"fulltextStatus":22,"lidarModels":3216,"equipmentCount":736},"locus2021","Palieri et al., 2021","LOCUS","LOCUS: A Multi-Sensor Lidar-Centric Solution for High-Precision Odometry and 3D Mapping in Real-Time","LOCUS 是以 LiDAR 為主的里程計：每顆 LiDAR 的點先依 IMU 或其他里程計做運動畸變校正，再依已知外參合併，經體素與隨機降採樣後，以多執行緒 GICP 依序做掃描對掃描與掃描對子地圖配準。其他感測來源（VIO、KIO、輪式慣性里程計或 IMU 旋轉）不做緊耦合，而是由健康監測依固定優先順序挑出仍健康者，只提供 GICP 的初始值；所有來源失效時退回純 LiDAR 里程計。系統另可依情境啟用平地假設以抑制 Z 向與俯仰、滾轉誤差，並曾是 CoSTAR 團隊贏得 DARPA SubT Urban Circuit 解決方案的關鍵元件。",[3207,3208,3209],"one or more 360-degree 3D LiDARs (two Velodyne VLP16 on Husky, one flat and one pitched forward 30 deg; one VLP16 on Spot)","IMU (Vector Nav 100 on Husky), used for rotation priors and motion distortion correction","external odometry used loosely: wheel-inertial (WIO), visual-inertial (VIO) and kinematic-inertial (KIO) odometry","global point-cloud map stored in an octree (minimum resolution 0.001 m) accumulated every 1 m of translation or 30 deg of rotation (Sec. II-B3)","none inside LOCUS; loop closures of all compared methods were disabled; in the competition the LOCUS output fed a robust odometry aggregator and a separate back-end SLAM (footnote 9, Sec. III-C2)","loosely coupled: a health monitor (in this implementation a message-rate check above 1 Hz) selects the highest-priority healthy source from a static priority queue (Spot: VIO, KIO, IMU, none; Husky: VIO if present, WIO, IMU, none); its relative motion interpolated at LiDAR timestamps seeds multithreaded GICP scan-to-scan and then scan-to-submap registration; the odometry is the integration of the incremental transforms (Sec. II-B)","dense GICP on point clouds filtered by a 0.1 m voxel grid and a random downsampling filter (90%); scan-to-submap against a local region of the global map (Sec. II-A, II-B)","motion distortion correction of each point informed by the IMU or an external odometry source, before multi-LiDAR merging (Sec. II-A)","6-DoF odometry and an accumulated 3D point-cloud map (Figs. 1, 5)",[143],{"id":3218,"label":3219,"shortName":3220,"title":3221,"year":249,"era":10,"cluster":151,"scope":132,"keyIdeaZh":3222,"sensors":3223,"mapRepresentation":3225,"loopClosure":3226,"estimator":3227,"association":3228,"deskew":3229,"outputGeometry":3230,"fulltextStatus":22,"lidarModels":3231,"equipmentCount":377},"mulls2021","Pan et al., 2021","MULLS","MULLS: Versatile LiDAR SLAM via Multi-metric Linear Least Square","MULLS 不依賴掃描線或距離影像，直接把每幀點雲分類為地面、立面、屋頂、柱、梁與頂點等幾何特徵點，因而可用於不同線數與配置的 LiDAR。前端以「多度量線性最小平方」ICP 在各類別內同時最小化點到點、點到面與點到線距離，並以殘差、方向平衡與強度一致性加權；後端以子地圖為單位，透過 TEASER 全域配準與 MULLS-ICP 精修建立迴圈邊，再做階層式位姿圖最佳化。",[3224],"3D LiDAR (seven types: Velodyne HDL-64E, VLP-32C, HDL-32E; Hesai Pandar QT Lite, XT, 64, 128); no IMU required","local map of static classified feature points cropped to a radius; periodically stored submaps (Sec. III-D, III-E)","submap-to-submap global registration with TEASER using neighborhood-category-context (NCC) features, refined by MULLS-ICP; edges rejected by posterior std. and overlap thresholds (Sec. III-E)","multi-metric linear least-squares ICP (small-angle linearization, Gauss-Markov estimation) with residual, direction-balance and intensity weights; hierarchical inter-\u002Finner-submap pose-graph optimization (Sec. III-C, III-E)","ring\u002Frange-image independent classification into ground, facade, roof, pillar, beam and vertex points (dual-threshold ground filter + PCA); category-constrained nearest neighbors with point-to-point, point-to-plane and point-to-line metrics (Sec. III-B, III-C)","optional uniform-motion correction with slerp when point-wise timestamps are available and no IMU (Sec. III-A)","point cloud map and trajectory; compared with a TLS point cloud on MIMAP (Sec. IV-B1, Fig. 9)",[161,926,225,3232,3233,3234,3235],"Hesai Pandar128","Hesai Pandar64","Hesai PandarXT","Hesai PandarQT Lite",{"id":3237,"label":3238,"shortName":3239,"title":3240,"year":661,"era":10,"cluster":755,"scope":132,"keyIdeaZh":3241,"sensors":3242,"mapRepresentation":3250,"loopClosure":3251,"estimator":3252,"association":3253,"deskew":3254,"outputGeometry":3255,"fulltextStatus":22,"lidarModels":3256,"equipmentCount":1593},"pinslam2024","Pan et al., 2024","PIN-SLAM","PIN-SLAM: LiDAR SLAM Using a Point-Based Implicit Neural Representation for Achieving Global Map Consistency","PIN-SLAM 以稀疏可最佳化的神經點編碼局部符號距離場（SDF），里程計採不需最近點配對的點對隱式 SDF 配準，並以局部地圖產生的描述子偵測迴圈、做位姿圖最佳化。因神經點隨所屬影格一起移動，迴圈修正後隱式地圖可保持全域一致並輸出網格。論文另在 Newer College 以毫米級 TLS 參考地圖評估網格精度，並在 Hilti-21（含營建工地序列）報告軌跡誤差。",[3243,3244,3245,3246,3247,3248,3249],"Velodyne HDL64 (KITTI)","Ouster OS1-64 (MulRAN; Newer College long sequences; IPB-Car 2020)","OS1-128 (IPB-Car 2023)","OS0-128 (Newer College shorter sequences)","OS0-64 (Hilti-21, handheld)","32-beam LiDAR on a Spot robot (Nebula, qualitative)","synthetic RGB-D (Replica)","sparse optimizable neural points indexed by voxel hashing with a shared decoder to SDF","distance-based local loop check plus polar context descriptors computed from the local neural map; verification by scan-to-map registration","correspondence-free scan-to-implicit-SDF registration with second-order (Levenberg-Marquardt) optimization and robust weights; pose graph optimization after loop closure","point-to-implicit SDF (no closest-point association)","constant-velocity prediction with per-point timestamps before odometry, then re-deskew with the odometry estimate (Sec. III-B); motion compensation disabled on KITTI because those scans are already deskewed (Sec. V-B1)","mesh via marching cubes from the SDF; compact neural point map",[161,223,222,224,45],{"id":3258,"label":3259,"shortName":3260,"title":3261,"year":207,"era":10,"cluster":755,"scope":132,"keyIdeaZh":3262,"sensors":3263,"mapRepresentation":3267,"loopClosure":3268,"estimator":3269,"association":3270,"deskew":121,"outputGeometry":3271,"fulltextStatus":22,"lidarModels":3272,"equipmentCount":736},"pings2025","Pan et al., 2025","PINGS","PINGS: Gaussian Splatting Meets Distance Fields within a Point-Based Implicit Neural Map","PINGS 在 PIN-SLAM 的神經點上同時編碼連續 SDF 與高斯潑濺輻射場，並加上兩者之間的幾何一致性約束，使影像的稠密光度線索回饋改善距離場，距離場則約束高斯分布。作者在 Oxford Spires 以 Leica RTC360 地面雷射掃描（TLS）參考地圖評估表面重建，但該評估關閉定位模組、全部使用真值位姿，因此量到的是建圖元件品質。",[3264,3265,3266],"Ouster OS1-128 LiDAR, 128 beams, 45 deg vertical FOV, 10 Hz, mounted horizontally (in-house car dataset)","four Basler Ace cameras giving 360 deg coverage at 10 Hz (in-house car dataset)","Oxford Spires handheld rig: 64-beam LiDAR and three global-shutter cameras (dataset)","neural points jointly encoding a signed distance field and a Gaussian-splatting radiance field","loop closure detection running in parallel (inherited from PIN-SLAM)","LiDAR odometry by Gauss-Newton alignment of each scan to the SDF zero level set using only SDF values and gradients (no explicit correspondences); camera poses initialised from LiDAR odometry and extrinsics and refined by gradient descent during radiance-field training to absorb imperfect camera and LiDAR synchronization; loop closure detection and pose graph optimization run in parallel","point-to-implicit SDF for odometry; photometric Gaussian-splatting loss with SDF-radiance geometric consistency for mapping","SDF mesh via marching cubes; rendered RGB and depth from the Gaussian radiance field",[1591,45],{"id":3274,"label":3275,"shortName":3276,"title":3277,"year":150,"era":10,"cluster":151,"scope":132,"keyIdeaZh":3278,"sensors":3279,"mapRepresentation":3284,"loopClosure":3285,"estimator":3286,"association":3287,"deskew":3288,"outputGeometry":3289,"fulltextStatus":22,"lidarModels":3290,"equipmentCount":109},"elasticlidarfusion2018","Park et al., 2018","Elastic LiDAR Fusion","Elastic LiDAR Fusion: Dense Map-Centric Continuous-Time SLAM","Elastic LiDAR Fusion 把連續時間（continuous-time）SLAM 與 ElasticFusion 的「以地圖為中心」（map-centric）概念結合：局部仍以滑動視窗的連續時間軌跡處理手持旋轉 LiDAR 的運動畸變，但全域一致性不靠整條軌跡的批次最佳化，而是在迴圈發生時對整張面元地圖做非剛性變形（deformation graph）。因此迴圈閉合的計算量取決於迴圈前探索的空間大小，而非運作時間。多次觀測以機率式面元融合（surfel fusion）合併，論文報告可降低重建表面的雜訊。",[3280,3281,3282,3283],"rotating 2D LiDAR (Hokuyo UTM-30LX spinning, with encoder)","IMU (Microstrain 3DM-GX3)","Grasshopper3 2.8 MP colour camera with fisheye lens used only for colourisation","Optris PI 450 thermal-infrared camera (382 x 288 pixels) on the device but not used","sparse multi-resolution ellipsoid surfel map plus dense 2D disk surfel map with probabilistic surfel fusion (Sec. III)","two detection sources: (1) rigid ICP between active and inactive sparse-surfel maps detects moderate misalignment on the fly (Algorithm 1); (2) for large misalignment, 3D point-cloud place recognition by keypoint voting (ref. [17], Bosse and Zlot 2013) with descriptors computed every frame and compared with stored scene keys; loop constraints are applied as map deformation, not trajectory optimisation","local sliding-window continuous-time trajectory optimisation solved by iterative nonlinear least squares over subsampled trajectory elements Q, IMU biases and an additional state d that the paper does not define (x = [Q, b_omega, b_alpha, d]), combining surfel-to-surfel, surfel-to-map-prior and IMU acceleration and angular-velocity constraints (Eq. 3-8); global consistency by Gauss-Newton optimisation of an ElasticFusion-style deformation graph with loop, pinning and regularisation terms (Eq. 16-19)","multi-resolution 3D ellipsoidal sparse surfels (from Bosse and Zlot 2009, ref. [9]) matched pairwise and to the global-map prior in their averaged normal direction (Eq. 4-5); dense 2D disk surfels associated with a sensor-noise model that searches deeper along the beam direction under a resolution threshold, then fused by Bayesian fusion (Sec. V-A)","handled by the continuous-time trajectory representation (Sec. I, VIII)","dense fused surfel map (e.g., 20 mm surfels at 10 mm resolution in Fig. 1); no global trajectory is maintained (Sec. VII-B)",[375],{"id":3292,"label":3293,"shortName":3294,"title":3295,"year":97,"era":10,"cluster":151,"scope":132,"keyIdeaZh":3296,"sensors":3297,"mapRepresentation":3305,"loopClosure":3306,"estimator":3307,"association":3308,"deskew":3309,"outputGeometry":3310,"fulltextStatus":22,"lidarModels":3311,"equipmentCount":1593},"elasticity_ct2022","Park et al., 2022","Map-centric dense 3D LiDAR SLAM (ElasticLiDAR++)","Elasticity Meets Continuous-Time: Map-Centric Dense 3D LiDAR SLAM","本文是 Elastic LiDAR Fusion 的期刊延伸，正式版將系統命名為 ElasticLiDAR++，把以地圖為中心的變形式 SLAM 推廣到旋轉單線與多線 3D LiDAR，並融合 IMU 與相機。局部以 100 Hz 線性內插的連續時間軌跡處理運動畸變，再以 B 樣條控制點在 SE(3) 上估計修正量，同時線上估計 LiDAR 與相機的時間延遲；全域則以變形圖讓整張面元地圖產生彈性變形來閉合迴圈，不保存整條軌跡。面元以常態逆 Wishart 模型遞迴融合，搭配保持表面解析度的匹配規則，使多次掃描融合成不重複的稠密地圖。迴圈的錯位估計結合面元點對面約束與 3D 特徵點對點約束，並在多個位置序列式融合，直到不確定度低於門檻。作者以 CT-SLAM 的全域最佳化軌跡為參考，軌跡差異為 0.173 至 0.552 m，平面補丁雜訊最多約降為三分之一；實作需約 2.1 秒處理 1 秒資料，尚非即時。",[3298,3299,3300,3301,3302,3303,3304],"rotating 2D LiDAR (Hokuyo UTM-30LX with encoder, rotor at 1 rotation\u002Fs, hand-held)","3D LiDAR (Velodyne VLP-16, hand-held and robot-mounted)","IMU (Microstrain 3DM-GX3 in hand-held payloads","model not stated for the robot payload)","RGB camera on the single-beam device","independent GoPro without common clock on the multi-beam hand-held device","camera on robot payload (model not stated)","multi-resolution sparse ellipsoid surfels plus fixed-size hexagonal dense surfels with Wishart-based fusion (Sec. III)","detection: the overview describes 2D visual features compared with stored key frames (Sec. III); Sec. VII-D calls the trigger a 'visual place voting method' but cites [52], which is Bosse and Zlot's 3D LiDAR keypoint voting paper; Sec. VII-A states that experiments used a combined 3D and 2D detector [12], [55]; active-inactive sparse-surfel ICP detects moderate misalignment (Sec. V-D, VI-A4). Misalignment is estimated from LiDAR only by tightly combining surfel point-to-plane and 3D sparse-feature (for example FPFH) point-to-point constraints, sequentially fused at several places until the covariance meets a threshold; residual monitoring rejects false positives (Sec. VI-A5, Appendix B)","local composition-type continuous-time trajectory optimisation: discrete poses at 100 Hz corrected by cubic B-spline control points with SE(3) update, minimising surfel-to-surfel, surfel-to-map-prior and IMU residuals over control points, IMU biases and two time lags (LiDAR and camera) (Eq. 3-9); Gauss-Newton deformation-graph optimisation for loop closure (Eq. 21-24); sequential SE(3) pose fusion for metric localisation (Appendix B)","surfel-to-surfel and surfel-to-map-prior constraints; surface-resolution-preservative surfel matching for non-pinhole sensors (Sec. I, III)","continuous-time trajectory representation (abstract)","dense fused surfel map (Sec. III, Fig. 1)",[375,143],{"id":3313,"label":3314,"shortName":3315,"title":3316,"year":661,"era":10,"cluster":755,"scope":132,"keyIdeaZh":3317,"sensors":3318,"mapRepresentation":3320,"loopClosure":3321,"estimator":3322,"association":3323,"deskew":57,"outputGeometry":3324,"fulltextStatus":22,"lidarModels":3325,"equipmentCount":46},"rtgslam2024","Peng et al., 2024","RTG-SLAM","RTG-SLAM: Real-time 3D Reconstruction at Scale using Gaussian Splatting","RTG-SLAM 是以 RGB-D 相機即時重建大範圍室內場景的三維高斯 SLAM。每個高斯只能是不透明或近乎透明：不透明高斯被視為橢圓圓盤，深度以射線與圓盤交點計算，使單一高斯即可貼合一塊局部表面，透明高斯只補足殘餘顏色，因此所需高斯數量與記憶體大幅減少。系統只對新觀測、顏色誤差大或深度誤差大的像素新增高斯，並只最佳化尚未穩定的高斯與其覆蓋的像素，追蹤則採用傳統的影格對模型 ICP，並以沿用 ORB-SLAM2 的後端做特徵地標圖最佳化。",[3319],"RGB-D camera (Microsoft Azure Kinect for the self-scanned dataset)","compact 3D Gaussians forced to be opaque (alpha 0.99, fitting surface and dominant colour, depth rendered by ray intersection with the Gaussian's ellipsoid disc) or nearly transparent (alpha 0.1, residual colour); stable and unstable states with confidence counts; spherical harmonics colour (Sec. 3.1, 3.2)","not_reported (the back end is inherited from ORB-SLAM2, but loop detection is not described in the paper)","multi-level frame-to-model point-to-plane ICP against depth and normals rendered from the Gaussians (front end), plus an ORB-SLAM2-derived back-end graph optimization over 3D ORB landmarks in a separate C++ thread; mapping optimizes only unstable Gaussians with L1 colour and depth losses (Sec. 3.2, Supp. C)","projective point-to-plane ICP correspondences between the current depth frame and the rendered model; ORB feature landmarks in the back end (Sec. 3.2)","Gaussian map rendered to colour, depth and normals; geometry evaluated from points sampled uniformly from the Gaussians (Sec. 4.2)",[],{"id":3327,"label":3328,"shortName":3329,"title":3330,"year":3331,"era":52,"cluster":31,"scope":32,"keyIdeaZh":3332,"sensors":3333,"mapRepresentation":3334,"loopClosure":57,"estimator":57,"association":57,"deskew":57,"outputGeometry":3335,"fulltextStatus":22,"lidarModels":3336,"equipmentCount":77},"pfister2000surfels","Pfister et al., 2000","Surfels","Surfels: surface elements as rendering primitives",2000,"作者把面元（surfel）定義為帶有形狀與著色屬性、可局部近似物體表面的零維 n 元組，沒有顯式連接關係。前處理時以光線追蹤沿三個正交方向取樣，得到三張層狀深度影像組成的層狀深度立方體（LDC），並以八元樹組成 LDC 樹，每個面元儲存深度、法向量索引、材質索引與三層預先濾波的紋理色彩（surfel mipmap）；渲染時以階層式前向投影把區塊投到 z 緩衝，以可見性潑濺（visibility splatting）偵測孔洞，再進行著色與影像重建。此表示後來被即時稠密重建與 LiDAR SLAM 借用作地圖元素。",[],"LDC tree: octree whose nodes (blocks, typically b = 16) store layered depth cubes of surfels, each surfel with 32-bit depth, normal-table index, material-table index and three prefiltered texture mipmap colors (about 20 bytes); optional 3-to-1 reduction to single LDIs","rendered images; surfel set (no explicit connectivity)",[],{"id":3338,"label":3339,"shortName":3340,"title":3341,"year":661,"era":10,"cluster":98,"scope":99,"keyIdeaZh":3342,"sensors":3343,"mapRepresentation":3345,"loopClosure":72,"estimator":3346,"association":3347,"deskew":3348,"outputGeometry":3349,"fulltextStatus":22,"lidarModels":3350,"equipmentCount":24},"coinlio2024","Pfreundschuh et al., 2024","COIN-LIO","COIN-LIO: Complementary Intensity-Augmented LiDAR Inertial Odometry","COIN-LIO 以 FAST-LIO2 的點到平面配準為基礎，將 LiDAR 強度回波投影為強度影像並做亮度一致化濾波，再把影像區塊的光度誤差（photometric error）一併放入迭代擴展卡爾曼濾波。作者偵測點雲配準中資訊不足的方向，刻意挑選能在該方向提供互補資訊的影像區塊，以改善隧道、平坦場地等幾何退化場景的穩健性。作者並公開以全測站量測地面真值的 ENWIDE 退化場景資料集。",[3344,36],"high-resolution 3D LiDAR with intensity (Ouster OS0-128)","ikd-Tree point map inherited from FAST-LIO2 plus a feature map of tracked 5 x 5 intensity patches whose pixels are each initialised at their own global 3D position","iterated extended Kalman filter built on FAST-LIO2, fusing IMU, point-to-plane registration and photometric error on intensity-image patches","point-to-plane geometry plus photometric patch residuals on brightness-filtered intensity images; patches selected to be informative in directions where geometry is uninformative","IMU-propagated undistortion of each point to the scan-end time as in FAST-LIO2 (Sec. III-B); for photometric residuals, tracked points are projected into the distorted LiDAR frame using a projection-based undistortion map that returns each pixel's point index and timestamp (Sec. III-G)","odometry and point map; export format not_reported",[2787],{"id":3352,"label":3353,"shortName":3354,"title":3354,"year":115,"era":52,"cluster":527,"scope":99,"keyIdeaZh":3355,"sensors":3356,"mapRepresentation":3358,"loopClosure":121,"estimator":3359,"association":3360,"deskew":121,"outputGeometry":3361,"fulltextStatus":22,"lidarModels":3362,"equipmentCount":144},"pomerleau2014_icpmapper","Pomerleau et al., 2014","Long-term 3D map maintenance in dynamic environments","本文提出以單一 3D 雷射進行長期定位與建圖的系統，重點是地圖隨時間的維護。新點雲先以 libpointmatcher 的 ICP 配準到全域稀疏點雲地圖，系統再依可見性假設逐點更新地圖點為動態的貝氏機率：若新讀值出現在原地圖點後方（1 度圓錐內），代表雷射穿過了該點位置，該點較可能已移動；更新時同時考慮兩射線夾角、入射角與距離雜訊。被判為動態的點再以雙向非剛性 ICP 估計逐點速度，不需物件模型或分群；P(Dyn) 也可作為 ICP 權重，以減少動態點造成的定位漂移。",[2568,3357],"wheel odometry (prior alignment for registration)","sparse global point cloud with surface normals, timestamps and per-point Bayesian probability of being dynamic; a new point is added only if its nearest map point is farther than 0.3 m; time history of per-point velocities kept (Sec. III-V)","ICP registration of each point cloud to the global map with libpointmatcher, wheel odometry as prior; P(Dyn) can weight points in ICP to discount dynamic points (Sec. II; Sec. V.A)","ICP nearest neighbours via libnabo k-d trees; for dynamic inference, map points are associated with each new reading inside a 1 deg cone in spherical coordinates (Sec. III)","static-scene point map (P(Dyn) \u003C 0.5), dynamic points with velocity vectors, and maps of dynamic-element occurrence, average speed and heading (Figs. 5, 10, 11)",[225],{"id":3364,"label":3365,"shortName":3366,"title":3367,"year":249,"era":10,"cluster":98,"scope":132,"keyIdeaZh":3368,"sensors":3369,"mapRepresentation":3372,"loopClosure":3373,"estimator":3374,"association":3375,"deskew":3376,"outputGeometry":3377,"fulltextStatus":22,"lidarModels":3378,"equipmentCount":46},"rflio2021","Qian et al., 2021","RF-LIO","RF-LIO: Removal-First Tightly-coupled Lidar Inertial Odometry in High Dynamic Environments","RF-LIO 以 LIO-SAM 為基礎，處理大量移動物體時「先要準確位姿才能移除動態點、但動態點又破壞配準」的循環問題。新關鍵影格到達時先不做掃描配準，而是以 IMU 預積分取得初始位姿，並依預測的平移與旋轉誤差決定距離影像的角解析度；將目前掃描與周邊特徵子地圖投影成同解析度距離影像，依可見性差異移除子地圖中的動態點，再做 LOAM 特徵配準。若以邊緣點距離計算的收斂分數未達門檻，就以新解析度重複移除與配準；收斂並完成圖最佳化後，再以細解析度移除目前關鍵影格殘留的動態點。",[3370,3371],"3D LiDAR at 10 Hz (model not named)","IMU at 400 Hz (model not named)","keyframe-based edge and planar feature map from which moving points are removed; global point map without ghost tracks (Fig. 5)","yes; same loop detection as LIO-SAM (Sec. IV-A)","LIO-SAM factor graph in GTSAM (IMU preintegration, LiDAR odometry and loop closure factors) with an added removal-first loop: IMU prior, dynamic-point removal, scan matching, convergence check and repeated removal at a new resolution (Sec. III-A, IV-A)","LOAM edge and planar features with nearest-neighbour point-to-line and point-to-plane distances to a feature submap, after removing submap points flagged as dynamic (Sec. III-E)","IMU-based motion compensation of each scan as in LIO-SAM (Sec. III-A, Fig. 2)","trajectory and a static point cloud map with moving-object points removed",[45],{"id":3380,"label":3381,"shortName":3382,"title":3383,"year":61,"era":10,"cluster":151,"scope":99,"keyIdeaZh":3384,"sensors":3385,"mapRepresentation":3387,"loopClosure":3388,"estimator":3389,"association":3390,"deskew":3391,"outputGeometry":3392,"fulltextStatus":22,"lidarModels":3393,"equipmentCount":144},"aloam_software","Qin & Cao, n.d.","A-LOAM","A-LOAM: Advanced implementation of LOAM (GitHub repository HKUST-Aerial-Robotics\u002FA-LOAM)","A-LOAM 是 HKUST 空中機器人組對 LOAM 的重新實作，README 說明以 Eigen 與 Ceres Solver 簡化程式結構並移除繁複推導，定位為學習用的精簡版本。程式分為特徵擷取、掃描對掃描里程計與掃描對地圖精修三個節點，只使用 3D LiDAR，不讀取 IMU，也沒有迴圈閉合。依原始碼判讀，預設設定關閉了單幀內的運動畸變校正（DISTORTION 設為 0），地圖以 50 m 立方格保存邊緣與平面特徵點。在本群集已核對的論文中，F-LOAM 的授權檔說明其程式碼由 A-LOAM 修改而來，Loam_livox 與 LiLi-OM 以 A-LOAM 作為 LOAM 基準；Feng 等人的施工現場評估也提及 A-LOAM 並指出其缺少迴圈閉合。（推論）後續比較中的「LOAM」可能多指 A-LOAM；由於 A-LOAM 預設關閉單幀內畸變校正，其結果不宜直接當成原始 LOAM 論文所述方法的表現。原始 LOAM 程式碼本次未核對，兩者是否等價仍屬未驗證。",[3386],"3D spinning LiDAR only: launch files for Velodyne VLP-16, HDL-32 and HDL-64 (README examples name 'Velodyne VLP-16' and 'Velodyne HDL-64'; 'HDL-64E' is not written); no IMU subscription in the code","global edge and planar feature point clouds held in a 21 x 21 x 11 array of 50 m cubes (4851 cells) re-centred around the sensor; voxel-grid downsampling with defaults 0.4 m (line) and 0.8 m (plane), 0.2 m and 0.4 m in the VLP-16 launch file","none: no loop-closure or place-recognition code in scanRegistration.cpp, laserOdometry.cpp or laserMapping.cpp (code inspection), consistent with feng2025_construction_lidar_eval Sec. 4.1.1","LOAM-style two-stage estimation re-implemented with Eigen and Ceres Solver: scan-to-scan odometry minimising point-to-edge and point-to-plane residuals (2 Ceres solves, Huber loss 0.1) followed by scan-to-map refinement (2 outer iterations, max 4 solver iterations) (README; laserOdometry.cpp; laserMapping.cpp; lidarFactor.hpp)","curvature-based features per scan-line segment: up to 2 sharp and 20 less-sharp edge points (curvature > 0.1) and 4 flat points (curvature \u003C 0.1), remaining less-flat points downsampled at 0.2 m; scan-to-scan correspondences by KD-tree nearest neighbours on adjacent scan lines (distance threshold 5 m); scan-to-map uses 5 nearest map points, accepts a line when the largest eigenvalue exceeds 3 times the second and fits a plane otherwise; only 16, 32 or 64 scan lines are supported","not applied in the default code: '#define DISTORTION 0' makes TransformToStart use the full scan-to-scan relative pose for every point (s = 1.0), and the TransformToEnd re-projection block is wrapped in 'if (0)'; no IMU input is used","odometry streams (\u002Flaser_odom_to_init at scan rate, \u002Faft_mapped_to_init after scan-to-map refinement, \u002Faft_mapped_to_init_high_frec), registered full-resolution cloud (\u002Fvelodyne_cloud_registered), local surround map every 5 frames and full feature map every 20 frames",[143,225,161],{"id":3395,"label":3396,"shortName":3397,"title":3398,"year":150,"era":10,"cluster":236,"scope":132,"keyIdeaZh":3399,"sensors":3400,"mapRepresentation":3401,"loopClosure":3402,"estimator":3403,"association":3404,"deskew":57,"outputGeometry":3405,"fulltextStatus":22,"lidarModels":3406,"equipmentCount":1764},"vinsmono2018","Qin et al., 2018","VINS-Mono","VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator","VINS-Mono 以單眼相機加低成本 IMU 估計具公制尺度的六自由度狀態，先以僅視覺 SfM 與視覺慣性對齊完成初始化（陀螺儀偏差、速度、重力方向與尺度），再以滑動視窗緊耦合融合 IMU 預積分（pre-integration）與特徵觀測。迴圈偵測後進行緊耦合重定位，並因 roll、pitch 可由 VIO 觀測，只對 x、y、z 與偏航角做 4-DOF 位姿圖最佳化。作者將稠密建圖列為未來工作。",[37,36],"sparse features in sliding window; keyframe pose graph","DBoW2 loop detection with temporal and geometric checks, tightly-coupled relocalization, and loop edges in the pose graph with Huber norm (Sec. VII, VIII-B)","tightly-coupled sliding-window nonlinear optimisation with IMU pre-integration and marginalisation; 4-DOF pose-graph optimisation for global consistency (Sec. III, VI, VIII)","KLT tracking keeping 100 to 300 uniformly spaced corners, RANSAC fundamental-matrix rejection; loop detection with DBoW2 on 500 extra BRIEF corners, then BRIEF matching with 2D-2D fundamental and 3D-2D PnP RANSAC checks before relocalization (Sec. IV-A, VII-A, VII-B)","metric 6-DOF trajectory and sparse features; dense mapping named as future work (Sec. X)",[],{"id":3408,"label":3409,"shortName":3410,"title":3411,"year":9,"era":10,"cluster":236,"scope":99,"keyIdeaZh":3412,"sensors":3413,"mapRepresentation":3414,"loopClosure":3415,"estimator":3416,"association":3417,"deskew":57,"outputGeometry":3418,"fulltextStatus":22,"lidarModels":3419,"equipmentCount":487},"vinsfusion2019","Qin et al., 2019","VINS-Fusion (local odometry framework)","A General Optimization-based Framework for Local Odometry Estimation with Multiple Sensors","此預印本提出以最佳化為核心的通用區域里程計框架，把每種感測器量測視為一個因子（factor），共享狀態變數的因子相加組成最佳化問題。論文以雙目、單眼加 IMU、雙目加 IMU 三種組合示範。其目標為區域精度，全域感測器如 GPS 被列為後續工作。",[2101,37,36],"sparse visual landmarks","none in this preprint (local odometry)","sliding-window nonlinear least squares (MAP) solved with Ceres; camera reprojection factors and IMU preintegration factors; ten spatial camera frames kept, older states marginalised by Schur complement into a prior (Sec. IV-A to IV-D)","Shi-Tomasi corner features tracked between frames by KLT and matched left to right for stereo; each feature forms a reprojection factor from its first observation, parameterised by its depth in the first observing frame (Sec. IV-A-1, IV-B-1)","local odometry trajectory",[],{"id":3421,"label":3422,"shortName":3423,"title":3424,"year":396,"era":10,"cluster":151,"scope":99,"keyIdeaZh":3425,"sensors":3426,"mapRepresentation":3429,"loopClosure":72,"estimator":3430,"association":3431,"deskew":3432,"outputGeometry":3433,"fulltextStatus":22,"lidarModels":3434,"equipmentCount":554},"lins2020","Qin et al., 2020","LINS","LINS: A Lidar-Inertial State Estimator for Robust and Efficient Navigation","LINS 以機器人中心（robocentric）表述的迭代誤差狀態卡爾曼濾波器（iterated ESKF）緊耦合 6 軸 IMU 與 3D LiDAR：每次迭代都重新尋找點到邊、點到面的特徵對應，以降低錯誤匹配造成的線性化誤差。狀態以上一時刻的局部座標表示，再組合成全域位姿，以避免長時間運作時不確定性增長導致濾波發散。論文聚焦於里程計模組，建圖直接沿用 LeGO-LOAM 的建圖演算法。",[3427,3428],"3D LiDAR (Velodyne VLP-16 on a car in the port test; RS-LiDAR-16 on a bus in the indoor parking lot and urban tests)","6-axis IMU (Xsens MTi-G-710 in the port test; the IMU placed inside the bus is not specified)","global feature map from the LeGO-LOAM mapping algorithm (Sec. III-A, IV)","robocentric iterated error-state Kalman filter (iterated ESKF), re-finding feature correspondences at each iteration (Sec. III-C)","LOAM\u002FLeGO-style edge and planar features; point-to-edge and point-to-plane residuals against the previous scan only (Sec. III-B, III-C, IV-C)","raw features undistorted with the relative transformation estimated after the iterated update (Sec. III-C3)","global 3D map (1 Hz) and fused odometry (400 Hz) (Fig. 2)",[143,3435],"RS-LiDAR-16",{"id":3437,"label":3438,"shortName":3439,"title":3440,"year":51,"era":10,"cluster":68,"scope":69,"keyIdeaZh":3441,"sensors":3442,"mapRepresentation":3444,"loopClosure":72,"estimator":3445,"association":3446,"deskew":57,"outputGeometry":389,"fulltextStatus":22,"lidarModels":3447,"equipmentCount":77},"qin2023geotransformer","Qin et al., 2023","GeoTransformer","GeoTransformer: Fast and Robust Point Cloud Registration With Geometric Transformer","GeoTransformer 屬學習式、免關鍵點的配準：先在降採樣的超點（superpoint）間比對，再傳播到稠密點。其幾何 Transformer 編碼點對距離與三點角度，使特徵對剛體變換不變，並在低重疊情形下保持穩健；摘要指出匹配精度高到不需 RANSAC 即可估計轉換。",[3443],"none of its own; public benchmarks: 3DMatch and 3DLoMatch indoor RGB-D scene fragments, KITTI odometry LiDAR scans (pairs at least 10 m apart, ground truth refined by ICP), ModelNet40 synthetic CAD points, Augmented ICL-NUIM synthetic RGB-D with a noise model, 4DMatch and 4DLoMatch non-rigid animations","superpoints and dense points","local-to-global registration (LGR): weighted SVD on the point correspondences of each superpoint match gives candidate transforms, the candidate with most inliers within an acceptance radius is kept and re-estimated on inliers for Nr = 5 iterations; RANSAC-50k and plain weighted SVD are also evaluated","KPConv-FPN backbone; geometric self-attention (pair-wise distance and triplet-wise angle embeddings) interleaved three times with feature-based cross-attention; Gaussian correlation with dual normalisation selects the top Nc superpoint matches (256 at test time); an optimal-transport layer (Sinkhorn) with mutual top-k extracts dense point correspondences inside matched patches",[],{"id":3449,"label":3450,"shortName":3451,"title":3452,"year":97,"era":10,"cluster":151,"scope":132,"keyIdeaZh":3453,"sensors":3454,"mapRepresentation":3457,"loopClosure":3458,"estimator":3459,"association":3460,"deskew":3461,"outputGeometry":3462,"fulltextStatus":22,"lidarModels":3463,"equipmentCount":1764},"wildcat2022","Ramezani et al., 2022","Wildcat","Wildcat: Online Continuous-Time 3D Lidar-Inertial SLAM","Wildcat 是 CSIRO 的線上 3D LiDAR 慣性 SLAM，其里程計是 Zebedee 等離線連續時間方法概念的即時實作：在固定長度的滑動時間視窗內，把點雲依位置與時間聚成多解析度橢球面元（surfel），以面元對面元的點到面型成本與 IMU 成本共同修正取樣位姿，再以三次 B-spline 內插得到高頻軌跡並重投影面元，藉此處理運動畸變。後端以六秒子地圖為節點做位姿圖最佳化，加入重力方向項並合併重疊節點，使計算量隨探索空間而非任務時間成長，亦支援多機去中心化建圖。",[3455,3456],"3D LiDAR Velodyne VLP-16 in two configurations: servo-spun at 0.5 Hz on an inclined mount for 120 deg vertical FoV with the measurement rate set to 20 Hz (SpinningPack), or fixed with the native 30 deg vertical FoV (FlatPack; rate not stated); Ouster OS1-64 at 10 Hz with 120 m range in MulRan DCC03","IMU (9-DoF 3DM-CV5 at 100 Hz in the SpinningPack; FlatPack IMU model not stated)","local multi-resolution surfel maps bundled into six-second submaps (Sec. V-A)","candidates by Mahalanobis-distance search or place recognition (e.g., Scan Context); point-to-plane ICP between surfel submaps, with global initialization when uncertainty is large; Mahalanobis gating (Sec. V-B2)","sliding-window continuous-time trajectory optimization (Gauss-Newton with Cauchy IRLS) over sample-pose corrections, followed by pose-graph optimization with a gravity-direction term (Cauchy IRLS) (Sec. IV-V)","multi-resolution surfels (voxel clustering by position and time, planarity-filtered ellipsoids); reciprocal kNN matching in a 7-D descriptor space; point-to-plane-type surfel cost weighted by lidar noise and surfel thickness (Sec. IV-A to IV-C)","surfels re-projected with the continuously updated IMU-rate trajectory (Sec. IV-B2)","globally optimized multi-agent point cloud map (e.g., submitted to DARPA SubT) and trajectory (Sec. VI-B)",[143,223],{"id":3465,"label":3466,"shortName":3467,"title":3468,"year":1461,"era":52,"cluster":397,"scope":1035,"keyIdeaZh":3469,"sensors":3470,"mapRepresentation":57,"loopClosure":57,"estimator":3475,"association":3476,"deskew":57,"outputGeometry":3477,"fulltextStatus":22,"lidarModels":3478,"equipmentCount":109},"rehder2016spatiotemporal","Rehder et al., 2016","General spatiotemporal calibration","A General Approach to Spatiotemporal Calibration in Multisensor Systems","作者把感測器時間戳記與實際量測時刻之間的固定偏移視為確定性誤差，在連續時間 B 樣條批次最大概似估計中與空間外參一起求解。文中推導相機與 IMU、相機與 IMU 與 2D 雷射測距儀、以及立體相機與雷射測距儀等多種估計器；雷射部分以自動平面偵測與沿光束方向的距離模型建立約束。實驗顯示時間偏移可估到遠小於最短取樣間隔，且經軟體去除抖動與時鐘偏斜後的結果接近硬體同步。",[3471,3472,3473,3474],"Aptina MT9V034 WVGA global-shutter cameras at 20 Hz (one in Setup I, two in Setup II)","Analog Devices ADIS16488 (Setup I) and ADIS16448 (Setup II) IMUs at 200 Hz","Hokuyo UTM-30LX 2D laser range finder (Setup II), 270 deg scans at 40 Hz","FPGA-based visual-inertial sensor assigning hardware timestamps","continuous-time batch maximum-likelihood estimation solved by Levenberg-Marquardt, with a sixth-order B-spline IMU pose and cubic B-spline biases; constant time offsets folded into measurement times with analytic Jacobians; five estimators (J, G, A, L, C) for different sensor subsets; estimator J released in kalibr","camera: checkerboard corner reprojection; LRF: range points associated with planes found by RANSAC (threshold 60 mm), region growing and eigenvalue checks, using a beam-direction range model with a cumulative range bias and a Blake-Zisserman robust cost","temporal offsets, camera-IMU and LRF-IMU transforms, gravity direction, IMU bias trajectories, plane parameters and a cumulative LRF range bias",[375],{"id":3480,"label":3481,"shortName":3482,"title":3483,"year":396,"era":10,"cluster":236,"scope":132,"keyIdeaZh":3484,"sensors":3485,"mapRepresentation":3490,"loopClosure":3491,"estimator":3492,"association":3493,"deskew":3494,"outputGeometry":3495,"fulltextStatus":22,"lidarModels":3496,"equipmentCount":736},"voxgraph2020","Reijgwart et al., 2020","Voxgraph","Voxgraph: Globally Consistent, Volumetric Mapping Using Signed Distance Function Submaps","Voxgraph 以一組相互重疊的符號距離函數（SDF）子地圖表示環境。