SLAM 點雲圖鑑Point clouds for construction
M-01 · 2019

ReFusion

ReFusion 是以 TSDF 為模型的 RGB-D 稠密 SLAM，目標是在有多個移動物體的室內場景中只重建靜態部分。位姿估計不渲染合成視圖，而是把目前影格的點直接帶入 TSDF，以內插得到的符號距離作為殘差，並加入體素色彩的光度誤差。第一次配準後，殘差超過門檻的像素視為動態區域，再以考慮深度的區域成長（flood fill）擴展成遮罩，排除後重新配準並整合。另把相機視錐內確定為空的體素標記為自由空間，之後落在自由空間的量測即視為動態物體而拒絕。方法純幾何、不依賴語意偵測，體素以雜湊配置並在 GPU 平行處理；作者同時發布以動作捕捉提供軌跡、以 Leica BLK360 地面雷射掃描提供靜態場景真值的 Bonn RGB-D 動態資料集。

里程計＋局部建圖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 1905.02082v3 (2019-08-29), accepted IROS 2019 version; IEEE version of record not compared
全文取得途徑
arXiv 預印本
本頁內容

Dynamic-scene RGB-D TSDF SLAM: direct point-to-TSDF plus voxel-colour photometric tracking, dynamic pixels found from large registration residuals and grown by depth-aware flood fill, and free-space carving that rejects measurements in voxels previously observed empty; class agnostic, GPU voxel hashing, released with the Bonn RGB-D Dynamic Dataset (motion-capture trajectories, TLS static-scene reference).

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RGB-D camera (ASUS Xtion Pro LIVE in the Bonn RGB-D Dynamic Dataset; TUM RGB-D dynamic sequences)
原文測試平台handheld (TUM RGB-D dynamic sequences, camera carried by a cameraman)、原文未報告 (carrier of the ASUS Xtion in the Bonn dataset is not described)
狀態估計Frame-to-model direct alignment: points of the current frame are transformed into the TSDF and the interpolated SDF value is the geometric residual, plus a photometric residual against voxel colours (weight 0.025); Levenberg-Marquardt on three coarse-to-fine levels, GPU-parallel; a second registration is run after masking dynamic pixels
資料關聯Correspondence-free point-to-implicit residuals; dynamic pixels are those whose residual exceeds t = gamma times tau squared (gamma 0.5, tau 0.1 m), grown by depth-aware flood fill (threshold 0.007) and dilation
時間表示discrete poses
去畸變不適用 (RGB-D input)
迴圈閉合none
全域最佳化none
地圖表示TSDF with weight and colour per voxel in dynamically allocated voxel-hashed blocks (1 cm voxels, 0.1 m truncation); voxels seen empty in the camera frustum are marked as free space (SDF set to the truncation distance)
先驗資訊none
可輸出幾何mesh of the static part of the scene, trajectory
計算需求GPU-parallel voxel and pixel processing; GPU model and runtime not reported; registration takes up to twice as long as without dynamics handling because of the second pass (Sec. III-C)

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
地面雷射掃描儀（TLS）Leica BLK360參考或真值量測Bonn RGB-D Dynamic Dataset (this paper)high-resolution point cloud of the static part of the test environment(Palazzolo et al., 2019, Sec. IV-B, IV-C; Fig. 12b)
RGB-D 相機ASUS Xtion Pro LIVE歸入：Asus Xtion Pro Live方法輸入Bonn RGB-D Dynamic Dataset (this paper)recorded depth always within the valid sensor range for the Bonn sequences(Palazzolo et al., 2019, Sec. IV-B; Sec. III-E)
其他Optitrack Prime 13 motion capture system參考或真值量測Bonn RGB-D Dynamic Dataset (this paper)ground-truth sensor trajectories(Palazzolo et al., 2019, Sec. IV-B)
其他tilt and turn targets參考或真值量測Bonn RGB-D Dynamic Dataset (this paper)located by both the laser scanner and the motion capture system to align the TLS cloud to the motion-capture frame(Palazzolo et al., 2019, Sec. IV-C; Fig. 12c)

