SLAM 點雲圖鑑Point clouds for construction
M-01 · 2023

ImMesh

ImMesh 以 VoxelMap 的機率平面與迭代卡爾曼濾波估計位姿，並把經空間降採樣、配準後的 LiDAR 點當成網格頂點（以 ikd-Tree 維持頂點最小間距）；每個有新點的體素將其頂點投影到該體素主平面上，以二維 Delaunay 三角化建立三角面，再以類似 git 的 pull、commit、push 步驟增量合併到全域網格。整體在一般桌上型 CPU 上即時執行。

里程計＋局部建圖近十年已讀全文同儕審查已出版地圖表示與表面重建

閱讀版本
arXiv v3 (2023-11-11) read in full; IEEE T-RO version of record 39(6):4312-4331 obtained via NTU access and its Sec. VII-VIII text compared (matches v3); the VoR tables are embedded images, so table values were transcribed from arXiv v3
全文取得途徑
arXiv 預印本
本頁內容

Uses VoxelMap-based localization and treats registered LiDAR points as vertices, meshing each active voxel by projecting to its dominant plane and applying 2D Delaunay triangulation with incremental pull/commit/push updates.

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 and solid-state)、IMU (optional)
原文測試平台handheld、vehicle、UAV、simulation
狀態估計iterated Kalman filter with probabilistic planes (built on VoxelMap)
資料關聯point-to-plane registration against voxel plane features (VoxelMap)
時間表示discrete poses
去畸變in-frame motion distortion compensated by IMU backward propagation (method of FAST-LIO) before registration (Sec. V-A)
迴圈閉合none (stated limitation, Sec. IX)
全域最佳化none
地圖表示spatially downsampled, registered LiDAR points kept as mesh vertices with a minimum spacing (0.15 m for mechanical, 0.10 m for solid-state LiDAR) enforced by an ikd-Tree, stored in hashed voxels (0.60 m or 0.40 m) and hashed regions (15 m or 10 m); triangle facets stored per region and indexed in a facet hash table
先驗資訊none
可輸出幾何triangle mesh published at scan rate; also point cloud reinforcement via mesh rasterization
計算需求real-time on CPU (handheld mini-computer, Intel i9-10900, 64 GB RAM); per-dataset mean per-scan time 9.8 to 31.3 ms for meshing and 11.9 to 42.2 ms for localization at 10 Hz input (Table IV); the Experiment-3 comparison ran on an Intel i7-9700K with 64 GB RAM, where only the TSDF baseline used the Nvidia 2080 Ti GPU

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 Avia方法輸入未標示solid-state, Risley prism; FoV 70.4 x 77.2 deg; 240,000 points/s single return; USD 1,599 (Table I)(Lin et al., 2023, Sec. VIII-A1, Table I, Fig. 6)
LiDARVelodyne HDL-64E資料集感測器KITTImechanical spinning 64-line; FoV 360.0 x 26.8 deg; 1,333,312 points/s; USD 75,000(Lin et al., 2023, Table I)
LiDARVelodyne HDL-32E資料集感測器NCLTmechanical spinning 32-line; FoV 360.0 x 41.3 deg; 695,000 points/s; USD 8,800(Lin et al., 2023, Table I)
LiDAROuster OS1-16 Gen1歸入：Ouster OS1-16資料集感測器NTU VIRALmechanical spinning 16-line; FoV 360.0 x 33.2 deg; 327,680 points/s; USD 3,500(Lin et al., 2023, Table I)
LiDARLivox Avia資料集感測器R3LIVE datasetsolid-state, Risley prism; FoV 70.4 x 77.2 deg; 240,000 points/s(Lin et al., 2023, Table I)
LiDARLivox Avia方法輸入未標示原文未報告 beyond model(Lin et al., 2023, Sec. VIII-E, Fig. 11)
相機Hikvision CA-050-11UC方法輸入未標示global shutter RGB camera; images used for mesh texturing with R3LIVE++ poses(Lin et al., 2023, Sec. VIII-E)
載具平台handheld data-collection device (custom rig, Fig. 6a)方法輸入未標示mini-computer, Livox Avia LiDAR and a preview RGB camera(Lin et al., 2023, Sec. VIII-A1, Fig. 6)
載具平台DJI M300方法輸入未標示drone carrying Livox Avia and Hikvision camera(Lin et al., 2023, Sec. VIII-E, Fig. 11)
運算硬體Intel i9-10900執行運算平台未標示mini-computer CPU with 64 GB RAM(Lin et al., 2023, Sec. VIII-A1)
運算硬體Intel i7-9700K執行運算平台未標示desktop CPU with 64 GB RAM(Lin et al., 2023, Sec. VIII-C3)
運算硬體Nvidia 2080 Ti執行運算平台未標示12 GB graphics memory; used only by the TSDF baseline(Lin et al., 2023, Sec. VIII-C3)
其他Microsoft AirSim simulated depth camera資料集感測器AirSim synthetic (Urban city, Cluttered field)depth images with FoV 120 x 80 deg at 640x480, 320x240 and 160x120, unprojected to simulate LiDAR(Lin et al., 2023, Sec. VIII-C2, Fig. 7)

