SLAM 點雲圖鑑Point clouds for construction
M-01 · 2020

Voxgraph

Voxgraph 以一組相互重疊的符號距離函數（SDF）子地圖表示環境。前端依固定時間間隔把連續點雲以 voxblox 光線投射整合成 TSDF 子地圖，子地圖完成後再計算歐氏符號距離場（ESDF），並以 marching cubes 取出等值面點。後端以位姿圖最佳化各子地圖的位置與偏航角，約束包括里程計、外部迴圈閉合（例如 DBoW2）以及作者提出的免對應配準約束：把一個子地圖的等值面點轉入相鄰子地圖的 ESDF，直接讀取距離值作為殘差，並依體素權重隨機只取 5% 殘差以降低計算量。系統以地圖為中心，不重新估計完整軌跡，並在搭載 Intel i7-8650U 的六旋翼無人機上以 Ouster OS1 光達或 RealSense D415 即時運作。

具全域修正的完整 SLAM近十年已讀全文同儕審查已出版視覺、視覺慣性與 RGB-D 稠密 SLAM

閱讀版本
arXiv 2004.13154v1 (2020-04-27), header 'IEEE Robotics and Automation Letters. Preprint version. Accepted October, 2019'; IEEE version of record not compared
全文取得途徑
arXiv 預印本
本頁內容

Map-centric SDF-submap SLAM: overlapping TSDF/ESDF submaps are aligned by a pose graph over x, y, z and yaw using odometry, external loop closures and correspondence-free ESDF registration residuals (iso-surface points of one submap looked up in its neighbour's ESDF), with weight-proportional residual sub-sampling that keeps global optimization real time on an MAV CPU.

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 Ouster OS1 (64-beam) in the outdoor MAV field experiment、RGB-D camera Intel RealSense D415 (pointclouds) in the indoor MAV experiment、Odometry input from a time-synchronized camera-IMU running ROVIO visual-inertial odometry; VI-sensor stereo and IMU in the indoor dataset、RTK-GNSS used only as trajectory ground truth
原文測試平台UAV (hexacopter MAV; four outdoor flights of about 400 m at a search and rescue training site)、UAV (MAV with RGB-D camera and VI-sensor in an underground industrial space at ETH Zurich; dataset from C-blox)、simulation (RotorS, LiDAR-equipped MAV flying around a multi-story building)
狀態估計Pose-graph nonlinear least squares over submap poses in R3 x SO(2) (x, y, z, yaw; roll and pitch taken from gravity-aligned odometry) with odometry and loop-closure terms (Mahalanobis) and registration terms (weighted squared ESDF distance); per-scan poses come from the external visual-inertial odometry (ROVIO) and are kept fixed relative to their submap
資料關聯Correspondence-free submap-to-submap registration: iso-surface points of one submap are transformed into the overlapping submap and its trilinearly interpolated ESDF value is the residual; overlapping pairs found with axis-aligned bounding boxes; residuals randomly sub-sampled per solver iteration with probability proportional to voxel weight (5% sampling ratio used)
時間表示discrete poses; per-scan odometry poses stored relative to their submap
去畸變LiDAR undistortion runs as a component in the field experiment (listed in the CPU breakdown, Fig. 6); the method is not described
迴圈閉合External, source-agnostic loop closures between two sensor frames are converted into constraints between the submaps that contain them (DBoW2 place recognition on VI-sensor images in the RGB-D experiment); registration constraints link overlapping submaps, while wide loops rely on the external loop closures (Secs. VI-B, VIII-B2)
全域最佳化Pose graph over all submap poses, re-solved when each new submap is added; maximum global optimization time about 4 s in the field experiments
地圖表示collection of overlapping TSDF submaps built at a fixed frequency by ray casting into spatially hashed voxel blocks (voxblox), each with an ESDF and marching-cubes iso-surface points; submaps can be fused into one global map
先驗資訊none (assumes gravity direction observable from the odometry front end)
可輸出幾何globally aligned SDF submap collection and fused global volumetric map; trajectory obtained by projecting submap-relative odometry with the optimised submap poses
計算需求CPU only; Intel NUC Core i7-8650U carried by the MAV; fewer than 3 CPU threads (230 to 305% where 100% is one hyper-thread); global optimization uses 44% of one core; RGB-D dataset processed on a desktop CPU with 5 cm voxels (Sec. VIII-B)

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
LiDAROuster OS1方法輸入未標示64 beam (Fig. 5 caption)(Reijgwart et al., 2020, Sec. VIII-B; Fig. 5)
慣性量測單元（IMU）time-synchronized IMU (model not reported)方法輸入未標示feeds ROVIO visual-inertial odometry(Reijgwart et al., 2020, Sec. VIII-B; Fig. 5)
GNSS 接收器RTK-GNSS system (model not reported)參考或真值量測未標示attached to the MAV; used for trajectory evaluation only(Reijgwart et al., 2020, Sec. VIII-B1; Fig. 5)
相機monocular camera (model not reported)方法輸入未標示time-synchronized with IMU; feeds ROVIO visual-inertial odometry(Reijgwart et al., 2020, Sec. VIII-B; Fig. 5)
雙目相機VI-sensor (visual-inertial sensor of Nikolic et al. [36])方法輸入indoor MAV industrial dataset of C-blox [26]images used for DBoW2 loop closures and odometry in the indoor dataset(Reijgwart et al., 2020, Sec. VIII-B2)
RGB-D 相機Intel RealSense D415 Depth Camera歸入：Intel RealSense D415方法輸入indoor MAV industrial dataset of C-blox [26]produces RGB-D data; indoor dataset processed with 5 cm voxels(Reijgwart et al., 2020, Sec. VIII-B, VIII-B2)
載具平台hexacopter MAV方法輸入未標示carries 64-beam LiDAR, monocular camera, synchronized IMU and RTK-GNSS(Reijgwart et al., 2020, Sec. VIII-B; Fig. 5)
運算硬體Intel NUC Core i7-8650U執行運算平台未標示carried by the MAV; all calculations on board; 800% maximum load (8 hyper-threads)(Reijgwart et al., 2020, Abstract; Sec. VIII-B)
運算硬體desktop CPU (model not reported)執行運算平台未標示used to process the RGB-D industrial dataset(Reijgwart et al., 2020, Sec. VIII-B2)

