VoxelMap
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。
本頁內容
An adaptive, coarse-to-fine voxel map (hash table plus per-voxel octree) storing plane features with covariances derived from LiDAR noise and pose error, used for probabilistic point-to-plane registration in an IEKF.
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 | 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 |
|---|---|
| 原文測試平台 | KITTI odometry urban sequences (vehicle platform per KITTI, not described in this paper)、handheld device with Intel L515 in laboratory and warehouse、handheld device with Livox Avia in a park、UAV carrying the Livox Avia facing the ground over a mountain、UGV data of the SSL_SLAM dataset (re-run only) |
| 狀態估計 | 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 |
| 資料關聯 | 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 |
| 時間表示 | discrete poses |
| 去畸變 | 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) |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | 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 |
| 先驗資訊 | none |
| 可輸出幾何 | 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 原文未報告 |
| 計算需求 | desktop Intel i7-10700 @ 2.9 GHz, 16 GB RAM, CPU only (GTX 730 2 GB used only by the GPU baseline SuMa); KITTI mean 26.45 ms (std 9.87 ms) per scan (Table III); L515 6.38 to 7.26 ms per scan (Table IV); Livox Avia 7.7 to 8.3 ms per scan in the park and 34.2 ms on the mountain, averaged over 5 runs (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 odometry | mechanical, 10 Hz, repetitive scan, FoV 360°×32°; KITTI in-frame motion pre-compensated; 0.22° vertical angle correction applied as in IMLS-SLAM | (Yuan et al., 2022, Table I; Sec. IV-A) |
| LiDAR | Realsense L515歸入：Intel RealSense L515 | 方法輸入 | 未標示 | solid-state, 30 Hz, repetitive scan, FoV 70°×55° (Table I); written as 'Intel L515' in Sec. IV-B; point colors in the supplementary warehouse map are provided by the L515 (Supp. Fig. 1 caption) | (Yuan et al., 2022, Table I; Sec. IV-B; Supp. Fig. 1) |
| LiDAR | Livox Avia | 方法輸入 | 未標示 | solid-state, 10 Hz, non-repetitive scan, FoV 70°×77° | (Yuan et al., 2022, Table I; Sec. IV-C) |
| 慣性量測單元（IMU） | Livox Avia built-in IMU | 方法輸入 | 未標示 | 200 Hz; used for motion-distortion compensation and as motion prior (LiDAR-inertial mode) | (Yuan et al., 2022, Sec. IV-C) |
| 載具平台 | handheld device (L515) | 方法輸入 | 未標示 | three sequences in laboratory and warehouse; hand-carried with much faster motion (especially rotation) than the slow UGV data of SSL_SLAM; routes start and end at the same place | (Yuan et al., 2022, Sec. IV-B) |
| 載具平台 | handheld device (Livox Avia) | 方法輸入 | 未標示 | park with trees; two 485 m loops and one 815 m loop starting and ending at the same place | (Yuan et al., 2022, Sec. IV-C1) |
| 載具平台 | UAV | 方法輸入 | 未標示 | carries the Livox Avia facing the ground over a mountain, 3490 m trajectory; outdoor aerial experiment helped by Ambit-Geospatial (Acknowledgment) | (Yuan et al., 2022, Sec. IV-C2; Table V; Acknowledgment) |
| 載具平台 | UGV | 資料集感測器 | SSL_SLAM dataset [33] | moving quite stably and slowly (sensor not restated in this paper) | (Yuan et al., 2022, Sec. IV-B) |
| 運算硬體 | Intel i7-10700 | 執行運算平台 | 未標示 | desktop computer, 2.9 GHz, 16 GB RAM; used for all experiments | (Yuan et al., 2022, Sec. IV (intro); Sec. IV-C) |
| 運算硬體 | Nvidia GeForce GTX 730 | 執行運算平台 | 未標示 | 2 GB RAM; used only when a method needs a GPU (SuMa) | (Yuan et al., 2022, Sec. IV (intro); Sec. IV-A Runtime) |
作者報告的優勢與限制
優勢