論文圖片

只收錄原文以開放授權（open license）釋出的圖片，並依授權條件標示出處、圖號、授權與修改方式。

  • ImMesh 線上重建的三角網格與估計位姿，並以 R3LIVE 相機位姿貼上影像紋理的結果

    Fig. 1ImMesh 線上重建的三角網格與估計位姿，並以 R3LIVE 相機位姿貼上影像紋理的結果

    出處：Lin et al., 2023，Fig. 1。授權：CC BY 4.0 (arXiv v3)。原始圖檔。修改：縮小至寬度不超過 1400 px，並轉存為 WebP 格式。

  • 系統架構：接收、定位、地圖結構、建網格與廣播模組

    Fig. 2系統架構：接收、定位、地圖結構、建網格與廣播模組

    出處：Lin et al., 2023，Fig. 2。授權：CC BY 4.0 (arXiv v3)。原始圖檔。修改：縮小至寬度不超過 1400 px，並轉存為 WebP 格式。

  • 手持資料收集裝置（迷你電腦、Livox Avia 光達、預覽相機）與實驗影片畫面

    Fig. 6手持資料收集裝置（迷你電腦、Livox Avia 光達、預覽相機）與實驗影片畫面

    出處：Lin et al., 2023，Fig. 6。授權：CC BY 4.0 (arXiv v3)。原始圖檔。修改：縮小至寬度不超過 1400 px，並轉存為 WebP 格式。

  • DJI M300 無人機載具與山區場景的網格、高程著色及無損紋理重建成果

    Fig. 11DJI M300 無人機載具與山區場景的網格、高程著色及無損紋理重建成果

    出處：Lin et al., 2023，Fig. 11。授權：CC BY 4.0 (arXiv v3)。原始圖檔。修改：縮小至寬度不超過 1400 px，並轉存為 WebP 格式。

作者報告的優勢與限制

優勢

  • Real-time meshing on a CPU across KITTI, NCLT, NTU VIRAL and R3LIVE datasets with bounded per-scan time (Sec. VIII-B, Tables II and IV).
  • In quantitative tests with ground-truth poses, triangle quality ranked second only to offline Delaunay/graph-cut and above TSDF and Poisson (Sec. VIII-C4).
  • Handheld trials closed loops after 957 m and 391 m without loop detection (Sec. VIII-A).
  • On three Complex Urban sequences (3.06 to 11.72 km) ImMesh needed 3 to 6 min against about 5 to 10 h for Poisson and scored higher recall, precision and F-score (Table V).
  • In the UAV texture application a 975 m, 325 s flight was meshed and textured in 686 s in total (328 s ImMesh, 330 s R3LIVE++, 28 s texturing) (Sec. VIII-E).

限制

  • Lacks scalability in spatial resolution: large planes get many small facets and objects smaller than a voxel are poorly reconstructed (Sec. IX).
  • No loop correction, so drift can make revisited areas inconsistent (Sec. IX).
  • TSDF baseline could not run on over-3 km sequences due to GPU memory; Delaunay baseline failed on Complex Urban data (Sec. VIII-C4).
  • Textured UAV mesh shows isolated facets where scans are missing and blurry textures at large viewing angles (Sec. VIII-E).
  • (reviewer check) Table V reports Poisson completeness 0.0070 m and accuracy 0.0059 m on Urban03, better than ImMesh, which conflicts with the text; the stated 0.93% to 1.06% runtime ratio does not match Urban02 (about 0.86%).

營建工程相關證據

以校園、城市、公開資料集與 AirSim 模擬驗證，未涉及營建；作者指出固定體素尺寸無法重建小於體素的細節，對小型構件與設備量測是限制（推論）。

原文驗證環境：模擬、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 6 個比較組，合計 100 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 2 組列在最後，並連到性能比較頁。

Wang et al., 2025a · Table I 本方法 27 筆

表格設定（擷取紀錄原文）：Oxford Spires; each method meshes individual scans with ground-truth poses (every undistorted scan registered to the TLS map); meshes sampled to the raw scan point count; distances to the TLS map after pre-filtering areas not seen by both; precision, recall and F-score at 0.1 m; OctoMap voxel 0.05 m, ImMesh and VDBFusion 0.1 m, baselines configured for about 1 Hz on one core, PlanarMesh on all 28 cores; file size as PLY binary; OctoMap has no faces or vertices (N/A) (Wang et al., 2025a, Table I)

Per-Scan Time (s)，Oxford Spires · Christ Church 03 (about 307 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 Wang et al., 2025a 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：s；場景：existing buildings, indoor and outdoor (walking survey)

資料來源作者報告值(Wang et al., 2025a, Table I)