前端依固定時間間隔把連續點雲以 voxblox 光線投射整合成 TSDF 子地圖，子地圖完成後再計算歐氏符號距離場（ESDF），並以 marching cubes 取出等值面點。後端以位姿圖最佳化各子地圖的位置與偏航角，約束包括里程計、外部迴圈閉合（例如 DBoW2）以及作者提出的免對應配準約束：把一個子地圖的等值面點轉入相鄰子地圖的 ESDF，直接讀取距離值作為殘差，並依體素權重隨機只取 5% 殘差以降低計算量。系統以地圖為中心，不重新估計完整軌跡，並在搭載 Intel i7-8650U 的六旋翼無人機上以 Ouster OS1 光達或 RealSense D415 即時運作。",[3486,3487,3488,3489],"3D LiDAR Ouster OS1 (64-beam) in the outdoor MAV field experiment","RGB-D camera Intel RealSense D415 (pointclouds) in the indoor MAV experiment","Odometry input from a time-synchronized camera-IMU running ROVIO visual-inertial odometry; VI-sensor stereo and IMU in the indoor dataset","RTK-GNSS used only as trajectory ground truth","collection of overlapping TSDF submaps built at a fixed frequency by ray casting into spatially hashed voxel blocks (voxblox), each with an ESDF and marching-cubes iso-surface points; submaps can be fused into one global map","External, source-agnostic loop closures between two sensor frames are converted into constraints between the submaps that contain them (DBoW2 place recognition on VI-sensor images in the RGB-D experiment); registration constraints link overlapping submaps, while wide loops rely on the external loop closures (Secs. VI-B, VIII-B2)","Pose-graph nonlinear least squares over submap poses in R3 x SO(2) (x, y, z, yaw; roll and pitch taken from gravity-aligned odometry) with odometry and loop-closure terms (Mahalanobis) and registration terms (weighted squared ESDF distance); per-scan poses come from the external visual-inertial odometry (ROVIO) and are kept fixed relative to their submap","Correspondence-free submap-to-submap registration: iso-surface points of one submap are transformed into the overlapping submap and its trilinearly interpolated ESDF value is the residual; overlapping pairs found with axis-aligned bounding boxes; residuals randomly sub-sampled per solver iteration with probability proportional to voxel weight (5% sampling ratio used)","LiDAR undistortion runs as a component in the field experiment (listed in the CPU breakdown, Fig. 6); the method is not described","globally aligned SDF submap collection and fused global volumetric map; trajectory obtained by projecting submap-relative odometry with the optimised submap poses",[624],{"id":3498,"label":3499,"shortName":3500,"title":3501,"year":97,"era":10,"cluster":151,"scope":99,"keyIdeaZh":3502,"sensors":3503,"mapRepresentation":3507,"loopClosure":3508,"estimator":3509,"association":3510,"deskew":3511,"outputGeometry":3512,"fulltextStatus":22,"lidarModels":3513,"equipmentCount":736},"locus2_2022","Reinke et al., 2022","LOCUS 2.0","LOCUS 2.0: Robust and Computationally Efficient Lidar Odometry for Real-Time 3D Mapping","LOCUS 2.0 是以 LiDAR 為核心、可鬆耦合其他里程計的多階段 GICP 里程計，針對算力與記憶體受限的地下探勘機器人設計。它把 GICP 所需的點共變異數改由預先計算的法向量直接構成，地圖點不必重算共變異數；以自適應體素濾波把每幀點數維持在設定值附近，使運算時間不隨環境大小或 LiDAR 數量劇烈變動；地圖只保留以機器人為中心 50 m 的滑動視窗，可用多執行緒八元樹或 ikd-Tree 儲存，以限制記憶體用量。系統本身沒有迴圈閉合。",[3504,3505,3506],"one or more 3D LiDARs merged in the body frame (three Velodyne VLP16 on Husky; one lidar on Spot)","IMU for per-lidar motion distortion correction","optional non-lidar odometry (wheel-inertial, kinematic-inertial or visual-inertial) as initial guess through the sensor integration module (LiDAR-centric, loosely coupled)","sliding-window robot-centred point map (window 50 m) stored in a multi-threaded octree (two threads alternately box-filter and rebuild) or an ikd-tree; normals are stored with map points (Sec. III-C; Sec. IV-D)","none (odometry system; no loop closure described)","multi-stage GICP registration, scan-to-scan then scan-to-submap, with point covariances built from precomputed normals (GICP from normals); health-aware loosely coupled use of optional non-lidar odometry as the scan-to-scan initial guess (Sec. III; Sec. III-A)","GICP nearest-neighbour correspondences (maximum correspondence distance 0.3, 20 iterations) on points reduced by an adaptive voxel grid filter that holds the point count near a set value (1000 to 10000 tested) (Sec. III-B; Sec. IV-C)","IMU-based motion distortion correction of each lidar stream in the preprocessor (Sec. III)","6-DoF odometry and a local 3D point cloud map",[143,45],{"id":3515,"label":3516,"shortName":3517,"title":3518,"year":396,"era":10,"cluster":236,"scope":132,"keyIdeaZh":3519,"sensors":3520,"mapRepresentation":3521,"loopClosure":3522,"estimator":3523,"association":3524,"deskew":57,"outputGeometry":3525,"fulltextStatus":22,"lidarModels":3526,"equipmentCount":24},"kimera2020","Rosinol et al., 2020","Kimera","Kimera: an Open-Source Library for Real-Time Metric-Semantic Localization and Mapping","Kimera 是模組化的開源度量語意（metric-semantic）視覺慣性 SLAM 函式庫，包含以 GTSAM iSAM2 固定延遲平滑器實作的 VIO、以 PCM 剔除錯誤迴圈的強健位姿圖最佳化、低延遲 3D 網格生成器，以及用雙目稠密匹配（SGM）與 Voxblox TSDF 產生全域語意網格的模組。各模組可獨立或組合執行，並在 CPU 上即時運作。作者以 EuRoC 地面真值點雲評估網格的精度與完整度，但評估前先以 ICP 將估計點雲對齊至真值。",[37,2101,36],"per-frame mesh from 2D Delaunay triangulation of tracked features back-projected with VIO landmark estimates; multi-frame mesh over the VIO horizon, coupled back to VIO through regularity factors when planar surfaces are detected in the mesh; global Voxblox TSDF built at keyframes from semi-global-matching dense stereo with bundled raycasting (fast option), meshed by marching cubes, with Bayesian per-voxel semantic label updates within the truncation distance","DBoW2 putative loops, geometric verification, outlier rejection with a modified PCM before GTSAM PGO (Sec. II-B)","keyframe-based MAP visual-inertial estimator run as full or fixed-lag smoothing (fixed-lag typically used to bound estimation time), with on-manifold IMU preintegration and structureless vision factors solved by iSAM2 in GTSAM and marginalisation of states leaving the horizon; Kimera-RPGO keeps odometry and loop edges separately, selects the largest consistent loop set with a modified incremental PCM (odometry chi-squared check plus pairwise consistency, fast maximum clique) and optimises the pose graph with Gauss-Newton in GTSAM","Shi-Tomasi corners tracked by Lucas-Kanade, left-right stereo matching, 5-point mono and 3-point stereo RANSAC verification (optional 2-point and 1-point variants using IMU rotation) at keyframes; structureless vision factors triangulated by DLT with degenerate and high-reprojection-error points removed; DBoW2 bag-of-words loop candidates verified with the same mono and stereo checks","trajectory, low-latency local mesh, and global semantically annotated mesh from TSDF (Sec. II)",[],{"id":3528,"label":3529,"shortName":3530,"title":3531,"year":51,"era":10,"cluster":755,"scope":99,"keyIdeaZh":3532,"sensors":3533,"mapRepresentation":3534,"loopClosure":3535,"estimator":3536,"association":3537,"deskew":57,"outputGeometry":3538,"fulltextStatus":22,"lidarModels":3539,"equipmentCount":77},"nerfslam2023","Rosinol et al., 2023","NeRF-SLAM","NeRF-SLAM: Real-Time Dense Monocular SLAM with Neural Radiance Fields","NeRF-SLAM 把稠密單眼 SLAM 與即時雜湊式神經輻射場串接：追蹤端直接採用 DROID-SLAM 的學習式光流與稠密光束法平差，並依 σ-Fusion 的做法由 Hessian 結構計算每個深度與位姿的邊際共變異數。建圖端以 Instant-NGP 表示場景，損失同時包含顏色誤差與以共變異數加權的深度誤差，使雜訊大的單眼深度不致把幾何拉偏，並在同一執行緒中微調關鍵影格位姿。兩個執行緒平行運作，在單張 RTX 2080 Ti 上約每秒 10 影格，但系統沒有迴圈閉合。",[37],"Instant-NGP hash-based hierarchical volumetric neural radiance field (density and colour), supervised by RGB and depth weighted by its marginal covariance (Sec. III-B, Sec. III-D)","none described; tracking uses only a sliding window of keyframes (Sec. III-C)","DROID-SLAM dense bundle adjustment over a sliding window of at most 8 keyframes (Schur complement and Cholesky solve), with marginal covariances of dense depths and poses computed as in sigma-Fusion; the mapping thread minimizes photometric plus covariance-weighted depth loss jointly over poses and radiance-field parameters (Sec. III-A to III-C)","dense learned optical flow with per-measurement weights from a RAFT-style ConvGRU (DROID-SLAM) (Sec. III-A)","keyframe poses, dense keyframe depth maps with uncertainty and a radiance field rendered to colour and depth; no mesh or point-cloud accuracy evaluation, geometry assessed by rendered Depth L1 (Sec. IV-C)",[],{"id":3541,"label":3542,"shortName":3543,"title":3544,"year":396,"era":10,"cluster":151,"scope":12,"keyIdeaZh":3545,"sensors":3546,"mapRepresentation":3548,"loopClosure":3549,"estimator":3550,"association":3551,"deskew":3552,"outputGeometry":3553,"fulltextStatus":22,"lidarModels":3554,"equipmentCount":46},"lol2020","Rozenberszki & Majdik, 2020","LOL","LOL: Lidar-only Odometry and Localization in 3D point cloud maps","LOL 只用 LiDAR 在既有 3D 點雲地圖中做里程計與定位：以 LOAM 連續估計位姿，並把最近數幀點雲累積成局部地圖、切成片段，以 SegMatch 或 SegMap 描述子與預先切割描述的目標地圖片段比對。為減少誤匹配，只在里程計位置附近搜尋候選，再以質心平移一致性與 RANSAC 過濾；通過後先以匹配片段質心的平均位移做初值，再以 ICP 對齊片段點雲求出修正量，送入 SegMap 的增量位姿圖。它是在先驗地圖中定位，而不是線上迴圈閉合。",[3547],"3D LiDAR only (KITTI raw drives; sensor model not named in the paper); no IMU, wheel encoder or GPS (Sec. I)","LOAM map voxelized at 5 cm; online local cloud densified from the last k scans and segmented; offline target point cloud segmented and described into a segment database searchable by centroid position (Sec. III-A; Sec. III-B)","not used; drift is cancelled by relocalization against the prior target map instead of loop closure in the online map (Sec. I)","LOAM odometry and mapping (mapping at one tenth of the odometry rate) supplies poses; relocalization updates from segment matches are inserted into SegMap's incremental pose-graph mapping module, modified to incorporate relocalization in the global map (Sec. III-A; Sec. III-B; Fig. 1)","SegMatch eigenvalue (1x7) or SegMap CNN (1x64) segment descriptors; target segments searched only within a distance threshold of the odometry position; pairwise centroid-translation consistency, RANSAC alignment of matched centroids, a mean centroid-shift prior and final ICP between matched segment point clouds (Sec. III-C; Sec. III-D)","not described (LOAM front end used as is)","trajectory relocalized in the target map frame",[45],{"id":3556,"label":3557,"shortName":3558,"title":3559,"year":51,"era":10,"cluster":31,"scope":99,"keyIdeaZh":3560,"sensors":3561,"mapRepresentation":3562,"loopClosure":3563,"estimator":3564,"association":3565,"deskew":121,"outputGeometry":3566,"fulltextStatus":22,"lidarModels":3567,"equipmentCount":185},"ruan2023slamesh","Ruan et al., 2023","SLAMesh","SLAMesh: Real-time LiDAR Simultaneous Localization and Meshing","SLAMesh 將掃描點分入體素格，在每格內以高斯過程（Gaussian process）回歸局部表面，於規則分布的位置預測頂點座標與不確定性，再直接連接相鄰頂點形成網格。新掃描同樣重建後，依頂點位置快速建立點對網格（point-to-mesh）對應進行配準；地圖更新只需修正頂點的一維預測值，因此可在 CPU 上即時同時定位與建網格。",[35],"hash map of voxel cells (1.6 m; 1.5 m for the mesh evaluation), each holding up to three Gaussian-process layers of 6 x 6 regularly located vertices with a predicted coordinate and variance, adjacent or diagonal valid vertices connected into triangles","no explicit loop closure; optional implicit alignment when registering to revisited areas improves map consistency but was disabled for KITTI odometry metrics (Sec. IV-C)","iterative point-to-mesh registration with constraint combination (per-layer residual averaging), solved by Levenberg-Marquardt in Ceres with analytic Jacobians; constant-velocity initial guess","location-based matching of GP-reconstructed vertices to mesh faces in same or adjacent cells, with smoothed face normals","triangle mesh map with vertex uncertainty",[161],{"id":3569,"label":3570,"shortName":3571,"title":3572,"year":9,"era":10,"cluster":68,"scope":69,"keyIdeaZh":3573,"sensors":3574,"mapRepresentation":3577,"loopClosure":72,"estimator":3578,"association":3579,"deskew":57,"outputGeometry":389,"fulltextStatus":22,"lidarModels":3580,"equipmentCount":77},"rusinkiewicz2019symmetric","Rusinkiewicz, 2019","Symmetric ICP","A symmetric objective function for ICP","此文提出對稱化的 ICP 目標函數：以對應點兩側法向量的和作為誤差方向，並把旋轉拆成兩半，以相反方向分別作用於兩個表面。只要兩點與其法向量落在同一個局部二次曲面（2D 為圓弧）上，殘差即為零，因此比只在平面上為零的點對平面誤差保留更多沿表面滑動的自由度，卻不需估計曲率。作者以 Rodrigues 公式推導線性化，每次迭代仍只解一個線性最小平方問題，對應正確時結果是精確的。在 dragon 模型、部分重疊的 bunny 掃描與一組 TUM RGB-D 室內掃描上，每次迭代的誤差下降都快於點對點、點對平面、雙平面與二次近似法；bunny 收斂範圍實驗中，Levenberg-Marquardt 版本在 20 次迭代時收斂範圍最寬。論文沒有報告執行時間，也沒有使用 LiDAR 或大尺度場景資料。",[3575,3576],"[\"range scans and meshes (dragon model, Curless and Levoy 1996","bunny range scans bun000, bun090 and other overlapping pairs, Turk and Levoy 1994)\", \"RGB-D (Kinect scans from TUM RGB-D freiburg1_xyz)\"]","3D models \u002F surfaces","E_symm = sum_i [(R p_i - R^-1 q_i + t) . (n_p,i + n_q,i)]^2: the rotation is split so that each surface is rotated by half the angle in opposite directions, with normals kept fixed (the rotated-normals variant E_symm-RN behaves similarly and is dropped for simplicity); linearized from the Rodrigues formula with a~ = a tan(theta) into one linear least-squares solve per iteration, interpreted as a Gauss-Newton step with Gibbs rotation parameters, exact when correspondences are exact; a Levenberg-Marquardt variant (LM-Symmetric) is also evaluated","both normals enter the residual (Sec. 3.2; if oriented normals are unavailable they are flipped to agree, footnote 1); in the convergence-basin test all variants sample points from both meshes and take Euclidean closest points on the other mesh (Sec. 4.4); the dragon self-alignment test uses no outlier rejection, while the partially overlapping bunny test and the basin test reject pairs with negative normal dot product and pairs farther than 2.5 sigma (sigma = 1.4826 x median distance) at each iteration (Sec. 4.3, 4.4)",[],{"id":3582,"label":3583,"shortName":3584,"title":3585,"year":1217,"era":52,"cluster":31,"scope":1035,"keyIdeaZh":3586,"sensors":3587,"mapRepresentation":3588,"loopClosure":57,"estimator":57,"association":57,"deskew":57,"outputGeometry":3589,"fulltextStatus":22,"lidarModels":3590,"equipmentCount":185},"rusu2011pcl","Rusu & Cousins, 2011","PCL","3D is here: Point Cloud Library (PCL)","PCL 是以 C++ 模板實作、採 BSD 授權的開源點雲處理函式庫，分成濾波、特徵、輸入輸出、分割、表面重建、配準、關鍵點與距離影像等可獨立編譯的模組，底層以 Eigen、FLANN 與 OpenMP 或 TBB 支援線性代數、近鄰搜尋與多核心平行化，並以 ROS nodelet 在同一行程內串接處理圖，避免資料複製。文中示範統計離群值移除（k = 50、1.0σ）與 RANSAC 平面分割（1 cm 門檻）。PCL 的體素格網濾波常用於 SLAM 前處理降採樣，此點屬推論，本文沒有討論，也未評估降採樣對幾何的影響。",[],"point clouds (templated n-D point types)","library operations incl. filtering (e.g., voxel grid), features, registration, surface reconstruction",[45],{"id":3592,"label":3593,"shortName":3594,"title":3595,"year":345,"era":52,"cluster":68,"scope":69,"keyIdeaZh":3596,"sensors":3597,"mapRepresentation":3598,"loopClosure":72,"estimator":3599,"association":3600,"deskew":57,"outputGeometry":3601,"fulltextStatus":22,"lidarModels":3602,"equipmentCount":61},"rusu2009fpfh","Rusu et al., 2009","FPFH \u002F SAC-IA","Fast Point Feature Histograms (FPFH) for 3D registration","作者先整理點特徵直方圖（PFH）：在查詢點半徑內的鄰點兩兩建立 Darboux 座標框，統計三個角度特徵，並刪去原本的距離特徵；再以快取與點重新排序縮短實際計算時間。快速點特徵直方圖（FPFH）只計算每點與其鄰點的簡化直方圖（SPFH），再依距離加權鄰點的 SPFH，使複雜度由 O(n·k²) 降為 O(n·k)，並把三個特徵拆成獨立直方圖串接。SAC-IA 從特徵相似的候選點隨機抽取對應來求剛體轉換，以 Huber 懲罰評分，最後用 Levenberg-Marquardt 細化。在一組重疊約 45% 的 Ljubljana 都市戶外資料上，SAC-IA 以 1000 次迭代、10462 點在 34 秒內完成，先前的貪婪初始對齊只用 200 點就超過 17 分鐘（Table I）。論文沒有提供配準精度的量化數據，驗證場景僅 Stanford bunny 與這組戶外資料。",[],"point clouds with local descriptors","SAC-IA: select s sample points with pairwise distances above d_min, pick for each a random correspondence among points with similar histograms, compute the rigid transform and score it with a Huber penalty; repeat (1000 iterations in Sec. V) and keep the best transform, then refine with Levenberg-Marquardt non-linear optimization","FPFH(p) = SPFH(p) + (1\u002Fk) sum_i (1\u002Fw_i) SPFH(p_i): each point first gets a Simplified PFH from the three angular features (alpha, phi, theta) of a Darboux uvn frame between the point and its neighbours, then neighbour SPFHs are weighted by distance; the fourth (distance) feature of earlier PFH is dropped and the three features are binned as separate concatenated histograms rather than a 5x5x5 = 125-bin joint histogram; complexity O(n k) vs O(n k^2) for PFH; a multi-radius persistence analysis keeps salient points; SAC-IA matches each sample point to one of the points with similar histograms","initial rigid alignment",[],{"id":3604,"label":3605,"shortName":3606,"title":3607,"year":51,"era":10,"cluster":755,"scope":99,"keyIdeaZh":3608,"sensors":3609,"mapRepresentation":3610,"loopClosure":3611,"estimator":3612,"association":3613,"deskew":57,"outputGeometry":3614,"fulltextStatus":22,"lidarModels":3615,"equipmentCount":144},"pointslam2023","Sandström et al., 2023","Point-SLAM","Point-SLAM: Dense Neural Point Cloud-based SLAM","Point-SLAM 將神經特徵錨定在隨輸入逐步生成的點雲上，並依影像梯度動態調整點密度，細節處加密、平坦處稀疏；追蹤與建圖共用同一個以 RGB-D 重渲染誤差最佳化的點式表示。其網格評估在計算精確率與召回率前先以 ICP 對齊，因此量到的是局部形狀品質而非全域位置精度。",[772],"neural point cloud: features anchored at points with density adapted to image-gradient information","none (authors note a gap to traditional methods with loop closures)","gradient-based minimization of an RGB-D re-rendering loss for tracking and mapping, run as separate alternating processes; tracking initialised with a constant-speed assumption; mapping iterations adapted to the number of newly added points; optional exposure-compensation MLP for ScanNet","direct colour and depth re-rendering loss","mesh produced by rendering depth and colour every fifth frame along the estimated trajectory, TSDF Fusion at 1 cm voxels and marching cubes; evaluated with precision, recall and F-score at a 1 cm threshold after ICP alignment to the GT mesh; rendered depth and colour images",[],{"id":3617,"label":3618,"shortName":3619,"title":3620,"year":97,"era":10,"cluster":11,"scope":12,"keyIdeaZh":3621,"sensors":3622,"mapRepresentation":3624,"loopClosure":3625,"estimator":3626,"association":3627,"deskew":121,"outputGeometry":3628,"fulltextStatus":22,"lidarModels":3629,"equipmentCount":109},"schaub2022pc2bim","Schaub et al., 2022","Point cloud to BIM registration (SLAM tracking)","Point cloud to BIM registration for robot localization and Augmented Reality","作者以 Kudan LiDAR SLAM 追蹤 Ouster OS0-128 光達（含感測器 IMU 資料），將關鍵影格累積的點雲配準到以 IfcOpenShell 解析並體素化（0.1 m）的 BIM 點雲：先以法向量角度直方圖做軸向對齊，再以只保留垂直於投影面點的法向過濾樣板匹配（每 1° 測試）粗對位，最後以隨機子取樣 ICP 精對位，得到感測器在 BIM 座標中的位姿，供擴增實境檢查與 Boston Dynamics Spot 遠端操作使用。評估只用手持錄製資料，且以每段錄製起點人工量測的初始位置作為唯一參考：28 m 走廊（只含走廊的 BIM）各關鍵影格中位數平均 XY 誤差 0.03 m、Z 0.035 m，全部成功；TU Wien 圖書館六樓含多個相似房間的環形走廊，前 10 個關鍵影格成功率只有 30%，到第 36 個關鍵影格前平均 XY 0.19 m、Z 0.24 m，之後 XY 低於 0.3 m、Z 低於 0.4 m。誤差同時包含配準誤差與 SLAM 漂移。",[3623],"Ouster OS0-128 Gen 2 LiDAR (128 channels, 90 deg VFOV; 512x20 in env. 1, 1024x20 in env. 2; up to 131,072 points per frame) with IMU data from the sensor","keyframe-accumulated LiDAR point cloud from Kudan SLAM; BIM converted to a voxel point cloud (0.1 m) with IfcOpenShell and the Voxelization Toolkit for one manually selected floor (env. 2 floor: 1,035,690 points)","may occur inside Kudan SLAM, but neither loop-closure accuracy nor its influence on registration was evaluated","Kudan LiDAR SLAM (commercial SDK; tracking voxel 0.25 m) for pose tracking; point cloud to BIM registration of the accumulated keyframe map gives the SLAM-to-BIM transform, repeatable when drift grows","PCA normals and normal-angle histograms for axis alignment; normal-filtered template matching (points kept only when the dot product of point normal and plane normal is at most 0.1): rotation from XY projections tested at each 1 deg, first at lower resolution and then at the 0.1 m resolution, height matched with projections along the X and Y axes; ICP with random sub-sampling, 100 iterations split into 10 random 10% subsets","sensor pose in BIM frame",[222],{"id":3631,"label":3632,"shortName":3633,"title":3634,"year":150,"era":10,"cluster":397,"scope":32,"keyIdeaZh":3635,"sensors":3636,"mapRepresentation":3641,"loopClosure":57,"estimator":3642,"association":3643,"deskew":57,"outputGeometry":3644,"fulltextStatus":22,"lidarModels":3645,"equipmentCount":77},"schauer2018peopleremover","Schauer & Nuchter, 2018","Peopleremover","The Peopleremover, Removing Dynamic Objects From 3-D Point Cloud Data by Traversing a Voxel Occupancy Grid","本法以已配準的多站或多切片點雲建立全域體素網格，每個體素只記錄有哪些掃描在其中量到點。從每個感測器原點沿視線走訪到各量測點，若某體素被其他掃描看穿為空，體素內的點即判定為動態並移除；為避免斜掃表面與取樣不均造成誤判，以最近點的「陰影」與法向量限制走訪距離，再以叢集過濾孤立誤判，並可依掃描編號做次體素移除。演算法對掃描數與點數呈線性複雜度，只有體素大小一個參數。",[3637,3638,3639,3640],"3D laser scans only: Riegl VZ-400 TLS for the authors' lecturehall, campus and wrzburg datasets","third-party datasets of Underwood et al. (sim, lab, carpark; sensor not stated in this paper)","mobile-mapping scan slices from an automotive production line (authors' earlier work)","stated as compatible in principle with RADAR, RGB-D or stereo point clouds (not tested)","global regular voxel grid storing only the set of scan identifiers per voxel (no point coordinates); single parameter voxel size (0.1 to 0.6 m in Table I); C++ standard library containers","not_applicable (post-registration cleaning)","Amanatides-Woo voxel traversal (made stricter and free of floating-point accumulation) from each sensor origin to each point; traversal aborts at voxels holding the same scan identifier; search distances clipped by point 'shadows' using normals and per-scan sphere quadtrees; small clusters of free voxels reset to static; optional sub-voxel removal by scan identifier","cleaned point cloud without dynamic objects",[],{"id":3647,"label":3648,"shortName":3649,"title":3650,"year":51,"era":10,"cluster":599,"scope":32,"keyIdeaZh":3651,"sensors":3652,"mapRepresentation":3653,"loopClosure":57,"estimator":3654,"association":3655,"deskew":121,"outputGeometry":3656,"fulltextStatus":22,"lidarModels":3657,"equipmentCount":144},"dynablox2023","Schmid et al., 2023","Dynablox","Dynablox: Real-Time Detection of Diverse Dynamic Objects in Complex Environments","Dynablox 延伸 Voxblox 的雜湊區塊體素地圖，在機器人運作中逐步估計「高信心自由空間」，並同時建模感測雜訊與稀疏性、狀態估計漂移及地圖不完整；落入高信心自由空間的點即判定為移動點，再以其為種子擴張叢集。方法不假設物體外觀或類別，可偵測搬運物品的人、擺動的門等多樣動態物。",[35],"Voxblox hashed voxel blocks (16^3 voxels per block, voxel size 0.2 m) holding a TSDF; each voxel additionally stores the last occupied frame, the occupancy duration and a high-confidence free flag; voxels in dynamic clusters are overwritten rather than averaged at the next update (Sec. IV-A, IV-D)","not_applicable (uses external state estimate, e.g., FAST-LIO2 in new sequences)","a voxel is occupied if its TSDF distance is below 1.5 voxel sizes or a current point falls in it, with a sparsity compensation of 2 frames; it becomes high-confidence free only after 5 frames unoccupied for itself and all observed neighbours; points in or next to free voxels seed clusters grown over connected voxels, clusters under 20 voxels discarded; slowly re-occupied voxels are reset after τr frames derived from the expected drift rate (Sec. IV-C, IV-D)","per-point dynamic labels during online mapping; static volumetric map",[223,2787],{"id":3659,"label":3660,"shortName":3661,"title":3662,"year":9,"era":10,"cluster":236,"scope":132,"keyIdeaZh":3663,"sensors":3664,"mapRepresentation":3665,"loopClosure":3666,"estimator":3667,"association":3668,"deskew":3669,"outputGeometry":3670,"fulltextStatus":22,"lidarModels":3671,"equipmentCount":1289},"badslam2019","Schöps et al., 2019","BAD SLAM","BAD SLAM: Bundle Adjusted Direct RGB-D SLAM","BAD SLAM 提出可即時執行的直接式 BA，以面元表示地圖，同時使用深度的幾何約束與影像梯度的光度約束，並交替最佳化地圖與相機位姿。作者另建立以同步全域快門（global shutter）RGB 與深度相機錄製、經精確校正的 ETH3D SLAM 基準，指出直接式 RGB-D SLAM 對捲簾快門、RGB 與深度不同步及校正誤差高度敏感。在此基準上方法排名與既有資料集不同，顯示資料集設定本身會左右比較結論。",[772],"surfels (oriented discs with centre, normal, radius and a scalar gradient descriptor) created on 4x4 pixel cells of each keyframe, plus keyframes storing raw RGB-D; about 335,000 surfels for the Fig. 1 scene","bag-of-words detection with binary features; relative pose from keypoint matches refined by direct alignment and checked for consistency against neighbouring keyframes m-1 and m+1; then pose-graph optimisation (Sec. 4; Fig. 2)","Alternating direct BA (Alg. 1): Gauss-Newton on a cost of Tukey-weighted point-to-plane residuals and Huber-weighted photometric gradient residuals (photometric weight 1e-2), each normalised by a stereo depth-noise model or an empirical sigma; surfels move only along their normals (joint 2x2 solve of offset and descriptor per surfel); independent per-keyframe SE(3) pose updates; optional intrinsics and per-pixel depth-deformation calibration solved with the Schur complement; interleaved discrete surfel creation, merging, outlier deletion and radius update. A PCG Gauss-Newton solver was slightly worse in the ablation.","Surfel to pixel correspondences by projecting surfel centres into every keyframe, kept only if the pixel has depth, the Tukey weight of the geometric residual is positive and normals agree; photometric residual compares the surfel descriptor with the gradient magnitude sampled at the surfel centre and two disc boundary points. Front-end odometry: direct photometric and geometric SE(3) alignment of each frame to the last keyframe using intensity gradients.","not_applicable (rolling shutter avoided by global-shutter, synchronised benchmark cameras rather than modelled)","surfel map and keyframe trajectory",[],{"id":3673,"label":3674,"shortName":3675,"title":3676,"year":396,"era":10,"cluster":236,"scope":32,"keyIdeaZh":3677,"sensors":3678,"mapRepresentation":3680,"loopClosure":3681,"estimator":3682,"association":3683,"deskew":684,"outputGeometry":3684,"fulltextStatus":22,"lidarModels":3685,"equipmentCount":144},"surfelmeshing2020","Schöps et al., 2020","SurfelMeshing","SurfelMeshing: Online Surfel-Based Mesh Reconstruction","SurfelMeshing 假設相機已校正且位姿由外部 SLAM 提供，不把深度融合進體素體積，而是融合成稠密面元（surfel）雲，再在背景非同步地對平滑後的面元做局部三角化，產生頂點即為面元的網格。作者在 ElasticFusion 式的面元重建上加入兩個去雜訊步驟：沿法向與鄰近面元的正則化，以及在觀測邊界漸進混合深度差以避免表面斷裂；並提出只在失效三角形附近重新三角化的增量演算法，使網格能隨迴圈閉合造成的面元變形快速更新。由於面元依輸入影像解析度建立，網格與色彩解析度會隨觀測距離調整，也能重建體素法難以保留的薄物體。",[3679],"RGB-D camera (mainly Microsoft Kinect v1 sequences of the TUM RGB-D benchmark; pre-registered CoRBS sequences; synthetic ICL-NUIM depth with simulated noise)","dense surfel cloud at input image resolution (position, normal, colour, confidence, radius, creation and update timestamps, denoised position, four neighbours) indexed by a lazily updated compressed octree; a triangle mesh whose vertices are the surfels","Adopts ElasticFusion's loop-closure handling: the surfel cloud is deformed based on surfel timestamps, offsets are averaged among neighbours for 100 iterations, and affected mesh regions are remeshed; the public code excludes loop closure","No pose estimation of its own: calibrated camera and poses from an external SLAM system (ElasticFusion in the implementation; BAD SLAM poses for the ETH3D examples)","Projective association of each surfel with the pixel it projects to and the nearest neighbouring pixel; surfels classified as conflicting, occluded or supported using a depth uncertainty interval of plus or minus 5% of the measured depth and normal checks","coloured triangle mesh with vertices at the surfels, updated online; not guaranteed manifold",[],{"id":3687,"label":3688,"shortName":3689,"title":3690,"year":150,"era":10,"cluster":236,"scope":99,"keyIdeaZh":3691,"sensors":3692,"mapRepresentation":3694,"loopClosure":3695,"estimator":3696,"association":3697,"deskew":684,"outputGeometry":3698,"fulltextStatus":22,"lidarModels":3699,"equipmentCount":185},"staticfusion2018","Scona et al., 2018","StaticFusion","StaticFusion: Background Reconstruction for Dense RGB-D SLAM in Dynamic Environments","StaticFusion 是針對動態環境的 RGB-D 稠密 SLAM，同時估計相機運動與影像中哪些區域靜止。每張影像先以 K-means 依三維座標分成幾何群集，再與由靜態面元地圖渲染出的預測影像做光度與幾何直接對齊；每個群集有一個 0 至 1 的靜態分數，用來加權其殘差，分數則由殘差門檻、相鄰群集平滑與深度差先驗共同決定，並以 IRLS 交替求解位姿與分數。融合時只寫入靜態部分，每個面元以對數勝算累積被靜態點重複觀測的可信度，持續被動態點匹配的面元會被移除，因此地圖只保留背景結構，並把這個背景作為下一影格分割所需的時間資訊。",[3693],"RGB-D camera, registered RGB-D images at QVGA 320x240 (TUM Freiburg sequences and two hand-held recordings; hand-held camera model not named)","surfel map (Keller et al. model through the ElasticFusion implementation) in which each surfel carries a viability value accumulated as log-odds of matches with static input points; surfels with viability below 0.5 for more than 10 consecutive frames are removed and free-space violations are cleaned","none described","Joint minimization over the camera twist and per-cluster static scores b in [0, 1]: Cauchy-robust photometric and geometric residuals between the current RGB-D frame and a prediction rendered from the static surfel map, weighted by b, plus a residual-threshold term, spatial regularization between contiguous clusters and a depth-difference prior; IRLS for the twist with closed-form b after each iteration, coarse-to-fine","Dense direct alignment (warping current pixels into the rendered model prediction); scene split into K geometric clusters by K-means on 3D coordinates; per-pixel segmentation derived from cluster scores","static-background surfel map (coloured for visualization), camera trajectory and per-frame static\u002Fdynamic segmentation",[],{"id":3701,"label":3702,"shortName":3703,"title":3704,"year":345,"era":52,"cluster":68,"scope":69,"keyIdeaZh":3705,"sensors":3706,"mapRepresentation":3710,"loopClosure":72,"estimator":3711,"association":3712,"deskew":121,"outputGeometry":183,"fulltextStatus":22,"lidarModels":3713,"equipmentCount":24},"segal2009gicp","Segal et al., 2009","GICP","Generalized-ICP","GICP 將點對點與點對平面 ICP 納入同一機率框架：兩片點雲的每個點都被視為來自高斯分布，最小化步驟以最大概似估計計算位姿。作者依局部平面假設，令每點沿表面法向量的共變異數很小、沿平面方向很大，形成「平面對平面」配準；點對點與點對平面皆可視為其特例。對應點仍以歐氏距離與 kd-tree 搜尋，因此保留 ICP 的速度與簡潔。實驗使用射線追蹤模擬的 SICK 旋轉掃描（室內走廊與建物周邊戶外，1 cm 雜訊）及車頂 Velodyne 的郊區環線記錄；Velodyne 測試掃描對彼此相距 15 至 20 m 以上，各方法的初始誤差在每軸 ±1.5 m 與 ±15° 內隨機產生。真實資料的參考位姿來自結合 GPS 與 IMU 的成對約束 SLAM，作者自承並非完美的真值。論文結果只以圖呈現，沒有數值表。",[3707,3708,3709],"[\"3D LiDAR (simulated SICK scanner on a rotating joint","roof-mounted Velodyne on an instrumented car)\", \"GPS and IMU (car logs","used only to build ground truth)\"]","point clouds with per-point covariance from PCA of 20 nearest neighbours (Sec. III-B)","maximum-likelihood estimate over Gaussian point models; minimized with conjugate gradients in the experiments (Sec. III-IV)","Euclidean nearest neighbour via kd-tree with max match distance d_max; plane-to-plane covariance weighting (Sec. III)",[717,3714],"Velodyne range finder",{"id":3716,"label":3717,"shortName":3718,"title":3719,"year":150,"era":10,"cluster":151,"scope":132,"keyIdeaZh":3720,"sensors":3721,"mapRepresentation":3724,"loopClosure":3725,"estimator":3726,"association":3727,"deskew":3728,"outputGeometry":3729,"fulltextStatus":22,"lidarModels":3730,"equipmentCount":109},"legoloam2018","Shan & Englot, 2018","LeGO-LOAM","LeGO-LOAM: Lightweight and Ground-Optimized Lidar Odometry and Mapping on Variable Terrain","LeGO-LOAM 針對地面載具，先把點雲投影為距離影像（range image），分離地面點，並以影像式分割剔除少於 30 點的小群集（如樹葉）；邊緣特徵只取自非地面點，以避開草地造成的不穩定特徵，再依 LOAM 的粗糙度指標擷取邊緣與平面特徵。其兩步驟 Levenberg-Marquardt 最佳化先以地面平面特徵求 [tz, roll, pitch]，再以邊緣特徵求 [tx, ty, yaw]，以降低計算量。地圖改為儲存每次掃描的特徵集合與對應位姿，並可選擇性地接上以 ICP 建立迴圈約束、iSAM2 最佳化的位姿圖（pose graph）。",[3722,3723],"3D LiDAR (Velodyne VLP-16; HDL-64E via KITTI)","IMU (low-cost CH Robotics UM6, used only for initial guess)","per-scan edge\u002Fplanar feature sets stored with sensor poses; local map assembled from sets within 100 m or the k most recent sets (Sec. III-E)","optional: ICP between current and earlier feature sets adds pose-graph constraints, optimized by iSAM2; used only in the KITTI seq. 00 test (Sec. III-E, IV-D)","two-step Levenberg-Marquardt: ground planar features estimate [tz, roll, pitch], then edge features estimate [tx, ty, yaw]; optional pose graph optimized with iSAM2 (Sec. III-D, III-E, IV-D)","range-image ground separation and image-based segmentation (clusters \u003C 30 points discarded); LOAM-style roughness-based edge\u002Fplanar features; label-consistent point-to-edge \u002F point-to-plane matching (Sec. III-B to III-D)","not_reported in the paper (feature and matching details deferred to LOAM [20]); IMU provides the initial guess (Sec. IV-A)","feature-set point cloud map and 6-DoF poses (Sec. III-E); export of full-resolution map not described in paper",[143,161],{"id":3732,"label":3733,"shortName":3734,"title":3735,"year":396,"era":10,"cluster":151,"scope":132,"keyIdeaZh":3736,"sensors":3737,"mapRepresentation":3740,"loopClosure":3741,"estimator":3742,"association":3743,"deskew":3744,"outputGeometry":3745,"fulltextStatus":22,"lidarModels":3746,"equipmentCount":487},"liosam2020","Shan et al., 2020","LIO-SAM","LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping","LIO-SAM 把 LiDAR 慣性里程計建構在因子圖（factor graph）上，以 iSAM2 增量最佳化 IMU 預積分、LiDAR 里程計、GNSS 與迴圈閉合四種因子，形成緊耦合（tightly-coupled）系統。