LIPS
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。
本頁內容
Graph-based LiDAR-inertial plane SLAM that parameterizes infinite planes by their closest point to the frame origin, anchors each plane in its first observing frame to avoid the zero-distance singularity, compresses RANSAC-segmented points into local closest-point measurements with covariance, and fuses them with continuous IMU preintegration in iSAM2.
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 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) |
|---|---|
| 原文測試平台 | simulation、sensor unit (Quanergy M8 with IMU attached underneath) moved in front of planar boards in a small indoor scene; carrier not described (Sec. VI-D, Fig. 6) |
| 狀態估計 | 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) |
| 資料關聯 | 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) |
| 時間表示 | discrete IMU states at LiDAR times linked by closed-form continuous preintegration (Sec. IV) |
| 去畸變 | none in the reported experiments; the authors note preintegration could unwarp clouds at high speed but did not use it (Sec. III-B) |
| 迴圈閉合 | implicit through re-observation of previously estimated plane landmarks (no separate place recognition) (Sec. VI-B) |
| 全域最佳化 | full smoothing of the IMU state history and plane landmarks with iSAM2 (Sec. III-B) |
| 地圖表示 | sparse landmark map of infinite planes, each stored as a closest-point vector anchored in the frame of its first observation (Sec. V-B) |
| 先驗資訊 | LiDAR-IMU extrinsic estimated manually for the real test (Sec. VI-D); simulator uses the known extrinsic of Table I; no map prior |
| 可輸出幾何 | IMU trajectory and a set of plane parameters; no dense point cloud output is described |
| 計算需求 | measurement compression and iSAM2 estimator ran in real time in the reported tests, but RANSAC plane extraction was run offline and the authors state it must be replaced or accelerated for real use; hardware not reported (Sec. VI-B, VI-D) |
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 |
|---|---|---|---|---|---|
| LiDAR | Quanergy M8 | 方法輸入 | 未標示 | eight-channel LiDAR operating at 10 Hz | (Geneva et al., 2018, Sec. VI-D) |
| LiDAR | Quanergy M8 (simulated) | 資料集感測器 | LIPS simulator (extruded floor plan) | simulator modelled on it: 0.25 deg angular resolution, 8 zenith angles from 3.2 to -18.3 deg, 1 cm and 3 cm point deviation, 5 Hz | (Geneva et al., 2018, Sec. VI-A; Table I) |
| 慣性量測單元（IMU） | Microstrain 3DM-GX3-25 | 方法輸入 | 未標示 | attached to the bottom of the LiDAR, 500 Hz | (Geneva et al., 2018, Sec. VI-D) |
| 慣性量測單元（IMU） | ADIS16448 (simulated) | 資料集感測器 | LIPS simulator (extruded floor plan) | gyro noise density 0.005 rad/s/sqrt(Hz), accel noise density 0.01 m/s2/sqrt(Hz), 800 Hz | (Geneva et al., 2018, Sec. VI-A; Table I) |
作者報告的優勢與限制
優勢
- Closest-point planes gave lower average RMSE than the relative quaternion plane factor of Kaess in 80 Monte-Carlo runs at 1 cm and 3 cm LiDAR noise (Table II)
- Authors observed faster convergence than the quaternion form in full batch optimization (Sec. VI-C)
- Re-observing previously seen hallway planes after 300 s rapidly reduced the estimation error, acting as loop closure (Sec. VI-B, Fig. 4)
- Real test: 1.5 cm start-to-end difference after a 30 m trajectory, 0.05% of path length (Sec. VI-D, Fig. 7)
限制
- Real-world validation is a small proof of concept with planar boards placed around the sensor to ease RANSAC extraction and avoid degenerate motion (Sec. VI-D)
- RANSAC plane extraction ran offline and needs replacement or acceleration for real-world use (Sec. VI-D)
- Poses may be under-constrained when extracted planes do not constrain all degrees of freedom (Sec. VII)
- Simple Mahalanobis-distance plane matching was only shown sufficient for small-scale experiments (Sec. VI-D)
- LiDAR-IMU extrinsic estimated manually; no scan deskewing used (Sec. VI-D, III-B)
營建工程相關證據
論文以室內人造環境為目標，模擬為 2D 平面圖垂直拉伸的曼哈頓式建物，真實測試僅在人工擺放平板的小場景中進行，未在營建工地或完成建物中以獨立參考驗證。平面地標的最小參數化對牆、樓板等大平面主導的室內施工場景有參考價值，也與平面型 BIM 元件的對應概念相近，但平面擷取與對應在真實雜亂工地的可行性未經檢驗（推論）。
原文驗證環境：模擬、受控實驗
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 2 個比較組，合計 5 筆紀錄。
Geneva et al., 2018 · Table II 本方法 4 筆
表格設定（擷取紀錄原文）：Average RMSE over 80 Monte-Carlo simulations on the 180 m simulated indoor trajectory; known plane correspondences; iSAM2; closest-point (CP) versus relative quaternion plane factor (Geneva et al., 2018, Table II)
Average RMSE (position)，LIPS simulator (extruded floor plan) · LiDAR noise 1 cm
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。
按 Tab 進入圖表後，用上下方向鍵逐一瀏覽各類別，Esc 關閉提示框；也可開啟表格檢視閱讀全部數值。
這些是 Geneva et al., 2018 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。
資料來源作者報告值(Geneva et al., 2018, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Closest Point本方法原文提出 | 0.005 m | (Geneva et al., 2018, Table II) |
| Quaternion [6] (Kaess relative quaternion factor) | 0.016 m | (Geneva et al., 2018, Table II) |
Geneva et al., 2018 · Text Sec.VI-D 本方法 1 筆
指標difference between the start and end poses
資料集與序列authors' real-world test (planar boards) · 30 m loop
表格設定（擷取紀錄原文）：Real sensor unit moved in front of planar boards and returned to the start; difference between start and end poses after a 30 m path (Geneva et al., 2018, Text Sec.VI-D)
difference between the start and end poses，authors' real-world test (planar boards) · 30 m loop
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 Geneva et al., 2018 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。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LIPS本方法原文提出 | 1.5 cm | (Geneva et al., 2018, Sec. VI-D; Fig. 7) |
來源
Geneva et al., 2018
（2018）LIPS: LiDAR-Inertial 3D Plane SLAM2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 123-130
同儕審查已出版已讀全文近十年查證後修正
相關版本
- accepted manuscript：Author copy on first author's website (IEEE pagination pp. 123-130) https://pgeneva.com/downloads/papers/Geneva2018IROS.pdf
- technical report supplement：Technical report RPNG-2018-LIPS with Jacobians (last updated July 31, 2018) https://udel.edu/~ghuang/papers/tr_lips.pdf
- 資料集：rpng/lips LiDAR-inertial 3D plane simulator (MIT License); no estimator code https://github.com/rpng/lips