IMLS-SLAM
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。
本頁內容
LiDAR-only scan-to-model odometry that registers observability-sampled points to an implicit moving least squares surface built from the last n localized sweeps, with constant-velocity de-skewing and size-based small-object removal; no loop closure and not real time.
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 only (Velodyne HDL32 and HDL64 in the experiments); no IMU, GPS or camera |
|---|---|
| 原文測試平台 | vehicle |
| 狀態估計 | 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) |
| 資料關聯 | 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) |
| 時間表示 | one discrete pose per sweep end; poses inside a sweep linearly interpolated between the previous and current end pose, current pose predicted by constant relative motion (Sec. III) |
| 去畸變 | 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) |
| 迴圈閉合 | none (drift reported without any loop closure) |
| 全域最佳化 | none |
| 地圖表示 | 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) |
| 先驗資訊 | none |
| 可輸出幾何 | trajectory and accumulated de-skewed point cloud (Figs. 4-6) |
| 計算需求 | C++ with FLANN and Eigen on one CPU core at 4 GHz, less than 1 GB RAM; not real time: 1.25 s per scan on KITTI (0.2 s normals, 1 s k-d tree, 0.05 s matching) (Sec. VI; Sec. VI.C) |
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 | Velodyne HDL32歸入：Velodyne HDL-32E | 方法輸入 | 未標示 | 32 laser beams, spinning at 10 Hz (100 ms per scan); mounted vertically on a car roof (Paris, 12951 scans) and tilted 60 degrees in pitch (Lille, 1500 scans) | (Deschaud, 2018, Sec. VI; Sec. VI-A) |
| LiDAR | Velodyne HDL64歸入：Velodyne HDL-64E | 資料集感測器 | KITTI odometry | vertical, on the roof of a car, 64 beams, 10 Hz; scans provided de-skewed | (Deschaud, 2018, Sec. VI; Sec. VI-B) |
| GNSS 接收器 | GPS+IMU navigation system (model not reported) | 參考或真值量測 | KITTI odometry | ground truth for 11 training sequences; author notes errors above 5 m at the start of sequence 8 | (Deschaud, 2018, Sec. VI-B) |
| 載具平台 | car (vehicle roof mount, model not reported) | 方法輸入 | 未標示 | vehicle driven through Paris (two 2 km loops) and a square in Lille | (Deschaud, 2018, Sec. VI-A; Fig. 1) |
| 運算硬體 | single CPU core at 4 GHz (processor model not reported) | 執行運算平台 | 未標示 | one core, less than 1 GB RAM | (Deschaud, 2018, Sec. VI) |
作者報告的優勢與限制
優勢
- 0.40% drift (16 m) over a 4 km, two-loop drive in Paris with a vertical HDL32, without loop closure (Sec. VI.A; Fig. 1).
- KITTI training average 0.55% translational drift and 0.0015 deg/m; test set 0.69% and 0.0018 deg/m; lower drift than LOAM's published values on every training sequence (Table I; Sec. VI.B).
- Ablations favour the proposed sampling (0.55% versus 0.64% random and 0.57% geometric stable sampling), a 100-scan model (versus 1.41% for scan-to-scan, n = 1) and object removal (0.55% versus 0.58%) (Tables II-IV).
- Only about 7% of a scan (900 samples) is used for matching, so matching iterations take 0.05 s per scan (Sec. IV; Sec. VI.C).
限制
- Not real time: 1.25 s per scan on KITTI, dominated by normal computation on the 3D cloud and per-scan k-d tree rebuilding (Sec. VI.C).
- KITTI drift is worse than on the HDL32 data; the author attributes this to intrinsic calibration distortion (a 0.22 deg vertical-angle correction was applied), GPS ground-truth errors above 5 m at the start of sequence 8, and more varied environments (Sec. VI.B).
- LOAM's later KITTI website result (0.64%) was slightly better than IMLS-SLAM's test score (Sec. VI.B).
- Dynamic handling removes every object whose bounding box is below 14 m x 14 m x 4 m, which is size-based rather than true motion detection (Sec. III).
- The Lille run with the HDL32 tilted 60 degrees is evaluated only qualitatively (Sec. VI.A).