IMU 積分的運動用來對點雲去畸變並提供掃描配準初值，LiDAR 里程計結果再回饋估計 IMU 偏差。為維持即時性，新關鍵影格只與固定數量的近期子關鍵影格（sub-keyframes）組成的局部體素地圖配準，而非與整張全域地圖配準；迴圈以歐氏距離搜尋候選並以掃描配準建立約束。",[1201,3738,3739],"IMU (MicroStrain 3DM-GX5-25)","GNSS (Reach M, optional)","keyframe edge\u002Fplanar feature clouds; local voxel maps downsampled at 0.2 m (edge) and 0.4 m (planar) (Sec. III-C)","Euclidean-distance candidate search within 15 m, scan matching of the new keyframe to +\u002F-12 sub-keyframes around the candidate, added as a loop factor; compatible with descriptor-based place recognition (Sec. III-E)","factor graph smoothing with iSAM2: IMU preintegration, lidar odometry, GPS and loop-closure factors; scan matching solved by Gauss-Newton (Sec. III)","LOAM-style edge\u002Fplanar features matched to a local voxel map built from n=25 recent sub-keyframes (point-to-edge \u002F point-to-plane) (Sec. III-C)","IMU-estimated nonlinear motion de-skews each scan and gives the initial guess for scan matching (Sec. I, abstract)","global feature map assembled from keyframe edge and planar feature clouds plus the keyframe trajectory; lidar frames between keyframes (1 m or 10 deg pose change) are discarded; maps are shown aligned with Google Earth imagery (Figs. 4 to 7); dense map export not described",[143],{"id":3748,"label":3749,"shortName":3750,"title":3751,"year":249,"era":10,"cluster":263,"scope":132,"keyIdeaZh":3752,"sensors":3753,"mapRepresentation":3756,"loopClosure":3757,"estimator":3758,"association":3759,"deskew":3760,"outputGeometry":3761,"fulltextStatus":22,"lidarModels":3762,"equipmentCount":487},"lvisam2021","Shan et al., 2021","LVI-SAM","LVI-SAM: Tightly-coupled Lidar-Visual-Inertial Odometry via Smoothing and Mapping","LVI-SAM 以因子圖（factor graph）為核心，將視覺慣性子系統（VIS）與光達慣性子系統（LIS）緊密耦合：LIS 提供位姿與 IMU 偏差協助 VIS 初始化，VIS 的視覺里程計則作為光達掃描配準（scan matching）的初始猜測。視覺特徵可由累積的光達點取得深度，迴圈（loop closure）候選先由 DBoW2 視覺詞袋產生，再以光達配準精修後加入 iSAM2 全域最佳化。任一子系統偵測到失效時可被繞過，以提升在無紋理或幾何退化場景的穩健性。",[1201,3738,3754,3755],"monocular camera (FLIR BFS-U3-04S2M-CS)","GPS (Reach RS+, ground-truth reference only)","global LiDAR feature map built from keyframes; sparse visual landmarks in the VIS","DBoW2 with BRIEF descriptors proposes candidates in the VIS; candidates refined by LiDAR scan matching in the LIS (Sec. II-B4, II-C)","factor graph smoothing with iSAM2 in the lidar-inertial subsystem (IMU preintegration, visual odometry, lidar odometry and loop factors); VINS-Mono-style sliding-window bundle adjustment in the visual-inertial subsystem","LiDAR edge and planar features matched scan-to-map against a sliding-window keyframe feature map (adapted from LIO-SAM); KLT-tracked corner features with depth from stacked LiDAR points (3 nearest points on a unit sphere, plane intersection, 2 m consistency check)","LiDAR point clouds de-skewed with IMU measurements before feature extraction (Fig. 1, Sec. I)","not_reported (paper evaluates trajectories; no colored map or export format described)",[143],{"id":3764,"label":3765,"shortName":3766,"title":3767,"year":9,"era":10,"cluster":263,"scope":132,"keyIdeaZh":3768,"sensors":3769,"mapRepresentation":3773,"loopClosure":3774,"estimator":3775,"association":3776,"deskew":3777,"outputGeometry":3778,"fulltextStatus":22,"lidarModels":3779,"equipmentCount":487},"vilslam2019","Shao et al., 2019","VIL-SLAM","Stereo Visual Inertial LiDAR Simultaneous Localization and Mapping","VIL-SLAM 把三個模組串接：緊耦合的雙目視覺慣性里程計以固定滯後的位姿圖平滑器估計運動，並以 IMU 頻率輸出位姿；LiDAR 建圖模組用這些位姿為每個點去畸變，再以 LOAM 式邊緣與平面特徵做掃描對地圖配準；迴圈閉合先以視覺詞袋偵測候選並用 EPnP 求初始約束，再以稀疏 LiDAR 特徵點的 ICP 精修，最後用 iSAM2 增量最佳化全域位姿圖，並把修正後的位姿即時回饋給建圖模組重新定位。作者以 Faro 地面雷射掃描為參考評估地圖，並在隧道與走廊等 LiDAR 退化場景顯示視覺慣性先驗的幫助。",[3770,3771,3772],"stereo camera pair (two megapixel cameras; model not reported)","16 scan-line 3D LiDAR (model not reported)","IMU at 400 Hz (model not reported)","sparse LiDAR feature map (all previous edge and surface feature points) for registration; dense map produced in post-processing by stitching dewarped scans with the best estimated poses, reported as 1 cm voxel dense maps near real time (Sec. III; abstract)","yes; visual Bag-of-Words detection (DBoW3) of key images within a time threshold, descriptor matching to reject false positives, EPnP initial constraint, then ICP refinement on sparse LiDAR feature points of the key scans (LibPointMatcher) (Sec. VII-A, VII-B)","loosely coupled chain: tightly coupled stereo VIO as a fixed-lag smoother over the most recent N stereo frames (IMU pre-integration factors and structureless vision factors, Levenberg-Marquardt, Schur-complement marginalization, GTSAM) provides IMU-rate motion priors to LOAM-style LiDAR mapping; a global pose graph of LiDAR mapping poses with LiDAR odometry and loop constraint factors is optimized incrementally with iSAM2 (Secs. V-VII)","visual: KLT tracking of stereo matches, Shi-Tomasi corners with ORB descriptors and brute-force stereo matching; LiDAR: edge and planar feature points registered scan-to-map by point-to-line (two closest edge points) and point-to-plane (three closest surface points) distances as in LOAM (Secs. IV, VI-B)","each LiDAR point dewarped to the end-of-scan time using the closest IMU-rate VIO poses (Sec. VI-A, Eq. 6)","loop-closure-corrected 6-DoF LiDAR poses in real time and a dense point-cloud map near real time (abstract; Fig. 5)",[45],{"id":3781,"label":3782,"shortName":3783,"title":3784,"year":30,"era":10,"cluster":263,"scope":676,"keyIdeaZh":3785,"sensors":3786,"mapRepresentation":3791,"loopClosure":1451,"estimator":3792,"association":3793,"deskew":121,"outputGeometry":3794,"fulltextStatus":22,"lidarModels":3795,"equipmentCount":487},"litgs2026","Shi et al., 2026","LIT-GS","LIT-GS: LiDAR-Inertial-Thermal Gaussian Splatting for Illumination-Robust Mapping","LIT-GS 以熱影像取代易受光照影響的 RGB 光度監督，建立光達、慣性與熱影像的高斯潑濺（Gaussian Splatting）地圖。它以上游 FAST-LIVO2 的具不確定度視覺地圖點作為跨模態錨點建立熱影像與光達的對應，並把加權的光達點到平面殘差放入光束法平差，再於高斯最佳化中加入光達平面正則化，以抑制表面增厚與結構漂移。此方法為離線流程，作者僅報告 Car 場景的訓練時間為 70.6 分鐘。",[3787,3788,3789,3790],"3D LiDAR (Livox Avia, 10 Hz per Fig. 2)","IMU (built into the Livox Avia, 200 Hz per Fig. 2)","visible-light camera (MV-CA013-21UC)","long-wave thermal imager (MV-CI003-GL-N15, 10 Hz per Fig. 2)","3D Gaussians (thermal), initialized from LiDAR voxel map and refined structure","upstream FAST-LIVO2 LIV estimate, then offline LiDAR-plane-constrained bundle adjustment (extension of COLMAP-PCD) and differentiable Gaussian optimization","learned thermal feature matching; uncertainty-tagged LIV visual map points as cross-modal anchors; weighted LiDAR point-to-plane residuals in BA and as a splatting regularizer","thermal Gaussian map with rendered images; geometric consistency evaluated with Earth Mover's Distance to reference point clouds (Sec. IV-B2)",[437],{"id":3797,"label":3798,"shortName":3799,"title":3800,"year":332,"era":52,"cluster":53,"scope":54,"keyIdeaZh":3801,"sensors":3802,"mapRepresentation":3803,"loopClosure":3804,"estimator":3805,"association":3806,"deskew":57,"outputGeometry":3807,"fulltextStatus":22,"lidarModels":3808,"equipmentCount":185},"sibley2010swf","Sibley et al., 2010","Sliding window filter","Sliding window filter with application to planetary landing","本文以延遲狀態邊際化（delayed state marginalization）提出滑動視窗濾波器（SWF），用於提升行星著陸時長距離立體視覺的地表結構估計精度。方法在 k 個位姿的視窗內以含 Huber 核的穩健 Gauss-Newton 同時最佳化位姿與地標，並以 Schur 補把最舊位姿與不再被觀測的地標邊際化為先驗資訊；視窗涵蓋全部時間時等同完整 BA，只保留一個時間步時等同 EKF 的時間更新，逐步邊際化時則成為固定時間的方法。作者在實驗室以兩台 Point Grey Flea 相機組成的立體相機朝貼有 HiRISE 影像的平面牆移動，以 1:10 比例模擬著陸，結果顯示 3 至 5 影格的視窗已接近批次解；作者另指出在 10 個影格內，SWF 的誤差比視覺里程計低約 76%。但過早邊際化會鎖住線性化誤差，且邊際化使迴圈閉合難以處理。",[2101],"sparse 3D point landmarks (tracked surface features) with a possibly dense prior information block created by marginalization","not handled: the authors state that marginalization makes re-observing landmarks difficult, which effectively precludes loop closures (not an issue for descent and landing)","sliding window filter: robust Gauss-Newton (Huber kernel, typically 4 to 10 iterations) over all measurements of a k-pose window with a kinematic process model and a prior information term; the oldest poses and landmarks without active support are marginalized by the Schur complement into the prior (k = 1 reproduces the first-order EKF time step, a window over all time equals full BA or full SLAM)","sum-of-absolute-differences patch matching of Harris corners, Lucas-Kanade subpixel refinement with projective patch warping (local plane with normal along the first optical axis), Moravec's rigid-consistency check (greedy maximal clique) against gross outliers, then Huber M-estimation over the whole window","sparse 3D landmark positions on the target surface; map error evaluated as the shortest distance from each landmark to a plane fitted to the wall (plane from the batch solution with fiducials)",[],{"id":3810,"label":3811,"shortName":3812,"title":3813,"year":3814,"era":52,"cluster":527,"scope":54,"keyIdeaZh":3815,"sensors":3816,"mapRepresentation":3817,"loopClosure":3818,"estimator":3819,"association":3820,"deskew":57,"outputGeometry":3821,"fulltextStatus":22,"lidarModels":3822,"equipmentCount":61},"smith_cheeseman1986","Smith & Cheeseman, 1986","Smith-Cheeseman spatial uncertainty","On the Representation and Estimation of Spatial Uncertainty",1986,"本文以「近似轉換（approximate transformation, AT）」表示座標框架之間不確定的相對位姿，每個 AT 由平均關係與共變異數矩陣組成。作者定義兩個基本運算：串接（compounding）以一階泰勒展開與 3×6 雅可比矩陣傳遞共變異數，把一連串 AT 合成一個，不確定性隨之變大；合併（merging）以靜態卡爾曼濾波公式加權平均平行的 AT，使不確定性變小，並以電阻串並聯作類比。對惠斯登電橋這類無法以串並聯化簡的網路，作者提出刪除形成迴路的 AT（未用上全部資訊，並非最佳）或以 Delta-Y 轉換改寫網路（可用上全部資訊，但無法化簡所有網路），並表示以共同參考框架做遞迴狀態估計的通用方法仍在研究中。移動機器人範例與蒙地卡羅模擬顯示，平均值與共變異數的相對誤差通常小於 1%，但角度誤差大時分布呈新月形而非高斯分布。",[],"network of approximate transformations (relational map), each with a mean relation and covariance; the robot keeps the original ATs from motions and sensings and computes composite ATs on demand","implicit: when the robot observes its start frame from a later pose, the sensed AT is merged with the compounded chain as a parallel relation; no separate loop-closure module","first-order (linearized) mean and covariance propagation for compounding and reversal of ATs; merging of parallel ATs with static-state Kalman filter equations (optimal for Gaussian variables and linear mappings, optimal-linear otherwise); extended Kalman filter update named for nonlinear coordinate mappings","no data-association algorithm; the sensing procedure rejects an observation whose probability, given the prior AT estimate and the sensor error, is below a threshold (e.g., the camera viewed the wrong object), which the authors also describe as detecting sensor glitches","mean and covariance of the relative pose (x, y, theta) between any two frames; confidence ellipses derived from the covariance",[],{"id":3824,"label":3825,"shortName":3826,"title":3827,"year":3828,"era":52,"cluster":527,"scope":54,"keyIdeaZh":3829,"sensors":3830,"mapRepresentation":3831,"loopClosure":3832,"estimator":3833,"association":3834,"deskew":57,"outputGeometry":3835,"fulltextStatus":22,"lidarModels":3836,"equipmentCount":61},"smith_self_cheeseman1990","Smith et al., 1990","Stochastic map","Estimating Uncertain Spatial Relationships in Robotics",1990,"本文提出「隨機地圖（stochastic map）」：把機器人與各物件之間的空間關係組成一個狀態向量，同時保存其平均值與完整共變異數矩陣，以描述關係之間的相依性。作者以一階線性化推導位姿複合（compounding）與反轉運算的共變異數傳遞，並以（擴展）卡爾曼濾波器（Kalman filter）在新量測加入時遞迴更新整張地圖。文中明示兩項前提：角度誤差須夠小以支持線性化，且只估計前兩階動差即足以支援決策（Sec. 6）。",[],"stochastic map: vector of object\u002Frobot frame relations with full covariance matrix","implicit: re-sensing a previously mapped object is incorporated as a constraint that reduces the uncertainty of the robot and all correlated objects (Sec. 2.3 example); no separate loop-closure module","extended Kalman filter on a joint state of robot and object frames (iterated EKF also given)","no dedicated data-association algorithm; the running example uses the stochastic map to decide that a newly sensed object cannot be the previously mapped object #1 (Sec. 2.3), Sec. 6 mentions ignoring sensor results that are too improbable, and the developed example assumes the sensor identifies the re-observed object as object #1, noting that in practice the new object would first be compared with the old ones (Sec. 5, Step 4)","mean and covariance of relative spatial relationships between frames (2D in text; 6-DoF Jacobians in Appendix A)",[],{"id":3838,"label":3839,"shortName":3840,"title":3841,"year":1217,"era":52,"cluster":397,"scope":1035,"keyIdeaZh":3842,"sensors":3843,"mapRepresentation":57,"loopClosure":57,"estimator":3846,"association":57,"deskew":57,"outputGeometry":3847,"fulltextStatus":22,"lidarModels":3848,"equipmentCount":46},"soudarissanane2011scanninggeometry","Soudarissanane et al., 2011","TLS scanning geometry","Scanning geometry: Influencing factor on the quality of terrestrial laser scanning points","作者由簡化的雷達距離方程式推導，指出地面雷射掃描的訊噪比隨入射角餘弦與距離平方下降，並提出入射角係數 cos α 與距離係數；以總體最小平方擬合平面後，把沿雷射束方向的殘差換算為垂直於平面的殘差，藉此分離掃描幾何對單點雜訊的貢獻。實驗先以 Leica HDS6000 掃描 1 m 見方的白色合板（固定 20 m 並旋轉 0° 至 80°，再於 5 m 至 50 m 重複），再以 FARO LS880 HE 從房間中央與角落兩站掃描空房間。房間平均標準差由 3.23 mm 降為去除入射角效應後的 2.55 mm，約 20% 的雜訊來自非零入射角；作者據此建議先以 CAD 圖與初步低解析度掃描評估各測站的入射角與距離，再規劃測站位置。",[3844,3845],"terrestrial laser scanner Leica HDS6000 (reference board experiments)","terrestrial laser scanner FARO LS880 HE (room experiment, 1\u002F4 of full resolution)","total least squares plane fitting per segment or per 5°x5° spherical patch; beam-direction residuals converted to orthogonal residuals with the incidence-angle coefficient c_I(α) = cos α and a range coefficient c_R(ρ) derived from a simplified radar range equation under a Lambertian assumption","per-point noise levels (beam direction and orthogonal) and per-patch standard deviation maps shown as net-views; no new point cloud product",[],{"id":3850,"label":3851,"shortName":3852,"title":3853,"year":332,"era":52,"cluster":527,"scope":99,"keyIdeaZh":3854,"sensors":3855,"mapRepresentation":3859,"loopClosure":3860,"estimator":3861,"association":3862,"deskew":3863,"outputGeometry":3864,"fulltextStatus":22,"lidarModels":3865,"equipmentCount":144},"tinyslam2010","Steux & Hamzaoui, 2010","tinySLAM (CoreSLAM)","tinySLAM: A SLAM algorithm in less than 200 lines C-language program","tinySLAM 以少於 200 行 C 程式實作雷射 SLAM，核心只有兩個函式：計算掃描與地圖的距離，以及更新地圖。地圖是 1 cm 解析度的網格，每個障礙點不是畫成單一點，而是以修改的 Bresenham 演算法畫出以障礙為尖端的「洞」形函數，使匹配較容易收斂；距離函數則直接加總轉換後掃描端點所在格值。單機版以簡單的蒙地卡羅搜尋做掃描對地圖匹配；也可把距離函數當作粒子濾波的似然函數，以處理歧義、重新定位與里程打滑。",[3856,3857,3858],"2D laser scanner (Hokuyo URG-04LX)","wheel odometry (two free odometry wheels with 2000-point encoders)","GPS and compass optionally fused in the particle filter (the compass without good success)","2D grid of 2048 x 2048 16-bit cells at 1 cm per cell; each obstacle hit is drawn as a 'hole' function with its tip at the obstacle using a modified Bresenham ray update and an integration-speed (quality) parameter (Sec. IV; Algorithms 1, 3, 4)","none explicit","stand-alone: simple Monte Carlo search of the pose that best matches the scan to the map; or particle filter in which the scan-to-map distance is each particle's likelihood, with a slippage model (10% of particles stay in place with high noise) (Sec. IV)","no explicit correspondences: the scan-to-map distance sums the map values under the transformed scan endpoints (Algorithm 2)","each scan corrected with a constant longitudinal and rotational speed during the sweep (Sec. III)","2D grid map and robot trajectory",[1576],{"id":3867,"label":3868,"shortName":3869,"title":3870,"year":363,"era":52,"cluster":68,"scope":69,"keyIdeaZh":3871,"sensors":3872,"mapRepresentation":3874,"loopClosure":72,"estimator":3875,"association":3876,"deskew":121,"outputGeometry":3877,"fulltextStatus":22,"lidarModels":3878,"equipmentCount":185},"stoyanov2012d2dndt","Stoyanov et al., 2012","D2D-NDT","Fast and accurate scan registration through minimization of the distance between compact 3D NDT representations","此研究把固定與移動兩片掃描都轉成三維常態分布轉換（3D-NDT）模型，也就是在規則網格的每個格子以一個高斯分布描述局部表面，再直接最小化兩個模型之間的 L2 距離（分布對分布，D2D），不像點對分布（P2D）或 ICP 那樣逐點計算。目標函數只在兩模型彼此最近的高斯成分之間評估，具有解析梯度與 Hessian，以牛頓法搭配 More-Thuente 線搜尋求解，並在 4、2、1、0.5 m 的網格上逐層配準；每次迭代直接轉換整個 NDT 模型並累積齊次轉換矩陣，以處理 SE(3) 的結構。作者另以 3D-NDT 直方圖對齊主要平面法向量來估計初始旋轉，並依 Censi 的方法推導封閉形式共變異數。在 AASS 室內迴圈與 Hannover2 戶外資料上，D2D 的精度與 ICP、P2D 相當，在 AASS 上快將近一個數量級；模擬中直方圖初始化平均約 150 ms，FPFH 約 15 秒。結果多以箱形圖呈現且未報告硬體；在稀疏的戶外資料上各方法都有大量失敗，共變異數估計也偏大。",[3873],"[\"rotating SICK laser (AASS loop and Hannover2 data sets)\", \"simulated 3D range sensor in ROS\u002FGazebo with SICK LMS 200 error models, 180 x 120 deg field of view\"]","3D-NDT: one Gaussian (mean and covariance) per occupied cell of a regular grid, built for both scans at cell sizes 4, 2, 1 and 0.5 m; about 1,500 components at 0.5 m for a typical AASS scan","minimizes an L2 distance between the fixed and moving 3D-NDT models, written as a sum of negative Gaussian terms over component pairs (Eq. 18); analytic gradient and Hessian; Newton's method with More-Thuente line search; parameters d1 = 1, d2 = 0.05; optimization repeated at grid sizes of 4, 2, 1 and 0.5 m; each pose increment is applied by transforming the whole moving NDT model and increments are accumulated in a homogeneous matrix so derivatives are always taken at the zero pose","distribution-to-distribution: each Gaussian component of the moving 3D-NDT is paired only with the closest component of the fixed 3D-NDT; no point-level correspondences; baselines ICP (PCL) and NDT-P2D run on point sets sub-sampled on a 0.1 m grid","6-DoF rigid transformation (homogeneous matrix) plus a closed-form covariance derived after Censi (2007) with the NDT component covariances as the measurements; in the simulation test the estimate over-estimated the sample covariance, which the authors consider acceptable",[3879,2744],"rotating SICK laser",{"id":3881,"label":3882,"shortName":3883,"title":3884,"year":115,"era":52,"cluster":236,"scope":132,"keyIdeaZh":3885,"sensors":3886,"mapRepresentation":3888,"loopClosure":3889,"estimator":3890,"association":3891,"deskew":684,"outputGeometry":3892,"fulltextStatus":22,"lidarModels":3893,"equipmentCount":46},"mrsmap2014","Stückler & Behnke, 2014","MRSMap","Multi-resolution surfel maps for efficient dense 3D modeling and tracking","MRSMap 把每張 RGB-D 影像轉成八元樹多解析度面元地圖：各層節點都以單次掃描累加的充分統計量，保存點位置與 Lαβ 色彩的六維常態分布，並依最多六個觀測方向分開保存面元；最細解析度隨深度平方放寬，以反映 RGB-D 深度雜訊。配準時由最細解析度開始，在鄰域中尋找形狀與紋理描述子相符的面元配對，先以 Levenberg-Marquardt 初始化、再以牛頓法最大化配對似然，並以閉式近似估計位姿共變異。SLAM 以關鍵視角為節點，每一影格隨機抽選一個鄰近關鍵視角嘗試配準以發現迴圈，再以 g2o 最佳化位姿圖，全程在 CPU 上即時執行，同一方法也用於物件建模與追蹤。",[3887],"RGB-D camera at VGA 640x480 and 30 Hz (TUM Freiburg benchmark sequences and the authors' object dataset; sensor model not named); QVGA used on the robot","octree multi-resolution surfel map: every node stores sufficient statistics of a 6D Gaussian of position and L-alpha-beta colour, up to six surfels per node for orthogonal view directions; finest node size adapted to squared depth (0.0125 m limit); border and occluded-background surfels excluded","Randomized hypothesis-and-test: per frame one key view, sampled with probability decreasing with distance and angle from the reference, is registered; the constraint is accepted if its bidirectional matching likelihood is at least a fraction of that of the key view's initial constraint","Key-view SLAM: each frame is registered to the current reference key view by maximizing the surfel-match likelihood (approximate Levenberg-Marquardt initialization, then Newton's method with trilinear interpolation, typically 10 to 20 LM and 5 Newton iterations); key views linked by relative-pose constraints with closed-form covariance; pose graph solved by sparse Cholesky in g2o, one iteration per frame","Multi-resolution surfel association starting at the finest resolution with a cubic volume query around the transformed surfel mean (side twice the node resolution), bootstrapped from previous associations via the 26-neighbourhood; accepted only if shape-texture descriptors (surfel-pair angle histograms and luminance and chrominance contrasts) differ by at most 0.1 and contour flags agree","multi-view multi-resolution surfel map of a scene or an object model (visualized by sampling the surfel distributions; object models about 54 MB for a chair and 19 MB for a humanoid)",[],{"id":3895,"label":3896,"shortName":3897,"title":3898,"year":207,"era":10,"cluster":11,"scope":12,"keyIdeaZh":3899,"sensors":3900,"mapRepresentation":3904,"loopClosure":3905,"estimator":3906,"association":3907,"deskew":3908,"outputGeometry":3909,"fulltextStatus":22,"lidarModels":3910,"equipmentCount":1289},"stuhrenberg2025liobim","Stührenberg & Smarsly, 2025","LIO-BIM","LIO-BIM - Coupling lidar inertial odometry with building information modeling for robot localization and mapping","作者指出僅依 BIM 導出地圖定位需要高發展程度（LOD）模型，且非結構物件常與模型不符。LIO-BIM 以光達慣性里程計持續建立現況地圖，並將機器人周邊的局部地圖與 BIM 做掃描匹配，以同時取得相對 BIM 的定位與現況建圖。系統實作於四足機器人並在辦公室環境與 ConSLAM 工地資料集驗證，程式碼以開源釋出。",[1201,3901,3902,3903],"9-DoF IMU (LORD MicroStrain 3DM-GX5-25; Xsens MTi-610 in ConSLAM)","camera for AprilTag detection (Intel RealSense D435i; Alvium U-319c 3.2 MP in ConSLAM)","reference: Faro Focus S 70 TLS (office); Leica RTC 360 TLS scans of ConSLAM","keyframe-based lidar feature point cloud map in the map frame, aligned to the BIM frame; the BIM is stored as sparse edge and planar feature point clouds (PCD) extracted from IFC geometry after filtering windows, doors and furniture","no loop closure module is described; long-term drift is corrected by unary BIM factors from accepted local-map-to-BIM matches","LIO-SAM-based factor graph; BIM factors added when local-map-to-BIM scan matching converges with inlier RMSE below and fitness above thresholds, with noise variance set to the inlier RMSE (Sec. 3.4)","feature-based ICP with edge (point-to-line) and planar (point-to-plane) correspondences solved by Levenberg-Marquardt, used both for LIO against preceding scans and for matching a 5 m local keyframe feature map to edge and planar feature clouds sampled from the IFC meshes (500 points\u002Fm2, curvature thresholds 0.6 and 0.1)","the latest lidar scan is deskewed in the LIO-SAM-derived front end before feature extraction (Sec. 3.3)","keyframe trajectory (TUM format) and point cloud map aligned with the BIM model; evaluated by APE against ConSLAM and SLAM2REF ground truth and by inlier RMSE (0.3 m) against TLS",[143],{"id":3912,"label":3913,"shortName":3914,"title":3915,"year":249,"era":10,"cluster":755,"scope":99,"keyIdeaZh":3916,"sensors":3917,"mapRepresentation":3919,"loopClosure":72,"estimator":3920,"association":3921,"deskew":57,"outputGeometry":3922,"fulltextStatus":22,"lidarModels":3923,"equipmentCount":185},"imap2021","Sucar et al., 2021","iMAP","iMAP: Implicit Mapping and Positioning in Real-Time","iMAP 首次以單一 MLP 作為即時 RGB-D SLAM 的唯一地圖表示，追蹤執行緒對固定網路最佳化目前位姿，建圖執行緒同時最佳化網路與關鍵影格位姿。以資訊導向的像素取樣與關鍵影格重播緩解遺忘。作者強調 MLP 能對未觀測區域做平滑且合理的補洞，這對工程量測而言代表部分幾何並非量測所得。",[3918],"RGB-D (hand-held Microsoft Azure Kinect for real recordings; rendered Replica RGB-D sequences; TUM RGB-D sequences)","single MLP with Fourier-feature embedding (occupancy + colour)","Adam gradient descent: pose-only tracking against a frozen MLP; mapping jointly optimizes MLP weights and keyframe poses","direct photometric (L1) + depth-variance-normalized geometric rendering losses on actively sampled pixels","mesh by marching cubes on queried occupancy (for visualization\u002Fevaluation only, not part of SLAM)",[],{"id":3925,"label":3926,"shortName":3927,"title":3928,"year":9,"era":10,"cluster":236,"scope":132,"keyIdeaZh":3929,"sensors":3930,"mapRepresentation":3933,"loopClosure":3934,"estimator":3935,"association":3936,"deskew":57,"outputGeometry":3937,"fulltextStatus":22,"lidarModels":3938,"equipmentCount":24},"openvslam2019","Sumikura et al., 2019","OpenVSLAM (stella_vslam)","OpenVSLAM: A Versatile Visual SLAM Framework","OpenVSLAM 是設計成可被第三方程式呼叫的視覺 SLAM 程式庫，演算法沿用 ORB-SLAM 類的間接法：追蹤模組以 ORB 特徵匹配估計每張影格位姿，建圖模組三角化新點並做局部光束法平差，全域模組負責迴圈偵測、位姿圖最佳化與全域光束法平差。其特點是同一架構支援單目、立體與 RGB-D 輸入，以及透視、魚眼與等距柱狀（360 度）相機模型，並可把地圖以 MessagePack 格式儲存與載入，在既有地圖上定位。",[3931,3932],"monocular, stereo or RGB-D camera","camera models: perspective, fisheye, equirectangular (360-degree)","keyframes and sparse 3D landmarks; map database stored and loaded in MessagePack format for reuse and localisation on prebuilt maps (Sec. 3.3)","loop detection in the global optimisation module followed by pose-graph optimisation that also removes scale drift for monocular input; detection method not detailed in the paper (Sec. 3.1)","graph-based indirect SLAM following ORB-SLAM and ProSLAM: tracking by keypoint matching and pose optimisation; mapping module with triangulation and local bundle adjustment; global optimisation module with loop detection, pose-graph optimisation (g2o) and global bundle adjustment (Secs. 3, 3.1; Fig. 2)","ORB features matched to the local map, with an additional robust-matching frame-tracking method (Secs. 3.1, 4.1)","camera trajectory and a sparse 3D point map (Figs. 1, 7-9), exportable as a MessagePack map database",[],{"id":3940,"label":3941,"shortName":3942,"title":3943,"year":150,"era":10,"cluster":236,"scope":346,"keyIdeaZh":3944,"sensors":3945,"mapRepresentation":3947,"loopClosure":72,"estimator":3948,"association":3949,"deskew":57,"outputGeometry":3950,"fulltextStatus":22,"lidarModels":3951,"equipmentCount":554},"smsckf2018","Sun et al., 2018","S-MSCKF (msckf_vio)","Robust Stereo Visual Inertial Odometry for Fast Autonomous Flight","S-MSCKF 把多狀態約束卡爾曼濾波（MSCKF）擴充到立體相機，目標是在微型飛行器的筆電等級電腦上以低運算量穩健估計位姿。前端以 FAST 角點與 KLT 光流同時做時間追蹤與左右影像匹配，並以 2 點 RANSAC 與環狀匹配剔除離群；後端以 4 維立體量測更新，不需影像校正，並採可觀測性約束 EKF（OC-EKF）維持一致性。為平均運算負載，每隔一次更新移除兩個相機狀態。",[3946,36],"stereo camera","none in the filter; features are marginalised by nullspace projection (MSCKF)","stereo multi-state-constraint EKF: IMU state with camera-IMU extrinsics plus a window of left-camera poses; RK4 propagation; 4-D stereo measurement that does not require rectification; nullspace projection of feature errors; observability-constrained EKF (OC-EKF) for consistency; two camera states removed every other update, chosen by a two-way keyframe rule (Sec. III)","FAST corners tracked temporally by KLT optical flow and matched across the stereo pair also by KLT; 2-point RANSAC for temporal outliers and circular matching between consecutive stereo pairs (Sec. III-E)","IMU pose and velocity; in the field test the poses were used to register a laser point cloud (laser used for mapping only) (Sec. IV-C; Fig. 6)",[45],{"id":3953,"label":3954,"shortName":3955,"title":3956,"year":191,"era":52,"cluster":527,"scope":132,"keyIdeaZh":3957,"sensors":3958,"mapRepresentation":3962,"loopClosure":3963,"estimator":3964,"association":3965,"deskew":3966,"outputGeometry":3967,"fulltextStatus":22,"lidarModels":3968,"equipmentCount":24},"surmann2003_kurt3d","Surmann et al., 2003","AIS 3D laser robot for indoor digitalization","An autonomous mobile robot with a 3D laser range finder for 3D exploration and digitalization of indoor environments","本文提出一套不需人工介入的室內 3D 數位化系統。Ariadne 輪式機器人上裝有以伺服馬達俯仰轉動 2D 雷射而成的 AIS 3D 雷射測距儀，停車後掃描水平 180 度、垂直 120 度的範圍。各次 3D 掃描以里程計為初值，用加上 k-d 樹與縮減點的 ICP 做六自由度配準，再以「同步匹配」把每筆掃描對所有重疊的鄰近掃描重新配準，直到不再移動，以分散累積誤差。下一個掃描位置由近似藝廊問題的最佳視點規劃決定，機器人再以全域穩定的馬達控制器與 3D 物體外框避開桌面等突出障礙物，最後輸出 DXF、VRML 與八元樹網格。",[3959,3960,3961],"AIS 3D laser range finder: a 2D laser range finder pitched by a servo, 180 deg (h) x 120 deg (v), with reflectance (2D scanner model not reported)","wheel encoders (odometry)","two 2D safety laser scanners used as bumpers and for dynamic collision avoidance","registered 3D point clouds; octree for visualization and meshing; horizontal-slice polygons with seen and unseen edges for planning; object bounding boxes (Sec. 4, 5.2, 6.1)","no explicit detection; revisits are handled through neighbour overlap in simultaneous matching","ICP with Horn's quaternion closed form, k-d trees and reduced points, initialized by odometry; 'simultaneous matching' re-registers every scan against the union of its overlapping neighbours through a queue until no scan moves, distributing the global error (after Pulli) (Sec. 3.1-3.2)","closest-point correspondences on reduced points; two scans overlap if more than 250 corresponding point pairs exist (Sec. 3.2)","not_applicable (the robot stands still during each 3D scan; a 181 x 256 scan takes 3.4 s)","2D point and line map, 3D volumetric model in DXF and VRML, 3D grid for an OpenGL viewer, octree-based mesh (Sec. 6)",[3969],"2D safety laser scanners",{"id":3971,"label":3972,"shortName":3973,"title":3974,"year":363,"era":52,"cluster":53,"scope":54,"keyIdeaZh":3975,"sensors":3976,"mapRepresentation":3977,"loopClosure":3978,"estimator":3979,"association":3980,"deskew":57,"outputGeometry":3981,"fulltextStatus":22,"lidarModels":3982,"equipmentCount":77},"sunderhauf2012switchable","Sünderhauf & Protzel, 2012","Switchable Constraints","Switchable constraints for robust pose graph SLAM","作者主張 SLAM 後端應能在最佳化過程中自行辨識錯誤的迴圈閉合，而非完全依賴前端資料關聯（data association）。作法是為每條可能出錯的迴圈約束加入一個切換變數（switch variable），以介於 0 與 1 的線性切換函數縮放該約束的權重，並以切換先驗（switch prior）將其錨定在初始值 1；如此位姿圖的拓樸本身成為最佳化對象。作者在 g2o 中實作，於合成與真實位姿圖資料上人工加入最多 1000 條錯誤迴圈，以相對位姿誤差（RPE）與精確率-召回率評估。",[],"not_applicable (pose graph only)","consumes front-end loop closures; each loop edge is multiplied by a switch function of a switch variable in [0,1] (linear function preferred over sigmoid) so that its information can be driven toward zero (Sec. II-A)","nonlinear least-squares pose-graph optimization over odometry factors, switchable loop-closure factors and switch-prior factors, implemented in g2o with Gauss-Newton (Sec. II, Sec. III-B, Fig. 7 caption)","not_applicable (operates on loop-closure constraints delivered by any front-end)","optimized pose graph (trajectory) only; no map geometry is produced or evaluated",[],{"id":3984,"label":3985,"shortName":3986,"title":3987,"year":249,"era":10,"cluster":98,"scope":346,"keyIdeaZh":3988,"sensors":3989,"mapRepresentation":3992,"loopClosure":3993,"estimator":3994,"association":3995,"deskew":121,"outputGeometry":3996,"fulltextStatus":22,"lidarModels":3997,"equipmentCount":24},"lion2021","Tagliabue et al., 2021","LION","LION: Lidar-Inertial Observability-Aware Navigator for Vision-Denied Environments","LION 是 CoSTAR 團隊參加 DARPA 地下挑戰賽所用的 LiDAR 慣性里程計。前端以廣義 ICP 做相鄰掃描配準並先以 IMU 旋轉對齊重力，後端在 GTSAM 中以 3 秒固定延遲滑動視窗平滑器融合 IMU 預積分與掃描間相對位姿，同時線上估計 LiDAR 與 IMU 外參；作者明確說明它是鬆耦合架構，且不建地圖、不做迴圈閉合。另以點到平面 ICP Hessian 平移區塊的條件數作為可觀性指標，條件數過大時通知監督邏輯 HeRO 改用其他里程計來源，例如輪式慣性里程計。",[3990,3991],"3D LiDAR (model not named)","IMU (model not named)","none (LION builds no map) (Sec. 4)","none within LION; drift is compensated by the separate LAMP mapping system (Sec. 4)","fixed-lag sliding-window smoother (3 s window) in GTSAM solved with iSAM2, fusing IMU preintegration factors with relative-pose factors from LiDAR odometry; loosely coupled because points or scans are not in the state (Sec. 2, 3.1, 4)","scan-to-scan Generalized ICP between consecutive clouds, each pre-rotated into a gravity-aligned frame with the IMU rotation as initial guess; no feature extraction; LOCUS can replace the front-end (Sec. 2, 4)","gravity-aligned odometry at up to 200 Hz plus an observability (condition-number) score for a supervisory switching logic (HeRO) (Sec. 2, 3.1)",[45],{"id":3999,"label":4000,"shortName":4001,"title":4002,"year":51,"era":10,"cluster":98,"scope":99,"keyIdeaZh":4003,"sensors":4004,"mapRepresentation":4007,"loopClosure":4008,"estimator":4009,"association":4010,"deskew":4011,"outputGeometry":4012,"fulltextStatus":22,"lidarModels":4013,"equipmentCount":487},"fflins2023","Tang et al., 2023","FF-LINS","FF-LINS: A Consistent Frame-to-Frame Solid-State-LiDAR-Inertial State Estimator","FF-LINS 認為把掃描配準到自建全域地圖（frame-to-map）會讓 LiDAR 慣性估計器把原本不可觀的全域偏航與位置錯誤地當成可觀，造成不一致。