作者報告的優勢與限制

優勢

  • Globally consistent volumetric map computed on the MAV CPU; ATE RMSE 0.83, 0.59, 0.94 and 0.52 m on four 400 m flights, lower than its ROVIO input, VINS-Mono and LOAM (Table I, Sec. VIII-B1)
  • In simulation, sub-sampling registration residuals down to about 5% showed no observable change in ESDF or trajectory RMSE while solver time fell linearly (Sec. VIII-A, Fig. 4)
  • External loop closures removed gross map distortion in the RGB-D industrial dataset where submap registration alone could not (Sec. VIII-B2, Fig. 9)

限制

  • Higher CPU load than the compared trajectory estimators (230 to 305% versus 125 to 159%), which the authors attribute to computing the volumetric map (Table I, Sec. VIII-B1)
  • Wide-baseline loops cannot be corrected by submap registration alone and require an external place-recognition source (Sec. VIII-B2)
  • Field evaluation uses RTK-GNSS trajectory error as a proxy for map quality; reconstruction error against reference geometry is reported only in simulation (Sec. VIII-A, VIII-B1)
  • Submap poses are optimized in 4 DoF, relying on gravity-aligned odometry (Sec. III)

營建工程相關證據

論文未在施工現場測試。室外實驗在瑞士 Wangen an der Aare 的搜救訓練場，場景含倒塌結構瓦礫堆與穿越建築物的室內外轉換（Fig. 1、Sec. VIII-B1）；室內 RGB-D 實驗在 ETH Zürich 的地下工業空間。場地幾何精度只以 RTK-GNSS 軌跡誤差作為代理指標。其 OS1-64 無人機資料集後來被 Blanco-Claraco, 2025 與 Millane et al., 2024 用作評估資料。

原文驗證環境：模擬、獨立參考量測

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 2 個比較組，合計 10 筆紀錄。

Reijgwart et al., 2020 · Table I 本方法 8 筆

表格設定（擷取紀錄原文）：Four MAV flights (about 400 m each) at the Wangen an der Aare search and rescue training site; RTK-GNSS ground truth; each system run 10 times and averages reported; 4-DoF alignment for visual-inertial systems and 6-DoF for LOAM; '-' means the estimator diverged (Reijgwart et al., 2020, Table I)

RMSE (m)，Voxgraph MAV field dataset (this paper) · t0

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 Reijgwart et al., 2020 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：outdoor search and rescue training site with rubble and indoor-outdoor transitions, hexacopter MAV

資料來源作者報告值(Reijgwart et al., 2020, Table I)

數值與出處
方法（原文寫法）報告值出處
ROVIO (odometry input to Voxgraph)4.55 m(Reijgwart et al., 2020, Table I)
Voxgraph本方法原文提出0.83 m(Reijgwart et al., 2020, Table I)
Vins-Mono5.51 m(Reijgwart et al., 2020, Table I)
Loam2.64 m(Reijgwart et al., 2020, Table I)

Reijgwart et al., 2020 · Text Sec.VIII-B1 本方法 2 筆

資料集與序列Voxgraph MAV field dataset (this paper) · t0 to t3

表格設定（擷取紀錄原文）：Global optimization time over 10 trials of each of 4 trajectories; text gives only the maximum (Reijgwart et al., 2020, Text Sec.VIII-B1)

maximum runtime for global optimization，Voxgraph MAV field dataset (this paper) · t0 to t3

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 Reijgwart et al., 2020 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。

統計量：最大值（max）；對齊方式：原文未報告；單位：s；場景：outdoor search and rescue training site with rubble and indoor-outdoor transitions, hexacopter MAV

數值與出處
方法（原文寫法）報告值出處
Voxgraph本方法原文提出硬體：Intel NUC Core i7-8650U (on board)4 s僅報告範圍註記（擷取紀錄）：other: approximate value ('~4 s' in Sec. VIII-B1; the abstract says less than 4 s)(Reijgwart et al., 2020, Sec. VIII-B1; Fig. 8; Abstract (120x80 m map optimized in less than 4 s))

來源

  • Reijgwart et al., 2020

    Victor Reijgwart, Alexander Millane, Helen Oleynikova, Roland Siegwart, Cesar Cadena, Juan Nieto（2020）Voxgraph: Globally Consistent, Volumetric Mapping Using Signed Distance Function SubmapsIEEE Robotics and Automation Letters, 5(1):227-234

    DOI 10.1109/lra.2019.2953859arXiv 2004.13154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/ethz-asl/voxgraph（授權：BSD-2-Clause (LICENSE file checked)）。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