SLAM 點雲圖鑑Point clouds for construction
M-01 · 2024

VoxelMap++

VoxelMap++ 延伸 VoxelMap：每個 0.5 m 體素只以三自由度參數（a、b、d）與其共變異數表示平面，並以可累加的和式遞增最小平方擬合，降低計算與記憶體。體素內平面在累積 50 點收斂後即丟棄原始點，並以 union-find 與鄰近體素平面做共面檢定（馬氏距離配合卡方 95% 門檻），將多個「子平面」視為同一「父平面」的量測，以共變異數跡加權融合，使整面牆或地板共享一個更準確、共變異數更小的平面。狀態估計沿用 FAST-LIO 與 VoxelMap 的迭代誤差狀態卡爾曼濾波。

里程計＋局部建圖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 v1 (2023-08-05; only arXiv version); RA-L version of record not read, and its author list differs
全文取得途徑
arXiv 預印本
本頁內容

VoxelMap extension that stores a 3DOF plane with covariance per 0.5 m hashed voxel, fitted incrementally by least squares, and merges converged coplanar voxel planes through union-find with a Mahalanobis test and trace-weighted fusion, so large walls and floors share one lower-uncertainty plane used in a FAST-LIO style iterated error-state Kalman filter.

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, spinning or non-repetitive solid-state (Velodyne VLP-32C in M2DGR; Livox HAP in the authors' data)、IMU (Realsense D435i IMU in M2DGR; ZED 2i built-in IMU in the authors' data)
原文測試平台ground robot (M2DGR)、trolley (pushed cart with Livox HAP and ZED 2i)
狀態估計iterated error-state Kalman filter as in FAST-LIO and VoxelMap, with point-to-plane observations whose noise combines point and merged-plane covariances (Sec. III-A, III-D)
資料關聯each point is hashed to its 0.5 m voxel and matched point-to-plane to the root (father) plane of that voxel's union-find node (Sec. III-C, III-D)
時間表示discrete scan poses with FAST-LIO style IMU propagation (Sec. III-A)
去畸變FAST-LIO style preprocessing of raw points (backward propagation implied by 'similar to FAST-LIO'; not detailed) (Sec. III-A)
迴圈閉合none
全域最佳化none
地圖表示hash table of 0.5 m voxels, each holding a 3DOF plane (a, b, d) with 3x3 covariance fitted incrementally by least squares; converged planes (after 50 points, raw points discarded) are merged with coplanar neighbours by union-find using a Mahalanobis chi-square test and trace-weighted fusion, so many voxels share one father plane (Sec. III-B, III-C)
先驗資訊LiDAR-IMU extrinsic calibrated with LI-Init; sensors synchronized by IEEE 1588-2008 in the authors' data (Sec. IV)
可輸出幾何odometry and a compact plane map (merged planes plus unmerged voxel planes); raw points are not kept after voxel convergence
計算需求CPU: average 4.88 ms and 126 MB on a 203 m corridor loop and 13.83 ms and 196 MB on a 329 m unstructured loop, lower than VoxelMap, Faster-LIO and FAST-LIO2, on a laptop with a 2.9 GHz 8-core CPU and 16 GiB memory (Table IV, Sec. IV)

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
LiDARLivox HAP歸入：Livox Hap方法輸入VoxelMap++ own datasets (UESTC forest, grassland and corridors)solid-state non-repetitive LiDAR, 120 x 25 deg FOV, 10 Hz; described as the first automotive-grade LiDAR for serial production(Wu et al., 2024b, Sec. IV; Fig. 4)
LiDARVelodyne VLP-32C資料集感測器M2DGR360 x 40 deg FOV, 10 Hz(Wu et al., 2024b, Sec. IV)
慣性量測單元（IMU）ZED 2i camera built-in IMU (written 'built-in Next-Gen IMU')方法輸入VoxelMap++ own datasets (UESTC forest, grassland and corridors)IMU with gyroscope, accelerometer, barometer and magnetometer; 500 Hz; synchronized with the LiDAR by IEEE 1588-2008(Wu et al., 2024b, Sec. IV; Fig. 4)
慣性量測單元（IMU）Realsense D435i IMU (written 'VI-sensor Realsense d435i')資料集感測器M2DGR200 Hz(Wu et al., 2024b, Sec. IV)
GNSS 接收器GNSS-IMU system with real-time kinematic signals (model not stated)參考或真值量測M2DGRground truth(Wu et al., 2024b, Sec. IV-A)
載具平台trolley (pushed cart)方法輸入VoxelMap++ own datasets (UESTC forest, grassland and corridors)sensors strapped down on the trolley(Wu et al., 2024b, Sec. IV; Fig. 4)
運算硬體laptop, 2.9 GHz 8-core CPU (model not stated)執行運算平台未標示16 GiB memory(Wu et al., 2024b, Sec. IV)