它採以 INS 為中心的架構：先以 INS 機械編排的高頻位姿去畸變並選取關鍵影格，再把兩關鍵影格間所有非重複掃描的影格累積成較稠密的關鍵影格點雲地圖；最新關鍵影格的點直接與滑動視窗內其他關鍵影格地圖做點到平面關聯，形成相對位姿約束，與 IMU 預積分一起在因子圖中最佳化，並線上估計 LiDAR 與 IMU 外參及時間延遲。",[4005,4006],"solid-state non-repetitive LiDAR (Livox Mid-70 in the Robot dataset; Livox Horizon and Livox AVIA in public datasets)","MEMS IMU (ADI ADIS16465 in the Robot dataset; built-in Livox IMUs in public datasets)","per-keyframe point-cloud maps accumulated from all frames since the previous keyframe with INS poses, voxel-downsampled at 0.5 m; no global map is used for state estimation (Sec. III-A-3)","none (authors state it could be added for large-scale mapping, Sec. V)","sliding-window factor graph (10 IMU preintegration intervals) solved with Levenberg-Marquardt in Ceres, tightly coupling LiDAR frame-to-frame point-to-plane factors with IMU preintegration and marginalization prior; LiDAR-IMU extrinsics and time delay estimated online; INS-centric update only at LiDAR keyframes (Sec. II, III-C)","direct frame-to-frame: each point of the newest keyframe is projected into the accumulated keyframe point-cloud maps of the other keyframes in the window; plane fitted to 5 nearest points and accepted if all lie within 0.1 m; Huber loss plus chi-square rejection between two optimizations (Sec. III-B, III-C-4)","interpolated INS poses from mechanization undistort each frame before 0.5 m voxel downsampling (Sec. III-A-2)","continuous INS-rate poses between keyframes, keyframe states, online LiDAR-IMU extrinsics and time delay; keyframe point clouds",[505,1482,437],{"id":4015,"label":4016,"shortName":4017,"title":4018,"year":30,"era":10,"cluster":263,"scope":99,"keyIdeaZh":4019,"sensors":4020,"mapRepresentation":4024,"loopClosure":4025,"estimator":4026,"association":4027,"deskew":4028,"outputGeometry":4029,"fulltextStatus":22,"lidarModels":4030,"equipmentCount":1764},"palvio2026","Tang et al., 2026","PA-LVIO","PA-LVIO: Real-Time LiDAR-Visual-Inertial Odometry and Mapping with Pose-Only Bundle Adjustment","PA-LVIO 提出僅含位姿的光束法平差（pose-only bundle adjustment），把光達與視覺的多幀幾何約束轉為幀間位姿約束，在滑動視窗因子圖中與 IMU 預積分緊密融合，以降低計算量。另加入不需邊緣化的幀對地圖（frame-to-map）光達位姿約束抑制漂移，並以 IMU 為中心線上估計光達、相機的時空參數，達到像素級對齊後產生 RGB 上色點雲地圖。",[4021,4022,4023],"3D LiDAR (Livox AVIA 10 Hz on MARS-LVIG, R3LIVE and HandNav data; Hesai AT128 10 Hz on i2Nav-Robot)","IMU (MEMS; BMI088 on MARS-LVIG, R3LIVE and HandNav; ADIS16465 on i2Nav-Robot; 200 Hz)","RGB camera (models not reported; 1280x1024 to 2448x2048 at 10 to 15 Hz)","global point cloud map in ikd-Tree; RGB-rendered point-cloud map","none (authors note place recognition or loop closure could be added; Sec. I)","INS-centric sliding-window factor graph optimization with pose-only bundle adjustment factors for LiDAR and visual measurements, IMU preintegration, and a marginalization-free frame-to-map LiDAR pose factor","same-plane LiDAR points associated across keyframes (following BA-LINS); visual features tracked with INS prior; frame-to-map LiDAR pose optimization against a global ikd-Tree map","point clouds undistorted using high-rate INS pose (Sec. II)","RGB-rendered point-cloud map (abstract, Sec. III end)",[437,4031],"Hesai AT128",{"id":4033,"label":4034,"shortName":4035,"title":4036,"year":769,"era":10,"cluster":755,"scope":132,"keyIdeaZh":4037,"sensors":4038,"mapRepresentation":4039,"loopClosure":4040,"estimator":4041,"association":4042,"deskew":57,"outputGeometry":4043,"fulltextStatus":22,"lidarModels":4044,"equipmentCount":144},"cnnslam2017","Tateno et al., 2017","CNN-SLAM","CNN-SLAM: Real-Time Dense Monocular SLAM with Learned Depth Prediction","CNN-SLAM 以 LSD-SLAM 的直接法關鍵影格架構為基礎，只在建立關鍵影格時用卷積網路預測稠密深度，並依目前相機與訓練相機的焦距比例調整尺度，再以後續影格的小基線立體匹配依不確定度加權修正深度。低紋理區保留網路預測、高梯度區由立體量測主導，因此單眼 SLAM 可取得絕對尺度，在純旋轉運動下也能重建。關鍵影格以位姿圖最佳化，深度圖可再融合成含語意標籤的三維模型。",[37],"dense per-key-frame depth and uncertainty maps fused into a global 3D model with optional semantic labels using the fusion scheme of its reference [27] (Sec. 3.5)","pose-graph edges added between a new key-frame and existing key-frames with a similar field of view (small relative pose); no appearance-based place recognition described (Sec. 3.3)","LSD-SLAM-style direct key-frame tracking: weighted Gauss-Newton minimization of Huber-weighted photometric residuals on high-gradient pixels against the nearest key-frame; key-frame depth from a CNN (ResNet-50 fully convolutional network of Laina et al.) scaled by the focal-length ratio and refined by uncertainty-weighted fusion of small-baseline stereo depth; key-frame pose-graph optimization (Sec. 3.1-3.4)","direct photometric alignment; per-frame depth from 5-pixel epipolar matching for refinement (Sec. 3.1, Sec. 3.4)","camera trajectory, dense key-frame depth maps and a fused, optionally semantically labelled, 3D reconstruction (Sec. 3.5, Figs. 1 and 6)",[],{"id":4046,"label":4047,"shortName":4048,"title":4049,"year":249,"era":10,"cluster":755,"scope":132,"keyIdeaZh":4050,"sensors":4051,"mapRepresentation":4052,"loopClosure":4053,"estimator":4054,"association":4055,"deskew":57,"outputGeometry":4056,"fulltextStatus":22,"lidarModels":4057,"equipmentCount":109},"droidslam2021","Teed & Deng, 2021","DROID-SLAM","DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras","DROID-SLAM 以卷積 GRU 反覆預測稠密光流修正，並透過可微分稠密光束調整（dense bundle adjustment, DBA）同步更新相機位姿與逐像素反深度。前端做局部光束調整，後端對全部關鍵影格做全域光束調整，回訪時加入長距離邊以形成迴圈。網路僅以合成 TartanAir 單目影片訓練，可在測試時使用雙目或 RGB-D。作者明言 SLAM 重建缺乏標準幾何評估協定，因此未評估稠密點雲精度。",[37,2101,772],"per-keyframe dense inverse-depth maps","long-range co-visibility edges added to the frame graph when revisiting mapped regions","recurrent learned update operator + differentiable dense bundle adjustment (Gauss-Newton, block-sparse Cholesky) over a keyframe frame-graph","learned dense optical flow via correlation volumes","camera trajectory and per-pixel inverse depth of keyframes (dense points by back-projection); authors did not evaluate 3D reconstructions",[],{"id":4059,"label":4060,"shortName":4061,"title":4062,"year":51,"era":10,"cluster":755,"scope":346,"keyIdeaZh":4063,"sensors":4064,"mapRepresentation":4065,"loopClosure":4066,"estimator":4067,"association":4068,"deskew":57,"outputGeometry":4069,"fulltextStatus":22,"lidarModels":4070,"equipmentCount":77},"dpvo2023","Teed et al., 2023","DPVO","Deep Patch Visual Odometry","DPVO 是深度學習式單眼視覺里程計，把 DROID-SLAM 的稠密光流改為只追蹤稀疏影像區塊（patch）。每張影格隨機取樣區塊，循環更新網路依相關特徵、時間向卷積與訊息傳遞預測區塊軌跡修正量與信心權重，再由可微分光束法平差在滑動視窗內更新位姿與區塊逆深度。整個網路只以合成資料 TartanAir 訓練，在 RTX-3090 上平均每秒 60 影格、約 4.9 GB 記憶體，但不含迴圈閉合，輸出僅為軌跡與稀疏點。",[37],"patch graph of sparse fronto-parallel patches with inverse depth; sparse 3D reconstruction (Sec. 3, Fig. 1, Appendix E)","none (visual odometry only; DPV-SLAM later adds loop closure)","recurrent update operator (correlation, 1D temporal convolution, softmax aggregation, transition block, factor head) predicts 2D patch-trajectory revisions and confidences; a differentiable bundle adjustment layer applies two Gauss-Newton iterations with the Schur complement to camera poses and patch inverse depths; poses of all but the last 10 keyframes are fixed (Sec. 3.1, Sec. 3.3, Appendix F)","sparse image patches at random locations (96 per frame by default, 48 in the fast setting) tracked by learned correlation features against frames within distance r in a bipartite patch graph (Sec. 3, Sec. 4)","camera trajectory and sparse 3D points of tracked patches (Fig. 1, Fig. B)",[],{"id":4072,"label":4073,"shortName":4074,"title":4075,"year":3331,"era":52,"cluster":527,"scope":132,"keyIdeaZh":4076,"sensors":4077,"mapRepresentation":4080,"loopClosure":4081,"estimator":4082,"association":4083,"deskew":121,"outputGeometry":4084,"fulltextStatus":22,"lidarModels":4085,"equipmentCount":24},"thrun2000_3dmapping","Thrun et al., 2000","Thrun-Burgard-Fox real-time 2D and 3D laser mapping","A real-time algorithm for mobile robot mapping with applications to multi-robot and 3D mapping","本文把增量式雷射掃描匹配與以樣本表示的位姿後驗結合：後驗的計算方式與蒙地卡羅定位相同，每次掃描以多個樣本作為爬山搜尋的起點，找到最可能的位姿後把掃描加入地圖。當後驗推得的位姿與單純增量估計不一致時，系統判定發生迴圈閉合，先把位姿差按比例分配到迴圈內各位姿，再以梯度下降反覆修正，於兩次量測之間完成反向校正。同一後驗機制讓第二台機器人先在領隊機器人的地圖中全域定位，再共同建圖；另以向前雷射做 2D 定位、向上雷射擷取 3D 資料，產生經多邊形簡化的精簡建物 3D 模型。",[4078,4079],"2D laser range finders: forward-looking for 2D mapping and localization, upward-pointed for 3D (model not reported)","wheel odometry (optional; also run with odometry removed)","collection of 2D scans with poses; 3D model as polygons from the upward laser, outlier-filtered and simplified with a computer-graphics polygon simplification (Sec. 2.6)","detected when the posterior-based pose differs from the incremental maximum-likelihood pose; the loop is identified from the scan that caused the adjustment (Sec. 2.4)","incremental maximum-likelihood scan alignment by gradient ascent, combined with a sample-based posterior over the current pose computed like Monte Carlo localization; every sample seeds a hill-climbing search (Sec. 2.2-2.3)","no explicit correspondences: perceptual likelihood penalizes obstacles in space previously seen as free (ray-traced), with a map made of past scans and poses (Sec. 2.1)","2D scan map and a simplified 3D polygonal model viewed in a standard virtual reality tool (VRweb) (Sec. 3.5)",[45],{"id":4087,"label":4088,"shortName":4089,"title":4090,"year":97,"era":10,"cluster":599,"scope":132,"keyIdeaZh":4091,"sensors":4092,"mapRepresentation":4096,"loopClosure":4097,"estimator":4098,"association":4099,"deskew":57,"outputGeometry":4100,"fulltextStatus":22,"lidarModels":4101,"equipmentCount":46},"kimeramulti2022","Tian et al., 2022","Kimera-Multi","Kimera-Multi: Robust, Distributed, Dense Metric-Semantic SLAM for Multi-Robot Systems","Kimera-Multi 是分散式多機器人度量語意 SLAM：各機器人以 Kimera-VIO（雙目與 IMU）估計軌跡並建立語意網格；相遇時交換詞袋描述子並做幾何驗證取得跨機迴圈；以分散式漸進非凸（D-GNC）穩健位姿圖最佳化剔除感知混淆造成的錯誤迴圈；最後以變形圖（deformation graph）依最佳化軌跡校正網格。",[4093,4094,4095],"stereo images and IMU (Kimera-VIO input)","depth images from an RGB-D camera or from stereo matching, plus 2D semantic segmentation (Kimera-Semantics input)","outdoor experiments: forward-facing RealSense D435i RGBD camera and IMU on a Clearpath Jackal UGV","metric-semantic 3D mesh","intra- and inter-robot visual loop closures with outlier rejection by D-GNC","per-robot Kimera-VIO; robust distributed pose-graph optimization via distributed graduated non-convexity (D-GNC) with RBCD solver","DBoW2 bag-of-words place recognition with geometric verification (five-point \u002F three-point RANSAC)","semantically labelled 3D mesh and optimized trajectories",[],{"id":4103,"label":4104,"shortName":4105,"title":4106,"year":51,"era":10,"cluster":11,"scope":12,"keyIdeaZh":4107,"sensors":4108,"mapRepresentation":4111,"loopClosure":4112,"estimator":4113,"association":4114,"deskew":4115,"outputGeometry":4116,"fulltextStatus":22,"lidarModels":4117,"equipmentCount":46},"vegatorres2023ogm2pgbm","Torres et al., 2023","OGM2PGBM","OGM2PGBM: Robust BIM-based 2D-LiDAR localization for lifelong indoor navigation","作者提出從 BIM 產生適合 2D LiDAR 定位的地圖，並比較不同定位器在 Scan-BIM 偏差下的表現。首先以 IfcConvert 在指定高度切出只含結構構件的 SVG 剖面，再用 OpenCV 輪廓階層區分室外、室內與障礙物，得到佔據格地圖；接著以骨架化與波前覆蓋路徑在可通行區域產生路點，不經 Gazebo 直接以光線投射模擬雷射掃描與里程計，組成 Cartographer 與 SLAM Toolbox 可讀的位姿圖地圖。作者在 Gazebo 中以空房間、依 TLS 資料建立的真實辦公室與災後雜亂三種情境（有無行人各一），比較 AMCL、GMCL 兩種粒子濾波與 Cartographer、SLAM Toolbox 兩種圖式定位的 RMSE 與全域定位收斂時間。",[4109,4110],"2D LiDAR (simulated Hokuyo UST-10LX)","simulated IMU and wheel odometry","2D occupancy grid map extracted from IFC (IfcConvert SVG section, OpenCV contour hierarchy separating outdoor, indoor and obstacles) and a pose-graph map (.pbstream for Cartographer, .posegraph for SLAM Toolbox) built from ray-cast virtual scans along a skeleton-based wavefront coverage path (Sec. 4.1-4.2)","not_applicable for pure localization; Cartographer and SLAM Toolbox constraints to the prior pose graph","Compared localizers: particle filters AMCL and GMCL, and graph-based Cartographer (pure localization) and SLAM Toolbox using the generated pose-graph maps; proposed use: GMCL for global localization until covariance below 0.05, then switch to a graph-based localizer for pose tracking (Sec. 4.3)","scan matching of the respective localizers against OGM or pose-graph submaps (not modified by the authors)","not_applicable (2D LiDAR in simulation)","2D robot pose in the BIM frame; no point cloud produced",[4118],"Hokuyo UST-10LX (simulated)",{"id":4120,"label":4121,"shortName":4122,"title":4123,"year":661,"era":10,"cluster":68,"scope":69,"keyIdeaZh":4124,"sensors":4125,"mapRepresentation":4128,"loopClosure":72,"estimator":4129,"association":4130,"deskew":4131,"outputGeometry":4132,"fulltextStatus":22,"lidarModels":4133,"equipmentCount":554},"tuna2024xicp","Tuna et al., 2024","X-ICP","X-ICP: Localizability-Aware LiDAR Registration for Robust Localization in Extreme Environments","X-ICP 針對 LiDAR 在幾何資訊不足環境（隧道、開放平面、狹窄走廊）中 ICP 沿弱約束方向發散的問題，先利用掃描與地圖的對應，分析各最佳化主方向的對齊強度，細緻判定可定位性（localizability）。再將此分析整合進掃描對地圖的點對平面 ICP，以約束最佳化控制或凍結退化方向的位姿更新。作者以 ANYmal 足式機器人在地下礦坑、營建工地與城市公園實測，並以 Leica RTC360 地面掃描作為礦坑與公園的參考地圖。",[4126,36,4127],"3D LiDAR (Velodyne VLP-16; Ouster OS0-128)","leg joint encoders (leg odometry prior)","point cloud map in a libpointmatcher-based registration framework (Sec. VII-A)","scan-to-map point-to-plane ICP with localizability-driven constrained optimization (Lagrange multipliers) that fixes or limits updates along degenerate directions (abstract, Sec. VI)","scan-to-map correspondences analysed against principal optimization directions for fine-grained localizability (abstract)","point cloud motion compensation done at the LiDAR driver level with the leg-odometry pose estimates in the transformation tree; the pose prior is used to transform and undistort the input cloud","pose updates; resulting point cloud maps compared with TLS reference maps (Sec. VII)",[143,222],{"id":4135,"label":4136,"shortName":4137,"title":4138,"year":4139,"era":52,"cluster":1344,"scope":514,"keyIdeaZh":4140,"sensors":4141,"mapRepresentation":57,"loopClosure":57,"estimator":4142,"association":1616,"deskew":57,"outputGeometry":57,"fulltextStatus":22,"lidarModels":4143,"equipmentCount":61},"umeyama1991least","Umeyama, 1991","Umeyama alignment","Least-squares estimation of transformation parameters between two point patterns",1991,"作者針對 m 維空間中已知對應關係的兩組點，推導使均方誤差最小的相似轉換（旋轉 R、平移 t、尺度 c）閉式解：先求兩組點的平均向量、變異數與交叉共變異矩陣，再對共變異矩陣做奇異值分解，並在其行列式為負時把對角符號矩陣 S 的最後一項設為 -1，以確保得到真正的旋轉而非反射；尺度與平移再由 S、奇異值與平均向量直接算出。作者指出 Arun 與 Horn 的解相當於不論行列式正負都取 S 為單位矩陣，在資料嚴重受擾時可能給出反射；本文解法適用任意維度，而四元數法僅適用三維。數值例中舊解以反射達成零誤差，新解則回傳誤差 0.533 的正常旋轉。",[],"closed-form least squares: SVD of the cross-covariance matrix Sigma_xy = U D V^T; R = U S V^T with S = diag(1,...,1,-1) when det(Sigma_xy) \u003C 0 (or det(U)det(V) = -1 when rank = m-1), c = tr(DS)\u002Fsigma_x^2, t = mu_y - c R mu_x; minimum error sigma_y^2 - tr(DS)^2\u002Fsigma_x^2",[],{"id":4145,"label":4146,"shortName":4147,"title":4148,"year":396,"era":10,"cluster":236,"scope":132,"keyIdeaZh":4149,"sensors":4150,"mapRepresentation":4151,"loopClosure":4152,"estimator":4153,"association":4154,"deskew":57,"outputGeometry":4155,"fulltextStatus":22,"lidarModels":4156,"equipmentCount":144},"basalt2020","Usenko et al., 2020","Basalt","Visual-Inertial Mapping With Non-Linear Factor Recovery","Basalt 採兩層架構整合視覺慣性里程計與全域一致建圖。下層為立體視覺 VIO，以 KLT 光流追蹤 FAST 角點，在滑動視窗中聯合最佳化重投影與 IMU 預積分誤差，並以首次估計 Jacobian 做部分邊際化。當關鍵影格被邊際化時，作者用非線性因子還原（NFR）把邊際化先驗近似成關鍵影格間的相對位姿因子與橫滾俯仰因子。上層以 ORB 特徵在關鍵影格間匹配並做光束法平差，結合這些因子得到重力對齊的全域地圖，不需估計每個關鍵影格的速度與偏差。",[3946,36],"keyframe poses and ORB landmark positions (inverse distance with stereographic bearing parameterisation) in a gravity-aligned global map; VIO landmarks hosted in keyframes (Secs. IV-B, V-A)","implicit, through ORB keypoint matching between keyframes in the global bundle adjustment (Sec. V)","two layers: (1) stereo VIO as fixed-lag smoother (Gauss-Newton over 7 pose-only keyframes and 3 latest states with velocity and biases) combining reprojection and preintegrated IMU terms, Schur-complement partial marginalisation with first-estimate Jacobians; (2) visual-inertial mapping as keyframe bundle adjustment of ORB landmarks plus relative-pose and roll-pitch factors recovered from the VIO marginalisation prior by non-linear factor recovery (KL-divergence minimisation) (Secs. IV, V)","VIO: FAST corners in a 50 pixel grid (80 to 120 features) tracked by pyramidal inverse-compositional KLT with SE(2) patch warp and locally scaled SSD, forward-backward consistency check; mapping: ORB features detected and matched between keyframes (Secs. IV-A, V-A, VI-a)","VIO pose for every frame; globally consistent, gravity-aligned keyframe trajectory and sparse ORB landmark map (Fig. 1)",[],{"id":4158,"label":4159,"shortName":4160,"title":4161,"year":150,"era":10,"cluster":599,"scope":600,"keyIdeaZh":4162,"sensors":4163,"mapRepresentation":4168,"loopClosure":4169,"estimator":57,"association":4170,"deskew":121,"outputGeometry":57,"fulltextStatus":22,"lidarModels":4171,"equipmentCount":109},"pointnetvlad2018","Uy & Lee, 2018","PointNetVLAD","PointNetVLAD: Deep Point Cloud Based Retrieval for Large-Scale Place Recognition","PointNetVLAD 結合 PointNet 的逐點特徵與 NetVLAD 聚合層，將去除地面並下採樣為 4096 點的子地圖映射為固定長度全域描述子，以最近鄰檢索完成地點辨識；並提出 lazy triplet 與 quadruplet 損失做度量學習。作者同時以 Oxford RobotCar 與三個自建區域建立點雲檢索基準。",[4164,4165,4166,4167],"2D LiDAR SICK LMS-151 scans accumulated into 3D reference maps and submaps using GPS\u002FINS (Oxford RobotCar benchmark)","Velodyne-64 LiDAR, as written (in-house U.S., R.A., B.D. sets)","GPS\u002FINS (reference maps in UTM frame and training labels)","stereo camera centre images used only for the NetVLAD image baseline","database of fixed-size (4096-point) ground-removed, normalized submaps","retrieval of structurally similar submaps; no metric pose output","learned global descriptor (PointNet features + NetVLAD aggregation + fully connected layer), nearest-neighbour retrieval; lazy triplet\u002Fquadruplet metric-learning loss",[2125,4172],"Velodyne-64 LiDAR",{"id":4174,"label":4175,"shortName":4176,"title":4177,"year":661,"era":10,"cluster":11,"scope":12,"keyIdeaZh":4178,"sensors":4179,"mapRepresentation":4182,"loopClosure":4183,"estimator":4184,"association":4185,"deskew":4186,"outputGeometry":4187,"fulltextStatus":22,"lidarModels":4188,"equipmentCount":109},"vegatorres2024slam2ref","Vega-Torres et al., 2024","SLAM2REF","SLAM2REF: advancing long-term mapping with 3D LiDAR and reference map integration for precise 6-DoF trajectory estimation and map extension","SLAM2REF 把行動 LiDAR 與 IMU 資料和既有 BIM 或點雲參考圖整合，用於室內無 GPS 環境的長期建圖。流程先由參考圖產生佔據網格與模擬 LiDAR 掃描作為「參考工作段」，再以 DLIO 去除實測掃描的運動畸變，接著用室內版 Scan Context 描述子與 YawGICP 找跨工作段對應，透過多工作段錨定（multi-session anchoring）位姿圖最佳化把漂移的 SLAM 結果對齊參考圖，最後逐幀以點對點 ICP 對齊 1 cm 密度的參考點雲。對齊後以 OctoMap 分析新增與移除的構件並網格化，並允許地圖延伸到參考圖範圍之外。",[4180,4181],"3D LiDAR of the ConSLAM handheld system (model not reported; the ISC descriptor requires a 360-degree horizontal FoV)","9-axis IMU of the ConSLAM handheld system (model not reported; used for DLIO deskewing; LiDAR-IMU extrinsics from OA-LICalib)","point cloud; OctoMap for dynamic-object removal and free-space reasoning; voxel-cube meshes for positive and negative differences (Sec. 4.3)","inter-session loops to reference-map sessions (ISC, then KNN); intra-session loops optional from the SLAM front end (Sec. 4.2)","DLIO front end for deskewing and odometry; multi-session anchoring pose graph in GTSAM (iSAM2, batch) with odometry, Indoor Scan Context and KNN inter-session constraints; final point-to-point ICP of each scan to a 1 cm dense reference cloud (Sec. 4.2, 5.2)","Indoor Scan Context descriptors (binary occupancy per bin, 60 sectors x 20 rings, at least 40 points, 10 m radius) matched against simulated reference scans: 100 top candidates from a nanoflann KD-tree of rotation-invariant 1D descriptors, column-wise cosine score with threshold 0.3 and yaw shifts limited to 36 deg, then YawGICP (built on Open3D GICP) against 3-scan reference submaps; KNN submap loops (K = 5) with adaptive covariance, omitted for BIM references; final point-to-point ICP to a 1 cm dense reference cloud with fitness at 1 cm and 3 cm computed on points within 30 cm (Sec. 4.2.2, 4.2.3, 5.2.2)","IMU-based point-wise motion correction using DLIO; bags replayed at half speed to avoid deskew errors (Sec. 4.2.1.1, 5.2.2.1)","6-DoF poses in the reference-map frame with per-scan alignment classes (perfect, good, bad, outside map); aligned and extended point cloud map; meshes of positive and negative differences (Sec. 4.2.3, 4.3)",[45],{"id":4190,"label":4191,"shortName":4192,"title":4193,"year":150,"era":10,"cluster":236,"scope":99,"keyIdeaZh":4194,"sensors":4195,"mapRepresentation":4197,"loopClosure":72,"estimator":4198,"association":4199,"deskew":1976,"outputGeometry":4200,"fulltextStatus":22,"lidarModels":4201,"equipmentCount":144},"supereight2018","Vespa et al., 2018","supereight","Efficient Octree-Based Volumetric SLAM Supporting Signed-Distance and Occupancy Mapping","supereight 提出以八元樹（octree）為空間索引的稠密體積 SLAM 框架。最底層以 8×8×8 體素區塊為單位，並以 Morton 編碼排序與逐層遮罩做免鎖的平行配置，再預先計算三線性內插的查詢順序，使八元樹在 CPU 上的融合與射線投射效率接近 InfiniTAM 的雜湊表。同一資料結構可存放 TSDF，也可存放機率佔據（occupancy）地圖；佔據地圖改寫 Loop 等人的 b-spline 雜訊模型，改用對數勝算累加、機率截斷與依時間遺忘的更新，使其適用於增量 SLAM，並明確表示已觀測的空區。追蹤採 KinectFusion 式只用深度的點對面 ICP，作者並示範佔據地圖可直接供 Informed RRT* 路徑規劃查詢。",[4196],"RGB-D camera, depth only (TUM RGB-D real sequences and ICL-NUIM synthetic sequences; sensor models not named)","Morton-coded octree whose leaves are 8 x 8 x 8 voxel blocks allocated in parallel from a memory pool; the field type is generic: TSDF, or log-odds occupancy using a quadratic b-spline depth-noise model (sigma proportional to range squared, 4 cm at 2 m), clamping to [0.03, 0.97] and a time-windowed forgetting update (tau = 5 s)","KinectFusion-style frame-to-model point-to-plane ICP solved by Gauss-Newton on depth only, against vertex and normal maps ray-cast from the octree map","Projective data association between the current depth frame and the ray-cast model prediction","TSDF or occupancy octree map at 1 cm finest resolution (surfaces as zero crossings), usable directly for sampling-based path planning",[],{"id":4203,"label":4204,"shortName":4205,"title":4206,"year":249,"era":10,"cluster":31,"scope":99,"keyIdeaZh":4207,"sensors":4208,"mapRepresentation":4209,"loopClosure":4210,"estimator":4211,"association":4212,"deskew":121,"outputGeometry":4213,"fulltextStatus":22,"lidarModels":4214,"equipmentCount":46},"vizzo2021puma","Vizzo et al., 2021","PUMA","Poisson Surface Reconstruction for LiDAR Odometry and Mapping","PUMA 把最近 N 次掃描累積成局部點雲，以 Poisson 表面重建生成三角網格，並依頂點密度修剪 10% 低支持頂點，移除 Poisson 在無資料處外插的表面；新掃描以射線投射求與網格三角面的交點作為對應，進行點對面 ICP（frame-to-mesh）。局部網格每 M 次掃描併入全域網格，全域網格不參與估計，也未實作迴圈修正。",[35],"local Poisson surface reconstruction mesh (octree depth 10) rebuilt from the last N = 30 scans after every registered scan, with the 10% lowest-density vertices trimmed; global mesh aggregated every M = 30 scans with duplicate-triangle removal and used only for visualisation and output (Sec. III-B, III-C, IV)","none (listed as future work, Sec. V)","iterative frame-to-mesh point-to-plane ICP with Huber kernel, initialised with the previous pose increment; during the first N = 30 scans, before a mesh exists, standard point-to-plane ICP is used (Sec. III-A, III-C, IV-D)","rays from the current sensor origin through every scan point are intersected with the local triangle mesh (Embree); the hit point and triangle normal form the correspondence, and pairs farther apart than 1 m are rejected (Sec. III-A, IV)","global triangle mesh",[45,4215,4216],"virtual 64-beam LiDAR sensor model (Mai City scans)","virtual 320-beam LiDAR sensor model (Mai City ground truth)",{"id":4218,"label":4219,"shortName":4220,"title":4221,"year":97,"era":10,"cluster":31,"scope":32,"keyIdeaZh":4222,"sensors":4223,"mapRepresentation":4224,"loopClosure":57,"estimator":4225,"association":4226,"deskew":4227,"outputGeometry":4228,"fulltextStatus":22,"lidarModels":4229,"equipmentCount":1289},"vizzo2022vdbfusion","Vizzo et al., 2022","VDBFusion","VDBFusion: Flexible and Efficient TSDF Integration of Range Sensor Data","VDBFusion 以 OpenVDB 的階層稀疏體積結構儲存 TSDF，提供 C++ 與 Python 介面；權重函數可在執行時以 lambda 傳入，空間雕刻（space carving）可選擇開關，網格擷取可設最小權重門檻。作者說明截斷距離大有助抑制雜訊，但會造成薄面增厚；關閉空間雕刻較快，但會留下動態物體殘影。",[35,772],"TSDF stored as two OpenVDB sparse grids (signed distance and weight); VDB leaf blocks are typically 8 x 8 x 8 voxels in a fixed-depth tree; truncation distance 3 voxels in the experiments; optional space carving; no occupancy probabilities (Sec. 3, 4.2, 4.3, 5)","not_applicable (poses supplied; points assumed in the global frame)","ray casting of points into a TSDF within the truncation distance","delegated to dataset-specific data loaders (system section)","triangle mesh via marching cubes adapted from Open3D to VDB, with optional hole filling (Curless and Levoy) and a runtime min_weight threshold that also removes dynamic objects; TSDF and weight grids saved as VDB files with lossless compression (Sec. 4.6, 5.3)",[45,161],{"id":4231,"label":4232,"shortName":4233,"title":4234,"year":51,"era":10,"cluster":98,"scope":99,"keyIdeaZh":4235,"sensors":4236,"mapRepresentation":4237,"loopClosure":72,"estimator":4238,"association":4239,"deskew":4240,"outputGeometry":4241,"fulltextStatus":22,"lidarModels":4242,"equipmentCount":77},"kissicp2023","Vizzo et al., 2023","KISS-ICP","KISS-ICP: In Defense of Point-to-Point ICP - Simple, Accurate, and Robust Registration If Done the Right Way","KISS-ICP 回歸最基本的點到點（point-to-point）ICP，僅保留等速運動預測與逐點去畸變、體素雙重降採樣、依運動模型偏差自適應的對應距離門檻，以及穩健核函數等少數元件。地圖為雜湊表中的降採樣體素點雲，並保留原始點座標以避免離散化誤差。作者主張在同一組參數下可用於車載、無人機、Segway 與手持 LiDAR，且不需 IMU，也不含迴圈或位姿圖。",[1355],"voxelized, downsampled local point cloud stored in a hash table with a maximum number of points per voxel; voxels beyond maximum range removed","robust point-to-point ICP (Gauss-Newton with robust kernel) frame-to-local-map, constant-velocity motion prediction","point-to-point nearest neighbour with adaptive correspondence threshold derived from observed deviation from the motion model","constant-velocity model applied with per-point relative timestamps (Sec. III-A); IMU or wheel odometry can replace the velocity source","odometry and local voxel point map; original point coordinates retained within voxels (no centroid snapping)",[45],{"id":4244,"label":4245,"shortName":4246,"title":4247,"year":769,"era":10,"cluster":755,"scope":346,"keyIdeaZh":4248,"sensors":4249,"mapRepresentation":4250,"loopClosure":72,"estimator":4251,"association":4252,"deskew":57,"outputGeometry":4253,"fulltextStatus":22,"lidarModels":4254,"equipmentCount":77},"deepvo2017","Wang et al., 2017","DeepVO","DeepVO: Towards end-to-end visual odometry with deep Recurrent Convolutional Neural Networks","DeepVO 是早期的端到端單眼視覺里程計：把相鄰兩張 RGB 影像疊合後送入以 FlowNet 預訓練權重初始化的卷積網路擷取運動特徵，再以兩層 LSTM 建模時間序列，直接迴歸每一時刻的六自由度位姿。方法不需特徵擷取、匹配、光束法平差，甚至不需相機校正，絕對尺度由訓練資料隱式學得。論文只在 KITTI 上驗證，平移漂移優於單眼 LIBVISO2，但仍明顯不如立體 LIBVISO2，且不產生地圖。",[37],"none (poses only)","end-to-end regression: a 9-layer CNN initialized from pretrained FlowNet extracts features from two stacked consecutive RGB frames, and two stacked LSTM layers (1000 hidden units each) output a 6-DoF pose per time step; trained with MSE on positions and Euler angles (orientation weight 100) (Sec. III)","none explicit; motion is learned implicitly from stacked image pairs without feature matching (Sec. III)","6-DoF camera trajectory only; no map or point cloud",[],{"id":4256,"label":4257,"shortName":4258,"title":4259,"year":9,"era":10,"cluster":236,"scope":32,"keyIdeaZh":4260,"sensors":4261,"mapRepresentation":4265,"loopClosure":4266,"estimator":4267,"association":4268,"deskew":4269,"outputGeometry":4270,"fulltextStatus":22,"lidarModels":4271,"equipmentCount":46},"densesurfelmapping2019","Wang et al., 2019","Dense Surfel Mapping (Wang, Gao, Shen)","Real-time Scalable Dense Surfel Mapping","作者提出只用 CPU 的面元（surfel）稠密建圖系統，相機位姿、參考關鍵影格與位姿圖都由外部稀疏視覺 SLAM（ORB-SLAM2 或 VINS-Mono）提供。每張影像先以擴充的 SLIC 依強度、深度與位置分割超像素，可處理無效深度，並以 Huber 損失求穩健的平均深度，再由超像素建立面元，因此能融合 RGB-D、立體相機或單眼深度預測等品質較差的深度圖。每個面元附屬於一個關鍵影格，只有在位姿圖上與參考關鍵影格相距少於 G_delta 條邊的面元才參與融合，使每一影格的融合時間與場景規模無關；位姿圖最佳化後，依各關鍵影格的位姿變化移動其面元，使整張地圖非剛性變形並保持全域一致。",[4262,4263,4264],"RGB-D (synthetic ICL-NUIM input with ORB-SLAM2 in RGB-D mode)","Stereo camera (KITTI odometry; depth from PSMNet stereo matching, ORB-SLAM2 stereo mode)","Monocular camera with learned depth (KITTI left images with monocular depth prediction; handheld camera with MVDepthNet depth and VINS-Mono tracking)","superpixel-based surfels (position, normal, intensity, weight, radius, update count, attached keyframe index) stored in a map database organized by keyframe","Provided by the localization system (ORB-SLAM2 or VINS-Mono); after loop closure previously built surfels are reused through new pose-graph edges","No pose estimation in the mapper: an external sparse visual SLAM (ORB-SLAM2 or VINS-Mono) provides each frame's pose, the reference keyframe and the optimized keyframe pose graph","Local surfels are projected into the current frame and matched to the surfel initialized from the superpixel at that pixel when depths agree within a disparity-derived bound and normals agree (dot product > 0.8); only surfels attached to keyframes within G_delta pose-graph edges of the reference keyframe (breadth-first search) are fused","not_applicable (camera input)","globally consistent surfel map (shown as point clouds and meshes)",[],{"id":4273,"label":4274,"shortName":4275,"title":4276,"year":396,"era":10,"cluster":151,"scope":132,"keyIdeaZh":4277,"sensors":4278,"mapRepresentation":4281,"loopClosure":4282,"estimator":4283,"association":4284,"deskew":4281,"outputGeometry":4285,"fulltextStatus":22,"lidarModels":4286,"equipmentCount":487},"iscloam2020","Wang et al., 2020","ISC-LOAM (Intensity Scan Context)","Intensity Scan Context: Coding Intensity and Geometry Relations for Loop Closure Detection","強度掃描脈絡（Intensity Scan Context，ISC）是一種同時編碼幾何與 LiDAR 強度的全域描述子：先以距離校正強度，再把 50 m 內的點依方位角與半徑分格，每格保留最大強度，形成一張二維矩陣。檢索分兩階段，先用二值化佔用矩陣做 XOR 比對並估計欄位平移，以處理反向重訪；再以餘弦相似度比對強度結構，最後以時間一致性與 FPFH 加 ICP 的幾何檢查確認迴圈。論文本身只提出迴圈偵測；官方 ISC-LOAM 程式碼再把它接上 LOAM 系列前端與後端最佳化，成為完整的 LiDAR SLAM。",[4279,4280],"3D LiDAR with intensity (Velodyne VLP-16 on the warehouse AGV; Velodyne HDL-64E in KITTI)","wheel odometry fused with LiDAR odometry for the front-end trajectory in the warehouse test (Sec. IV-B)","not described in the paper","Intensity Scan Context descriptor with two-stage hierarchical re-identification, temporal consistency over N = 5 neighbouring frames, and FPFH plus ICP geometric verification (Sec. III-C; Sec. III-D; Table I)","paper: loop candidates verified by FPFH-based initial alignment followed by ICP (Sec. III-D); the pose-graph back end is not described in the paper. Repository: front end based on LOAM, A-LOAM and F-LOAM and back end based on ISC, with Ceres and GTSAM listed as dependencies (README; LICENSE)","global descriptor: points within Lmax = 50 m binned into 20 sectors and 60 rings (Table I) with the maximum calibrated intensity per bin; two-stage retrieval: XOR-based binary geometry similarity over column shifts, then column-wise cosine similarity of intensity (Sec. III-B; Sec. III-C)","loop-closure pairs; corrected trajectory and map shown qualitatively for the warehouse test (Fig. 5)",[143,161],{"id":4288,"label":4289,"shortName":4290,"title":4291,"year":249,"era":10,"cluster":151,"scope":99,"keyIdeaZh":4292,"sensors":4293,"mapRepresentation":4295,"loopClosure":72,"estimator":4296,"association":4297,"deskew":4298,"outputGeometry":4299,"fulltextStatus":22,"lidarModels":4300,"equipmentCount":736},"floam2021","Wang et al., 2021a","F-LOAM","F-LOAM : Fast LiDAR Odometry and Mapping","F-LOAM 以 LOAM 為基礎，著眼於降低計算量：運動畸變校正改為非迭代的兩階段方法，先以等速模型預測並校正，待位姿最佳化後再重算一次畸變並更新地圖。配準時把特徵點直接對齊全域邊緣地圖與平面地圖，並以局部平滑度作為權重，偏重跨幀較穩定的特徵。系統不含迴圈閉合，地圖以關鍵影格更新並做體素降採樣。",[4294],"3D LiDAR only as method input: Velodyne HDL-64 as written (KITTI, whose cameras and GPS are not used), Velodyne VLP-16 on the warehouse AGV, virtual Velodyne VLP-16 in Gazebo; VICON motion capture used only as indoor ground truth","global edge map and global planar map updated at keyframes and voxel-grid downsampled (Sec. III-D)","Gauss-Newton minimization of smoothness-weighted point-to-edge and point-to-plane distances to global feature maps (Sec. III-C)","LOAM-style smoothness features; global line\u002Fplane estimated from nearby map points by covariance eigen-analysis (KD-tree) (Sec. III-A, III-C)","non-iterative two-stage compensation: constant-velocity prediction before matching, then recomputation with the optimized pose before map update (Sec. III-B, III-D)","global edge\u002Fplanar feature maps and poses (Sec. III-D)",[161,143],{"id":4302,"label":4303,"shortName":4304,"title":4305,"year":249,"era":10,"cluster":151,"scope":99,"keyIdeaZh":4306,"sensors":4307,"mapRepresentation":4309,"loopClosure":72,"estimator":4310,"association":4311,"deskew":637,"outputGeometry":4312,"fulltextStatus":22,"lidarModels":4313,"equipmentCount":109},"sslslam2021","Wang et al., 2021b","SSL_SLAM","Lightweight 3-D Localization and Mapping for Solid-State LiDAR","SSL_SLAM 是針對小視野、高頻率固態 LiDAR（Intel L515）設計的輕量 LiDAR 建圖定位。它先把點雲依垂直與水平角度分格並取格內平均，再以鄰域平滑度擷取邊緣與平面特徵，使特徵在大幅旋轉下仍較一致；位姿以掃描對滑動視窗局部地圖的點到邊、點到面殘差，在李群上以高斯牛頓法求解。全域地圖只用關鍵影格更新成八元樹佔據機率地圖。