- Best overall KITTI ATE among compared odometry methods with loop closure off, especially on long sequences 00, 02 and 08
- average 1.2 deg / 2.9 m (Table II)
- ablation shows the probabilistic plane module matters more than adaptive voxelization (averages 1.8/4.5 without it vs 1.3/3.4 without adaptive voxels, Table II)
- lowest KITTI time per scan, 26.45 ms vs 29.11 ms for FAST-LIO2 (Table III) thanks to near O(1) hash lookup
- handheld L515 end-to-end error 0.01 to 0.02 m vs 0.71 to 5.80 m for SSL_SLAM (Table IV)
- lower end-to-end error than Faster-LIO and FAST-LIO2 on all four Livox Avia sequences with KITTI parameters (Table V)
- works across spinning, solid-state and non-repetitive LiDARs (abstract, Sec. V)
限制
- No loop closure
- KITTI baselines compared with loop closure disabled (Sec. IV-A)
- only plane features, adding edge features left to future work (Sec. V)
- extra probability computation per point and plane costs time relative to Faster-LIO, e.g. 34.2 ms vs 32.0 ms on the mountain (Sec. IV-C3, Table V)
- gain on the slow UGV SSL_SLAM dataset is small, 0.12 m vs 0.15 m (Sec. IV-B)
- (reviewer observation) own datasets are evaluated only by end-to-end error without ground-truth trajectories or map accuracy
營建工程相關證據
論文未涉及營建場域，也未評估地圖幾何精度；最接近的情境是手持 L515 在實驗室與倉庫的室內資料（Table IV），以及無人機載 Livox Avia 朝地面掃描山區地形（Sec. IV-C2）。每個平面附帶共變異數，可作為工程點雲品質指標的候選概念，此點屬推論。
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 13 個比較組，合計 76 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 9 組列在最後，並連到性能比較頁。
Yuan et al., 2022 · Table II 本方法 13 筆
表格設定（擷取紀錄原文）：KITTI odometry training sequences 00 to 10, LiDAR odometry with loop closure off for all methods; each cell printed as rotation [deg] / translation [m] (unit label '[deg]/[m]' printed in the header of the average column; caption says ATE in meters). All VoxelMap variants use identical parameters on all sequences (max voxel 3 m, 3 octree layers). Ours (w/o adaptive) = ablation with fixed 2 m voxels; Ours (w/o prob.) = ablation with zero plane uncertainty. FAST-LIO2 run by the authors with a constant-velocity prior (no IMU in KITTI); MULLS run by the authors with default parameters; LiTAMIN2, SuMa and LeGO-LOAM values copied from the LiTAMIN2 paper [24]. Rotation part kept only for the average column (truncation, see notes). (Yuan et al., 2022, Table II)
ATE (translation part) [m]，KITTI odometry (training) · 00 (3724 m)
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 Yuan et al., 2022 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(Yuan et al., 2022, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Ours (full)本方法原文提出 | 2.8 m | (Yuan et al., 2022, Table II) |
| Ours (w/o adaptive) | 3.1 m | (Yuan et al., 2022, Table II) |
| Ours (w/o prob.) | 3.9 m | (Yuan et al., 2022, Table II) |
| LiTAMIN2 [24] | 5.8 m | (Yuan et al., 2022, Table II) |
| MULLS [30] | 6.1 m | (Yuan et al., 2022, Table II) |
| Suma [14] | 2.9 m | (Yuan et al., 2022, Table II) |
| FAST-LIO2 [13] | 3.8 m | (Yuan et al., 2022, Table II) |
| Lego-Loam [7] | 6.3 m | (Yuan et al., 2022, Table II) |
Liu et al., 2023b · Table V 本方法 12 筆
指標RMSE of the ATE, translation part of deg/m pair
表格設定（擷取紀錄原文）：KITTI with MULLS poses without loop closure as HBA input; RMSE of ATE printed as rotation (deg)/translation (m); only translation extracted; baselines run without loop closure; the '(loops)' sequence labels follow the asterisks of arXiv v1 Table V (the version-of-record table image carries no asterisks); version of record adds CT-ICP and LOAM rows (Liu et al., 2023b, Table V)