營建工程相關證據
未在營建場域驗證；實驗為巴黎與里爾的市區車載資料以及 KITTI。其掃描對隱式曲面模型的配準，以及以尺寸門檻剔除小物體的做法，和工地中機具、人員造成的點雲污染問題相關；但以尺寸刪除物體也可能刪去工地上小型而靜態的構件（推論）。後續 LiDAR 里程計論文常以它作為 KITTI 比較基準，例如 KISS-ICP、CT-ICP、MULLS、SuMa++、GenZ-ICP 與 MOLA-LO 的比較表（Vizzo et al., 2023）（Dellenbach et al., 2022）（Pan et al., 2021）（Chen et al., 2019）（Lee et al., 2025a）（Blanco-Claraco, 2025）。
原文驗證環境：公開基準、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 8 個比較組，合計 33 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 4 組列在最後，並連到性能比較頁。
Deschaud, 2018 · Table I 本方法 11 筆
指標drift (%)
表格設定（擷取紀錄原文）：KITTI odometry training sequences 00-10, HDL64; translation drift (%) with the KITTI metric; LOAM values copied from the LOAM Autonomous Robots paper [7] (Deschaud, 2018, Table I)
drift (%)，KITTI odometry · 00
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 Deschaud, 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(Deschaud, 2018, Table I)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LOAM [7] (results taken from paper) | 0.78% | (Deschaud, 2018, Table I) |
| Our SLAM (IMLS-SLAM)本方法原文提出 | 0.5% | (Deschaud, 2018, Table I) |
Deschaud, 2018 · Table IV 本方法 4 筆
指標Drift on KITTI training dataset
資料集與序列KITTI odometry · training 00-10 (overall)
表格設定（擷取紀錄原文）：Ablation of number of scans n kept in the model; drift on the whole KITTI training dataset (Deschaud, 2018, Table IV)
Drift on KITTI training dataset，KITTI odometry · training 00-10 (overall)
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 Deschaud, 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(Deschaud, 2018, Table IV)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| IMLS-SLAM, n = 1 scan本方法原文提出 | 1.41% | (Deschaud, 2018, Table IV) |
| IMLS-SLAM, n = 5 scans本方法原文提出 | 0.58% | (Deschaud, 2018, Table IV) |
| IMLS-SLAM, n = 10 scans本方法原文提出 | 0.56% | (Deschaud, 2018, Table IV) |
| IMLS-SLAM, n = 100 scans本方法原文提出 | 0.55% | (Deschaud, 2018, Table IV) |
Deschaud, 2018 · Text Sec. VI-B 本方法 4 筆
表格設定（擷取紀錄原文）：KITTI odometry, overall values stated in text (training set with ground truth; test set from KITTI website) (Deschaud, 2018, Text Sec. VI-B)
drift in translation，KITTI odometry · training 00-10 (overall)
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 Deschaud, 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 |
|---|---|---|
| IMLS-SLAM本方法原文提出 | 0.55% | (Deschaud, 2018, Sec. VI-B) |
Deschaud, 2018 · Text Sec. VI-C 本方法 4 筆
資料集與序列KITTI odometry · all
表格設定（擷取紀錄原文）：Processing time per scan on KITTI (normals from 3D points since raw range images are unavailable) (Deschaud, 2018, Text Sec. VI-C)
total time per scan，KITTI odometry · all
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 Deschaud, 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 |
|---|---|---|
| IMLS SLAM本方法原文提出硬體：one CPU core at 4 GHz, less than 1 GB RAM | 1.25 s | (Deschaud, 2018, Sec. VI-C) |
其他比較組
來源
Deschaud, 2018
（2018）IMLS-SLAM: Scan-to-Model Matching Based on 3D Data2018 IEEE International Conference on Robotics and Automation (ICRA), Brisbane, Australia, pp. 2480-2485
DOI 10.1109/icra.2018.8460653arXiv 1802.08633
同儕審查已出版已讀全文近十年
相關版本
- 預印本：IMLS-SLAM: scan-to-model matching based on 3D data (arXiv v1) https://arxiv.org/abs/1802.08633
- repository copy：HAL hal-01959570 (submitted version) https://hal.science/hal-01959570