SLAM 點雲圖鑑Point clouds for construction
M-01 · 2021

IN2LAAMA

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。

具全域修正的完整 SLAM近十年已讀全文同儕審查已出版當代 LiDAR（慣性）里程計與 SLAM

閱讀版本
arXiv v3 (2020-10-22), the accepted T-RO version with IEEE copyright notice; IEEE version of record not read
全文取得途徑
arXiv 預印本
本頁內容

Offline full-batch LiDAR-inertial localisation, mapping and targetless extrinsic and time-shift calibration that uses Gaussian-process-upsampled IMU preintegration at every point timestamp (UPMs) to remove motion distortion without a motion model, jointly optimizing point-to-line, point-to-plane, IMU, bias and time-shift factors with periodic feature recomputation and proximity loop closures.

技術屬性

欄位內容為文獻擷取紀錄的原文用語（英文），以原文為據；「未查證」表示本研究尚未讀到該資訊，不代表該方法不具備此能力。

感測輸入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)
原文測試平台authors' self-contained VLP-16 and Xsens MTi-3 sensor suite, carrier not stated (indoor lab, staircase and calibration runs)、vehicle (car, MC2SLAM campus drive)、simulation
狀態估計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)
資料關聯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)
時間表示discrete states at the start of each LiDAR frame; within frames, Upsampled Preintegrated Measurements (UPMs) from per-axis Gaussian-process regression of IMU readings give the pose at every point timestamp without an explicit motion model (Sec. III-C)
去畸變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)
迴圈閉合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)
全域最佳化the whole trajectory, biases, time-shifts and optionally extrinsics are estimated in one batch problem including loop-closure factors (Sec. III-B, VI-A)
地圖表示dense point cloud reconstructed by projecting all points with the final trajectory and UPMs; no persistent map structure (Sec. VI-E)
先驗資訊no trajectory prior; Gaussian prior on inter-sensor time-shift; temporary zero prior on the initial accelerometer bias to handle translation-only starts (Sec. III-B, VI-D)
可輸出幾何full 6-DoF trajectory, IMU biases, time-shift, LiDAR-IMU extrinsic calibration and dense motion-corrected point cloud map (Figs. 1, 11-13)
計算需求offline, not real time: 950 s for a 41 s lab sequence, 1306 s for a 35 s staircase sequence and 4699 s for an 85 s outdoor HDL-32 sequence, with 4.2 to 6.45 GiB data memory plus up to 5.77 GiB optimizer memory; UPMs need more than 340 MB per second of VLP-16 data; hardware not reported (Table VII, Sec. VI-E)

使用設備

原文使用的感測器、運算硬體與載具（equipment）。型號保留原文寫法，連結到設備頁中同一型號的歸併名稱；角色依原文用途分為方法輸入、資料集感測器、執行運算平台、參考或真值量測（reference or ground truth）與比較對象設備。

類別型號（原文寫法）角色資料集原文規格出處
LiDARVelodyne VLP-16方法輸入IN2LAAMA UTS datasets (lab, staircase, calibration)16 channels (plus or minus 15 deg), 10 Hz, 300k points per second, noise plus or minus 3 cm(Le Gentil et al., 2021, Sec. VII; Fig. 1)
LiDARVelodyne HDL-32歸入：Velodyne HDL-32E資料集感測器MC2SLAM dataset (campus drive)about four times the data of the VLP-16; roof of a car(Le Gentil et al., 2021, Sec. VII-E-2; Table VII)
慣性量測單元（IMU）Xsens MTi-3方法輸入IN2LAAMA UTS datasets (lab, staircase, calibration)100 Hz; noise 0.02 m/s2 and 0.097 deg/s; low cost(Le Gentil et al., 2021, Sec. VII; Sec. VII-E)
慣性量測單元（IMU）Velodyne HDL-32 built-in IMU資料集感測器MC2SLAM dataset (campus drive)原文未報告(Le Gentil et al., 2021, Sec. VII-E-2)
相機Intel Realsense D435歸入：Intel RealSense D435比較對象設備IN2LAAMA UTS datasets (lab, staircase, calibration)RGB camera used only for the chained IMU-camera-lidar calibration baseline(Le Gentil et al., 2021, Sec. VII-F)
載具平台car資料集感測器MC2SLAM dataset (campus drive)driven around a university campus(Le Gentil et al., 2021, Sec. VII-E-2)