作者報告的優勢與限制

優勢

  • Lowest ATE on 5 of the 7 M2DGR sequences (all except Street07 and the failed Lift04), e.g. Street02 1.16 m versus 1.74 m for VoxelMap and 2.32 m for FAST-LIO2 (Table I)
  • End-to-end error 0.03 to 0.07 m on five unstructured forest and grassland loops of 284 to 520 m (Table II)
  • End-to-end error 0.25 to 1.45 m on five indoor corridor loops where other methods reached up to 19.7 m, attributed to merging floor and ceiling into large planes that constrain pitch drift (Table III, Fig. 8)
  • Lowest computation time and memory among FAST-LIO2, Faster-LIO, VoxelMap and VoxelMap++ on the tested loops (Table IV)

限制

  • Fails in dynamic scenes: on M2DGR lift04 the converged voxels of the elevator doors do not update when the doors close and the estimate diverges (Sec. IV-A, V)
  • Street07 ATE (12.85 m) was worse than FAST-LIO2, Faster-LIO and LINS (Table I)
  • On unstructured loopH VoxelMap had a lower end-to-end error than VoxelMap++ (0.0492 vs 0.0734 m) (Table II)
  • Own datasets have no RTK ground truth; only start-to-end errors of loops that return to the origin are reported (Sec. IV-B)
  • Baseline values for A-LOAM, LeGO-LOAM, LIO-SAM and LINS are copied from the M2DGR paper (Sec. IV-A)
  • Voxels stop updating after 50 points, so later changes are not absorbed (Sec. III-C-2) (inference from the design and the lift04 failure)

營建工程相關證據

論文未在營建工地驗證；自建資料為 UESTC 校園森林草地與具長走廊的既有建物，只以回到原點的端點誤差評估。把整面地板、天花板與牆合併為大平面來抑制走廊漂移，對施工中建物的長走廊與樓板場景有直接參考價值；但作者指出場景變動（如電梯門關閉）會使已收斂的體素平面失效而發散，而施工現場的臨時構件、模板與材料堆置持續變動，這項限制可能更明顯（推論）。

原文驗證環境：公開基準、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 4 個比較組，合計 21 筆紀錄。

Wu et al., 2024b · Table I 本方法 7 筆

指標ATE

表格設定（擷取紀錄原文）：M2DGR structured urban sequences; ATE (m); A-LOAM, LeGO-LOAM, LIO-SAM and LINS copied from the M2DGR paper; FAST-LIO2, Faster-LIO, VoxelMap and VoxelMap++ computed with evo; same parameters for all sequences; X = failed (Wu et al., 2024b, Table I)

ATE，M2DGR · Street02 (1484.62 m, day, long-term)

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 Wu et al., 2024b 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：m；場景：ground robot; M2DGR structured urban sequences (streets by day and night, dark room, hall, door from outdoors to indoors, lift between floors) (Sec. IV-A, Table I)

資料來源作者報告值(Wu et al., 2024b, Table I)

數值與出處
方法（原文寫法）報告值出處
A-LOAM5.299 m(Wu et al., 2024b, Table I)
LeGO-LOAM20.021 m(Wu et al., 2024b, Table I)
LIO-SAM4.063 m(Wu et al., 2024b, Table I)
LINS5.636 m(Wu et al., 2024b, Table I)
FAST-LIO22.3236 m(Wu et al., 2024b, Table I)
Faster-LIO2.6667 m(Wu et al., 2024b, Table I)
VoxelMap1.7408 m(Wu et al., 2024b, Table I)
VoxelMap++本方法原文提出1.1608 m(Wu et al., 2024b, Table I)