RMSE of the ATE, translation part of deg/m pair，KITTI · Seq. 00 (loops)
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 Liu et al., 2023b 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(Liu et al., 2023b, Table V)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Proposed原文提出 | 1.2 m | (Liu et al., 2023b, Table V (version of record)) |
| CT-ICP | 4.5 m | (Liu et al., 2023b, Table V (version of record)) |
| MULLS | 6.1 m | (Liu et al., 2023b, Table V (version of record)) |
| Voxel Map本方法 | 2.8 m | (Liu et al., 2023b, Table V (version of record)) |
| SuMa | 2.9 m | (Liu et al., 2023b, Table V (version of record)) |
| LiTAMIN2 | 5.8 m | (Liu et al., 2023b, Table V (version of record)) |
| LOAM | 2.3 m | (Liu et al., 2023b, Table V (version of record)) |
Koide et al., 2024 · Table II 本方法 8 筆
指標Absolute Trajectory Error [m], reported as value ± (± not defined in the paper)
表格設定（擷取紀錄原文）：Eight real sequences (path 2.2-4.8 m) with a Livox Avia moved between two pillars while facing a flat wall; ground truth from AprilTag bundle adjustment and camera-IMU batch optimization; LIO-SAM gave no decent result (Koide et al., 2024, Table II)
Absolute Trajectory Error [m], reported as value ± (± not defined in the paper)，authors' flat-wall degeneration sequences · Seq. 01
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 Koide et al., 2024 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(Koide et al., 2024, Table II)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| FAST-LIO2 [5] | 0.815 m原文指標寫法：Absolute Trajectory Error [m], reported as value ± 0.5 (± not defined in the paper) | (Koide et al., 2024, Table II) |
| VoxelMap [61]本方法 | 0.577 m原文指標寫法：Absolute Trajectory Error [m], reported as value ± 0.157 (± not defined in the paper) | (Koide et al., 2024, Table II) |
| GLIM原文提出 | 0.118 m原文指標寫法：Absolute Trajectory Error [m], reported as value ± 0.047 (± not defined in the paper) | (Koide et al., 2024, Table II) |
Yuan et al., 2022 · Table V 本方法 8 筆
表格設定（擷取紀錄原文）：Livox Avia with built-in IMU (10 Hz LiDAR, 200 Hz IMU), LiDAR-inertial mode; same parameters as KITTI (max voxel 3 m, 3 layers); all three methods share the IKFoM framework and plane features and differ only in map structure; default 1:3 downsampling of Faster-LIO and FAST-LIO2 turned off; results averaged over 5 runs. Park = handheld, loops start and end at the same place; Mountain = UAV with LiDAR facing the ground. Cells: end-to-end error (m) / average computation time (ms). (Yuan et al., 2022, Table V)
End-to-End error (m), mean of 5 runs，authors' Livox Avia datasets · Park1 (485 m)
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 Yuan et al., 2022 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(Yuan et al., 2022, Table V)
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| Ours (full)本方法原文提出 | 0.03 m | (Yuan et al., 2022, Table V) |
| Faster-LIO [35] | 0.88 m | (Yuan et al., 2022, Table V) |
| FAST-LIO2 [13] | 1.09 m | (Yuan et al., 2022, Table V) |
其他比較組
來源
Yuan et al., 2022
（2022）Efficient and Probabilistic Adaptive Voxel Mapping for Accurate Online LiDAR OdometryIEEE Robotics and Automation Letters, 7(3):8518-8525
DOI 10.1109/lra.2022.3187250arXiv 2109.07082程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- 預印本：arXiv:2109.07082 (v1 2021-09-15, up to v5) https://arxiv.org/abs/2109.07082
程式碼：https://github.com/hku-mars/VoxelMap（授權：GPL-2.0 (per repository README)）。有公開程式碼不等於已被重現，也不代表目前版本與論文版本相同。