系統沒有迴圈閉合，在嵌入式小電腦上可即時執行。",[4308],"solid-state LiDAR only (Intel Realsense L515, 70 x 55 deg FoV, 30 Hz)","sliding-window local edge and planar feature maps for odometry; global octree with per-cell occupancy probability updated from key frames selected by translation, rotation or elapsed-time thresholds (Sec. III-B; Sec. III-C)","Gauss-Newton minimization of point-to-edge and point-to-plane residuals against a sliding-window local map, with left-perturbation updates on the Lie group and a constant-velocity initial guess (Sec. III-B; Algorithm 1)","points binned into an M x N grid of vertical and horizontal angle cells (cell means); edge and planar features from a local smoothness over a neighbourhood lambda; 2 nearest edge points or 3 nearest planar points found in k-d trees of the local edge and planar maps of the last q frames (Sec. III-A; Sec. III-B)","trajectory and dense 3D probabilistic octree map (Figs. 1, 4 and 5)",[1592],{"id":4315,"label":4316,"shortName":4317,"title":4318,"year":249,"era":10,"cluster":151,"scope":346,"keyIdeaZh":4319,"sensors":4320,"mapRepresentation":4322,"loopClosure":72,"estimator":4323,"association":4324,"deskew":4325,"outputGeometry":4326,"fulltextStatus":22,"lidarModels":4327,"equipmentCount":185},"pwclonet2021","Wang et al., 2021c","PWCLO-Net","PWCLO-Net: Deep LiDAR Odometry in 3D Point Clouds Using Hierarchical Embedding Mask Optimization","PWCLO-Net 是直接以原始 3D 點雲學習的監督式 LiDAR 里程計。它借用光流網路的金字塔、變形與代價體（PWC）結構：兩幀點雲先經共享權重的點特徵金字塔，再以注意力代價體建立軟對應；可訓練的嵌入遮罩為每個點加權，以回歸整體位姿並壓低動態物體與雜草等不可靠點的影響。估得的位姿用來變形前一幀，再由粗到細逐層修正位姿與遮罩。它只做逐幀里程計，沒有地圖或迴圈閉合。",[4321],"3D LiDAR point coordinates only (KITTI Velodyne; reflectance not used) (Sec. 4.1)","none (frame-to-frame odometry without a map)","supervised end-to-end network: siamese point feature pyramid (set conv with farthest point sampling and kNN), attentive point cost volume, trainable embedding mask, and three pose warp-refinement modules that refine a quaternion and translation over four levels; multi-level supervised loss with learnable weighting (Sec. 3; Sec. 4.2)","soft correspondences from an attentive cost volume between two consecutive frames (8192 randomly sampled points each) instead of explicit point matching; the embedding mask down-weights dynamic and irregular points (Sec. 3.2; Sec. 3.3; Sec. 4.2; Sec. 5.3)","not described; KITTI clouds used as provided","relative 6-DoF pose per frame pair; trajectory by chaining",[45],{"id":4329,"label":4330,"shortName":4331,"title":4332,"year":51,"era":10,"cluster":755,"scope":99,"keyIdeaZh":4333,"sensors":4334,"mapRepresentation":4335,"loopClosure":4336,"estimator":4337,"association":4338,"deskew":57,"outputGeometry":4339,"fulltextStatus":22,"lidarModels":4340,"equipmentCount":24},"coslam2023","Wang et al., 2023a","Co-SLAM","Co-SLAM: Joint Coordinate and Sparse Parametric Encodings for Neural Real-Time SLAM","Co-SLAM 結合多解析度雜湊網格（hash grid）與 one-blob 座標編碼，兼顧收斂速度與表面連續補洞，並以隨機取樣所有關鍵影格光線進行全域光束調整。作者特別指出評估前的網格裁切（mesh culling）策略會改變重建指標，所有方法在其新裁切策略下指標都變好，顯示神經 SLAM 幾何數字高度依賴評估協定。",[772],"multi-resolution hash grid + one-blob coordinate encoding with shallow MLPs predicting TSDF and colour","none (authors suggest incorporating loop closure as future work)","gradient-based per-frame tracking (constant-speed initialization) + global bundle adjustment over rays sampled from all keyframes","direct colour and depth rendering losses + approximate SDF and smoothness losses","mesh via marching cubes, evaluated after mesh culling",[],{"id":4342,"label":4343,"shortName":4344,"title":4345,"year":51,"era":10,"cluster":98,"scope":132,"keyIdeaZh":4346,"sensors":4347,"mapRepresentation":4350,"loopClosure":4351,"estimator":4352,"association":4353,"deskew":4354,"outputGeometry":4355,"fulltextStatus":22,"lidarModels":4356,"equipmentCount":377},"dliom2023","Wang et al., 2023b","D-LIOM","D-LIOM: Tightly-Coupled Direct LiDAR-Inertial Odometry and Mapping","D-LIOM 把 Cartographer 式的直接配準改為與 IMU 緊耦合的 3D 版本：每個去畸變掃描不擷取特徵，直接以高斯牛頓法對齊到 3D 佔據機率子地圖，得到的位姿作為一元 LiDAR 因子，與 IMU 預積分及線上估計的重力先驗因子組成子地圖時間窗內的局部因子圖，同時更新 IMU 偏差並抑制橫滾與俯仰漂移。後端利用重力對齊，把 3D 子地圖投影成 2D 影像，以 SURF 特徵與 FLANN 比對偵測迴圈並由 RANSAC 求得平面位移與旋轉，再以分支定界搜尋高度差、做精細配準後最佳化全域位姿圖。系統也支援多 LiDAR 輸入與 6 軸 IMU 的靜態或動態初始化。",[4348,4349],"one or more 3D spinning LiDARs (16-line RoboSense in the authors' device; two Ouster OS1-16 in NTU VIRAL; two inclined 16-line Velodyne in Complex Urban)","6-axis IMU (built-in consumer-grade IMU at 400 Hz; VectorNav VN100 in VIRAL)","3D probability submaps stored as octrees of occupancy log-odds voxels (Cartographer-style), plus 2D projections for loop detection (Sec. III-E-2)","yes; gravity-aligned 3D submaps projected to 2D images, SURF keypoints matched with FLANN, RANSAC 3-DoF transform, then branch-and-bound search of the vertical offset and fine scan-to-submap registration (Sec. III-F)","front-end local factor graph in GTSAM over the time window of the current submap with LiDAR odometry unary factors from direct scan-to-submap matching, IMU preintegration factors and an online gravity-prior factor on roll and pitch; back-end global sparse pose graph in Ceres (Sec. III-E, IV-A)","direct: raw deskewed points are registered to a 3D probability (occupancy log-odds) submap by Gauss-Newton maximization of voxel probability using map gradients; no feature extraction (Sec. III-E-2)","IMU preintegration between scans deskews each point to the previous body frame; points of auxiliary LiDARs are merged by timestamp with the primary LiDAR (Sec. III-C-2)","trajectory and probability submaps; dense point cloud maps shown qualitatively (Figs. 5-6)",[45,504,4357],"two 16-line Velodyne LiDARs (inclined)",{"id":4359,"label":4360,"shortName":4361,"title":4362,"year":661,"era":10,"cluster":755,"scope":32,"keyIdeaZh":4363,"sensors":4364,"mapRepresentation":4366,"loopClosure":57,"estimator":4367,"association":4368,"deskew":57,"outputGeometry":4369,"fulltextStatus":22,"lidarModels":4370,"equipmentCount":77},"dust3r2024","Wang et al., 2024","DUSt3R","DUSt3R: Geometric 3D Vision Made Easy","DUSt3R 將雙視角三維重建改寫為以 Transformer 直接回歸兩張影像在同一座標系下的逐像素點圖（pointmap），不需要相機內參或位姿；多張影像時以全域對齊合併點圖。訓練時以平均距離正規化點圖，因此輸出沒有公制尺度。作者在 DTU 零樣本測試指出其以回歸取得的幾何精度低於使用真值相機並做三角化的 MVS 方法。",[4365],"monocular camera (unposed, uncalibrated images)","per-pixel pointmaps with confidence","feed-forward Transformer pointmap regression for image pairs; global alignment (not reprojection BA) for >2 views","implicit (regressed pointmaps in a common frame; matches recoverable from pointmaps)","dense point clouds, depth maps, relative\u002Fabsolute camera poses and intrinsics (scale-normalized)",[],{"id":4372,"label":4373,"shortName":4374,"title":4375,"year":207,"era":10,"cluster":31,"scope":32,"keyIdeaZh":4376,"sensors":4377,"mapRepresentation":4378,"loopClosure":4379,"estimator":4380,"association":4381,"deskew":4382,"outputGeometry":4383,"fulltextStatus":22,"lidarModels":4384,"equipmentCount":144},"wang2025planarmesh","Wang et al., 2025a","PlanarMesh","PlanarMesh: Building Compact 3D Meshes from LiDAR using Incremental Adaptive Resolution Reconstruction","PlanarMesh 以「平面網格」表示場景：每個元素由一個平面（位置與法向量，以增量 PCA 更新）和落在該平面上的三角網格組成，頂點半徑近似局部曲率。每個新點經兩棵可動態插入的包圍體階層樹查詢：面相交搜尋（FIS）找出射線穿過的網格面以引入自由空間資訊，反向半徑搜尋（RRS）找出半徑涵蓋該點的邊界頂點；再以 95% 信賴區間的 z 檢定判定點在平面前方、平面上或後方，據以執行更新、擴張、新增、刪除、收縮與合併。輸出前依頂點半徑重新取樣並以 Delaunay 重新三角化，因此大平面用少量大三角形、細部用小三角形。位姿由外部 LiDAR 里程計提供。",[35],"planar-mesh (plane models combined with adaptive-resolution mesh), stored with a bounding volume hierarchy","none (elastic deformation for future loop closure mentioned as future work)","not_applicable (poses from a separate LiDAR odometry such as FastLIO or VILENS; evaluation used ground-truth poses from registering each undistorted scan to the TLS map)","per-point Face Intersection Search and Reverse Radius Search on bounding volume hierarchies to find candidate planar-meshes","delegated to the upstream LiDAR odometry (motion correction assumed in the supplied poses\u002Fscans)","compact triangle mesh saved as binary PLY with 280-550 K faces (about 10 MB) on the tested sequences; before output each planar-mesh is resampled by vertex radius and re-triangulated with Delaunay, restoring concavity by discarding faces outside the original (Sec. III-F, IV-C)",[45],{"id":4386,"label":4387,"shortName":4388,"title":4389,"year":207,"era":10,"cluster":755,"scope":32,"keyIdeaZh":4390,"sensors":4391,"mapRepresentation":4393,"loopClosure":57,"estimator":4394,"association":4395,"deskew":57,"outputGeometry":4396,"fulltextStatus":22,"lidarModels":4397,"equipmentCount":77},"vggt2025","Wang et al., 2025b","VGGT","VGGT: Visual Geometry Grounded Transformer","VGGT 是前饋式 Transformer，可由一張到數百張影像直接推論相機參數、深度圖、點圖與點軌跡，不需後續幾何最佳化；若再加上選用的 BA 後處理，位姿精度還能提升。訓練時以平均點距離正規化真值，模型學到的是固定的標準尺度而非公制尺度。ETH3D 點雲評估前以 Umeyama 演算法對齊真值，但論文未說明是否同時估計尺度，DTU 與 ETH3D 的精度指標也未註明單位，因此這些數值不能直接視為公制幾何誤差。",[4392],"monocular camera (one to hundreds of views)","per-view depth and point maps in the first-camera frame","feed-forward Transformer with alternating frame\u002Fglobal attention predicting cameras, depth, point maps and tracks","implicit (learned); 3D point tracks","point clouds (point head or depth+camera unprojection), camera parameters, depth maps; normalized, non-metric scale",[],{"id":4399,"label":4400,"shortName":4401,"title":4402,"year":30,"era":10,"cluster":599,"scope":676,"keyIdeaZh":4403,"sensors":4404,"mapRepresentation":4406,"loopClosure":4407,"estimator":4408,"association":4409,"deskew":4410,"outputGeometry":4411,"fulltextStatus":22,"lidarModels":4412,"equipmentCount":24},"lemon2026","Wang et al., 2026","LEMON-Mapping","LEMON-Mapping: Loop-Enhanced Large-Scale Multi-Session Point Cloud Merging and Optimization for Globally Consistent Mapping","LEMON-Mapping 指出傳統多機位姿圖最佳化只把迴圈當作位姿節點間約束，忽略地圖幾何，導致重疊區發散與模糊。其框架包含：迴圈處理模組（剔除離群、分類迴圈並召回被誤刪的正確迴圈）；對多機地圖做空間 BA（孤立迴圈用 DBA、成群迴圈用 HBA）以消除重疊區不一致；再以兩步驟位姿圖最佳化，把 BA 精修後的局部精度傳遞到整張地圖。",[4405],"3D LiDAR: Velodyne (S3E), Avia and Ouster (GEODE), Avia (MARS-LVIG, R3LIVE), Mid360 (self-collected) (Tables I-II)","multi-session point-cloud submaps","robust loop outlier rejection with recall of wrongly removed loops","loop processing (outlier rejection, classification, recall) then spatial BA (DBA for isolated loops, HBA for clustered loops) and two-step pose-graph optimization","RING++ provides all intra- and inter-robot loop candidates; statistical outlier removal, GICP initialized by RING++ and RANSAC correspondence rejection with an inlier count and GICP fitness test verify loops; BFS region growing labels loops as clustered or isolated; rejected loops are recalled if their poses are within 2 m after the first PGO; spatial BA windows are spherical radius searches in a pose kd-tree around each loop; RING++ similarity (inter-robot) and a registration-based minimum-eigenvalue test (intra-robot) select sparse BA constraints for the last PGO (Sec. IV, V, VI-B).","not_reported (inputs from LiDAR-inertial odometry)","merged globally consistent point-cloud map and trajectories",[507,45,437,4413],"DJI L1 LiDAR sensor",{"id":4415,"label":4416,"shortName":4417,"title":4418,"year":207,"era":10,"cluster":599,"scope":676,"keyIdeaZh":4419,"sensors":4420,"mapRepresentation":4423,"loopClosure":4424,"estimator":4425,"association":4426,"deskew":4427,"outputGeometry":4428,"fulltextStatus":22,"lidarModels":4429,"equipmentCount":487},"lamm2025","Wei et al., 2025b","LAMM","Large-Scale Multi-Session Point-Cloud Map Merging","LAMM 是離線的多時段光達點雲地圖合併框架，輸入各代理人由前端 SLAM（如 FAST-LIO2）得到的掃描與初始位姿。先以 M-Detector 為基礎，在正向與反向時間序列各做一次遮擋測試的雙向濾波移除動態點；再以 BTC 描述子在各序列的資料庫中搜尋序列內與序列間迴圈；序列間迴圈以「用每個迴圈把另一序列起點投影到本序列座標」後做 RANSAC 聚類來剔除離群；最後依連通性把序列分組，對每組子位姿圖以 GTSAM 做位姿圖最佳化，輸出一或多張全域一致的點雲地圖。",[4421,4422],"3D LiDAR of different scanning patterns: Velodyne (KITTI, WildPlaces), Ouster OS2-128 and Livox Avia (HeLiPR), Hesai 128-line (Shenzhen)","four Hikvision cameras on the Shenzhen backpack, used with R3LIVE to produce colored point clouds","registered LiDAR scans with poses per sequence; output is one merged global point-cloud map per connected group of sequences (Sec. III-B)","Inner-sequence loops are rejected when the initial poses of the matched frames are geographically far apart; for inter-sequence loops, each loop projects the start position of one sequence into the other sequence's frame (Eq. 1) and RANSAC clustering of these projected points removes outliers (Sec. III-E1, Fig. 3).","Offline back end: bidirectional M-Detector dynamic removal per sequence, BTC place recognition, false-loop filtering, connectivity check that splits sequences into sub-pose graphs, and standard pose graph optimization of each sub-pose graph with the first node anchored to a reference pose, solved in GTSAM (Sec. III-B to III-E, Eq. 2).","BTC (Binary Triangle Combined) descriptors stored in one database per loaded sequence; each scan is searched against all databases in descending order for inner- and inter-sequence loops; rough detection by hash-table matching, fine detection by clustering transforms between triangle pairs, geometric verification by point-to-plane distance using fewer than 50 key points (Sec. III-D).","not performed by LAMM; inputs are registered scans and poses from a front-end SLAM such as FAST-LIO2 (Sec. III-A, III-B)","merged multi-session point-cloud map (abstract)",[4430,437,1762,45],"Hesai 128-line LiDAR",{"id":4432,"label":4433,"shortName":4434,"title":4435,"year":235,"era":52,"cluster":236,"scope":132,"keyIdeaZh":4436,"sensors":4437,"mapRepresentation":4438,"loopClosure":4439,"estimator":4440,"association":4441,"deskew":121,"outputGeometry":4442,"fulltextStatus":22,"lidarModels":4443,"equipmentCount":46},"elasticfusion2015","Whelan et al., 2015a","ElasticFusion","ElasticFusion: Dense SLAM Without A Pose Graph","ElasticFusion 以面元（surfel）表示稠密地圖，採用由目前影像對模型（frame-to-model）的稠密追蹤與時間視窗內的面元融合。系統盡量頻繁地做局部模型對模型迴圈閉合，並以隨機蕨（randomised fern）影像編碼偵測全域迴圈，再以非剛性變形直接校正地圖，而不使用位姿圖或事後處理。作者將適用範圍定為房間尺度。",[772],"unordered list of surfels (position, normal, colour, weight, radius, initialisation and last-update timestamps) split into active and inactive sets by a time window; up to 4.8 million surfels in the qualitative scans","local: each frame the active model prediction is registered to the inactive model prediction and accepted if residual, inlier count and covariance eigenvalue checks pass, then the map is deformed and the region reactivated; global: randomised fern database of predicted views at 80x60, matched views registered and accepted only if the resulting deformation is consistent with the map geometry; the fern database can also serve relocalisation, not needed in the evaluated data","Gauss-Newton on E_track = E_icp + 0.1 E_rgb over a three-level coarse-to-fine pyramid (6x6 normal equations via CUDA tree reduction, Cholesky on CPU); loop closures applied through an embedded deformation graph rebuilt each frame by systematic sampling of surfels, nodes connected by initialisation time (k = 4), optimised by Gauss-Newton with sparse Cholesky on the CPU using rotation, regularisation, constraint and pin terms (weights 1, 10, 100, 100)","frame-to-model: point-to-plane ICP with projective data association between the live depth map and the splatted active-model depth prediction, plus photometric intensity error between the live colour image and the splatted model colour prediction (weight 0.1); the same registration is used for model-to-model loop alignment","surfel map (oriented points with radius and colour)",[],{"id":4445,"label":4446,"shortName":4447,"title":4448,"year":235,"era":52,"cluster":236,"scope":132,"keyIdeaZh":4449,"sensors":4450,"mapRepresentation":4451,"loopClosure":4452,"estimator":4453,"association":4454,"deskew":4455,"outputGeometry":4456,"fulltextStatus":22,"lidarModels":4457,"equipmentCount":24},"kintinuous2015","Whelan et al., 2015b","Kintinuous","Real-time large-scale dense RGB-D SLAM with volumetric fusion","Kintinuous 以 GPU 上的循環緩衝（cyclical buffer）讓 TSDF 融合體積隨相機移動，使稠密融合可延伸到無界空間，並結合稠密幾何與光度約束估計位姿。偵測到迴圈後，以 as-rigid-as-possible 空間變形校正已建立的稠密地圖。作者報告可在數百公尺範圍內即時產生全域一致的表面重建。",[772],"GPU TSDF of 512^3 voxels (6 bytes each: truncated float16 distance, uint8 weight, RGB) addressed with modulo arithmetic as a cyclical buffer that shifts with the camera; surface leaving the volume is extracted by axis-aligned raycasts into voxel-grid-filtered cloud slices tied to the pose that caused the shift and incrementally triangulated with Greedy Projection Triangulation; revisited areas are not re-fused","Frames enter the DBoW (SURF) database when a combined rotation and translation motion metric exceeds 0.3; a candidate needs at least 35 FLANN SURF matches, a RANSAC 3-point transform with a 2.0 px reprojection threshold and at least 25% inliers refined by Levenberg-Marquardt, and a final ICP between voxel-downsampled clouds accepted when the mean squared correspondence error is below 0.01; the accepted constraint is added to iSAM and the dense map is corrected by as-rigid-as-possible embedded deformation","Frame-to-model point-to-plane ICP against the raycast TSDF combined with frame-to-frame dense photometric RGB-D alignment in a weighted sum (w_rgbd = 0.1), three-level pyramids, GPU tree reduction and CPU Cholesky solve; loop constraints enter an iSAM pose graph, whose optimised poses and matched SURF points constrain an embedded-deformation optimisation (weights 1, 10, 100, 100) solved by Gauss-Newton with CHOLMOD","dense geometric and photometric; projective data association (conclusion)","not_reported (rolling shutter not modelled; authors note that projective data association limits the camera motions the front-end can handle, which also limits motion blur and rolling-shutter effects, and that real-time correction would add computation; conclusion)","Large-scale dense coloured surface as cloud slices and a triangle mesh; the seven hand-held datasets span 30 to 318 m and about 0.9 to 6.2 million vertices",[],{"id":4459,"label":4460,"shortName":4461,"title":4462,"year":51,"era":10,"cluster":263,"scope":346,"keyIdeaZh":4463,"sensors":4464,"mapRepresentation":4470,"loopClosure":4471,"estimator":4472,"association":4473,"deskew":4474,"outputGeometry":4475,"fulltextStatus":22,"lidarModels":4476,"equipmentCount":471},"vilens2023","Wisth et al., 2023","VILENS","VILENS: Visual, Inertial, Lidar, and Leg Odometry for All-Terrain Legged Robots","VILENS 是針對足式機器人的里程計，以因子圖（factor graph）在固定時間窗內緊耦合（tightly coupled）融合 IMU、腿部運動學、相機與 LiDAR 四種感測器。其關鍵在於把腿部運動學換算的速度預積分成因子，並在狀態中加入可線上估計的線速度偏差，用來吸收足端滑移、地面變形與足部橡膠形變造成的系統性漂移；此偏差需靠與外部感測器的緊耦合才可觀測。LiDAR 部分同時使用高頻的平面與線特徵追蹤，以及低頻、以局部子地圖為目標的 ICP 相對位姿因子。系統只做里程計，不含迴圈閉合，也不輸出全域點雲地圖。",[4465,4466,4467,4468,4469],"IMU (Xsens MTi-100 on ANYmal B300; Epson G365 on ANYmal C100; 400 Hz)","leg kinematics (ANYdrive joint encoders and torque sensors, 400 Hz)","3D LiDAR (Velodyne VLP-16, 10 Hz)","stereo camera (RealSense D435i gray stereo 848x480 at 30 Hz, or Sevensense Alphasense gray stereo 720x540 at 30 Hz)","monocular fisheye camera (FLIR BFS-U3-16S2C-CS RGB, 1440x1080 at 30 Hz, 150 deg diagonal FoV; SUB configuration)","no global map; local ICP submap of scans registered over the last 5 m travelled; plane\u002Fline landmarks inside the factor graph (Sec. IV-F)","none (odometry only; authors state it can be integrated with an external SLAM system, Sec. VI-C)","fixed-lag factor-graph smoothing with iSAM2 in GTSAM (5 s lag), with preintegrated IMU factors and a preintegrated leg-velocity factor whose linear and angular velocity biases are estimated online (Sec. III, IV, V)","visual FAST\u002FKLT feature tracks with reprojection factors and lidar-derived depth; tracked lidar plane and line primitives with anchor-frame residuals; ICP (after Pomerleau et al.) against a local submap of the last 5 m of scans, added as a relative-pose factor at about 2 Hz; DCS robust cost on visual and lidar factors (Sec. IV-D to IV-F, V)","lidar points motion-compensated with the IMU-propagated state, referenced to the closest camera keyframe timestamp (Sec. V-A)","pose and velocity estimates: IMU-propagated at 400 Hz, factor-graph optimized at 10 Hz, ICP-optimized at 2 Hz (Table V); local elevation mapping is done by downstream modules; no exported global point cloud is reported",[143],{"id":4478,"label":4479,"shortName":4480,"title":4481,"year":661,"era":10,"cluster":98,"scope":99,"keyIdeaZh":4482,"sensors":4483,"mapRepresentation":4486,"loopClosure":72,"estimator":4487,"association":4488,"deskew":4489,"outputGeometry":4490,"fulltextStatus":22,"lidarModels":4491,"equipmentCount":109},"lioekf2024","Wu et al., 2024a","LIO-EKF","LIO-EKF: High Frequency LiDAR-Inertial Odometry using Extended Kalman Filters","LIO-EKF 把 KISS-ICP 的點對點配準與傳統誤差狀態擴展卡爾曼濾波結合成緊耦合 LiDAR 慣性里程計。預測步使用慣性導航領域的精確捷聯 INS 機械編排，作者認為 IMU 預測夠準，因此每個掃描只做一次卡爾曼修正，不需迭代。資料關聯的最大對應距離不以經驗設定，而是由 IMU 相對位姿不確定度（經無跡轉換投影到點距）、體素地圖離散化誤差與 LiDAR 測距雜訊三者組合，以三倍標準差自動求得，藉此減少需調整的參數。",[4484,4485],"3D LiDAR (dataset sensors; models not named in the paper)","consumer-grade MEMS IMU","voxel-grid local point map with a maximum number of points per voxel, maintained as in KISS-ICP (Sec. III-B)","classical error-state EKF with a single correction per scan (no iterations), strapdown INS mechanization for prediction and the Kalman-gain reformulation of FAST-LIO for the update (Sec. III-A, III-B)","KISS-ICP-style point-to-point: voxel-subsampled scan matched to a voxel-hashed local map by nearest neighbour within an adaptive threshold tau = 3 sqrt(sigma_p2p^2 + sigma_map^2 + sigma_range^2) that combines IMU relative-pose uncertainty (unscented transform), voxel discretization error and LiDAR range noise (Sec. III-B, III-C)","each scan deskewed with the IMU-predicted pose before registration (Sec. III-B)","odometry at close to IMU rate and a local voxel point map",[45],{"id":4493,"label":4494,"shortName":4495,"title":4496,"year":661,"era":10,"cluster":98,"scope":99,"keyIdeaZh":4497,"sensors":4498,"mapRepresentation":4501,"loopClosure":72,"estimator":4502,"association":4503,"deskew":4504,"outputGeometry":4505,"fulltextStatus":22,"lidarModels":4506,"equipmentCount":487},"voxelmappp2024","Wu et al., 2024b","VoxelMap++","VoxelMap++: Mergeable Voxel Mapping Method for Online LiDAR(-Inertial) Odometry","VoxelMap++ 延伸 VoxelMap：每個 0.5 m 體素只以三自由度參數（a、b、d）與其共變異數表示平面，並以可累加的和式遞增最小平方擬合，降低計算與記憶體。體素內平面在累積 50 點收斂後即丟棄原始點，並以 union-find 與鄰近體素平面做共面檢定（馬氏距離配合卡方 95% 門檻），將多個「子平面」視為同一「父平面」的量測，以共變異數跡加權融合，使整面牆或地板共享一個更準確、共變異數更小的平面。狀態估計沿用 FAST-LIO 與 VoxelMap 的迭代誤差狀態卡爾曼濾波。",[4499,4500],"3D LiDAR, spinning or non-repetitive solid-state (Velodyne VLP-32C in M2DGR; Livox HAP in the authors' data)","IMU (Realsense D435i IMU in M2DGR; ZED 2i built-in IMU in the authors' data)","hash table of 0.5 m voxels, each holding a 3DOF plane (a, b, d) with 3x3 covariance fitted incrementally by least squares; converged planes (after 50 points, raw points discarded) are merged with coplanar neighbours by union-find using a Mahalanobis chi-square test and trace-weighted fusion, so many voxels share one father plane (Sec. III-B, III-C)","iterated error-state Kalman filter as in FAST-LIO and VoxelMap, with point-to-plane observations whose noise combines point and merged-plane covariances (Sec. III-A, III-D)","each point is hashed to its 0.5 m voxel and matched point-to-plane to the root (father) plane of that voxel's union-find node (Sec. III-C, III-D)","FAST-LIO style preprocessing of raw points (backward propagation implied by 'similar to FAST-LIO'; not detailed) (Sec. III-A)","odometry and a compact plane map (merged planes plus unmerged voxel planes); raw points are not kept after voxel convergence",[4507,926],"Livox Hap",{"id":4509,"label":4510,"shortName":4511,"title":4512,"year":207,"era":10,"cluster":755,"scope":99,"keyIdeaZh":4513,"sensors":4514,"mapRepresentation":4517,"loopClosure":4518,"estimator":4519,"association":4520,"deskew":57,"outputGeometry":4521,"fulltextStatus":22,"lidarModels":4522,"equipmentCount":487},"livgs2025","Xiao et al., 2025","LiV-GS","LiV-GS: LiDAR-Vision Integration for 3D Gaussian Splatting SLAM in Outdoor Environments","LiV-GS 以點雲與高斯共有的共變異為橋樑，直接把稀疏 LiDAR 點與連續可微的高斯地圖對齊做前端追蹤，並對 LiDAR 視野外的高斯施加條件約束，使其貼近鄰近可靠高斯。軌跡評估以 R3LIVE 的軌跡作為參考真值，幾何精度只以毫米波雷達跨模態重定位做定性說明。",[4515,4516],"3D LiDAR (Livox Horizon in NTU4DRadLM; Livox Avia in R3LIVE hku_park_00)","monocular camera (640 x 480 in NTU4DRadLM; 1280 x 1024 at 30 Hz in hku_park_00)","3D Gaussians with normal-aware loss and conditional Gaussian constraints outside the LiDAR field of view","none (authors state LiV-GS lacks loop closure detection)","frame-to-map alignment of LiDAR points to Gaussians through shared covariance\u002Fnormal attributes; sliding-window back-end pose and map optimization","point-to-Gaussian (covariance-based plane) matching","Gaussian map and renderings; geometry accuracy assessed only qualitatively via cross-modal radar relocalization",[1482,437],{"id":4524,"label":4525,"shortName":4526,"title":4527,"year":207,"era":10,"cluster":755,"scope":32,"keyIdeaZh":4528,"sensors":4529,"mapRepresentation":4530,"loopClosure":4531,"estimator":4532,"association":4533,"deskew":57,"outputGeometry":4534,"fulltextStatus":22,"lidarModels":4535,"equipmentCount":423},"gslivm2025","Xie et al., 2025","GS-LIVM","GS-LIVM: Real-Time Photo-Realistic LiDAR-Inertial-Visual Mapping with Gaussian Splatting","GS-LIVM 以改良的 SR-LIVO（ESIKF 緊耦合 LiDAR、慣性與視覺里程計）提供位姿，在體素層級以高斯過程回歸（Voxel-GPR）把稀疏且分布不均的 LiDAR 點轉為均勻網格點，並以預測變異數加權計算三維高斯的初始位置與尺度（旋轉設為單位四元數），再以影像、深度差與結構相似損失持續最佳化，在 8 GB 筆電 GPU 上完成大型戶外場景的即時寫實建圖。論文只評估渲染品質與軌跡，作者指出高斯初始化完全依賴點雲，LiDAR 未覆蓋處會出現缺漏。",[35,36,37],"voxel-hashed dense map of 3D Gaussians (position, covariance, opacity, zero-degree SH colour) initialized per voxel subgrid from Voxel-GPR predictions: position as the inverse-variance weighted mean, scale from the diagonal of the weighted covariance, rotation set to the identity quaternion","none; the authors state the tracking method lacks a loop closure detection module and is less accurate than LVI-SAM on some Botanic Garden sequences","ESIKF LiDAR-inertial-visual odometry (tracking thread, output at IMU rate) adopted from SR-LIVO [48] with two changes: original sensor timestamps replace the time-sweep refinement (fixing crashes with spinning LiDAR) and large matrix products are CUDA-accelerated","colour point cloud from LIVO; photometric, SSIM, structure-similarity and delta-depth losses for Gaussian optimization","dense 3D Gaussian map and rendered images; no geometric accuracy metric of the map is reported (evaluation covers PSNR, SSIM, LPIPS, runtime, memory and trajectory RPE and ATE only)",[437,4536,143,507,4537],"Ouster-16","Hesai PandarXT-32",{"id":4539,"label":4540,"shortName":4541,"title":4542,"year":249,"era":10,"cluster":98,"scope":99,"keyIdeaZh":4543,"sensors":4544,"mapRepresentation":4547,"loopClosure":72,"estimator":4548,"association":4549,"deskew":4550,"outputGeometry":4551,"fulltextStatus":22,"lidarModels":4552,"equipmentCount":487},"fastlio2021","Xu & Zhang, 2021","FAST-LIO","FAST-LIO: A Fast, Robust LiDAR-Inertial Odometry Package by Tightly-Coupled Iterated Kalman Filter","FAST-LIO 以緊耦合迭代擴展卡爾曼濾波（iterated extended Kalman filter, iEKF）融合 LiDAR 特徵點與 IMU，並以 IMU 前向傳播與反向傳播（back-propagation）將掃描內每個點補償到掃描結束時刻，以處理運動畸變。作者提出與傳統等價、但計算量取決於狀態維度而非量測維度的卡爾曼增益公式，使大量特徵點可在機載電腦上即時融合。其前端仍沿用 LOAM 式邊緣與平面特徵，地圖為特徵點集合，無迴圈偵測。",[4545,4546],"3D LiDAR (solid-state Livox Avia; Velodyne VLP-16 in LINS data)","IMU (model on the authors' rig not reported; Xsens MTiG-710 in LINS data)","accumulated feature-point map (edge and plane points) organized by k-d tree","tightly-coupled iterated extended Kalman filter on manifold, with an equivalent Kalman-gain formula whose matrix inversion scales with state dimension rather than measurement dimension","LOAM-style edge and planar feature points; point-to-edge \u002F point-to-plane residuals to nearest map features found with a k-d tree","IMU forward propagation plus backward propagation that projects each feature point to the scan-end time","odometry and registered feature-point map; raw-point maps shown in figures; export format not_reported",[437,143],{"id":4554,"label":4555,"shortName":4556,"title":4557,"year":9,"era":10,"cluster":11,"scope":132,"keyIdeaZh":4558,"sensors":4559,"mapRepresentation":4562,"loopClosure":4563,"estimator":4564,"association":4565,"deskew":4566,"outputGeometry":4567,"fulltextStatus":22,"lidarModels":4568,"equipmentCount":109},"xu2019ogmvslam","Xu et al., 2019","OGM-enhanced visual SLAM (Xu et al. 2019)","An Occupancy Grid Mapping enhanced visual SLAM for real-time locating applications in indoor GPS-denied environments","作者以 ORB-SLAM2 的 RGB-D 模式為基礎，建立可用於室內即時定位系統（RTLS）的視覺 SLAM。除了 ORB-SLAM2 原有的稀疏特徵地圖外，系統把 Kinect 點雲在指定高度範圍內切成虛擬雷射掃描，搭配關鍵影格位姿投影成二維佔據格地圖；為避免早期位姿誤差污染地圖，只用關鍵影格建圖，並在固定數量關鍵影格後或迴圈閉合修正後，以最佳化過的全部歷史關鍵影格重建佔據格地圖。系統可在 SLAM 模式與只定位模式間切換，佔據格地圖則讓使用者互動、A* 路徑規劃、依位置觸發的資料蒐集與局部點雲更新成為可能。定位精度以地下室走廊 15 個 AprilTag 標記的位置與間距評估。",[4560,4561],"RGB-D camera (Microsoft Kinect, via ROS openni_launch; depth registered to RGB)","virtual 2D laser scan cut from the Kinect point cloud (pointcloud_to_laserscan)","sparse ORB feature map plus a 2D occupancy grid map (ROS OccupancyGrid); colored RGB-D point cloud built incrementally for the application example","ORB-SLAM2 loop detection and correction; OGM rebuilt with the corrected keyframe poses (Sec. 4.1.3)","ORB-SLAM2 RGB-D (ORB2 RGBD) for 3D camera poses in SLAM or localization-only mode; a separate occupancy-grid mapping node projects keyframe poses to 2D and integrates virtual laser scans with log-odds updates (lfree 5, locc 15), rebuilding the OGM from all optimized historical keyframes periodically and after loop correction (Sec. 4.1)","ORB features and bag-of-words relocalization of ORB-SLAM2; OGM cells updated by Bresenham ray tracing of virtual scan beams (Sec. 4.1.4, 4.2)","not_applicable (RGB-D camera)","2D occupancy grid with real-time 2D camera pose and virtual scan overlay; incremental colored point cloud from RGB-D frames (Sec. 6.3)",[],{"id":4570,"label":4571,"shortName":4572,"title":4573,"year":97,"era":10,"cluster":98,"scope":99,"keyIdeaZh":4574,"sensors":4575,"mapRepresentation":4577,"loopClosure":4578,"estimator":4579,"association":4580,"deskew":4581,"outputGeometry":4582,"fulltextStatus":22,"lidarModels":4583,"equipmentCount":229},"fastlio2_2022","Xu et al., 2022","FAST-LIO2","FAST-LIO2: Fast Direct LiDAR-Inertial Odometry","FAST-LIO2 延續 FAST-LIO 的緊耦合迭代卡爾曼濾波，但取消手工特徵擷取，直接以原始點對地圖中局部平面做點到平面（point-to-plane）配準，使系統較不依賴特定 LiDAR 掃描樣式。地圖以作者提出的增量式 k-d 樹（ikd-Tree）維護，支援逐點插入、刪除、樹上降採樣與平行重建，因而可在里程計頻率同步更新稠密點雲地圖。地圖只保留一個邊長 L 的立方體區域內的點，該區域初始以起點為中心，並在 LiDAR 偵測範圍觸及邊界時移動；系統不含迴圈偵測或全域修正。",[4576,36],"3D LiDAR (solid-state Livox Horizon\u002FAvia and spinning Velodyne VLP-16\u002FHDL-32E in the tested datasets)","dense point map in an incremental k-d tree (ikd-Tree) with on-tree downsampling and box-wise deletion; the map region is a cube of side L initialized around the start position and moved when the LiDAR detection area reaches its border (default L = 1000 m)","none (authors state FAST-LIO2 is an odometry without loop detection or correction, Sec. VI-C)","tightly-coupled iterated Kalman filter on manifold (IKFOM toolbox) inherited from FAST-LIO, state includes LiDAR-IMU extrinsic (dimension 24)","direct: raw (downsampled) points registered without feature extraction; point-to-plane residual to a local plane fitted from 5 nearest map points found in ikd-Tree","IMU forward\u002Fbackward propagation per point (inherited from FAST-LIO)","odometry and registered dense point map inserted at odometry rate; export format not_reported in paper",[437,1482,143,225],{"id":4585,"label":4586,"shortName":4587,"title":4588,"year":769,"era":10,"cluster":236,"scope":132,"keyIdeaZh":4589,"sensors":4590,"mapRepresentation":4592,"loopClosure":4593,"estimator":4594,"association":4595,"deskew":684,"outputGeometry":4596,"fulltextStatus":22,"lidarModels":4597,"equipmentCount":77},"psmslam2017","Yan et al., 2017","PSM SLAM (Probabilistic Surfel Map)","Dense Visual SLAM with Probabilistic Surfel Map","PSM SLAM 以機率面元地圖（Probabilistic Surfel Map）結合逐影格與對模型兩類 RGB-D 視覺 SLAM。地圖中每個點帶有三維位置與 3×3 共變異、強度與其變異量以及法向，新觀測以兩個高斯分布相乘的方式融合，正確關聯的點其不確定度會快速下降，不可靠的點則被移除，因此地圖只保留稀疏而可靠的點。前端沿用 σ-DVO 的光度與幾何混合權重，把每一影格對齊到由全域地圖可見點與新關鍵影格觀測組成的 Keyframe PSM；後端以 g2o 交替最佳化關鍵影格之間的位姿約束，以及以不確定度加權的位姿對地圖點約束（光度式光束法平差）。