作者報告的優勢與限制

優勢

  • All UPM variants were at least an order of magnitude more accurate than standard preintegration at the LiDAR point times, with GP regression best (Sec. VII-A, Fig. 9)
  • With IMU factors it handled fast simulated motion (average 125 deg/s) without failures where the authors' earlier IN2LAMA failed in 37 of 50 runs (Table I)
  • Staircase wall RMS point-to-plane distance 10 mm versus 31 mm for IN2LAMA and 60 mm for A-LOAM (Table VI)
  • Removing the constant-velocity assumption cut RMSE position error from 2.67 m to 0.087 m in fast simulated motion (Table IV)
  • Targetless extrinsic calibration reached about 1 cm and below 0.13 deg error in simulation, and on real data gave crisper maps (27 mm mean plane RMS) than a camera-chained calibration (58 mm) (Table V, Sec. VII-F)

限制

  • Offline only: execution times are 23 to 55 times the sequence duration with several GiB of memory (Table VII, Sec. VIII)
  • Front-end designed for structured geometry; outdoor scenes are challenging for the feature extraction (Sec. VII-E-2)
  • Loop closure uses simple pose proximity and does not handle large drift or kidnapped cases (Sec. V-D)
  • Accelerometer biases are unobservable on translation-only segments; an extra zero-bias factor is needed (Sec. VI-D, VII-E-2)
  • Calibration requires trajectories that excite all six IMU axes and observe at least three non-coplanar planes or non-collinear edges (Sec. VII-F)

營建工程相關證據

論文未在營建工地驗證，真實資料為 UTS 實驗室、跨樓層樓梯間與校園道路。其實驗以人工分割的牆面與樓板計算點到平面 RMS 距離，並把實驗室地圖疊合於數位化的施工前平面圖，作者也指出平面圖因結構變更與家具而與現況不符；這種以平面殘差評估點雲品質的方式與營建驗收的平整度檢查概念相近（推論）。離線全批次、可自動校正外參與時間偏移的特性，適合事後處理的點雲測繪服務（作者以 3D mapping as a service 為例，Sec. VIII），但計算成本高。同作者的元件研究見（Le Gentil et al., 2018）與（Le Gentil et al., 2020）。

原文驗證環境：模擬、已完工建築、跨場域

報告的性能數據

以下是原文作者報告的性能數值（author-reported results），不是本研究重新量測的結果。每張圖只並列同一個比較組（comparison group，同一張表、同一組實驗設定）內的方法；不同比較組之間的數值不可直接比較，也不構成排名。

本方法共出現在 10 個比較組，合計 64 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 6 組列在最後，並連到性能比較頁。

Le Gentil et al., 2021 · Table I 本方法 21 筆

表格設定（擷取紀錄原文）：Simulated odometry set-up, 50-run Monte Carlo, loop closure off; trajectories average 288.7 m at 4.85 m/s (max 7.35 m/s); errors on successful runs only (favours [5] in Fast); values are mean with plus-minus spread (Le Gentil et al., 2021, Table I)

Num. fails; as printed: 0，IN2LAAMA simulation (virtual room with 7 planes, VLP-16 and MTi-3 models) · Slow (avg 14.7, max 22.1 deg/s)

只並列這張表在相同設定下報告的方法；以「本方法：」開頭者為本頁方法。失敗、未執行與未報告以標記呈現，不是 0。

這些是 Le Gentil et al., 2021 在此表設定下報告的數值（author-reported results），只能在同一個比較組內對照，不代表方法在其他資料或設定下的表現。

