LiTAMIN2
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。
本頁內容
ICP/NDT-style LiDAR SLAM that approximates each large voxel (3 m) by one normal distribution and registers distribution to distribution with a symmetric KL-divergence cost (distance plus shape terms), reaching hundreds to over a thousand odometry frames per second on KITTI with SuMa-level accuracy; loop closure and pose graph as in LiTAMIN.
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 HDL-64E S2 in KITTI) |
|---|---|
| 原文測試平台 | vehicle (KITTI) |
| 狀態估計 | 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) |
| 資料關聯 | 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) |
| 時間表示 | discrete poses |
| 去畸變 | not described; KITTI clouds are already de-skewed and were fed directly to all methods (Sec. IV-B) |
| 迴圈閉合 | yes; implemented as in LiTAMIN, with the proposed ICP cost used to compute loop constraints; detection details are deferred to LiTAMIN (Sec. III-C) |
| 全域最佳化 | graph optimizer implemented as in LiTAMIN (details not given in this paper) (Sec. III-C) |
| 地圖表示 | voxel map of normal distributions (mean and covariance per voxel) (Sec. III-A; Sec. III-C) |
| 先驗資訊 | none |
| 可輸出幾何 | trajectory and voxelized normal-distribution map; Fig. 1 colours distributions by normal direction |
| 計算需求 | tracking and mapping in one thread; with 3 m voxels odometry runs at 510 FPS (ICP term only) or 239 FPS (ICP plus shape term) and all KITTI sequences take 58 s or 119 s including loop closing, on a desktop PC with Intel Core i9-9900K and 32 GB RAM (the PC also had an RTX 2080 Ti; GPU use by LiTAMIN2 not stated) (Sec. III-C; Sec. IV-A; Table II; Table V) |
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 HDL-64E S2歸入：Velodyne HDL-64E | 資料集感測器 | KITTI Vision Benchmark (odometry) | on-board KITTI sensor; point clouds provided already de-skewed | (Yokozuka et al., 2021, Sec. IV-B) |
| 運算硬體 | Intel Core i9-9900K | 執行運算平台 | 未標示 | desktop PC with 32 GB RAM | (Yokozuka et al., 2021, Sec. IV-A) |
| 運算硬體 | NVIDIA GeForce RTX 2080 Ti歸入：NVIDIA Geforce RTX2080Ti | 執行運算平台 | 未標示 | GPU in the same desktop PC; used for all experiments (use by LiTAMIN2 itself not stated) | (Yokozuka et al., 2021, Sec. IV-A) |
作者報告的優勢與限制
優勢
- Odometry at 510 FPS (ICP term) or 239 FPS (ICP plus shape term) with 3 m voxels, and above 1000 FPS for coarser voxels with the ICP term only (Table II; Sec. V)
- KITTI stats 0.33 deg/100m and 0.85% with loop closure, similar to SuMa frame-to-model (0.32 and 0.89) (Table III)
- Average ATE after loop closure 2.4 m versus 3.2 m for SuMa frame-to-model with loop closure (Table IV)
- Loop closing succeeded on all sequences for LiTAMIN2, LiTAMIN and SuMa, while loops were not detected for LeGO-LOAM and hdl_graph_slam in the authors' runs (Sec. IV-D)
限制
- Evaluated only on KITTI; the authors state the voxel size must be chosen for indoor or confined environments and that choosing it when the environment changes needs further study (Sec. V; Sec. VI)
- Adding the shape term roughly doubles the processing time compared with the ICP term alone (Sec. V)
- Finer voxels do not always give better accuracy; accuracy degrades for voxels above about 4 m (Table II; Sec. IV-C)
- Loop detection and graph optimization are only described by reference to LiTAMIN (Sec. III-C)
營建工程相關證據
論文只在 KITTI 車載道路資料上評估，沒有施工現場或建築室內資料。作者明確指出室內或狹窄環境需另選體素大小，因此其極低運算量是否能在施工室內維持精度尚未驗證。它在多篇後續論文中被列為比較基準，例如（Pan et al., 2021）、（Yuan et al., 2022）、（Liu et al., 2023b）、（Pan et al., 2024）。
原文驗證環境：公開基準