需要稠密網格時，再以地圖點為控制點調整各關鍵影格的深度圖後融合。",[4591],"RGB-D camera (TUM RGB-D real sequences; ICL-NUIM synthetic sequences with noise; sensor models not named)","Probabilistic Surfel Map: sparse points with 3D position and 3x3 covariance, intensity and intensity variance, and normal; updated by multiplying Gaussian distributions so that uncertainties of consistently associated points shrink; unreliable points pruned","Nearest-neighbour search of keyframe poses in a kd-tree (as in sigma-DVO) adds pose-pose constraints","Keyframe-based dense visual odometry with the sigma-DVO hybrid weighting (Student-t for photometric, sensor-noise model for geometric residuals), aligning each frame to a Keyframe PSM; back end in g2o that alternates pose-pose graph optimization with uncertainty-weighted pose-point constraints (photometric bundle adjustment on PSM points), usually converging within 10 iterations","PSM points projected into the keyframe define active points; the Keyframe PSM merges them with back-projected new observations (pruned by position uncertainty and sampled at 20%); frame residuals are photometric and depth differences at projected points","sparse globally consistent PSM; dense point cloud or mesh on request by deforming each keyframe depth map toward nearby PSM points (Gaussian-weighted KNN) and fusing",[],{"id":4599,"label":4600,"shortName":4601,"title":4602,"year":661,"era":10,"cluster":755,"scope":99,"keyIdeaZh":4603,"sensors":4604,"mapRepresentation":4605,"loopClosure":72,"estimator":4606,"association":4607,"deskew":57,"outputGeometry":4608,"fulltextStatus":22,"lidarModels":4609,"equipmentCount":185},"gsslam2024","Yan et al., 2024","GS-SLAM (Yan et al.)","GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting","GS-SLAM 將三維高斯潑濺（3D Gaussian Splatting）用於 RGB-D 稠密 SLAM：場景由帶不透明度與一階球諧係數的各向異性高斯表示，位姿則透過作者推導的潑濺解析梯度直接最佳化。建圖時依累積不透明度偏低或渲染深度與量測深度不符的像素新增高斯，並降低不在表面附近之漂浮高斯的不透明度；追蹤先以半解析度渲染取得粗略位姿，再只用與深度一致的可靠高斯做全解析度精修。系統沒有迴圈閉合，評估用的網格是由估計位姿與深度經 TSDF 融合而得。",[1405],"anisotropic 3D Gaussians with opacity and first-degree spherical harmonics (12 coefficients); adaptive expansion adds Gaussians at pixels with low cumulative opacity or depth mismatch and suppresses floaters by opacity decay (Sec. 3.1-3.2)","gradient-based pose optimization (Adam on quaternion and translation) through analytical derivatives of differentiable Gaussian splatting; constant-velocity initialization, coarse stage on half-resolution renders, fine stage on full-resolution renders from depth-consistent (reliable) Gaussians (Sec. 3.3, Supp. Sec. 1-2, 5)","direct: L1 photometric loss on rendered colour for tracking; mapping and BA use L1 depth and colour rendering losses (Eqs. 6, 10, 13)","3D Gaussian map with rendered RGB and depth; meshes for evaluation are produced by TSDF fusion of estimated poses and depth, since meshing the Gaussians directly was unsatisfactory (Supp. Sec. 5, Fig. 10)",[],{"id":4611,"label":4612,"shortName":4613,"title":4614,"year":30,"era":10,"cluster":263,"scope":99,"keyIdeaZh":4615,"sensors":4616,"mapRepresentation":4620,"loopClosure":4621,"estimator":4622,"association":4623,"deskew":4624,"outputGeometry":4625,"fulltextStatus":22,"lidarModels":4626,"equipmentCount":736},"yan2026tunnel","Yan et al., 2026a","Deep feature-enhanced LVIO (tunnel)","Deep feature-enhanced LiDAR-visual-inertial odometry for robust mapping in tunnel environment","此研究針對隧道幾何特徵稀疏、結構重複而導致光達里程計退化的問題，提出光達、視覺與慣性融合的里程計。光達端以曲率區分邊緣與平面特徵，採點對線與點對面配準；將資訊矩陣求逆得到共變異數後，分別對旋轉與平移子區塊做特徵分解，以經驗門檻判定六自由度中哪些方向退化。視覺端以 SuperPoint 搭配自適應門檻擷取特徵、以 LightGlue 匹配，並結合 IMU 預積分，但不做後端最佳化。條件式擴展卡爾曼濾波器只在偵測到退化時，以選擇矩陣保留退化方向上的視覺慣性資訊來更新狀態。系統只做里程計，不含迴圈閉合；驗證完全使用 MIT 校園隧道（KMCT）與 WHU-Helmet 隧道、地鐵兩組公開資料。",[4617,4618,4619],"3D LiDAR (mechanical spinning model assumed in the LO derivation; Velodyne on KMCT, Livox on WHU-Helmet)","camera (Intel RealSense D455 RGB-D on KMCT; helmet cameras on WHU-Helmet)","IMU (preintegrated in the VIO)","point cloud map built by accumulating registered scans, with global edge and planar feature maps used for matching (Secs. 3, 3.1.1)","none; the authors argue loop closure is often infeasible in tunnels and choose an odometry design (Secs. 2.3, 3.3.1)","Conditional EKF: LiDAR odometry runs continuously; when covariance-based detection flags degenerate rotation or translation directions, a selection matrix keeps only the visual-inertial information along those directions and the state is updated with a FAST-LIO style Kalman gain; the VIO itself has no back-end optimization (Secs. 3.2, 3.3)","LiDAR: curvature from five horizontal neighbours on each side separates edge and planar points, matched point-to-line and point-to-plane to global feature maps and solved by Gauss-Newton; degeneracy from eigen-decomposition of the rotation and translation blocks of the inverted information matrix against empirical thresholds; visual: SuperPoint features with an adaptive score threshold and LightGlue matching inside an ORB-SLAM3 style tracking thread (Secs. 3.1 to 3.2)","linear interpolation of the inter-scan transform across the sweep by point index (Eq. 3, Sec. 3.1.1)","dense LiDAR point cloud map evaluated by Mean Map Entropy (local consistency only) and, on one KMCT sequence, by voxel error against the ground-truth map (Secs. 4.2.2, 4.3, Fig. 9)",[45],{"id":4628,"label":4629,"shortName":4630,"title":4631,"year":30,"era":10,"cluster":755,"scope":132,"keyIdeaZh":4632,"sensors":4633,"mapRepresentation":4635,"loopClosure":4636,"estimator":4637,"association":4638,"deskew":57,"outputGeometry":4639,"fulltextStatus":22,"lidarModels":4640,"equipmentCount":487},"yan2026_underground3dgsslam","Yan et al., 2026b","Underground RGB-D 3DGS SLAM","RGB-D Perception-Enhanced 3D Gaussian Splatting SLAM: A Robust Framework for Mapping Underground Spaces","本研究提出只用低成本 RGB-D 相機的地下空間 3DGS SLAM。前處理以 MSRCR、側窗濾波與 HIS 色彩空間正規化 gamma 校正增強低照度影像，並以預訓練深度補全網路（非局部傳播架構，未在地下資料上微調）填補深度空洞；追蹤以渲染色彩與深度殘差最佳化位姿，並以混合歐氏距離（0.3 m 或 15 影格）與重疊度兩階段選取關鍵影格；建圖以不透明度門檻（τ0 = 0.15）與加權觀測次數低於 5 剔除高斯；偵測迴圈後以 Levenberg-Marquardt 求解位姿圖，並剛性更新高斯。資料由搭載 Kinect2 的 Autolabor-Pro1 四輪機器人在煤礦巷道、車庫與地下室共 9 個場景（38.4 至 101.3 m²）蒐集，另以 TUM RGB-D 三序列驗證。軌跡真值來自 LiDAR、IMU 與相機融合的 SLAM 流程，並非獨立測量；地圖只以新視角 PSNR、SSIM、LPIPS 評估，作者明言沒有稠密三維幾何真值。9 個場景中本方法 ATE 為 4.8 至 9.6 cm，有 6 個場景在稠密方法中最低或並列最低，但 ORB-SLAM2 在 4 個場景更低、2 個場景持平。",[4634],"RGB-D camera (Kinect2, 512x424 pixels, 10 Hz, 70 x 60 deg field of view)","3D Gaussian ellipsoids (following GSORB-SLAM [56]) with isotropic scale regularization; pruning by an opacity threshold (tau0 = 0.15; the text words it as transparency below zero or above 0.15) and by a weighted observation count below 5","Loop frames detected by cosine similarity and geometric overlap ratio (strategy similar to GLC-SLAM [58]); relative loop poses added as pose-graph edges","Per-frame gradient-based pose optimization against colour and depth rendered from the Gaussian map (pixels with silhouette S(p) > 0.99 and valid depth; weighted colour and depth residuals), then sliding-window Gaussian optimization with poses fixed","Direct photometric and geometric rendering residuals against enhanced RGB and completed depth; keyframes by hybrid Euclidean pose distance (0.3 m or 15 frames) then overlap ratio; loop candidates by cosine similarity and geometric overlap following GLC-SLAM","3D Gaussian map; no 3D geometric accuracy evaluated (authors state dense geometric ground truth is unavailable); map quality reported only by novel-view PSNR, SSIM and LPIPS on held-out non-keyframes",[45],{"id":4642,"label":4643,"shortName":4644,"title":4645,"year":1461,"era":52,"cluster":68,"scope":69,"keyIdeaZh":4646,"sensors":4647,"mapRepresentation":4649,"loopClosure":72,"estimator":4650,"association":4651,"deskew":57,"outputGeometry":389,"fulltextStatus":22,"lidarModels":4652,"equipmentCount":144},"yang2016goicp","Yang et al., 2016","Go-ICP","Go-ICP: A Globally Optimal Solution to 3D ICP Point-Set Registration","Go-ICP 在整個 SE(3) 空間以分支定界（BnB）搜尋點對點 ICP 之 L2 誤差的全域最佳解。旋轉以角軸向量表示於 [-π, π]³ 立方體，平移限定在 [-ξ, ξ]³，兩者都以八元樹細分；作者由旋轉與平移的不確定半徑推導每點殘差的上下界，並採外層旋轉、內層平移的巢狀 BnB。每當找到更好的解就以局部 ICP 精修並更新上界，加快收斂而不失全域最佳性；另以修剪（trimming）處理部分重疊與離群點。在 Stanford bunny 與 dragon 的 2000 次部分對完整配準中全部成功，使用距離轉換時平均約 1.5 至 1.6 秒、最長 28.9 秒（Intel i7 3.4 GHz，1000 個資料點）。作者建議用於不要求即時性的情境，或作為最佳性基準。",[4648],"[\"Kinect (bowl and loom point sets)\", \"structured light 3D scanner (denture point set)\", \"RGB-D depth images from public datasets (camera localization dataset [68], RGB-D Object Dataset [69])\"]","3D point sets","nested best-first branch-and-bound: an outer BnB over rotation (angle-axis cube [-pi, pi]^3 split by octree) calls an inner BnB over translation (cube [-xi, xi]^3); per-point residual bounds come from rotation and translation uncertainty radii; whenever a better cube is found, local ICP is run from it and its result tightens the upper bound; the search stops when the best error minus the lower bound is below epsilon; outliers handled with a trimmed L2 error using Introselect (O(N))","closest point under the L2 residual; for bound evaluation closest distances come either from a kd-tree or, more often in the experiments, from a precomputed 3-D Euclidean distance transform (300 x 300 x 300 grid, approximate); local ICP always uses a kd-tree",[],{"id":4654,"label":4655,"shortName":4656,"title":4657,"year":396,"era":10,"cluster":755,"scope":99,"keyIdeaZh":4658,"sensors":4659,"mapRepresentation":4661,"loopClosure":72,"estimator":4662,"association":4663,"deskew":57,"outputGeometry":4664,"fulltextStatus":22,"lidarModels":4665,"equipmentCount":77},"d3vo2020","Yang et al., 2020a","D3VO","D3VO: Deep Depth, Deep Pose and Deep Uncertainty for Monocular Visual Odometry","D3VO 在直接稀疏里程計（DSO）中三個層次加入深度網路：自監督的 DepthNet 預測深度使新點一開始就有公制尺度，並形成虛擬立體項；網路同時預測光度不確定度，用來取代傳統的殘差權重；PoseNet 預測的相對位姿則在前端追蹤當作先驗因子，在後端光度平差中當作位姿能量項。網路只以立體影片自監督訓練，並預測仿射亮度參數以處理曝光變化。系統仍是無迴圈閉合的單眼里程計，輸出軌跡與稀疏點雲。",[4660],"monocular camera at run time (stereo videos only for self-supervised network training)","sparse point set hosted in keyframes (DSO) with depths initialized from the network (Sec. 3.2, Fig. 2)","DSO-style windowed sparse photometric bundle adjustment (Gauss-Newton) with three learned inputs: DepthNet depth initializes points at metric scale and adds a virtual stereo term, learned photometric uncertainty sets residual weights, and PoseNet relative poses act as front-end tracking prior factors and as a pose energy term in the back end (Sec. 3.2, Supp. C)","direct photometric residuals on sparse points with an 8-pixel neighbourhood pattern and Huber norm, weighted by the learned uncertainty (Sec. 3.2)","camera trajectory and sparse point cloud (Fig. 2)",[],{"id":4667,"label":4668,"shortName":4669,"title":4670,"year":396,"era":10,"cluster":53,"scope":54,"keyIdeaZh":4671,"sensors":4672,"mapRepresentation":57,"loopClosure":4673,"estimator":4674,"association":4675,"deskew":57,"outputGeometry":4676,"fulltextStatus":22,"lidarModels":4677,"equipmentCount":77},"yang2020gnc","Yang et al., 2020b","GNC (GNC-GM \u002F GNC-TLS)","Graduated Non-Convexity for Robust Spatial Perception: From Non-Minimal Solvers to Global Outlier Rejection","作者把穩健估計與離群值過程（outlier process）之間的 Black-Rangarajan 對偶，結合漸進非凸化（graduated non-convexity, GNC），讓任何在無離群值情況下已有非最小解算器（non-minimal solver）的問題，都能延伸為不需初始猜測的穩健求解。每一輪外層迭代固定控制參數 µ，先以非最小解算器做一次加權最小平方的變數更新，再做一次封閉解的權重更新，並逐步把代價函數由凸的替代函數推回 Geman-McClure 或截斷最小平方（TLS）。作者另以平方和（SOS）鬆弛提出形狀對齊的可驗證最佳非最小解算器。實驗涵蓋點雲與網格配準、位姿圖最佳化（PGO）與形狀對齊，離群值皆為人工注入；作者報告可承受約 70 至 80% 離群值，但也明言無法保證全域最佳。",[],"in PGO tests, odometry is kept and loop closures are spoiled with random outliers (random pose pairs with random measurements); rejection works through the GNC weight updates of Sec. III-IV (weights driven toward 0 for inconsistent measurements), while Sec. V-B itself reports only trajectory error and CPU time, not per-edge weights","graduated non-convexity combined with Black-Rangarajan duality: each outer iteration fixes the control parameter mu and performs a single variable update (weighted least squares solved globally by a non-minimal solver: Horn's closed form for point-cloud registration, the certifiably optimal relaxation of Briales and Gonzalez-Jimenez for point-to-point, point-to-line and point-to-plane mesh registration, SE-Sync for PGO, the authors' SOS relaxation for shape alignment) and a single closed-form weight update, with all weights initialised to 1. GNC-GM initialises mu = 2 r_max^2 \u002F c^2 (r_max^2 = largest residual after the first variable update), divides mu by 1.4 per outer iteration and stops when mu falls below 1; GNC-TLS initialises mu = c^2 \u002F (2 r_max^2 - c^2), multiplies mu by 1.4 and stops when the sum of weighted residuals converges; c is set to the maximum error expected for inliers","given putative correspondences (3D point-to-point for point-cloud registration; point-to-point, point-to-line and point-to-plane for mesh registration; 2D-3D keypoint correspondences for shape alignment) or relative-pose measurements (PGO), with outliers allowed; correspondence search and loop detection are outside the method, and all outliers in the experiments are synthetically injected into the given measurement sets","estimated rigid transform (point-cloud and mesh registration), pose-graph trajectory (PGO), or object scale, rotation and 2D translation under weak perspective projection (shape alignment); no map geometry",[],{"id":4679,"label":4680,"shortName":4681,"title":4682,"year":249,"era":10,"cluster":68,"scope":69,"keyIdeaZh":4683,"sensors":4684,"mapRepresentation":4686,"loopClosure":4687,"estimator":4688,"association":4689,"deskew":57,"outputGeometry":4690,"fulltextStatus":22,"lidarModels":4691,"equipmentCount":144},"yang2021teaser","Yang et al., 2021","TEASER \u002F TEASER++","TEASER: Fast and Certifiable Point Cloud Registration","TEASER 以截斷最小平方（Truncated Least Squares, TLS）成本處理大量錯誤對應，並以旋轉平移不變量的圖論框架將尺度、旋轉與平移分解後依序求解。尺度與平移以自適應投票求解，旋轉以半正定鬆弛（TEASER）或逐步非凸化（TEASER++）求解，並用最大團剔除大量離群對應。TEASER++ 另以 Douglas-Rachford 分裂計算最佳性證書，可用來辨識不可靠的配準結果。在 3DMatch 以 3DSmoothNet 對應點測試時，TEASER++ 平均 0.059 秒，八個場景中除 MIT Lab 外成功率都不低於 RANSAC-10K；經證書篩選（CERT）後成功率更高，但證書計算平均需 238 秒。場景對稱或正確對應少於 3 組時仍會失敗。",[4685],"[\"RGB-D (3DMatch scans in experiments)\", \"RGB-D (large-scale hierarchical multi-view RGB-D object dataset [36], object pose tests)\", \"sensor-agnostic 3D correspondences\"]","3D point correspondences","authors suggest certified registration for loop-closure validation in SLAM (Sec. XI)","truncated least squares; adaptive voting for scale and translation; max-clique pruning on invariant measurements; SDP relaxation (TEASER) or GNC with Douglas-Rachford certification (TEASER++) for rotation (abstract, Sec. X)","putative correspondences from FPFH for object pose (Sec. XI-D) or 3DSmoothNet descriptors with nearest-neighbour matching on the 5,000 provided keypoints per 3DMatch scan (Sec. XI-E); all-to-all hypotheses (|A| x |B| about 10^4 for 100-point clouds) in the correspondence-free test (Sec. XI-C)","scale, rotation, translation with optimality certificate",[],{"id":4693,"label":4694,"shortName":4695,"title":4696,"year":97,"era":10,"cluster":755,"scope":99,"keyIdeaZh":4697,"sensors":4698,"mapRepresentation":4700,"loopClosure":4701,"estimator":4702,"association":4703,"deskew":57,"outputGeometry":4704,"fulltextStatus":22,"lidarModels":4705,"equipmentCount":144},"voxfusion2022","Yang et al., 2022","Vox-Fusion","Vox-Fusion: Dense Tracking and Mapping with Voxel-based Neural Implicit Representation","Vox-Fusion 將神經隱式表面與傳統體素融合結合：場景以八元樹（octree）與 Morton 編碼管理的稀疏體素表示，體素頂點存放共享的特徵向量，再由多層感知器解碼成 SDF 與顏色。新影格的深度點雲一旦落在既有體素之外就即時配置新體素，因此不必預先知道場景範圍，記憶體也只花在有觀測的表面附近。追蹤時固定地圖，只以可微分體積渲染最佳化相機位姿；建圖時則對隨機挑選的關鍵影格視窗聯合最佳化地圖與位姿，但系統沒有迴圈閉合或全域最佳化。",[4699],"RGB-D camera (synthetic Replica, ScanNet, iPhone 13 Pro and iPad Pro (2020) with onboard LiDAR depth)","sparse voxel grid (voxel size 0.2 m) in an octree with Morton coding, 16-D embeddings on voxel vertices shared by neighbours, decoded by an MLP into SDF and colour; voxels allocated on the fly from back-projected depth, so no scene bound is needed (Sec. 4.4, Sec. 5.1)","none (the authors list drift in long-time tracking as unresolved, Sec. 7)","gradient-based 6-DoF pose optimization in se(3) through differentiable SDF volume rendering against a frozen copy of the map (zero-motion initialization); mapping jointly optimizes decoder, voxel embeddings and poses of a random keyframe window (Sec. 4.2-4.3)","direct: rendered colour and depth losses plus free-space and SDF losses on sparsely sampled pixels whose rays hit allocated voxels (Sec. 4.1)","camera trajectory and SDF-based surface mesh (evaluated as mesh against ground-truth mesh); rendered colour and depth images (Sec. 5)",[],{"id":4707,"label":4708,"shortName":4709,"title":4710,"year":661,"era":10,"cluster":599,"scope":300,"keyIdeaZh":4711,"sensors":4712,"mapRepresentation":4713,"loopClosure":4714,"estimator":4715,"association":4716,"deskew":121,"outputGeometry":4717,"fulltextStatus":22,"lidarModels":4718,"equipmentCount":185},"yang2024lifelong","Yang et al., 2024","Lifelong 3D Mapping Framework (hand-held & robot-mounted)","Lifelong 3D Mapping Framework for Hand-Held & Robot-Mounted LiDAR Mapping Systems","此框架針對手持與機器人搭載光達建圖系統，串接四個模組：以 OctoMap 為基礎，加入子地圖多平面 RANSAC 回填、K 近鄰投票與半徑搜尋後處理的動態點移除；以 PCA-SHOT 特徵配對與 RANSAC 粗對齊、再以 NDT 精配準的多時段地圖對齊（六個參數以網格搜尋並取 Chamfer 距離最小者）；先以 k 近鄰半徑搜尋區分共存、重疊與非重疊區域，再比較沿平面法向投影的 2D 俯視最大高度描述子，找出正負變化；最後以類似 Git 的版本控制只保存一張基準地圖、各時段正負差異與邊界點，可重建任一時段乾淨地圖並查詢任兩時段差異。",[35],"clean static session maps; single base map plus stored positive\u002Fnegative changes and boundary points (version control)","none (map-to-map rigid alignment)","not_applicable (poses supplied by external SLAM or commercial device software)","PCA-SHOT keypoint descriptors with RANSAC for initial alignment, then NDT fine registration; grid search over six parameters selected by lowest Chamfer distance","clean static maps, aligned maps, positive\u002Fnegative change point sets, reconstructable past session maps",[],{"id":4720,"label":4721,"shortName":4722,"title":4723,"year":9,"era":10,"cluster":151,"scope":99,"keyIdeaZh":4724,"sensors":4725,"mapRepresentation":4727,"loopClosure":72,"estimator":4728,"association":4729,"deskew":4730,"outputGeometry":4731,"fulltextStatus":22,"lidarModels":4732,"equipmentCount":109},"liomapping2019","Ye et al., 2019","LIO-mapping (LIOM)","Tightly Coupled 3D Lidar Inertial Odometry and Mapping","LIO-mapping 在滑動視窗內以固定延遲平滑器（fixed-lag smoother）與邊緣化，將 IMU 預積分與 LiDAR 平面特徵的點到面殘差聯合最佳化，並同時線上估計 LiDAR-IMU 外參。里程計之後再做「旋轉約束」的全域地圖配準：利用里程計對 roll、pitch 的較佳估計修改最佳化，使地圖持續與重力方向對齊。論文指出此法需要足夠 IMU 激勵才能初始化。",[1201,4726],"IMU (Xsens MTi-100, 400 Hz)","global feature point cloud map as by-product of refinement (Sec. V)","fixed-lag smoother over a sliding window with marginalization (MAP, Gauss-Newton via Ceres) jointly optimizing IMU states and lidar-IMU extrinsics; followed by rotation-constrained refinement against the global map (Sec. IV-E, V)","LOAM-style features; only planar features used in odometry; KNN plane fitting in a local map built in the pivot frame (relative lidar measurements) (Sec. IV-B, IV-C)","IMU-propagated motion with a linear motion model interpolates each point to the sweep end (Sec. IV-B)","global point cloud map and poses at IMU rate (Sec. V, VII-C)",[143],{"id":4734,"label":4735,"shortName":4736,"title":4737,"year":51,"era":10,"cluster":11,"scope":12,"keyIdeaZh":4738,"sensors":4739,"mapRepresentation":4740,"loopClosure":4741,"estimator":4742,"association":4743,"deskew":121,"outputGeometry":4744,"fulltextStatus":22,"lidarModels":4745,"equipmentCount":185},"yin2023semanticbimloc","Yin et al., 2023","Semantic localization on BIM maps","Semantic localization on BIM-generated maps using a 3D LiDAR sensor","作者將 BIM 依樓層拆分，經 IfcOpenShell 轉為網格後取樣成帶有構件類別的語意點雲地圖，免除事先以 SLAM 建圖。定位時先做點對面 ICP，再依語意一致性篩選並加權的 ICP 精化位姿。實驗在新加坡國立大學六層校舍（已完工使用）進行，參考軌跡取自離線 2D Cartographer SLAM，非獨立測量。",[1201],"storey-wise semantic point cloud sampled from BIM meshes (IfcOpenShell -> OBJ -> sampling at 30 points\u002Fm3 as written), labelled with the 13 BIM categories extracted with Dynamo, using non-oriented (axis-aligned) Dynamo bounding boxes searched with a k-d tree; about 40% walls, 20% floors and 20% curtain panels (Sec. 3.2, 4.1, 4.2)","none (localization only)","frame-to-map registration: point-to-plane ICP then semantic-weighted point-to-plane ICP with Huber kernel; previous pose as initial guess; no odometry (Alg. 2, Sec. 3.3)","coarse point-to-plane ICP, then each scan point is labelled only if all K nearest map points share one BIM category (Eq. 7); only floors, walls and columns are kept (chosen from the Seq. 3-1 test in Table 3); weighted point-to-plane ICP with w = w_c w_rho, semantic weight mu = 0.8 and Huber threshold delta = 0.05 m; about 3% of raw points reach the final step (Sec. 3.3, 4.3, Table 5)","pose trajectory in BIM frame (no new map)",[143],{"id":4747,"label":4748,"shortName":4749,"title":4750,"year":249,"era":10,"cluster":151,"scope":132,"keyIdeaZh":4751,"sensors":4752,"mapRepresentation":4754,"loopClosure":4755,"estimator":4756,"association":4757,"deskew":4758,"outputGeometry":4759,"fulltextStatus":22,"lidarModels":4760,"equipmentCount":144},"litamin2_2021","Yokozuka et al., 2021","LiTAMIN2","LiTAMIN2: Ultra Light LiDAR-based SLAM using Geometric Approximation applied with KL-Divergence","LiTAMIN2 把每次 LiDAR 掃描的點投票到較大的體素（實驗採 3 m），每個體素只以一個常態分布近似，使參與配準的點數降到原始掃描的約 0.5%。為了在點數大減後維持精度，它在 ICP 成本中引入對稱 KL 散度：除了以共變異數加權的距離項，還加入比較兩個分布形狀的項，並以牛頓法求解。迴圈閉合與圖最佳化沿用前作 LiTAMIN，但迴圈約束改用新的成本計算。在 KITTI 上里程計可達每秒數百至上千幀，精度與 SuMa 相近。",[4753],"3D spinning LiDAR only (Velodyne HDL-64E S2 in KITTI)","voxel map of normal distributions (mean and covariance per voxel) (Sec. III-A; Sec. III-C)","yes; implemented as in LiTAMIN, with the proposed ICP cost used to compute loop constraints; detection details are deferred to LiTAMIN (Sec. III-C)","Newton's method with full Hessian (no Levenberg-Marquardt damping) on a Frobenius-normalized symmetric KL-divergence cost: covariance-weighted point distance term plus a distribution-shape term, each with a robust weight (sigma_ICP = 0.5, sigma_Cov = 3, lambda = 1e-6) (Sec. III-B; Sec. III-C)","input points voted into voxels (3 m in the experiments) and each voxel approximated by one normal distribution; distribution-to-distribution matching with correspondences found by k-d tree search in a voxel map of distributions (Sec. III-A; Sec. III-C; Sec. IV-C)","not described; KITTI clouds are already de-skewed and were fed directly to all methods (Sec. IV-B)","trajectory and voxelized normal-distribution map; Fig. 1 colours distributions by normal direction",[161],{"id":4762,"label":4763,"shortName":4764,"title":4765,"year":249,"era":10,"cluster":397,"scope":1035,"keyIdeaZh":4766,"sensors":4767,"mapRepresentation":57,"loopClosure":57,"estimator":4770,"association":4771,"deskew":57,"outputGeometry":4772,"fulltextStatus":22,"lidarModels":4773,"equipmentCount":46},"yuan2021lidarcameracalib","Yuan et al., 2021","livox_camera_calib","Pixel-Level Extrinsic Self Calibration of High Resolution LiDAR and Camera in Targetless Environments","本法不用棋盤格，而以自然場景中的邊緣特徵對齊 LiDAR 與相機。作者依 LiDAR 量測原理分析：深度不連續邊緣受前景與背景混合影響不可靠，因此改以體素切分與平面擬合取得深度連續邊緣，並分析邊緣分布對校正精度的敏感度。報告在多種室內外場景達到像素級精度。",[4768,4769],"Livox Avia solid-state LiDAR (non-repetitive scanning, 20 s accumulation) with Intel RealSense D435i camera (main suite)","Ouster OS2-64 spinning LiDAR with MV-CA013-21UC industrial camera (Sec. IV-C; detailed results in supplementary material)","rough calibration by alternating grid search (0.5° rotation, 2 cm translation) maximising the percentage of edge correspondences, then iterative maximum-likelihood point-to-edge reprojection estimation on SE(3) weighting LiDAR range and bearing noise and 1.5-pixel image edge noise; calibration covariance from the inverse Hessian","depth-continuous LiDAR edges from voxel cutting (e.g., 1 m outdoor, 0.5 m indoor), repeated RANSAC plane fitting and intersection of connected plane pairs at 30° to 150°; points sampled on each edge projected and matched to the κ nearest Canny edge pixels in a 2-D k-d tree, with a direction-orthogonality check","LiDAR-camera extrinsic for point cloud colorization",[437,4774],"Ouster OS2-64",{"id":4776,"label":4777,"shortName":4778,"title":4779,"year":97,"era":10,"cluster":98,"scope":99,"keyIdeaZh":4780,"sensors":4781,"mapRepresentation":4786,"loopClosure":72,"estimator":4787,"association":4788,"deskew":4789,"outputGeometry":4790,"fulltextStatus":22,"lidarModels":4791,"equipmentCount":1289},"yuan2022voxelmap","Yuan et al., 2022","VoxelMap","Efficient and Probabilistic Adaptive Voxel Mapping for Accurate Online LiDAR Odometry","VoxelMap 把空間切成以雜湊表索引的根體素，每個根體素再以八元樹由粗到細細分，直到內部點足以擬合一個平面；每個平面同時估計參數與共變異數，共變異數來自 LiDAR 測距與方位雜訊及位姿估計誤差的傳播。新點以考慮點與平面不確定性的點對面距離配準，並在迭代擴展卡爾曼濾波（IEKF）中形成最大後驗估計。",[4782,4783,4784,4785],"Velodyne HDL-64E S2 mechanical LiDAR (KITTI, 10 Hz, 360°×32°)","Intel RealSense L515 solid-state LiDAR (30 Hz, 70°×55°)","Livox Avia non-repetitive solid-state LiDAR (10 Hz, 70°×77°)","Livox Avia built-in IMU at 200 Hz, used only in the LiDAR-inertial experiment","adaptive coarse-to-fine voxel map: hash table of root voxels at the coarse resolution, each split as an octree until its points pass a planarity test (minimum eigenvalue below a threshold) or the maximum layer is reached; each (sub)voxel holds one plane (normal, center) with covariance; 3 m root voxels with 3 layers (minimum 0.375 m) for KITTI and Livox Avia, 0.5 m with 2 layers for L515; a simulation in Fig. 4 (point noise variance 0.1 m^2) shows the normal covariance converging once about 50 points are reached, and after convergence the method discards the historical points, keeps the plane parameters and covariance, and uses the latest 10 points to detect change and trigger reconstruction","iterated extended Kalman filter on the IKFoM framework, similar to FAST-LIO2, solved as a MAP problem; prior from a constant-velocity model (LiDAR-only: KITTI and L515) or IMU propagation (LiDAR-inertial: Livox Avia); observation noise of each residual propagated from plane covariance and raw point noise","point-to-plane: the predicted world point finds its root voxel by hash key, all sub-voxel planes are polled; a match is accepted if the point-to-plane distance lies within 3σ of its distribution (σ from plane covariance and point covariance), the most probable plane is chosen when several pass, and points passing no test are discarded","KITTI: in-frame motion already compensated in the dataset, plus a 0.22° vertical angle correction as in IMLS-SLAM (Sec. IV-A); L515: no motion compensation described, the constant-velocity model only provides the state prior (Sec. IV-B, III-D); Livox Avia: built-in IMU compensates motion distortion and provides the prior, similar to FAST-LIO2 (Sec. IV-C)","LiDAR poses and a probabilistic plane-feature voxel map; an aggregated colored point cloud of registered scans is shown for the L515 warehouse run (colors from the L515, Supp. Fig. 1); raw points inside a voxel are discarded once its plane converges, so the plane map does not keep all raw points; export format not_reported",[161,1592,437],{"id":4793,"label":4794,"shortName":4795,"title":4796,"year":51,"era":10,"cluster":263,"scope":99,"keyIdeaZh":4797,"sensors":4798,"mapRepresentation":4801,"loopClosure":465,"estimator":4802,"association":4803,"deskew":4804,"outputGeometry":4805,"fulltextStatus":22,"lidarModels":4806,"equipmentCount":109},"sdvloam2023","Yuan et al., 2023a","SDV-LOAM","SDV-LOAM: Semi-Direct Visual-LiDAR Odometry and Mapping","SDV-LOAM 把視覺與 LiDAR 分成前後兩個模組：視覺模組是半直接法深度增強視覺里程計，先以光度誤差直接估計位姿，再做帶傳播的點匹配與重投影修正，並以滑動視窗光束法平差最佳化，追蹤點的深度直接取自投影的 LiDAR 點；其位姿作為 LiDAR 模組的運動先驗。LiDAR 模組以 CT-ICP 為基礎，將 10 Hz 掃描重組成與 60 Hz 影像同步的片段，並依地面與垂直約束比例在 3 自由度與 6 自由度的掃描對地圖最佳化之間切換，以減少垂直方向漂移。",[4799,4800],"monocular grayscale camera","3D LiDAR (Velodyne HDL-64E on KITTI; VLP-16 on the authors' rig)","voxel map of LiDAR points for sweep-to-map registration; global point cloud map output (Fig. 8)","semi-direct depth-enhanced visual odometry built on DSO ideas: direct photometric pose estimate, then point matching with propagation, reprojection-based refinement and sliding-window bundle adjustment with marginalization; its pose is the motion prior for a CT-ICP-based LiDAR odometry with adaptive sweep-to-map optimization (Secs. V-VI)","visual tracking points take depth from projected LiDAR points without depth interpolation, plus extra high-gradient points without LiDAR depth; LiDAR sweep-to-map point-to-plane registration against a voxel map (20 points per voxel) switching between 3-DoF and 6-DoF optimization by the ratio of ground to vertical constraints (threshold 0.8) (Secs. V-VI)","not described as a separate step; the LiDAR module is based on CT-ICP (inference: intra-sweep motion is handled by its continuous-time model)","vehicle trajectory and global LiDAR point cloud map",[143,161],{"id":4808,"label":4809,"shortName":4810,"title":4811,"year":51,"era":10,"cluster":599,"scope":600,"keyIdeaZh":4812,"sensors":4813,"mapRepresentation":4814,"loopClosure":4815,"estimator":4816,"association":4817,"deskew":4818,"outputGeometry":57,"fulltextStatus":22,"lidarModels":4819,"equipmentCount":46},"std2023","Yuan et al., 2023b","STD","STD: Stable Triangle Descriptor for 3D place recognition","STD 在由數次掃描累積而成的關鍵影格上，先以體素共變異數矩陣的特徵值判斷平面並以區域成長擴展，再把平面邊界體素中的點投影到所屬平面形成影像，取 5×5 鄰域極大值作為關鍵點；每個關鍵點以 kd-tree 取 20 個近鄰組成三角形，三邊長與三個法向量內積共六個屬性對剛體變換不變，作為雜湊鍵投票檢索前 10 個候選關鍵影格；再以三角形頂點對應經 SVD 與 RANSAC 求相對位姿，並以平面重合比例做幾何驗證，可選擇以 STD-ICP 精化位姿。方法支援非重複掃描的固態光達。",[35],"per-keyframe triangle descriptor hash database plus extracted planes","hash-table voting selects the top-10 candidate keyframes; a candidate is accepted when the plane coincidence percentage after the RANSAC transform exceeds sigma_pc (0.5 to 0.6 suggested from KITTI08); returns a 6-DoF relative pose, with loop correction left to the host SLAM back-end","not_applicable for odometry; the loop relative pose is solved in closed form by SVD from the three matched triangle vertices inside RANSAC (maximizing correctly matched descriptors), optionally refined by STD-ICP, a Ceres optimization of plane normal difference and point-to-plane distance between coinciding planes","triangle descriptors from keypoints on plane boundaries; hash table on rotation\u002Ftranslation-invariant side lengths and normal dot products; top-10 candidates; RANSAC and plane-based geometric verification","not_reported (motion compensation is not discussed); the component receives scans already registered by an external LiDAR odometry and accumulates 10 scans per keyframe for spinning LiDARs and 20 for Livox LiDARs",[437,1482,45],{"id":4821,"label":4822,"shortName":4823,"title":4824,"year":661,"era":10,"cluster":263,"scope":99,"keyIdeaZh":4825,"sensors":4826,"mapRepresentation":4830,"loopClosure":4831,"estimator":4832,"association":4833,"deskew":4834,"outputGeometry":4835,"fulltextStatus":22,"lidarModels":4836,"equipmentCount":1289},"srlivo2024","Yuan et al., 2024","SR-LIVO","SR-LIVO: LiDAR-Inertial-Visual Odometry and Mapping With Sweep Reconstruction","SR-LIVO 以掃描重組（sweep reconstruction）把光達點流重新切段，使每段掃描的結束時間對齊影像擷取時間，讓較可靠的 LIO 直接估計每張影像當下的位姿。視覺模組因此不再負責狀態估計，只最佳化相機內參、外參與時間偏移，並沿用 R3LIVE 的方式為地圖點上色。",[4827,4828,4829],"3D LiDAR (16-channel OS1 gen1 on NTU-VIRAL; LiDAR written 'LiVOX AVAI' on the R3Live data)","IMU (internal IMU of each LiDAR; LiDAR-IMU extrinsics treated as exact, camera-IMU extrinsics optimized online)","camera (left grayscale camera on NTU-VIRAL; camera of the R3Live handheld rig)","hash voxel map (as in CT-ICP) with RGB-colored points","none; listed as future work (Sec. VII)","ESIKF in the LIO module for all state estimation; separate ESIKF in the vision module optimizing camera intrinsics, extrinsics and time offset only","LIO registration is identical to the authors' SR-LIO and is not re-described (Sec. IV); the vision module tracks map points with Lucas-Kanade optical flow and updates only camera parameters, first by PnP reprojection error and then by photometric error, and renders map-point colors with the R3LIVE rendering function (Sec. V-B)","LiDAR points motion-compensated with the IMU-integrated pose; camera images undistorted with offline-calibrated distortion parameters (Sec. III-B)","dense RGB-colored point cloud map (grayscale on NTU-VIRAL) (Sec. VI-E)",[4837,437],"16-channel OS1 gen1",{"id":4839,"label":4840,"shortName":4841,"title":4842,"year":30,"era":10,"cluster":755,"scope":99,"keyIdeaZh":4843,"sensors":4844,"mapRepresentation":4847,"loopClosure":4848,"estimator":4849,"association":4850,"deskew":57,"outputGeometry":4851,"fulltextStatus":22,"lidarModels":4852,"equipmentCount":46},"yuan2026_adaptive3dgsslam","Yuan et al., 2026","Adaptive 3DGS-SLAM (indoor digital twinning)","Adaptive 3DGS-SLAM-driven incremental online geometric digital twinning complex indoor built environments","本研究針對室內建成環境的線上幾何數位孿生更新，以 MonoGS 預先建立的基準 3DGS 模型為先驗，提出自適應 3DGS-SLAM：新進 RGB-D 影格先以渲染比對方式對齊基準模型求位姿，再以高斯模糊後的滑動視窗 SSIM 產生變化遮罩，變化像素比例超過 2% 且符合原關鍵影格準則者才成為更新用關鍵影格；接著在 CUDA 反向傳播中累計各高斯對變化像素的貢獻，將高貢獻高斯及其鄰域隨機軟剪除一半，並只由遮罩區的 RGB-D 像素加入新高斯。