統計量：原文未報告；對齊方式：原文未報告；單位：count；場景：simulated room

資料來源作者報告值(Le Gentil et al., 2021, Table I)

數值與出處
方法（原文寫法）報告值出處
[10] (A-LOAM implementation of LOAM)0 count(Le Gentil et al., 2021, Table I)
[5] IN2LAMA (no IMU factors)0 count(Le Gentil et al., 2021, Table I)
IN2LAAMA本方法原文提出0 count(Le Gentil et al., 2021, Table I)

Le Gentil et al., 2021 · Table VII 本方法 9 筆

表格設定（擷取紀錄原文）：Memory consumption and execution time of the offline batch optimisation on real data; Ne frames between optimisations; outdoor uses HDL-32 (about four times the VLP-16 data) and optional ICP loop tests (1588 s of the total) (Le Gentil et al., 2021, Table VII)

Mem. data/prog.，IN2LAAMA UTS datasets and MC2SLAM dataset · Lab (41 s)

這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Le Gentil et al., 2021 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。

統計量：原文未報告；對齊方式：不適用；單位：GiB；場景：offline processing

數值與出處
方法（原文寫法）報告值出處
IN2LAAMA本方法原文提出4.7 GiB(Le Gentil et al., 2021, Table VII)

Le Gentil et al., 2021 · Table II 本方法 8 筆

資料集與序列IN2LAAMA simulation · closed loops (50 runs)

表格設定（擷取紀錄原文）：50 simulated closed trajectories (mean 210 m, 3.53 m/s, 8.16 deg/s); IN2LAAMA with and without loop closure; mean with plus-minus spread (Le Gentil et al., 2021, Table II)

Final pose error (position); as printed: 0.110 m ± 0.043，IN2LAAMA simulation · closed loops (50 runs)

這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Le Gentil et al., 2021 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。

統計量：平均值（mean）；對齊方式：原文未報告；單位：m；場景：simulated room

數值與出處
方法（原文寫法）報告值出處
IN2LAAMA (Without loop closure)本方法原文提出0.11 m(Le Gentil et al., 2021, Table II)

Le Gentil et al., 2021 · Table V 本方法 8 筆

資料集與序列IN2LAAMA simulation · initial guess error 0.17 m, 1.74 deg

表格設定（擷取紀錄原文）：Simulated LiDAR-IMU extrinsic calibration accuracy for four initial-guess error levels; 50 runs of 19.6 s Fast trajectories; mean with plus-minus spread (Le Gentil et al., 2021, Table V)

calibration translation error; as printed: 10.5e-3 ± 6.34e-3，IN2LAAMA simulation · initial guess error 0.17 m, 1.74 deg

這張表在此指標與資料序列只列出本方法一筆，沒有可並列的其他方法，因此不畫圖，數值與出處見下表。這是 Le Gentil et al., 2021 在此表設定下報告的數值（author-reported results），不代表方法在其他資料或設定下的表現。

統計量：平均值（mean）；對齊方式：不適用；單位：m；場景：simulated room

數值與出處
方法（原文寫法）報告值出處
IN2LAAMA calibration本方法原文提出0.0105 m(Le Gentil et al., 2021, Table V)

其他比較組

列出其餘 6 個比較組

來源

  • Le Gentil et al., 2021

    Cedric Le Gentil, Teresa Vidal-Calleja, Shoudong Huang（2021）IN2LAAMA: Inertial Lidar Localization Autocalibration and MappingIEEE Transactions on Robotics, 37(1):275-290

    DOI 10.1109/tro.2020.3018641arXiv 1905.09517

    同儕審查已出版已讀全文近十年查證後修正

回到方法圖鑑

SLAM 點雲圖鑑

本站為綜述論文《SLAM-Based Point Clouds for Construction: A Review》研究過程的互動說明文件。互動圖為教學模擬，不代表任何特定系統的實測表現；方法描述、數值與結論以所引原始文獻為準。

內容基準日 2026-09-25 · 研究與網站製作使用 AI 輔助工具，見 證據標記與方法

選擇開啟Esc關閉