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 4 個比較組，合計 32 筆紀錄。
Yokozuka et al., 2021 · Table III 本方法 22 筆
指標KITTI stats translation [%]
表格設定（擷取紀錄原文）：KITTI odometry sequences 00-10; KITTI stats translation error (%) averaged over 100-800 m segments with the benchmark code; rows without loop closure; LiTAMIN2 with 3 m voxels; LOAM (from [10]), LO-Net and DeepLO values copied from their papers (Yokozuka et al., 2021, Table III)
KITTI stats translation [%]，KITTI odometry · 00 (4541 frames)
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 Yokozuka 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。
資料來源作者報告值(Yokozuka et al., 2021, Table III)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LiTAMIN2 (ICP+Cov), without loop closure本方法原文提出 | 0.78% | (Yokozuka et al., 2021, Table III) |
| LiTAMIN2 (ICP), without loop closure本方法原文提出 | 0.75% | (Yokozuka et al., 2021, Table III) |
| LiTAMIN [2], without loop closure | 0.91% | (Yokozuka et al., 2021, Table III) |
| SuMa (Frame-to-Frame) | 2.11% | (Yokozuka et al., 2021, Table III) |
| SuMa (Frame-to-Model), without loop closure | 0.72% | (Yokozuka et al., 2021, Table III) |
| LeGO-LOAM | 2.17% | (Yokozuka et al., 2021, Table III) |
| hdl graph slam | 3.92% | (Yokozuka et al., 2021, Table III) |
| LOAM (open source, run by authors) | 1.92% | (Yokozuka et al., 2021, Table III) |
| LOAM (from [10]) | 0.78% | (Yokozuka et al., 2021, Table III) |
| LO-Net (Frame-to-Frame) | 1.47% | (Yokozuka et al., 2021, Table III) |
| LO-Net (Frame-to-Model) | 0.78% | (Yokozuka et al., 2021, Table III) |
Yokozuka et al., 2021 · Table III (overall) 本方法 4 筆
指標KITTI stats [deg/100m] / [%], translation part
資料集與序列KITTI odometry · 00-10 overall
表格設定（擷取紀錄原文）：KITTI stats over all sequences 00-10 (final column), translation error %; LiTAMIN2 with 3 m voxels (Yokozuka et al., 2021, Table III (overall))
KITTI stats [deg/100m] / [%], translation part，KITTI odometry · 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 Yokozuka 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。
資料來源作者報告值(Yokozuka et al., 2021, Table III (overall))
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LiTAMIN2 (ICP+Cov), without loop closure本方法原文提出 | 0.88% | (Yokozuka et al., 2021, Table III) |
| LiTAMIN2 (ICP+Cov), with loop closure本方法原文提出 | 0.85% | (Yokozuka et al., 2021, Table III) |
| LiTAMIN2 (ICP), without loop closure本方法原文提出 | 0.95% | (Yokozuka et al., 2021, Table III) |
| LiTAMIN2 (ICP), with loop closure本方法原文提出 | 0.89% | (Yokozuka et al., 2021, Table III) |
| LiTAMIN [2], without loop closure | 1.6% | (Yokozuka et al., 2021, Table III) |
| LiTAMIN [2], with loop closure | 1.59% | (Yokozuka et al., 2021, Table III) |
| SuMa (Frame-to-Frame) | 2.19% | (Yokozuka et al., 2021, Table III) |
| SuMa (Frame-to-Model), without loop closure | 0.84% | (Yokozuka et al., 2021, Table III) |
| SuMa (Frame-to-Model), with loop closure | 0.89% | (Yokozuka et al., 2021, Table III) |
| LeGO-LOAM | 2.49% | (Yokozuka et al., 2021, Table III) |
| hdl graph slam | 9.57% | (Yokozuka et al., 2021, Table III) |
| LOAM (open source, run by authors) | 2.13% | (Yokozuka et al., 2021, Table III) |
Yokozuka et al., 2021 · Table IV 本方法 4 筆
指標Absolute trajectory error, Avg. of all frames [m]
資料集與序列KITTI odometry · 00-10, average of all frames
表格設定（擷取紀錄原文）：KITTI odometry 00-10; absolute trajectory error, average of all frames (translation part, m); statistic and alignment not stated; loop closure as marked (Yokozuka et al., 2021, Table IV)