ReplicaCAD 模擬中，相較 MonoGS 直接更新，每影格平均時間由 2.730 秒降至 1.162 秒（降低 57.4%），測試視角 SSIM 由 0.6537 升至 0.7542，ATE RMSE 約 0.105 m；在約 20 平方公尺實驗室以 Intel RealSense D435 實測，每影格時間降低 21.5%，SSIM 由 0.80 升至 0.84。全文未評估點雲或表面幾何精度，也沒有閉環與光束法平差。",[4845,4846],"RGB-D camera: Intel RealSense D435, 1280 x 720 px at 30 frames per second (real-world case)","simulated RGB-D frames from ReplicaCAD (computer experiment)","3D Gaussians (MonoGS map); each Gaussian's contribution to changed pixels (alpha times transmittance, accumulated in the modified CUDA backward pass) selects candidates; candidates and neighbours within radius r are soft-pruned at random with probability 0.5 (opacity set to 0.01); new Gaussians are added only from masked RGB-D pixels and are protected while their source keyframe is in the window buffer","none (authors state the framework has no loop closure detection)","frame-to-model tracking inherited from MonoGS: each frame's 6-DoF pose (unit quaternion and translation), initialised from the previous pose, is optimised by minimising a rendering-based loss between the captured image and the baseline 3DGS rendering (Eq. 4: L1 plus SSIM terms; Sec. 4.6 states RGB-D gives joint photometric and geometric supervision); no backend bundle adjustment","dense rendering-based alignment (no feature matching); change detection by sliding-window SSIM between Gaussian-blurred rendered and captured images, pixels below a threshold form a change mask, and a frame becomes an update keyframe only if the masked ratio exceeds 2% and the MonoGS keyframe criteria are also met","updated 3DGS model (about 128,251 Gaussians in the ReplicaCAD run, about 70% kept from the baseline); no point cloud or surface accuracy is evaluated; outputs assessed by rendering metrics (PSNR, SSIM) and ATE RMSE on simulation",[],{"id":4854,"label":4855,"shortName":4856,"title":4857,"year":150,"era":10,"cluster":397,"scope":1035,"keyIdeaZh":4858,"sensors":4859,"mapRepresentation":57,"loopClosure":57,"estimator":4861,"association":4862,"deskew":57,"outputGeometry":4863,"fulltextStatus":22,"lidarModels":4864,"equipmentCount":185},"yun2018reflection","Yun & Sim, 2018","Glass reflection removal","Reflection Removal for Large-Scale 3D Point Clouds","地面雷射掃描遇到玻璃時，同一雷射脈衝可能同時產生玻璃點、穿透點，以及經玻璃反射而落在玻璃後方的虛像點。本法利用 RIEGL VZ-400 的多回波特性，把單位球面切成約 3×3 個脈衝的面片並計算投影點數，以兩成分高斯混合模型與 EM 分出一般與玻璃面片，再以 RANSAC 擬合最靠近掃描儀的主要玻璃平面，並依距離計算可靠度。玻璃平面後方的點若經 Householder 鏡射後能找到位置相近且 FPFH 特徵相似的真實點，就判為虛像點，最後以資料項加鄰點平滑項組成的標記能量函數，用 ICM 求解後移除虛像點。以六個含玻璃的戶外場景點雲作定性驗證，每個模型處理時間約 55 至 135 s。",[4860],"RIEGL VZ-400 terrestrial laser scanner with multiple echo returns, angular resolution 0.06° x 0.06°","two-component Gaussian mixture (EM) on per-patch point counts; RANSAC glass-plane fit with distance-weighted reliability; per-point score from reflection-symmetry distance and FPFH Hellinger similarity; binary labelling by an energy with data and neighbour smoothness terms (MRF-type; the paper does not use the term) minimised with ICM","k-d tree nearest real point to the Householder-mirrored position of each candidate behind the glass plane; FPFH on 50 nearest neighbours; 48-neighbour smoothness term limited to 0.1% of the bounding-box diagonal","point cloud with virtual reflection points removed",[],{"id":4866,"label":4867,"shortName":4868,"title":4869,"year":249,"era":10,"cluster":236,"scope":99,"keyIdeaZh":4870,"sensors":4871,"mapRepresentation":4873,"loopClosure":4874,"estimator":4875,"association":4876,"deskew":684,"outputGeometry":4877,"fulltextStatus":22,"lidarModels":4878,"equipmentCount":77},"manhattanslam2021","Yunus et al., 2021","ManhattanSLAM","ManhattanSLAM: Robust Planar Tracking and Mapping Leveraging Mixture of Manhattan Frames","ManhattanSLAM 是只用 CPU 的室內 RGB-D SLAM。每一影格擷取 ORB 點、LSD 線段與深度圖中的平面；只要找到兩或三個相互垂直的平面就組成一個曼哈頓座標系（Manhattan Frame），並把場景視為多個曼哈頓座標系的混合。若目前觀測到的曼哈頓座標系先前已存入曼哈頓地圖，就直接由兩次觀測求出無漂移的旋轉，平移再由點、線、面特徵最佳化；若場景不符合曼哈頓假設，則以點、線、面及平行與垂直平面約束估計完整六自由度位姿。稠密建圖沿用 Dense Surfel Mapping 的超像素面元，但平面區域改以稀疏地圖中的平面點建立面元，以節省記憶體。",[4872],"RGB-D camera (synthetic ICL-NUIM, TUM RGB-D and TAMU RGB-D sequences; sensor models not named in the paper)","sparse map of point, line and plane landmarks with keyframe co-visibility graph and a Manhattan map of MF observations; dense surfel map in which planar regions reuse sparse-map plane points (surfel radius from the 0.2 m downsampling voxel) and non-planar regions use superpixel surfels as in Dense Surfel Mapping","none (adding a loop-closure module is listed as future work)","Keyframe-based feature SLAM with constant-velocity prediction and local-map refinement; in Manhattan scenes rotation is taken drift-free from a previously stored Manhattan-frame observation and only translation is optimized from point, line and plane errors; otherwise full 6-DoF Levenberg-Marquardt with Huber cost over point, line and plane reprojection errors plus parallel and perpendicular plane constraints","ORB points matched by projection and Hamming distance; LSD line segments matched with LBD descriptors; planes extracted by AHC from the downsampled point cloud and matched by normal angle and point-plane distance; Manhattan frames detected from two or three mutually perpendicular plane normals (SVD-orthogonalized) and matched through shared map-plane IDs","trajectory, sparse point-line-plane map, dense surfel point cloud",[],{"id":4880,"label":4881,"shortName":4882,"title":4883,"year":115,"era":52,"cluster":151,"scope":99,"keyIdeaZh":4884,"sensors":4885,"mapRepresentation":4889,"loopClosure":4890,"estimator":4891,"association":4892,"deskew":4893,"outputGeometry":4894,"fulltextStatus":22,"lidarModels":4895,"equipmentCount":736},"loam2014","Zhang & Singh, 2014","LOAM","LOAM: Lidar Odometry and Mapping in Real-time","LOAM 將 3D LiDAR 的同時定位與建圖拆成兩個並行、頻率不同的演算法：高頻（約 10 Hz）里程計（odometry）以掃描對掃描配準估計速度並校正運動畸變（motion distortion），低頻（約 1 Hz）建圖（mapping）再把去畸變點雲精細配準到地圖。兩者都只使用依局部平滑度挑出的邊緣點（edge）與平面點（planar），分別以點到線、點到面距離作為殘差，並以 Levenberg-Marquardt 最佳化求解。系統沒有迴圈閉合（loop closure），IMU 只是選用的前處理先驗，因此長距離漂移無法做全域修正。",[4886,4887,4888],"3D LiDAR (custom rotating Hokuyo UTM-30LX 2D scanner)","IMU (optional, Xsens MTi-10)","3D LiDAR (360 deg Velodyne lidar via KITTI, 10 Hz; the model is not named in this paper)","registered point cloud map Q_k built from the undistorted sweeps, stored in 10 m cubes; cubes intersecting the new sweep are loaded into a KD-tree; matching uses ten times more feature points than odometry with edge or plane neighbourhoods found by eigen-analysis; the map is downsized with a 5 cm voxel grid","none (stated in Sec. I and listed as future work in Sec. VIII)","Levenberg-Marquardt nonlinear least squares with bisquare robust weights; scan-to-scan odometry at about 10 Hz and scan-to-map mapping at about 1 Hz running in parallel (Sec. IV-B, V-C, VI)","edge and planar feature points selected by local smoothness per scan line; point-to-line and point-to-plane distances; mapping stage finds line\u002Fplane correspondences by eigen-analysis of local map point clusters (Sec. V-A, V-B, VI)","points reprojected with the linearly interpolated odometry pose; optional IMU preprocessing removes orientation change and part of acceleration-induced distortion (Sec. V-C, VII-B)","motion-corrected registered point cloud map (5 cm voxel-grid downsampled) and 6-DoF pose at about 10 Hz (Sec. VI); dense raw-point export not described in paper",[375,45],{"id":4897,"label":4898,"shortName":4899,"title":4900,"year":235,"era":52,"cluster":263,"scope":99,"keyIdeaZh":4901,"sensors":4902,"mapRepresentation":4907,"loopClosure":4908,"estimator":4909,"association":4910,"deskew":4911,"outputGeometry":4912,"fulltextStatus":22,"lidarModels":4913,"equipmentCount":554},"vloam2015","Zhang & Singh, 2015","V-LOAM","Visual-lidar odometry and mapping: low-drift, robust, and fast","V-LOAM 以單眼相機搭配掃描式 3D 光達（由馬達帶動的 Hokuyo 2D 雷射掃描儀），分成兩個依序運作的階段：視覺里程計以影像速率（60 Hz）估計相鄰影格間的運動，特徵點的深度取自光達深度圖或三角化，沒有深度的特徵也納入求解；光達里程計每次掃描（約 1 秒）執行一次，先以線性運動模型做掃描對掃描精修，消除視覺漂移造成的點雲畸變，再以邊緣與平面特徵做掃描對地圖配準並累積地圖。最後整合低頻光達位姿與高頻視覺運動，以影像速率輸出位姿。方法與硬體都沒有使用 IMU，也刻意不做迴圈閉合。",[4903,4904,4905,4906],"monocular camera (uEye monochrome at 60 Hz; wide-angle lens 76 deg or fisheye lens 185 deg horizontal FoV)","3D lidar built from a Hokuyo UTM-30LX 2D laser scanner rotated back-and-forth by a motor with encoder (1 s sweep)","KITTI configuration: single camera and Velodyne lidar","no IMU in the method or hardware","incrementally built map point cloud: each distortion-free sweep is matched to the existing map cloud, with correspondences found by eigenvalue analysis of local point clusters, and then merged into it; edge and planar points of the previous sweep are kept in two 3D KD-trees for sweep-to-sweep matching (Sec. VI, Fig. 6)","none; the authors intentionally omit loop closure to focus on odometry (Sec. I)","two sequential stages: frame-to-frame visual odometry solved by Levenberg-Marquardt in a robust-fitting framework, using features with depth from the lidar depthmap, depth from triangulation, or no depth; lidar odometry once per sweep with sweep-to-sweep refinement (linear drift model) and then sweep-to-map registration; transforms from both stages integrated into poses at image rate (Secs. IV to VI)","up to 300 Harris corners tracked by KLT over 5x6 image subregions; feature depth interpolated from the three nearest depthmap points found in a 2D KD-tree on angular coordinates; lidar edge and planar points selected by local curvature and matched to edge lines and planar patches with 3D KD-trees (sweep-to-sweep) or by eigenvalue analysis of local map clusters (sweep-to-map, ICP-style) (Secs. V to VII)","points registered with the visual odometry motion; residual distortion from visual drift removed by the linear-motion model in the sweep-to-sweep refinement (Sec. VI, Fig. 4)","registered 3D point cloud maps (Figs. 9 to 14) and 6-DoF poses at the image frame rate; export format not_reported",[375,4914],"Velodyne lidar of the KITTI setup",{"id":4916,"label":4917,"shortName":4918,"title":4919,"year":769,"era":52,"cluster":151,"scope":99,"keyIdeaZh":4920,"sensors":4921,"mapRepresentation":4927,"loopClosure":4928,"estimator":4929,"association":4930,"deskew":4931,"outputGeometry":4932,"fulltextStatus":22,"lidarModels":4933,"equipmentCount":2038},"loam2017_auro","Zhang & Singh, 2017","LOAM (journal version)","Low-drift and real-time lidar odometry and mapping","本記錄為 LOAM 的期刊版本（Autonomous Robots，2016-02-18 線上發表、2017 年卷期）。方法核心與 RSS 2014 版相同：高頻、低精度的里程計估計速度並去除點雲運動畸變，低頻（預設為里程計的十分之一）的建圖以更多特徵點精細配準；兩者都使用依局部平滑度挑出的邊緣點與平面點，以 Levenberg-Marquardt 搭配 bisquare 權重求解，IMU 僅為選用的前處理先驗，且不含迴圈閉合。相較 RSS 版，期刊版把驗證擴充到四種感測系統（含八旋翼機與 Velodyne HDL-32E 車載測試），建圖以 5 cm（邊緣）與 10 cm（平面）體素平均，並說明 KITTI 測試為求精度改為逐幀建圖、只達即時速度的約 10%。",[4922,4923,4924,4925,4926],"2-axis lidar: back-and-forth spinning Hokuyo UTM-30LX (Sec. 4.1)","continuously spinning Hokuyo on an octo-rotor (Sec. 7.3)","Velodyne HDL-32E (Sec. 7.4)","Velodyne HDL-64E via KITTI (Sec. 7.5)","IMU optional: Xsens MTi-10 (Sec. 7.2), Microstrain 3DM-GX3-45 (Sec. 7.3)","feature point cloud map stored in 10 m cubes with KD-tree search; voxel-grid averaging at 5 cm (edge) and 10 cm (planar); map truncated to a 500 m cube around the sensor (Sec. 6)","none; authors state they do not consider loop closure (Sec. 1) and list it as future work (Sec. 9)","Levenberg-Marquardt adapted to robust fitting with bisquare weights; odometry at scan rate and mapping at about one-tenth of that rate (ratio 10 preferred, Sec. 8); for KITTI the mapping ran every scan, giving 10% of real-time speed (Sec. 7.5)","edge and planar feature points selected by local smoothness; point-to-edge-line and point-to-planar-patch distances; KD-tree nearest neighbours (Sec. 5-6)","odometry-estimated velocity used to reproject points; optional IMU pre-processing removes orientation change and part of the acceleration effect (Sec. 7.2)","registered feature point cloud map and 6-DoF pose (Sec. 6); dense raw-point export not described in the text read",[375,4934,225,161],"Hokuyo laser scanner",{"id":4936,"label":4937,"shortName":4938,"title":4939,"year":150,"era":10,"cluster":263,"scope":99,"keyIdeaZh":4940,"sensors":4941,"mapRepresentation":4945,"loopClosure":4946,"estimator":4947,"association":4948,"deskew":4949,"outputGeometry":4950,"fulltextStatus":22,"lidarModels":4951,"equipmentCount":1764},"zhang2018lvio","Zhang & Singh, 2018","Zhang & Singh LVIO (JFR 2018)","Laser-visual-inertial odometry and mapping with high robustness and low drift","此研究以 3D 雷射掃描儀、相機與 IMU 建立多層次、依序執行的管線，由粗到細估計運動，而非卡爾曼濾波或因子圖：先以 IMU 機械編排（200 Hz）預測運動，再以關鍵影格式視覺慣性里程計（50 Hz）估計運動並為特徵點補上雷射深度，最後以掃描配準（5 Hz）精修位姿，並把點雲配準到以兩層體素管理的地圖。相機與雷射的結果回饋修正 IMU 的速度漂移與偏差；各模組以特徵值判斷退化方向，只在條件良好的方向更新，因此相機或雷射失效時可整段或部分略過該模組。論文另延伸到在既有地圖上定位。本文擴充自作者 2017 年的 ICRA 與 FSR 論文，並把 V-LOAM（ICRA 2015）當作另一個既有方法比較，兩者不是同一方法。",[4942,4943,4944],"3D LiDAR (Velodyne HDL-32E or Velodyne VLP-16 at 5 Hz; on the handheld Contour a Hokuyo UTM-30LX-EW spun at 1 Hz)","IMU (Xsens MTi-30 at 200 Hz; Xsens MTi-20 on Contour)","monochrome camera (uEye UI-1220SE, 752x480, 76 deg horizontal FoV, 50 Hz) on the two Velodyne suites; on Contour a 640x512 wide-angle camera for motion estimation and a 1600x1200 HD color camera for point colorization","two-level voxel map of edge and planar points truncated around the sensor, with a 3D KD-tree per voxel; map downsampled to constant density after each merge; scan matching on up to four CPU threads (Secs. 6.3 to 6.4, Table 1)","none; drift is measured at loop returns, e.g., a building registered twice at the start and end of Accuracy Test 2 (Sec. 10.1.1, Fig. 18)","sequential multilayer coarse-to-fine pipeline, not a Kalman filter or factor graph: IMU mechanization predicts motion; keyframe visual-inertial odometry solves a marginalized 6-DoF problem (landmarks not optimized) by Newton gradient descent with robust fitting; scan matching refines pose with prior-motion constraints; camera and lidar feedback correct IMU velocity drift and biases through a sliding-window average; degenerate directions are found from eigenvalues and only well-conditioned directions are updated, so failed modules are bypassed fully or partially (Secs. 3 to 8)","visual: up to 300 Harris corners tracked by KLT; depth from a lidar depthmap (three nearest points on a unit sphere in a 2D KD-tree, validity check, planar interpolation) or Bayesian triangulation, and features without depth also used; lidar: edge and planar points selected by local smoothness and matched to map point clusters verified by eigenvalue analysis, with point-to-line and point-to-plane distances (Secs. 5.3, 6.1 to 6.2, 10.1)","each scan is locally registered using visual-inertial odometry key-poses with IMU interpolation between them before feature extraction (Sec. 6.1)","dense registered 3D point cloud maps; Contour colorizes points with its HD camera (Sec. 10.2); export format not_reported",[225,143,375],{"id":4953,"label":4954,"shortName":4955,"title":4956,"year":115,"era":52,"cluster":263,"scope":346,"keyIdeaZh":4957,"sensors":4958,"mapRepresentation":4960,"loopClosure":465,"estimator":4961,"association":4962,"deskew":4963,"outputGeometry":4964,"fulltextStatus":22,"lidarModels":4965,"equipmentCount":487},"demo2014","Zhang et al., 2014","DEMO","Real-time depth enhanced monocular odometry","DEMO 以單眼相機為主，從 RGB-D 相機或 LiDAR 取得深度：先用估測的運動把深度點登錄成局部深度地圖並存進以兩個角度座標建立的 2D KD 樹，再以最近三點構成的小平面內插特徵深度；沒有深度的特徵改用前幾影格的運動三角化，仍無法取得時也保留並以較弱的約束參與求解。逐影格運動用具強健權重的 Levenberg-Marquardt 求解，另以 iSAM 對 8 張影像做低頻光束法平差，最後整合成高頻輸出。",[37,4959],"depth from an RGB-D camera (Xtion Pro Live) or from a 3D LiDAR (rotating Hokuyo UTM-30LX; Velodyne on KITTI)","local registered depth map (point cloud) around the camera used only for depth association","frame-to-frame motion by Levenberg-Marquardt robust fitting with bisquare weights, using features with depth (two equations each) and without depth (one equation each); sliding bundle adjustment with iSAM on 8 images (one of every five of 40 frames) at about 0.25 to 1.0 Hz; transform integration combines the high-rate and low-rate estimates (Secs. V-VI)","Harris corners tracked by KLT; depth map registered with the estimated motion, downsampled by angular interval and stored in a 2D KD-tree over two angular coordinates; feature depth interpolated from three nearest depth points (planar patch) or, if unavailable, triangulated from previous motion (Sec. V)","not described; LiDAR points are accumulated into the depth map using the estimated motion","camera trajectory; registered point clouds of the depth sensor (Fig. 7 shows maps built on KITTI)",[375,4966],"KITTI 360 deg Velodyne laser scanner",{"id":4968,"label":4969,"shortName":4970,"title":4971,"year":1461,"era":52,"cluster":68,"scope":69,"keyIdeaZh":4972,"sensors":4973,"mapRepresentation":57,"loopClosure":121,"estimator":4976,"association":4977,"deskew":4978,"outputGeometry":57,"fulltextStatus":22,"lidarModels":4979,"equipmentCount":24},"zhang2016degeneracy","Zhang et al., 2016","Degeneracy factor \u002F solution remapping","On degeneracy of optimization-based state estimation problems","本文把退化定義為解對約束擾動的剛度，並證明線性化系統的退化因子 D 等於 AᵀA 最小特徵值加一，對應的特徵向量即為最退化的方向。方法以門檻判定退化方向（門檻取自一組同時含良好與退化場景的樣本資料，設在兩群間隔的中點），再以解重映射（solution remapping）在退化方向保留預測值、只在條件良好方向更新，可作為 Levenberg-Marquardt 等求解器的外掛步驟，額外複雜度為 O(kn² + n³)。作者以 uEye 相機與馬達旋轉的 Hokuyo UTM-30LX 組成的手持視覺光達系統，在走廊、平坦地面以及 538 m 室內外路線測試，終點位置誤差為行進距離的 0.71%。",[4974,4975],"monocular camera (uEye monochrome, 60 Hz, 752 x 480, 76 deg horizontal FOV)","custom 3D lidar (Hokuyo UTM-30LX rotated by a motor, 0.25 deg encoder)","Plug-in to linear or nonlinear least-squares solvers: eigen-decomposition of A^T A at the first nonlinear iteration, eigenvalues below an empirical threshold mark degenerate directions, and solution remapping updates only well-conditioned directions (Algorithm 1); demonstrated inside three Levenberg-Marquardt modules (frame-to-frame visual odometry, sweep-to-sweep refinement, sweep-to-map registration)","not_applicable (generic to optimization-based estimation; demonstrated with scan matching and visual constraints)","not_applicable to the degeneracy method itself; the host system removes lidar distortion caused by visual odometry drift with a linear motion model within a sweep",[375],{"id":4981,"label":4982,"shortName":4983,"title":4984,"year":51,"era":10,"cluster":755,"scope":132,"keyIdeaZh":4985,"sensors":4986,"mapRepresentation":4987,"loopClosure":4988,"estimator":4989,"association":4990,"deskew":57,"outputGeometry":4991,"fulltextStatus":22,"lidarModels":4992,"equipmentCount":185},"goslam2023","Zhang et al., 2023b","GO-SLAM","GO-SLAM: Global Optimization for Consistent 3D Instant Reconstruction","GO-SLAM 以 DROID-SLAM 的學習式稠密光流與可微分稠密光束法平差（dense bundle adjustment）作為追蹤核心，在前端依光流估算的共視度偵測迴圈，並在獨立執行緒中對所有關鍵影格線上執行完整光束法平差，以抑制長序列的累積漂移。建圖端使用多解析度雜湊編碼的神經隱式 SDF，每次優先選取位姿變化最大的關鍵影格重新訓練，使重建隨全域最佳化後的位姿與深度同步更新。同一架構可接受單眼、立體或 RGB-D 影像，最後以 marching cubes 從 SDF 擷取網格。",[37,3946,1405],"neural implicit SDF and colour field with multi-resolution hash encoding (16 levels, Instant-NGP style) and shallow MLPs, rendered by NeuS-style unbiased volume rendering; mapping re-trains on selected keyframes (latest two, top-10 by pose change, 10 stratified) (Sec. 3.2, Supp. A)","flow-based: edges sampled from the unexplored part of the co-visibility matrix between local-window and historical keyframes; a loop is accepted after three consecutive candidates with mean flow below tau_co, then optimized by DBA (Sec. 3.1)","DROID-SLAM tracking extended with online global optimization: RAFT-based recurrent update operator predicts dense flow and confidence, and a differentiable dense bundle adjustment (DBA) layer solves poses and per-pixel inverse depths by damped Gauss-Newton; front end optimizes a local keyframe window with loop edges, back end runs online full BA over all keyframes in a separate thread (Sec. 3.1)","dense learned optical flow with per-pixel confidence between keyframe pairs; keyframe-graph edges chosen by co-visibility measured as mean rigid flow (threshold tau_co = 25) with neighbourhood suppression (Sec. 3.1, Sec. 4.1)","keyframe trajectory, per-keyframe depth and a triangle mesh extracted by marching cubes from the SDF (Sec. 4.1)",[],{"id":4994,"label":4995,"shortName":4996,"title":4997,"year":661,"era":10,"cluster":11,"scope":300,"keyIdeaZh":4998,"sensors":4999,"mapRepresentation":5002,"loopClosure":57,"estimator":5003,"association":5004,"deskew":57,"outputGeometry":5005,"fulltextStatus":22,"lidarModels":5006,"equipmentCount":24},"zhang2024globalbimreg","Zhang et al., 2024b","Global BIM-point registration and association","Global BIM-point cloud registration and association for construction progress monitoring","作者把 BIM 構件以構造實體幾何（CSG）拆解並以解析距離場表示，避免取樣造成資訊損失。粗配準以平面基元對 BIM 面在重力軸對齊下搜尋對應，並以剛體動力學模擬驗證幾何一致性；精配準則交替更新位姿與逐點對應權重，並以鄰近性、法向與構件存在與否截斷權重，使臨時材料與未施作構件不誤導配準。模擬採 ISPRS 室內建模基準（含手持與背包掃描）。",[5000,5001],"real site: handheld sensor suite with Ouster OS0-128 LiDAR (clouds built with FAST-LIO2)","simulation: ISPRS indoor modelling benchmark clouds from stationary, handheld and backpack scanners","BIM as CSG-decomposed analytic distance fields; input point cloud","primitive-level coarse registration (orientation hypotheses using gravity, correspondence tree, geometric constraint filter, rigid-body dynamics simulator) + point-level Gauss-Newton fine registration with Geman-McClure weights (Sec. 3.2-3.3)","plane primitives to BIM distance fields; per-point weights truncated by proximity, normal consistency and element existence (Sec. 3.3.2)","BIM-aligned point cloud with per-point association weights (progress existence check)",[222],{"id":5008,"label":5009,"shortName":5010,"title":5011,"year":207,"era":10,"cluster":755,"scope":132,"keyIdeaZh":5012,"sensors":5013,"mapRepresentation":5015,"loopClosure":5016,"estimator":5017,"association":5018,"deskew":57,"outputGeometry":5019,"fulltextStatus":22,"lidarModels":5020,"equipmentCount":185},"hislam2_2025","Zhang et al., 2025","HI-SLAM2","HI-SLAM2: Geometry-Aware Gaussian SLAM for Fast Monocular Scene Reconstruction","HI-SLAM2 是只用單眼 RGB 的三維高斯 SLAM：追蹤端沿用 DROID-SLAM 的學習式光流與稠密光束法平差，並以每張影像 2x2 的尺度網格把 Omnidata 單眼深度先驗對齊到估計深度，以修正先驗中隨位置變化的尺度失真。偵測到迴圈時以 Sim(3) 位姿圖平差同時修正位姿與尺度漂移，並依錨定關鍵影格的更新直接變形高斯，使地圖即時保持一致；離線階段再做完整光束法平差與位姿、高斯聯合最佳化。地圖以高斯表示並以射線與高斯交點計算無偏深度，最後由渲染深度經 TSDF 融合得到網格。",[5014],"monocular RGB camera","3D Gaussian splatting with RGB colours (no spherical harmonics), unbiased ray-Gaussian intersection depth, Gaussians anchored to keyframes and deformed with their Sim(3) updates; random downsampling (factor 32), densification and pruning; map grows without a predefined scene bound (Sec. III-D)","proximity-based: candidates with optical-flow distance below a threshold, orientation difference below a threshold and index gap beyond the local window; closed by online Sim(3) pose-graph bundle adjustment that also corrects scale drift (Sec. III-C)","DROID-SLAM-based recurrent optical flow and dense bundle adjustment (damped Gauss-Newton) on a keyframe graph, interleaved with joint depth and scale alignment (JDSA) that fits a 2x2 scale grid to each monocular depth prior; Sim(3) pose-graph BA on loop closure; offline full BA and joint pose and 3DGS refinement with Adam (Sec. III-B to III-E)","dense learned optical-flow correspondences with confidence weights between co-visible keyframes (Sec. III-B)","3DGS map and a mesh from TSDF fusion of rendered depth maps; rendered colour and depth (Sec. III)",[],{"id":5022,"label":5023,"shortName":5024,"title":5025,"year":30,"era":10,"cluster":5026,"scope":12,"keyIdeaZh":5027,"sensors":5028,"mapRepresentation":5031,"loopClosure":5032,"estimator":5033,"association":5034,"deskew":5035,"outputGeometry":5036,"fulltextStatus":22,"lidarModels":5037,"equipmentCount":487},"bimloc2026","Zhang et al., 2026","BIM-Loc (S05)","BIM-Loc: BIM-integrated discrepancy-aware LiDAR-based indoor localization","C11a","BIM-Loc 以設計階段 BIM 作為先驗，將受差異影響的定位問題拆為 BIM 輔助軌跡最佳化與階層式差異偵測兩個耦合子問題，並迭代求解。其以多次命中射線投射建立點雲與 BIM 面的資料關聯，於位姿圖中加入掃描間一致性與掃描對 BIM 一致性因子，並以貝氏核推論在 BIM 表面紋理空間中逐像素、面、構件更新差異狀態。作者在模擬、已完工辦公建物（SLABIM）與施工中工地（CityU）評估，並明言其差異偵測只判斷構件存在與否，無法量化偏差大小。",[5029,36,5030],"3D LiDAR (Velodyne VLP-16 in simulation; Livox Mid-360; Ouster OS0-128)","camera (visualisation only)","As-designed BIM meshes (LOD 300, IFC) plus 2D texture-space discrepancy maps","none (drift bounded by BIM constraints)","Pose graph optimization with BIM-integrated factors solved incrementally with iSAM2 (GTSAM); front-end odometry is DLO (front-end agnostic)","Multi-hit ray casting against BIM facets; point-cluster plane factors (eigenvalue-based inter-scan BA-style) and point-to-BIM-surface residuals","not_reported (delegated to front-end odometry)","BIM-aligned trajectory, aggregated scans, structure-level discrepancy labels (consistent, discrepant, unknown); no quantitative deviation magnitudes (Sec. 6)",[143,507,222],{"id":5039,"label":5040,"shortName":5041,"title":5042,"year":5043,"era":52,"cluster":68,"scope":69,"keyIdeaZh":5044,"sensors":5045,"mapRepresentation":5049,"loopClosure":72,"estimator":5050,"association":5051,"deskew":57,"outputGeometry":183,"fulltextStatus":22,"lidarModels":5052,"equipmentCount":46},"zhang1994icp","Zhang, 1994","Iterative point matching (Zhang)","Iterative point matching for registration of free-form curves and surfaces",1994,"本文提出迭代虛擬點匹配（iterative pseudo point matching）演算法，用於配準邊緣式立體視覺取得的三維曲線，或相關式立體視覺重建的稠密三維地圖。方法假設兩次觀測之間的運動很小，或已由里程計與慣性系統近似得知；每次迭代先以目前估計轉換第一組點，再以 k-D tree 搜尋第二組資料中的最近點，接著依配對距離的平均值與標準差動態設定最大容許距離 Dmax，剔除離群點、遮蔽以及出現或消失的點（曲線另以切線夾角不超過 60 度限制配對），最後以四元數或對偶四元數最小平方法求剛體運動並反覆至收斂。作者以合成曲線、移動載具上三目立體相機拍攝的椅子場景、相關式立體視覺重建的岩石場景以及頭像距離影像驗證。",[5046,5047,5048],"trinocular edge-based stereo (3-D curves)","correlation-based stereo (dense 3-D maps)","range images (head figure, Sec. 6.2; sensor not reported)","Raw 3-D point sets: chained points for curves and scattered points for dense 3-D maps, used without smoothing or primitive fitting","Closed-form least-squares rigid motion from the retained pairs (quaternion method and dual number quaternion method of Walker et al. 1991, both implemented with identical results), iterated until the relative changes of r and t are both below 1% or a maximum of 20 (curves) or 40 (surfaces) iterations is reached","Closest sample point in the second frame found with a 3-D tree whose search radius shrinks with Dmax; a pair is removed when its distance exceeds an adaptive Dmax computed each iteration from the mean mu and standard deviation sigma of pair distances (mu+3sigma, mu+2sigma, mu+sigma, or the histogram valley after the main peak, depending on mu relative to the user parameter D); for curves, pairs whose tangent angle exceeds 60 deg are also rejected",[],{"id":5054,"label":5055,"shortName":5056,"title":5057,"year":249,"era":10,"cluster":263,"scope":99,"keyIdeaZh":5058,"sensors":5059,"mapRepresentation":5062,"loopClosure":5063,"estimator":5064,"association":5065,"deskew":5066,"outputGeometry":5067,"fulltextStatus":22,"lidarModels":5068,"equipmentCount":554},"superodom2021","Zhao et al., 2021","Super Odometry","Super Odometry: IMU-centric LiDAR-Visual-Inertial Estimator for Challenging Environments","Super Odometry 以 IMU 為中心：IMU 里程計提供運動預測給視覺慣性與光達慣性子系統，後兩者回傳相對位姿約束來限制 IMU 偏差，形成由粗到細的估計流程，兼具鬆耦合的容錯與緊耦合的精度。光達端以 PCA 將點分類為點、線、面特徵並做多度量 ICP，地圖以動態八元樹（dynamic octree）組織以降低重建樹的成本。系統部署於 DARPA 地下挑戰賽的無人機與地面機器人。",[1201,5060,5061],"IMU (Xsens)","fisheye monocular camera","dynamic octree: hash table of voxels each holding its own octree","not used in reported experiments (Sec. V-B)","IMU-centric factor graph split into sub-graphs: IMU odometry (IMU preintegration plus relative-pose factors from VIO and LIO, weighted by covariances reflecting their reliability, with a marginalization prior) constrains the IMU biases; LIO solves scan-to-map registration by Levenberg-Marquardt with the IMU-odometry prior, and keyframe VIO minimizes reprojection, IMU and prior terms; estimates use historical frames in a sliding window and proceed coarse-to-fine (Secs. II-A, IV-A to IV-C)","PCA-based point, line and plane feature classification (linearity, planarity, curvature) with multi-metric point-to-point, point-to-line and point-to-plane ICP against the map through the dynamic octree; each correspondence weighted by how well its neighbours fit the assumed distribution, so noisy returns in dust are down-weighted or rejected (Sec. IV-B, Eqs. 5 to 8); monocular visual features tracked, with LiDAR points in the camera view providing feature depth (Sec. IV-C)","not_reported (full text read; no LiDAR motion-compensation step is described)","point cloud maps shown qualitatively (Fig. 6, Fig. 7); export not_reported",[143],{"id":5070,"label":5071,"shortName":5072,"title":5073,"year":661,"era":10,"cluster":397,"scope":1035,"keyIdeaZh":5074,"sensors":5075,"mapRepresentation":57,"loopClosure":57,"estimator":5079,"association":5080,"deskew":5081,"outputGeometry":5082,"fulltextStatus":22,"lidarModels":5083,"equipmentCount":144},"zhao2024deskew","Zhao et al., 2024a","Registration-based deskewing","Registration‐based point cloud deskewing and dynamic lidar simulation","作者以點對面 ICP 配準相鄰兩幀 LiDAR 點雲取得幀間運動，假設單幀掃描期間轉換參數的變化率固定、雷射發射間隔固定，依各點的發射順序線性內插出部分轉換，把每個點轉回該幀起始位姿，因此不需 IMU，也不需每點的實際時間戳記。作者另提出以參考平面為基準的去畸變位移指標 PRDT，處理點到平面距離在運動方向與目標表面平行時無法反映畸變的問題；並把同一模型反向用於動態 LiDAR 模擬器，以雷射發射頻率更新感測器位姿，產生帶運動畸變的合成點雲。室內驗證使用 UGV 搭載 Velodyne HDL-32E，並以 TLS 點雲作為參考。",[5076,5077,5078],"3D LiDAR Velodyne HDL-32E (UoM indoor data; up to 72,000 points per cloud, 1187 clouds)","No IMU used by the method","Evaluation reference: TLS point cloud (scanner model not reported); the dataset also includes a BIM of the same area, but no evaluation against the BIM is reported","Pairwise point-to-plane ICP between consecutive clouds gives the inter-frame transformation, followed by per-point linear interpolation of its rotation and translation parameters; no filter, smoother or pose-graph step","Point-to-plane ICP correspondences between consecutive point clouds (Low 2004); the authors state that any registration method could be substituted","IMU-free, registration-based: the ICP transformation between consecutive frames is linearly interpolated per point by firing order and each point is transformed back to the frame-start pose; the reverse process is used by the simulator to create distorted clouds","Deskewed point clouds; simulated skewed point clouds together with matching ideally deskewed clouds that serve as exact point-to-point ground truth",[225],{"id":5085,"label":5086,"shortName":5087,"title":5088,"year":9,"era":10,"cluster":11,"scope":12,"keyIdeaZh":5089,"sensors":5090,"mapRepresentation":5094,"loopClosure":5095,"estimator":5096,"association":5097,"deskew":5098,"outputGeometry":5099,"fulltextStatus":22,"lidarModels":5100,"equipmentCount":24},"zhen2019tunnellocalizability","Zhen & Scherer, 2019","Tunnel localizability with LiDAR and UWB","Estimating the Localizability in Tunnel-like Environments using LiDAR and UWB","作者把 LiDAR 在先驗地圖中定位的問題寫成一組點落在局部平面上的約束，計算量測距離對位置與姿態擾動的敏感度，分別堆疊成代表力的矩陣 F 與代表力矩的矩陣 T，並把特徵分解後各軸上累積的「虛擬力與力矩」大小定義為可定位性；這個觀點類比於操作力學中無摩擦的力封閉。UWB 測距只提供指向錨點方向的一個力，可補足隧道長軸方向的退化。定位部分以誤差狀態卡爾曼濾波融合 IMU、LiDAR 對地圖的位姿量測，以及由高斯粒子濾波把 UWB 距離轉成的位置量測。作者在卡內基美隆大學 35 m 長的 Smith Hall 隧道中，以改裝的 DJI M100 無人機定性比較有無 UWB 的定位結果。",[5091,5092,5093],"rotating 2D LiDAR (Hokuyo UTM-30LX-EW on a motor rotating 180 deg\u002Fs)","IMU (Microstrain, 100 Hz)","UWB ranging (Pozyx target board on the robot, one anchor in the tunnel)","prior point-cloud map of the tunnel built by aligning multiple local scans with ICP (Sec. IV-A, Fig. 8)","none (localization in a prior map)","Error-state Kalman filter: IMU propagation, 6D pose measurements from matching laser scans to the prior map, and UWB ranges converted to 3D position measurements by a Gaussian particle filter and inverse Kalman update (Sec. III-B)","laser scans matched to the prior map (details deferred to earlier work); localizability model uses plane normals fitted to 20 nearest neighbours of sampled points (Sec. III, IV-B)","laser scans projected into the robot body frame using motor encoder angles (Sec. IV-A); motion distortion handling not described","robot trajectory in the prior map; reconstructed map assembled from scans with estimated poses (qualitative)",[375],{"id":5102,"label":5103,"shortName":5104,"title":5105,"year":661,"era":10,"cluster":98,"scope":99,"keyIdeaZh":5106,"sensors":5107,"mapRepresentation":5109,"loopClosure":72,"estimator":5110,"association":5111,"deskew":5112,"outputGeometry":5113,"fulltextStatus":22,"lidarModels":5114,"equipmentCount":377},"trajlo2024","Zheng & Zhu, 2024","Traj-LO","Traj-LO: In Defense of LiDAR-Only Odometry Using an Effective Continuous-Time Trajectory","Traj-LO 把 LiDAR 量測視為高頻串流點，以由多段線性插值組成的連續時間軌跡描述感測器運動，並在滑動視窗內同時最小化點到平面幾何誤差與軌跡平滑（運動學）約束。