Wu et al., 2024b · Table II 本方法 5 筆

指標end-to-end error

表格設定（擷取紀錄原文）：Own Livox HAP data pushed on a cart around closed loops; end-to-end error between start and terminal point (no RTK available) (Wu et al., 2024b, Table II)

end-to-end error，VoxelMap++ own datasets · loopE (329.4 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。

這些是 Wu et al., 2024b 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：未對齊；單位：m；場景：unstructured forest and grassland at the UESTC library entrance

資料來源作者報告值(Wu et al., 2024b, Table II)

數值與出處
方法（原文寫法）報告值出處
LIO-Livox5.9141 m(Wu et al., 2024b, Table II)
FAST-LIO23.2134 m(Wu et al., 2024b, Table II)
Faster-LIO5.7331 m(Wu et al., 2024b, Table II)
VoxelMap1.6847 m(Wu et al., 2024b, Table II)
VoxelMap++本方法原文提出0.0336 m(Wu et al., 2024b, Table II)

Wu et al., 2024b · Table III 本方法 5 筆

指標end-to-end error

表格設定（擷取紀錄原文）：Own Livox HAP data pushed on a cart around closed loops; end-to-end error between start and terminal point (Wu et al., 2024b, Table III)

end-to-end error，VoxelMap++ own datasets · loop1 (202.9 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。

這些是 Wu et al., 2024b 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：未對齊；單位：m；場景：indoor building with long degenerate corridors

資料來源作者報告值(Wu et al., 2024b, Table III)

數值與出處
方法（原文寫法）報告值出處
LIO-Livox7.1192 m(Wu et al., 2024b, Table III)
FAST-LIO22.3812 m(Wu et al., 2024b, Table III)
Faster-LIO2.3488 m(Wu et al., 2024b, Table III)
VoxelMap9.3457 m(Wu et al., 2024b, Table III)
VoxelMap++本方法原文提出0.5305 m(Wu et al., 2024b, Table III)

Wu et al., 2024b · Table IV 本方法 4 筆

表格設定（擷取紀錄原文）：Resource usage: average computation time per scan and memory usage (Wu et al., 2024b, Table IV)

Avg. comp. time，VoxelMap++ own datasets · small scale (loop1, 202.9 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。

這些是 Wu et al., 2024b 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：平均值（mean）；對齊方式：不適用；單位：ms；場景：indoor corridor

資料來源作者報告值(Wu et al., 2024b, Table IV)

數值與出處
方法（原文寫法）報告值出處
FAST-LIO2硬體：laptop, 2.9 GHz 8 cores, 16 GiB11.5985 ms(Wu et al., 2024b, Table IV)
Faster-LIO硬體：laptop, 2.9 GHz 8 cores, 16 GiB7.3461 ms(Wu et al., 2024b, Table IV)
VoxelMap硬體：laptop, 2.9 GHz 8 cores, 16 GiB5.683 ms(Wu et al., 2024b, Table IV)
VoxelMap++本方法原文提出硬體：laptop, 2.9 GHz 8 cores, 16 GiB4.8763 ms(Wu et al., 2024b, Table IV)

來源

  • Wu et al., 2024b

    Chang Wu, Yuan You, Yifei Yuan, Xiaotong Kong, Ying Zhang, Qiyan Li, Kaiyong Zhao（2024）VoxelMap++: Mergeable Voxel Mapping Method for Online LiDAR(-Inertial) OdometryIEEE Robotics and Automation Letters, 9(1):427-434

    DOI 10.1109/lra.2023.3333736arXiv 2308.02799程式碼

    同儕審查已出版已讀全文近十年查證後修正

程式碼：https://github.com/uestc-icsp/VoxelMapPlus_Public（授權：not_stated (no LICENSE file found on main or master branch)）。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。

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