數值與出處
方法（原文寫法）報告值出處
VDBFusion硬體：28-core Intel i7 CPU, no GPU; PlanarMesh uses all cores, baselines one core each (Sec. IV-A)0.871 s(Wang et al., 2025a, Table I)
ImMesh本方法硬體：28-core Intel i7 CPU, no GPU; PlanarMesh uses all cores, baselines one core each (Sec. IV-A)0.724 s(Wang et al., 2025a, Table I)
PlanarMesh (Ours)原文提出硬體：28-core Intel i7 CPU, no GPU; PlanarMesh uses all cores, baselines one core each (Sec. IV-A)0.392 s(Wang et al., 2025a, Table I)
OctoMap硬體：28-core Intel i7 CPU, no GPU; PlanarMesh uses all cores, baselines one core each (Sec. IV-A)0.432 s(Wang et al., 2025a, Table I)

Lin et al., 2023 · Table V 本方法 24 筆

表格設定（擷取紀錄原文）：Complex Urban Dataset; ImMesh fed frame by frame with ground-truth poses (pose estimation disabled); Poisson (official implementation, octree level 12, facets with an edge over 15 cm removed) fed the accumulated cloud voxel-filtered at 1 cm; meshes sampled at 0.01 m, precision and recall at 5 cm; cost time is total processing time converted from h:min:s; Del crashed or gave no result after 3 days and TSDF exceeded 12 GB GPU memory, so both are absent (Lin et al., 2023, Table V)

Max-Min angle (deg), lower is better，Complex Urban Dataset · Urban01

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 Lin et al., 2023 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：未對齊；單位：deg；場景：urban driving, 11.72 km, 13846 frames

資料來源作者報告值(Lin et al., 2023, Table V)

數值與出處
方法（原文寫法）報告值出處
Poi60.1014 deg(Lin et al., 2023, Table V)
ImMesh (ours)本方法原文提出56.2941 deg(Lin et al., 2023, Table V)

Lin et al., 2023 · Table VI 本方法 24 筆

表格設定（擷取紀錄原文）：AirSim synthetic scenes (20 m x 10 m x 8 m) from depth images (FoV 120 x 80 deg) at three resolutions; ImMesh and TSDF (PCL, GPU, 0.2 m cells) given ground-truth poses; Del (OpenMVS Delaunay plus graph cut) and Poi given the accumulated cloud voxel-filtered at 1 cm; 5 cm threshold; cost time converted from min:sec (Lin et al., 2023, Table VI)

Completeness (m)，AirSim synthetic · Urban city, depth 640 x 480

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 Lin et al., 2023 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：未對齊；單位：m；場景：simulation (AirSim)

資料來源作者報告值(Lin et al., 2023, Table VI)

數值與出處
方法（原文寫法）報告值出處
Del0.0883 m(Lin et al., 2023, Table VI)
ImMesh (ours)本方法原文提出0.1002 m(Lin et al., 2023, Table VI)
Poi0.1094 m(Lin et al., 2023, Table VI)
TSDF0.1506 m(Lin et al., 2023, Table VI)

Tao et al., 2025 · Table 3 本方法 14 筆

指標RMS of ATE

表格設定（擷取紀錄原文）：ATE RMS (m) against LiDAR-to-TLS ground truth after SE(3) Umeyama alignment; online: VILENS-SLAM, Fast-LIO-SLAM, SC-LIO-SAM, ImMesh, Fast-LIVO2; offline: HBA (input VILENS-SLAM), COLMAP (images only). VILENS-SLAM = VILENS with pose-graph optimisation; Fast-LIO-SLAM and SC-LIO-SAM add Scan Context loop closures to Fast-LIO2 and LIO-SAM. 'x' in the table = failed or incomplete. Authors note methods could improve with further tuning. (Tao et al., 2025, Table 3)

RMS of ATE，Oxford Spires · Keble College 02 (290 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 Tao et al., 2025 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：均方根誤差（RMSE）；對齊方式：SE(3) 剛體對齊；單位：m；場景：historic site, outdoor and indoor parts (Keble College, Oxford)

資料來源作者報告值(Tao et al., 2025, Table 3)

數值與出處
方法（原文寫法）報告值出處
VILENS-SLAM0.06 m(Tao et al., 2025, Table 3)
Fast-LIO-SLAM0.25 m(Tao et al., 2025, Table 3)
SC-LIO-SAM1.26 m(Tao et al., 2025, Table 3)
ImMesh本方法0.08 m(Tao et al., 2025, Table 3)
Fast-LIVO20.95 m(Tao et al., 2025, Table 3)
HBA0.11 m(Tao et al., 2025, Table 3)
COLMAP0.05 m(Tao et al., 2025, Table 3)

其他比較組

列出其餘 2 個比較組

來源

  • Lin et al., 2023

    Jiarong Lin, Chongjian Yuan, Yixi Cai, Haotian Li, Yunfan Ren, Yuying Zou, Xiaoping Hong, Fu Zhang（2023）ImMesh: An Immediate LiDAR Localization and Meshing FrameworkIEEE Transactions on Robotics, 39(6):4312-4331

    DOI 10.1109/tro.2023.3321227arXiv 2301.05206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/hku-mars/ImMesh（授權：GPL-2.0, free for personal and academic use; commercial use requires contacting the authors (per 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。

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