由於每個點都用其時間戳查詢對應位姿，因此不需另外做運動補償。作者主張僅靠 LiDAR 也能在快速運動與 IMU 飽和情境下運作，並支援多 LiDAR。",[5108],"3D LiDAR only (single or multiple, spinning and non-repetitive)","Spatial hashing voxel map (following CT-ICP and KISS-ICP) storing up to 20 points per voxel, 7 nearest voxels searched, new points dropped when a voxel is full, points farther than 100 m removed; voxel size 0.4 m indoor, 0.8 m outdoor and 0.2 m for the Point-LIO indoor sequence","Sliding-window nonlinear least squares (Gauss-Newton with analytic SE(3) Jacobians) over K+1 control poses of a piecewise-linear continuous-time trajectory (K = 4 segments of 0.03 s in experiments), with point-to-plane terms, a smoothness term penalizing velocity change between consecutive segments, and a Schur-complement marginalization prior with first-estimate Jacobians; marginalization lowered ATE in all four ablation settings (Table IV)","point-to-plane to map neighbours, normals from PCA of five closest points; no feature selection","not required: each point is registered with the pose queried at its own timestamp","odometry and voxel point map; export format not_reported",[161,5115,224,505,437],"two 16-channel Ouster LiDARs (horizontal and vertical OS1-16)",{"id":5117,"label":5118,"shortName":5119,"title":5120,"year":97,"era":10,"cluster":263,"scope":99,"keyIdeaZh":5121,"sensors":5122,"mapRepresentation":5124,"loopClosure":72,"estimator":5125,"association":5126,"deskew":5127,"outputGeometry":5128,"fulltextStatus":22,"lidarModels":5129,"equipmentCount":423},"fastlivo2022","Zheng et al., 2022","FAST-LIVO","FAST-LIVO: Fast and Tightly-coupled Sparse-Direct LiDAR-Inertial-Visual Odometry","FAST-LIVO 的 LIO 與 VIO 皆採直接法（direct method）：光達原始點以點到平面殘差配準到地圖，視覺部分則把影像小區塊（patch）附掛在光達地圖點上，直接以稀疏光度誤差對齊新影像，不擷取、不三角化視覺特徵。兩者在 ESIKF 中緊密耦合，並以遮蔽與深度不連續檢測剔除不穩定地圖點，因此計算成本低且可在 ARM 處理器上即時執行。",[5123,36,1601],"3D LiDAR (Ouster OS1-16 in NTU-VIRAL; Livox Avia in private data)","LiDAR global map adopted from FAST-LIO2 (all past points in an ikd-Tree with internal downsampling) plus a separate visual global map of previously observed LiDAR points in equal-size hash-indexed voxels, each point storing several 8x8 patch pyramids with their camera poses","error-state iterated Kalman filter fusing LiDAR and visual updates (LIO adapted from FAST-LIO2)","raw LiDAR points with frame-to-map point-to-plane residual; sparse-direct frame-to-map alignment of 8x8 image patches attached to LiDAR map points, with occlusion and depth-discontinuity outlier rejection","backward propagation as in FAST-LIO2 (Sec. III)","real-time dense RGB-colored point cloud (Sec. VI-B4, Fig. 7)",[5130,437],"OS1 gen1 (16-channel)",{"id":5132,"label":5133,"shortName":5134,"title":5135,"year":207,"era":10,"cluster":263,"scope":99,"keyIdeaZh":5136,"sensors":5137,"mapRepresentation":5140,"loopClosure":5141,"estimator":5142,"association":5143,"deskew":5144,"outputGeometry":5145,"fulltextStatus":22,"lidarModels":5146,"equipmentCount":5148},"fastlivo2_2025","Zheng et al., 2025","FAST-LIVO2","FAST-LIVO2: Fast, Direct LiDAR-Inertial-Visual Odometry","FAST-LIVO2 以序列式更新的 ESIKF 先融合光達、再融合影像，解決兩種量測維度不匹配的問題；光達與視覺模組共用一個自適應體素（voxel）地圖，光達點同時作為視覺地圖點並附掛影像區塊。影像對齊利用光達平面先驗、動態更新參考區塊、按需射線投射（raycasting）處理近距盲區，並即時估計曝光時間。作者在 NTU-VIRAL 與 Hilti（含工地序列）等 25 個公開序列評估軌跡精度，並展示高精度彩色點雲、網格、貼圖與 3DGS 應用。",[5138,36,5139],"3D LiDAR (Livox Avia, Ouster OS1-16, Hesai PandarXT-32, Robosense BPearl across datasets)","camera (pinhole or fisheye)","Adaptive voxel map adapted from VoxelMap: hash table of 0.5 m root voxels, each an octree (max 3 layers) of plane leaf voxels with plane center, normal and covariance; mature planes stop accepting points; selected points carry 3-level patch pyramids (visual map points); local map of side L slid as a ring buffer when the detection sphere touches the boundary","none; authors note possible long-distance drift and list loop closure as future work (Sec. XI)","error-state iterated Kalman filter with sequential update (LiDAR update first, then image update)","raw LiDAR points, frame-to-map point-to-plane with per-point noise including beam divergence; sparse-direct patch photometric alignment using LiDAR plane priors, dynamic reference-patch update, on-demand voxel raycasting, outlier rejection","scan recombination with forward and backward IMU propagation (Fig. 3)","dense colored point map in real time; downstream TSDF mesh (VDBFusion), OpenMVS texture mapping and 3DGS initialization demonstrated (Sec. X-C)",[4837,4537,5147,437],"Robosense Bpearl",21,{"id":5150,"label":5151,"shortName":5152,"title":5153,"year":51,"era":10,"cluster":755,"scope":32,"keyIdeaZh":5154,"sensors":5155,"mapRepresentation":5156,"loopClosure":72,"estimator":5157,"association":5158,"deskew":57,"outputGeometry":5159,"fulltextStatus":22,"lidarModels":5160,"equipmentCount":144},"shinemapping2023","Zhong et al., 2023","SHINE-Mapping","SHINE-Mapping: Large-Scale 3D Mapping Using Sparse Hierarchical Implicit Neural Representations","SHINE-Mapping 以稀疏八元樹階層特徵格網搭配共用淺層 MLP，從已知位姿的 LiDAR 點雲學習符號距離場（SDF），並以正則化處理增量建圖的遺忘問題。它不做位姿估計，屬已知位姿下的建圖元件。作者用合成 MaiCity 與具 TLS 參考網格的 Newer College 評估，並以所有比較方法重建交集遮罩後的參考網格計算精度。",[35],"sparse octree-based hierarchical feature grid with a shared shallow MLP decoding SDF","not_applicable (mapping with known poses)","SDF supervision from range measurements (binary cross-entropy on samples along rays)","mesh via marching cubes on a fixed grid",[865,45],{"id":5162,"label":5163,"shortName":5164,"title":5165,"year":1461,"era":52,"cluster":68,"scope":69,"keyIdeaZh":5166,"sensors":5167,"mapRepresentation":5171,"loopClosure":72,"estimator":5172,"association":5173,"deskew":57,"outputGeometry":5174,"fulltextStatus":22,"lidarModels":5175,"equipmentCount":77},"zhou2016fgr","Zhou et al., 2016","FGR","Fast Global Registration","FGR 先以 FPFH 特徵的雙向最近鄰建立候選對應，再以互為最近鄰檢驗與三元組邊長比例檢驗（τ = 0.9）提高內點比例；之後對這組固定不變的對應直接最佳化單一穩健目標，同時對齊表面並使錯誤對應失效，內迴圈不更新對應，也不做最近點查詢。目標採縮放 Geman-McClure 穩健估計函數，藉 Black-Rangarajan 對偶交替更新線過程變數與位姿，並以逐步非凸化將 μ 從最大表面直徑的平方每四次迭代減半，直到真對應距離門檻的平方。作者報告在合成資料、UWA 與 Choi 等人的場景基準上，精度可比或優於既有全域配準流程；在單執行緒 Intel Core i7-5960X 上，合成距離影像每對平均 0.22 s，比最快的既有全域方法 CZK 快約 50 倍；此方法並可延伸為多片點雲聯合配準。",[5168,5169,5170],"synthetic range images (AIM@SHAPE Bimba, Dancing Children, Chinese Dragon; Berkeley Angel; Stanford Bunny) with added 3D Gaussian noise","UWA object and scene benchmark data (acquisition sensor not stated in the paper)","Choi et al. scene fragments whose high-frequency noise and low-frequency distortion simulate consumer depth camera scans; Augmented ICL-NUIM sequences (living room and office) used for multi-way registration","surfaces \u002F point clouds","scaled Geman-McClure robust objective optimized via Black-Rangarajan line-process duality with alternating updates and graduated non-convexity (Sec. 3.1-3.2); pose step is a Gauss-Newton solve on a locally linearized 6-vector mapped back to SE(3); mu starts at D^2 (largest surface diameter) and is halved every four iterations down to delta^2; alignment is validated once after convergence (Sec. 3.2, Algorithm 1)","FPFH nearest neighbours in feature space in both directions, filtered by a reciprocity test and a tuple test on three random pairs (edge-length ratios within tau = 0.9 and 1\u002Ftau); correspondences stay fixed during optimization (Sec. 3.3, Algorithm 1)","rigid transformation(s)",[],{"id":5177,"label":5178,"shortName":5179,"title":5180,"year":150,"era":10,"cluster":31,"scope":1035,"keyIdeaZh":5181,"sensors":5182,"mapRepresentation":5183,"loopClosure":57,"estimator":57,"association":57,"deskew":57,"outputGeometry":5184,"fulltextStatus":22,"lidarModels":5185,"equipmentCount":61},"zhou2018open3d","Zhou et al., 2018","Open3D","Open3D: A Modern Library for 3D Data Processing","Open3D 是提供 C++ 與 Python 介面的開源三維資料函式庫，核心資料結構為點雲、三角網格與 RGB-D 影像，內含體素降採樣、法向量估計、ICP 配準與體積整合（volumetric integration）等演算法，並以完整的 RGB-D 場景重建流程示範其功能；PUMA 即建構在此函式庫上。",[772],"point clouds, triangle meshes, RGB-D images; volumetric (TSDF) integration module","point clouds and meshes (e.g., via volumetric integration)",[],{"id":5187,"label":5188,"shortName":5189,"title":5190,"year":249,"era":10,"cluster":151,"scope":132,"keyIdeaZh":5191,"sensors":5192,"mapRepresentation":5194,"loopClosure":5195,"estimator":5196,"association":5197,"deskew":5198,"outputGeometry":5199,"fulltextStatus":22,"lidarModels":5200,"equipmentCount":24},"zhou2021planeadjust","Zhou et al., 2021","Plane-adjustment LiDAR SLAM (indoor)","LiDAR SLAM With Plane Adjustment for Indoor Environment","這個方法以平面作為室內 LiDAR SLAM 的地標，類比視覺 SLAM 的光束法平差，聯合最佳化關鍵影格位姿與平面參數，作者稱為平面平差。定位執行緒以前向 ICP 流把上一幀的平面點追蹤到目前幀，直接得到局部對全域平面的對應，並在配準中逐點內插掃描內運動；局部建圖在 8 個關鍵影格的滑動視窗內做平面平差，全域建圖則在舊平面被再次觀測時觸發全域平差，不必回到原地即可修正漂移。平面法向量一律指向感測器，用以區分牆與門的兩個表面，避免最近鄰搜尋把兩面混在一起。",[5193],"3D LiDAR only (Velodyne VLP-16 data recorded by a NavVis M6 device)","global plane landmarks (closest-point parameterization) with their supporting points and keyframe poses; a 4 x 4 integrated cost matrix per plane summarizes observations outside the sliding window (Sec. IV-2; Sec. VI-E; Sec. VI-F)","plane-revisit criterion: global plane adjustment is triggered when previously mapped planes are re-associated, which can happen far from where they were first seen, instead of when a place is revisited (Sec. I; Sec. VI; Sec. VII; Fig. 2)","three threads in an ORB-SLAM-like structure: real-time scan-to-global-plane registration linearized to first order and iterated up to 5 times (threshold 0.5 deg) with bisquare weights; local plane adjustment (LM) over a sliding window of 8 keyframes and their planes with older poses fixed; global plane adjustment (LM) over all keyframe poses and planes; all with point-to-plane costs (Sec. V-C; Sec. VI; Sec. VII)","forward ICP flow tracks each plane's points from scan k-1 into scan k (2 nearest neighbours, RANSAC plane fit, expansion within 5 cm, at least 30 points, normal change below 15 deg); new planes matched to global planes by normal angle below 10 deg and mean point distance below 5 cm, plus a geometric consistency check; plane normals oriented toward the sensor to separate the two sides of walls and doors (Sec. IV-2; Sec. V; Sec. VI-A)","per-point linear interpolation of the relative pose inside the registration cost; new keyframe scans are undistorted before plane detection; the first frame is assumed static (Sec. V-A; Sec. V-C; Sec. VI-A)","keyframe trajectory and plane-segmented point map in which both sides of planar objects such as walls and doors are reconstructed separately (Fig. 8)",[143,45],{"id":5202,"label":5203,"shortName":5204,"title":5205,"year":207,"era":10,"cluster":263,"scope":99,"keyIdeaZh":5206,"sensors":5207,"mapRepresentation":5211,"loopClosure":1451,"estimator":5212,"association":5213,"deskew":5214,"outputGeometry":5215,"fulltextStatus":22,"lidarModels":5216,"equipmentCount":736},"fastlivo2rc2025","Zhou et al., 2025","FAST-LIVO2 on Resource-Constrained Platforms","FAST-LIVO2 on Resource-Constrained Platforms: LiDAR-Inertial-Visual Odometry With Efficient Memory and Computation","此研究針對邊緣運算平台精簡 FAST-LIVO2：以光達退化評估決定何時需要影像更新，在光達約束充足時減少視覺幀，降低計算量；地圖改為小範圍的統一視覺光達局部地圖加上稀疏的長期視覺地圖，以限制記憶體。作者在 Hilti（資料集描述含工地序列）16 個序列與私人序列（礦坑隧道、黑暗樹林、野外公園等）上測試，並在約 100 美元的 RK3588 ARM 板上於地下停車場與夜間街道進行即時定位測試。",[5208,5209,5210],"3D LiDAR (Hilti dataset LiDARs","Livox Mid-360 on the authors' rig","a small-FoV AVIA LiDAR, written 'Aivia' in Sec. V-D2, in the degeneration-detection test of Fig. 9, sequence not named) | IMU | camera (B\u002FW fisheye on the authors' rig) | 15 W onboard illuminator for extremely dark scenes","Compact unified local visual-LiDAR voxel map (hash of 0.5 m root voxels with three-level octrees; typical edge 200 m, slid every 20 m of motion) plus a sparse long-term visual map (edge 800 m, slid every 100 m) that keeps visual points leaving the local map","ESIKF with sequential updates (from FAST-LIVO2) plus a LiDAR-degeneration-aware adaptive visual frame selector","LiDAR point-to-plane and patch photometric errors as in FAST-LIVO2; images used only when LiDAR constraints are weak or keyframe criteria met","Uses undistorted points of recombined scans (scan recombination); the undistortion step itself is inherited from FAST-LIVO2 and not re-described","sharp point cloud maps and colored point clouds (Sec. V-C1; Fig. 10)",[507,437],{"id":5218,"label":5219,"shortName":5220,"title":5221,"year":249,"era":10,"cluster":263,"scope":132,"keyIdeaZh":5222,"sensors":5223,"mapRepresentation":5227,"loopClosure":5228,"estimator":5229,"association":5230,"deskew":5231,"outputGeometry":5232,"fulltextStatus":22,"lidarModels":5233,"equipmentCount":487},"camvox2021","Zhu et al., 2021","CamVox","CamVox: A Low-cost and Accurate Lidar-assisted Visual SLAM System","CamVox 把低成本的 Livox Horizon 固態 LiDAR 當作 ORB-SLAM2 的深度感測器：LiDAR 點先以 IMU 依各點時間校正運動畸變並轉到相機觸發時刻，再投影成與彩色影像逐像素對應的深度圖，組成 RGB-D 影格交給 ORB-SLAM2 的追蹤、局部建圖與迴圈閉合；由於 LiDAR 可量到上百公尺，深度小於 130 m 的特徵都視為近點。系統另利用 Livox 非重複掃描在靜止數秒後即可累積成高密度影像的特性，比對相機影像與 LiDAR 反射強度及深度影像的邊緣，於機器人靜止時自動做無標靶外參校正。",[5224,5225,5226],"monocular rolling-shutter camera (MV-CE060-10UC)","solid-state non-repeating-scan LiDAR (Livox Horizon)","IMU (Inertial Sense uINS) used only for LiDAR motion distortion correction","ORB-SLAM2 keyframes and map points; dense colored point cloud (RGB-D map) reconstructed from the frames for visualization (Fig. 1)","yes, ORB-SLAM2 loop closing (Sec. III; Fig. 4; Table II)","ORB-SLAM2 in RGB-D mode: tracking, local mapping with local bundle adjustment, loop closing and full bundle adjustment of ORB-SLAM2, fed with RGB-D frames built from the camera image and a depth image projected from IMU-corrected Livox points; keypoints with LiDAR depth below 130 m are treated as close points (Sec. III)","ORB features on the camera image with depth from the projected LiDAR depth image; extrinsic calibration by edge matching between the camera image and LiDAR reflectivity and depth images (Canny edges, edges shorter than 200 pixels and cluttered interior edges removed), a K-D-tree ICP cost with a mismatch penalty, and coordinate descent over roll, pitch and yaw in Ceres (Sec. III-C, Eq. 3 in the IEEE version)","IMU-based correction of each LiDAR point to the trigger time of the camera image; IMU at 200 Hz synchronized with the trigger (Sec. III-A, III-B); the IEEE version describes the correction as partial","camera trajectory and a dense RGB-colored point cloud map (Fig. 1; Fig. 6 in the IEEE numbering)",[1482],{"id":5235,"label":5236,"shortName":5237,"title":5238,"year":97,"era":10,"cluster":755,"scope":99,"keyIdeaZh":5239,"sensors":5240,"mapRepresentation":5241,"loopClosure":5242,"estimator":5243,"association":5244,"deskew":57,"outputGeometry":5245,"fulltextStatus":22,"lidarModels":5246,"equipmentCount":144},"niceslam2022","Zhu et al., 2022a","NICE-SLAM","NICE-SLAM: Neural Implicit Scalable Encoding for SLAM","NICE-SLAM 以多層級特徵格網搭配預先訓練的小型解碼器取代單一 MLP，使地圖更新可局部進行，改善大型室內場景的可擴展性與過度平滑問題。追蹤與建圖以深度與顏色重渲染誤差交替最佳化。作者指出方法沒有迴圈閉合，且預測能力受限於粗網格尺度。",[772],"hierarchical coarse\u002Fmid\u002Ffine feature grids with pre-trained occupancy decoders plus a colour grid","none (authors list loop closure as future work)","alternating gradient-based optimization in parallel threads: staged mapping (mid-level grid, then mid and fine grids with the depth L1 loss) followed by a local bundle adjustment that jointly optimizes all feature grids, the colour decoder and the poses of K selected keyframes (Eq. 10); tracking optimizes only the current camera pose with a variance-weighted depth loss plus a photometric loss (Eq. 11 and 12)","direct depth (L1) and photometric re-rendering losses","mesh via marching cubes: fine-level decoder occupancy for observed points; for unseen points inside partially observed coarse voxels the coarse decoder predicts occupancy (shown in cyan); other points set to zero occupancy",[],{"id":5248,"label":5249,"shortName":5250,"title":5251,"year":97,"era":10,"cluster":397,"scope":1035,"keyIdeaZh":5252,"sensors":5253,"mapRepresentation":5256,"loopClosure":72,"estimator":5257,"association":5258,"deskew":5259,"outputGeometry":5260,"fulltextStatus":22,"lidarModels":5261,"equipmentCount":109},"zhu2022liinit","Zhu et al., 2022b","LI-Init","Robust Real-time LiDAR-inertial Initialization","LI-Init 在 LiDAR 慣性里程計啟動前，自動判斷資料激勵是否足夠，並線上估計 LiDAR 與 IMU 的時間偏移、外參、重力向量與 IMU 偏差。時間偏移先以互相關粗估，再與旋轉外參聯合最佳化。作者強調若時間偏移未知，依賴 IMU 的運動畸變補償就無法正確執行。",[5254,5255],"3D LiDAR: Livox Avia (small FoV), Livox Mid360 (non-repetitive scanning), Hesai PandarXT (mechanical spinning); all set to 10 Hz","6-axis IMU Bosch BMI088, both inside the Livox LiDARs (factory hardware-synchronized) and inside a Pixhawk flight controller (unsynchronized with PandarXT); raw data 200 Hz","Point map used for scan-to-map registration inside the FAST-LIO2-derived LiDAR odometry; the data structure is not described in this paper","LiDAR odometry with error-state iterated Kalman filter, then cross-correlation temporal alignment and joint temporal-spatial optimization","LiDAR-only odometry uses scan-to-map point-to-plane residuals (modified from FAST-LIO2), chosen over NDT scan-to-scan so that non-repetitive LiDARs are supported; LiDAR and IMU data are associated by cross-correlation of angular-velocity magnitudes, then least-squares alignment of angular velocity and acceleration","During initialization, points are deskewed without the IMU: each point is projected to the scan-end frame using the constant-velocity prediction, because the unsynchronized IMU cannot yet be used; Fig. 3 compares Mid360 maps with and without this compensation; after initialization FAST-LIO2 uses the calibrated offset","time offset, extrinsic, gravity, IMU bias as initial states for FAST-LIO2",[437,507,3234],{"id":5263,"label":5264,"shortName":5265,"title":5266,"year":661,"era":10,"cluster":755,"scope":99,"keyIdeaZh":5267,"sensors":5268,"mapRepresentation":5269,"loopClosure":5270,"estimator":5271,"association":5272,"deskew":57,"outputGeometry":5273,"fulltextStatus":22,"lidarModels":5274,"equipmentCount":185},"nicerslam2024","Zhu et al., 2024","NICER-SLAM","NICER-SLAM: Neural Implicit Scene Encoding for RGB SLAM","NICER-SLAM 是只用單眼 RGB 影像的神經隱式 SLAM，追蹤與建圖共用同一個階層式 SDF 表示：粗層為 32 立方的稠密特徵格網，細層以多解析度網格學習殘差 SDF，另以多解析度網格表示顏色。因為沒有深度量測，建圖時額外加入 Omnidata 的單眼深度與法向量、GMFlow 光流、影像扭曲與 Eikonal 等損失來消除歧義，並依各體素取樣次數局部調整 SDF 轉密度的參數。系統在 Replica 上的幾何品質接近 RGB-D 方法，但追蹤不如 DROID-SLAM，未做迴圈閉合，也遠非即時。",[5014],"hierarchical neural implicit SDF: coarse 32^3 dense feature grid plus 8-level fine residual grids (32-128) and a 16-level colour grid (16-2048) with small MLP decoders; VolSDF-style SDF-to-density with a locally adaptive beta from per-voxel sample counts (Sec. 3.1-3.2)","none (stated as a limitation, Sec. 5)","end-to-end optimization through differentiable volume rendering: tracking optimizes the current pose with an RGB rendering loss (100 iterations, 1024 pixels) with the map fixed; mapping runs a 3-stage optimization with RGB, warping, optical-flow, monocular depth, monocular normal and Eikonal losses, ending with local bundle adjustment over 16 selected frames of which half are frozen (Sec. 3.3; iteration counts from arXiv v1 Sec. 3.4)","direct photometric rendering loss plus dense correspondence cues: RGB warping between keyframes and optical flow from GMFlow (Sec. 3.3)","camera trajectory and a triangle mesh extracted by marching cubes at 512^3; rendered novel views (Sec. 3.4)",[],{"id":5276,"label":5277,"shortName":5278,"title":5279,"year":207,"era":10,"cluster":31,"scope":99,"keyIdeaZh":5280,"sensors":5281,"mapRepresentation":5282,"loopClosure":5283,"estimator":5284,"association":5285,"deskew":5286,"outputGeometry":5287,"fulltextStatus":22,"lidarModels":5288,"equipmentCount":487},"zhu2025meshloam","Zhu et al., 2025","Mesh-LOAM","Mesh-LOAM: Real-Time Mesh-Based LiDAR Odometry and Mapping","Mesh-LOAM 以隱式移動最小平方（IMLS）函數估計 SDF，但讓體素被動接收周圍點的 SDF 增量（passive voxel），避免逐體素搜尋近鄰，使每次掃描只需走訪各點一次；體素存於 GPU 平行空間雜湊表，並以 marching cubes 分區擷取網格。位姿以點對網格（point-to-mesh）里程計估計。",[35],"sparse passive voxels (0.1 m) storing position, normal, IMLS-based SDF, weight and frame index, updated by hybrid distance and normal weights over a 3-voxel influence cube; GPU spatial hash with linear probing; height-adaptive voxel blocks with length and width set to 2 (unit not stated) for partitioned marching cubes; expired voxels converted to mesh and deleted to bound memory; marching cubes at dynamic intervals (Sec. III-C, III-D, IV-A)","none described (odometry and mapping only)","scan-to-mesh odometry minimising point-to-facet-plane residuals by Gauss-Newton on SE(3) for a relative correction applied to a constant-velocity pose prediction (Sec. III-B.3)","planar points (PCA curvature below 0.1) are transformed by a constant-velocity prediction and matched to mesh facets by nearest-neighbour search; a match is kept by point-to-facet distance and only if the absolute cosine similarity of point and facet normals exceeds 0.98 (Sec. III-B, IV-A)","not_reported (LiDAR-only pipeline; no motion-compensation step described in arXiv v1)","triangle mesh",[161,224,5289,624],"virtual Velodyne HDL-64 LiDAR",{"id":5291,"label":5292,"shortName":5293,"title":5294,"year":51,"era":10,"cluster":263,"scope":132,"keyIdeaZh":5295,"sensors":5296,"mapRepresentation":5299,"loopClosure":5300,"estimator":5301,"association":5302,"deskew":637,"outputGeometry":5303,"fulltextStatus":22,"lidarModels":5304,"equipmentCount":109},"iriom4d2023","Zhuang et al., 2023","4D iRIOM","4D iRIOM: 4D Imaging Radar Inertial Odometry and Mapping","4D iRIOM 以 4D 成像雷達加 IMU 做里程計與建圖：每張雷達掃描先用漸進非凸（GNC）方法估計自身速度，排除移動物與多路徑造成的離群點，再把稀疏雷達點與局部子地圖的多個鄰近點以協方差加權配準；兩類量測都送入迭代擴展卡爾曼濾波器更新。最後以 scancontext 偵測回訪並建立位姿圖做迴圈閉合，得到全域一致的雷達點雲地圖。作者指出 LiDAR 與相機在雨、霧等劣化環境表現下降，毫米波雷達則能在霧中量測距離、方向與都卜勒速度。",[5297,5298],"4D imaging radar (Continental ARS548)","IMU (EPSON G345 inside the Bynav X1-5H GNSS\u002FINS)","radar point submap for scan-to-submap matching; global radar point map after pose graph optimization (Figs. 1, 5)","yes, scancontext place recognition with a descriptor threshold tuned for 4D radar; GICP relative constraints (Sec. III)","iterated extended Kalman filter with IMU propagation; state includes radar-IMU extrinsics and gravity; updates from radar ego-velocity and from scan-to-submap point matches; loop closure by scancontext detection with GICP constraints in a pose graph (Sec. III)","ego-velocity estimated from each radar scan by graduated non-convexity (GNC) with outlier relaxation and chi-square pruning at significance 0.05; scan-to-submap matching of sparse radar points against N = 5 nearest submap points weighted by covariance (distribution-to-multi-distribution) in an incrementally updated map (Sec. III)","robot trajectory and sparse radar point cloud map",[143],{"id":5306,"label":5307,"shortName":5308,"title":5309,"year":115,"era":52,"cluster":11,"scope":132,"keyIdeaZh":5310,"sensors":5311,"mapRepresentation":5315,"loopClosure":5316,"estimator":5317,"association":5318,"deskew":5319,"outputGeometry":5320,"fulltextStatus":22,"lidarModels":5321,"equipmentCount":109},"zlot_bosse2014_mine","Zlot & Bosse, 2014","CSIRO underground mine CT-SLAM (Northparkes)","Efficient Large‐scale Three‐dimensional Mobile Mapping for Underground Mines","本文把 CSIRO 的連續時間非剛性配準（原始版本出自 bosse_zlot2009_ctscan）擴展成完整的地下礦坑建圖流程：旋轉 SICK LMS 291 與 MEMS IMU 裝在皮卡車斗上，於澳洲 Northparkes 銅金礦以一般行車速度行駛 17.1 公里（含停車共 1 小時 53 分）。處理分三段：先以滑動視窗的非剛性配準（面元匹配加 IMU 約束，修正量以 B-spline 表示）產生開迴路軌跡；再以關鍵點投票的場所辨識找出回溯路段的迴圈，經穩健位姿圖最佳化粗對齊後，對整條軌跡做非剛性配準；最後把點雲配準到礦方既有的測量剖面，以定位到地理座標並修正殘餘漂移。作者報告總處理時間 53.9 分鐘，少於擷取時間一半；開迴路軌跡相對於「已配準到測量圖」的軌跡，前進方向偏差約 0.2%；配準後 95% 面元距測量面元 50 cm 內（RMS 25 cm）。須注意這些與測量圖的比較是在把點雲配準到同一份測量圖之後計算，不是獨立驗證。",[5312,5313,5314],"rotating SICK LMS 291 2D laser, one revolution every 2 s, spin axis pitched 65 deg from horizontal (Sec. 3.1)","MicroStrain 3DM-GX2 industrial-grade MEMS IMU on the non-spinning part of the mount (Sec. 3.1)","two fixed vertical SICK LMS 291 lasers in a pushbroom configuration, used only for surface reconstruction, not for trajectory estimation (Sec. 3.1, 4.5)","surfels for registration; output 3D point cloud (87 million georeferenced points) and a triangulated surface mesh from the vertical lasers, decimated to one-eighth resolution for the user (Sec. 4.5, 6)","place recognition by keypoint voting: 40 cm downsampled cloud, 20% random keypoints (planar ones discarded), 3D gestalt descriptors over 4 m reduced from 66 to 10 dimensions, places of 3,500 keypoints, RANSAC geometric verification; no false positives on this data set; the mine has no topological loops, so closures come only from backtracking (Sec. 2.2, 4.2)","continuous-time non-rigid registration: baseline trajectory plus low-bandwidth corrections parameterized as uniform B-splines (first-order, 0.1 s knots, 4 s window with 2 s shift for the open-loop stage; cubic, 5 s knots for whole-trajectory registration); stacked linearized constraints (surfel match, fixed-surfel, IMU accelerometer and gyro deviation, reference-velocity deviation, smoothness, initial conditions) solved by iteratively reweighted least squares with Cauchy weights and sparse Cholesky factorization (Sec. 2.1, 4.1, 4.2)","surfels from first and second moments in a multi-resolution voxel grid (0.4, 0.8, 1.6, 3.2 m; max 0.5 s time span; at least 15 points and two scans); approximate k-nearest-neighbour search (four neighbours at least 0.5 s apart, modified libnabo) in a weighted position-normal space, with removal of distant and non-reciprocal matches (Sec. 2.1.3, 4.1)","implicit in continuous-time non-rigid registration; surfel time spans capped (0.5 s in the open-loop stage) to limit uncorrected distortion (Sec. 4.1)","6-DoF trajectory (open-loop, closed-loop and survey-registered), georeferenced 3D point cloud and triangulated surface model (Sec. 4, 6)",[5322,5323],"SICK LMS 291 (rotating)","SICK LMS 291 (two fixed, vertical)",{"id":5325,"label":5326,"shortName":5327,"title":5328,"year":661,"era":10,"cluster":599,"scope":132,"keyIdeaZh":5329,"sensors":5330,"mapRepresentation":5331,"loopClosure":5332,"estimator":5333,"association":5334,"deskew":5335,"outputGeometry":5336,"fulltextStatus":22,"lidarModels":5337,"equipmentCount":109},"ltaom2024","Zou et al., 2024","LTA-OM","LTA‐OM: Long‐term association LiDAR-IMU odometry and mapping","LTA-OM 以 FAST-LIO2 作為光達慣性里程計、以 STD 作為迴圈偵測，整合迴圈校正、誤判迴圈剔除、長期關聯（long-term association, LTA）建圖與多時段定位建圖。其 LTA 建圖把校正後的歷史地圖直接作為 LIO 掃描對地圖配準的全域約束，使回到舊地點時里程計不再漂移；多時段模式可儲存校正後的地圖點、最佳化軌跡與描述子資料庫，供後續時段接續使用。",[35,36],"ikd-Tree live map holding recent scan points and corrected, on-tree downsampled history submap points loaded around the current position; stored multisession prior map with submap poses and STD descriptor database (Sec. 4.1, 4.4.1, 4.5)","STD-LCD loops on 20-scan submaps; only loops with overlap ratio above 0.5 are trusted; each loop is optimized after backing up the graph and rejected (graph restored) if key-point factor residuals exceed a threshold (2 in the benchmarks); the first loop or loops far from the current pose require two consecutive mutually consistent loops (Sec. 4.2, 4.3.2, 5.1)","FAST-LIO2 tightly coupled iterated Kalman filter LIO with long-term association (history map points reloaded into the ikd-Tree); separate loop optimization on a pose graph of submap poses with odometry factors plus neighbouring and loop-closure key-point factors (STD key points), solved with GTSAM and iSAM2; false-positive rejection by an optimize-and-recover graph consistency check (Sec. 4.1, 4.3)","Direct point-to-plane scan-to-map registration on the ikd-Tree (plane fitted to five neighbours); STD-LCD on submaps of 20 accumulated scans: triangle descriptors of key points, hash-table rough detection, transform clustering with more than 4 supporters, and plane-to-plane overlap verification; neighbouring key-point pairs associated by kd-tree radius search (Sec. 4.1, 4.2, 4.3.1)","FAST-LIO2 backward propagation with the IMU kinematic model compensates motion distortion of each scan (Sec. 4.1)","corrected point-cloud map and optimized trajectory; multi-session stitched map (abstract)",[437,45],{"id":5339,"label":5340,"shortName":5341,"title":5342,"year":9,"era":10,"cluster":263,"scope":346,"keyIdeaZh":5343,"sensors":5344,"mapRepresentation":5347,"loopClosure":5348,"estimator":5349,"association":5350,"deskew":5351,"outputGeometry":5352,"fulltextStatus":22,"lidarModels":5353,"equipmentCount":109},"licfusion2019","Zuo et al., 2019","LIC-Fusion","LIC-Fusion: LiDAR-Inertial-Camera Odometry","LIC-Fusion 在多狀態約束卡爾曼濾波器（MSCKF）架構中，緊密融合 IMU、稀疏視覺特徵，以及從光達掃描中擷取並追蹤的邊緣與平面特徵點。其特色是線上估計三種非同步感測器之間的空間外參與時間偏移，以因應低成本裝置的延遲與時鐘偏差。系統為純里程計，不維護全域地圖，也不使用迴圈。",[1201,2669,5345,5346],"monochrome global-shutter camera","GNSS RTK (reference only)","none (sliding window of cloned states; no global map maintained)","none (purely odometry, no global map; Sec. III)","MSCKF whose state holds the IMU state, camera-IMU and LiDAR-IMU extrinsics with time offsets (IMU clock as reference), and sliding windows of IMU clones at camera and LiDAR times; standard EKF update after measurement compression","LiDAR edge (high curvature) and surf (low curvature) points from scan rings tracked from the current scan to the previous scan by KD-tree nearest neighbours, giving point-to-line and point-to-plane distances with propagated covariance and chi-square Mahalanobis gating; FAST visual features tracked by KLT, triangulated from camera clones and used after MSCKF nullspace projection; all residuals compressed by Givens-rotation thin QR","Not described: the full text contains no LiDAR motion-distortion compensation step; edge and surf features are taken from raw scan rings","trajectory only (no map product reported)",[143],{"id":5355,"label":5356,"shortName":5357,"title":5358,"year":396,"era":10,"cluster":263,"scope":346,"keyIdeaZh":5359,"sensors":5360,"mapRepresentation":5363,"loopClosure":1451,"estimator":5364,"association":5365,"deskew":5366,"outputGeometry":5352,"fulltextStatus":22,"lidarModels":5367,"equipmentCount":24},"licfusion2_2020","Zuo et al., 2020","LIC-Fusion 2.0","LIC-Fusion 2.0: LiDAR-Inertial-Camera Odometry with Sliding-Window Plane-Feature Tracking","LIC-Fusion 2.0 將光達處理改為滑動視窗內的平面特徵追蹤：以 IMU 做運動補償後擷取低曲率平面點，跨多次掃描追蹤並初始化平面，且考慮幀間轉換不確定性來剔除錯誤匹配。論文同時分析光達慣性子系統在平面特徵下的可觀測性（observability），指出某些運動會使時空校正退化，並以蒙地卡羅模擬驗證估計一致性。",[4467,5361,5362],"IMU (Xsens, model not reported, 400 Hz)","monocular global-shutter camera (model not reported, 20 Hz, 1920x1200)","sliding-window plane features and SLAM plane landmarks; no global map reported","sliding-window MSCKF-type filter with online spatio-temporal calibration; visual pipeline based on OpenVINS","IMU-motion-compensated low-curvature planar points tracked across the sliding window as plane features, with an outlier test accounting for inter-frame transformation uncertainty; sparse visual features","IMU-propagated poses are buffered and interpolated to each LiDAR point time (SO(3) interpolation for orientation, linear for position); all points are transformed to the pose at the sweep start time (Sec. III-A)",[143],{"id":5369,"label":5370,"shortName":5371,"title":5372,"year":97,"era":10,"cluster":236,"scope":99,"keyIdeaZh":5373,"sensors":5374,"mapRepresentation":5375,"loopClosure":5376,"estimator":5377,"association":5378,"deskew":57,"outputGeometry":5379,"fulltextStatus":22,"lidarModels":5380,"equipmentCount":24},"dmvio2022","von Stumberg & Cremers, 2022","DM-VIO","DM-VIO: Delayed Marginalization Visual-Inertial Odometry","DM-VIO 是單目視覺慣性里程計，以 DSO 的直接光度光束法平差為核心，加入 IMU 預積分並把尺度與重力方向作為顯式變數持續最佳化。作者提出延遲邊際化：另外維護一個延遲 100 個關鍵影格才邊際化的因子圖，可在其中加入 IMU 因子做位姿圖光束法平差（PGBA），以完整的光度不確定度初始化 IMU，並重新推進該圖得到含 IMU 資訊的邊際化先驗；尺度大幅改變時也能替換邊際化先驗。另以動態光度權重在影像品質差時提高 IMU 比重。",[37,36],"sparse point cloud of inverse-depth points hosted in active keyframes (Fig. 1 point clouds)","none (odometry; loop closure and map reuse named as future work)","direct (DSO-based) photometric visual-inertial bundle adjustment over up to 8 active keyframes with IMU preintegration, dynamic photometric weight, and explicit scale and gravity-direction variables; Schur-complement partial marginalisation with FEJ plus a second delayed marginalisation graph (delay 100) used for pose graph bundle adjustment (PGBA) IMU initialisation and marginalisation replacement; Levenberg-Marquardt with SIMD photometric code and GTSAM (Sec. III)","direct: photometric residuals of sparse DSO points over a neighbourhood pattern with affine brightness and exposure; no feature matching (Sec. III-B)","metric-scale camera and IMU trajectory with sparse point cloud (Fig. 1)",[],1790510650038]