Absolute trajectory error, Avg. of all frames [m]，KITTI odometry · 00-10, average of all frames
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 Yokozuka 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。
資料來源作者報告值(Yokozuka et al., 2021, Table IV)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LiTAMIN2 (ICP+Cov), without loop closure本方法原文提出 | 5.1 m | (Yokozuka et al., 2021, Table IV) |
| LiTAMIN2 (ICP+Cov), with loop closure本方法原文提出 | 2.4 m | (Yokozuka et al., 2021, Table IV) |
| LiTAMIN2 (ICP), without loop closure本方法原文提出 | 6 m | (Yokozuka et al., 2021, Table IV) |
| LiTAMIN2 (ICP), with loop closure本方法原文提出 | 2.6 m | (Yokozuka et al., 2021, Table IV) |
| LiTAMIN [2], without loop closure | 8.3 m | (Yokozuka et al., 2021, Table IV) |
| LiTAMIN [2], with loop closure | 6.2 m | (Yokozuka et al., 2021, Table IV) |
| SuMa (Frame-to-Frame) | 14.1 m | (Yokozuka et al., 2021, Table IV) |
| SuMa (Frame-to-Model), without loop closure | 3.9 m | (Yokozuka et al., 2021, Table IV) |
| SuMa (Frame-to-Model), with loop closure | 3.2 m | (Yokozuka et al., 2021, Table IV) |
| LeGO-LOAM | 11.1 m | (Yokozuka et al., 2021, Table IV) |
| hdl graph slam | 76.7 m | (Yokozuka et al., 2021, Table IV) |
| LOAM (open source, run by authors) | 29.7 m | (Yokozuka et al., 2021, Table IV) |
Yokozuka et al., 2021 · Table V 本方法 2 筆
指標Avg. rate [FPS]
資料集與序列KITTI odometry · 00-10 (all frames)
表格設定（擷取紀錄原文）：KITTI odometry 00-10; average odometry frame rate over all frames; LiTAMIN2, LiTAMIN and SuMa include loop closing, others without; all run by the authors on the same PC (Yokozuka et al., 2021, Table V)
Avg. rate [FPS]，KITTI odometry · 00-10 (all frames)
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 Yokozuka 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。
資料來源作者報告值(Yokozuka et al., 2021, Table V)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| LiTAMIN2 (ICP), with loop closure本方法原文提出硬體：desktop PC, Intel Core i9-9900K, 32 GB RAM, NVIDIA GeForce RTX 2080 Ti | 508.9 FPS | (Yokozuka et al., 2021, Table V) |
| LiTAMIN2 (ICP+Cov), with loop closure本方法原文提出硬體：desktop PC, Intel Core i9-9900K, 32 GB RAM, NVIDIA GeForce RTX 2080 Ti | 238.8 FPS | (Yokozuka et al., 2021, Table V) |
| LiTAMIN [2], with loop closure硬體：desktop PC, Intel Core i9-9900K, 32 GB RAM, NVIDIA GeForce RTX 2080 Ti | 45.2 FPS | (Yokozuka et al., 2021, Table V) |
| SuMa (Frame-to-Model), with loop closure硬體：desktop PC, Intel Core i9-9900K, 32 GB RAM, NVIDIA GeForce RTX 2080 Ti | 58.4 FPS | (Yokozuka et al., 2021, Table V) |
| LeGO-LOAM硬體：desktop PC, Intel Core i9-9900K, 32 GB RAM, NVIDIA GeForce RTX 2080 Ti | 66.1 FPS | (Yokozuka et al., 2021, Table V) |
| hdl graph slam硬體：desktop PC, Intel Core i9-9900K, 32 GB RAM, NVIDIA GeForce RTX 2080 Ti | 4.8 FPS | (Yokozuka et al., 2021, Table V) |
| LOAM (open source)硬體：desktop PC, Intel Core i9-9900K, 32 GB RAM, NVIDIA GeForce RTX 2080 Ti | 10.7 FPS | (Yokozuka et al., 2021, Table V) |
來源
Yokozuka et al., 2021
（2021）LiTAMIN2: Ultra Light LiDAR-based SLAM using Geometric Approximation applied with KL-Divergence2021 IEEE International Conference on Robotics and Automation (ICRA), pp. 11619-11625
DOI 10.1109/icra48506.2021.9560947arXiv 2103.00784
同儕審查已出版已讀全文近十年
相關版本
- 預印本：LiTAMIN2 (arXiv v1) https://arxiv.org/abs/2103.00784
- 前身方法：LiTAMIN: LiDAR-based Tracking and MappINg by Stabilized ICP for Geometry Approximation with Normal Distributions (IROS 2020); folded into this entry, full text not read https://doi.org/10.1109/IROS45743.2020.9341341