作者報告的優勢與限制

優勢

  • Tracking on TUM dynamic sequences comparable to StaticFusion, and StaticFusion lost track on walking halfsphere (0.681 m vs 0.104 m) (Table II)
  • On 24 Bonn sequences the worst ATE (0.571 m) is on par with DynaSLAM N+G (0.575 m) and far below StaticFusion (3.586 m) and DynaSLAM G (1.217 m) (Table III)
  • Best geometric method on crowded scenes (crowd, crowd2, crowd3) (Table III)
  • Static model closer to the TLS reference than StaticFusion on crowd3 and removing nonobstructing box (Fig. 11)
  • Class agnostic; no object tracking, so the number and speed of moving objects are not limited (Sec. I)

限制

  • Feature-based DynaSLAM (geometric or with segmentation) outperforms it on the TUM dynamic sequences (Table II)
  • A dynamic object observed at a location that is never revisited stays in the model (walking xyz, Fig. 7)
  • Invalid (zero) depth prevents free-space reasoning; the virtual-depth workaround from 10 frames adds about 0.3 s delay and assumes nothing is closer than the minimum sensor range (Sec. III-E)
  • Marking all free voxels in the frustum is memory-inefficient (Sec. III-D)
  • No loop closure or global optimization (inference from Sec. III)

營建工程相關證據

論文未在施工現場測試，實驗為 TUM RGB-D 辦公室動態序列，以及作者在室內測試空間錄製、有人走動、搬箱與玩氣球的動態序列。以配準殘差與自由空間排除動態物體、只保留靜態結構的作法，符合施工中有人員與機具移動時的室內掃描需求。其 Bonn 資料集以 Leica BLK360 地面雷射掃描點雲作為靜態場景真值，並用傾斜與轉動標靶把掃描對齊到動作捕捉座標系（Sec. IV-C），此評估流程可作為工地動態場景掃描驗證的參考（推論）。

原文驗證環境：公開基準、獨立參考量測、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 2 個比較組，合計 30 筆紀錄。

Palazzolo et al., 2019 · Table III 本方法 24 筆

指標Absolute Trajectory Error (RMS) [m]

表格設定（擷取紀錄原文）：Bonn RGB-D Dynamic Dataset (24 highly dynamic scenes, ASUS Xtion Pro LIVE, OptiTrack Prime 13 ground truth); ATE RMS; 'o box' = obstructing box, 'no box' = nonobstructing box; all methods run by the authors with default parameters (Palazzolo et al., 2019, Table III)

Absolute Trajectory Error (RMS) [m]，Bonn RGB-D Dynamic Dataset · balloon

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 Palazzolo et al., 2019 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：原文未報告；單位：m；場景：indoor test room with people manipulating boxes, balloons or crowding the camera

資料來源作者報告值(Palazzolo et al., 2019, Table III)

數值與出處
方法（原文寫法）報告值出處
Ours (ReFusion)本方法原文提出0.175 m(Palazzolo et al., 2019, Table III)
SF (StaticFusion)0.233 m(Palazzolo et al., 2019, Table III)
DS (G) (DynaSLAM geometric)0.05 m(Palazzolo et al., 2019, Table III)
DS (N+G) (DynaSLAM neural network + geometric)0.03 m(Palazzolo et al., 2019, Table III)

Palazzolo et al., 2019 · Table II 本方法 6 筆

指標Absolute Trajectory Error (RMS) [m]

表格設定（擷取紀錄原文）：TUM RGB-D dynamic scenes; ATE RMS; ReFusion uses virtual depth from 10 frames (about 0.3 s delay) to fill invalid depth; ReFusion, SF and MF are dense, DynaSLAM is feature-based; StaticFusion lost track on walking halfsphere (Palazzolo et al., 2019, Table II)

Absolute Trajectory Error (RMS) [m]，TUM RGB-D · sitting static

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 Palazzolo et al., 2019 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：原文未報告；單位：m；場景：indoor office with seated or walking people, handheld RGB-D

資料來源作者報告值(Palazzolo et al., 2019, Table II)

數值與出處
方法（原文寫法）報告值出處
Ours (ReFusion)本方法原文提出0.009 m(Palazzolo et al., 2019, Table II)
SF (StaticFusion)0.014 m(Palazzolo et al., 2019, Table II)
DS (G) (DynaSLAM geometric)0.009 m(Palazzolo et al., 2019, Table II)
DS (N+G) (DynaSLAM neural network + geometric)0.007 m(Palazzolo et al., 2019, Table II)
MF (MaskFusion, values from its paper)0.021 m(Palazzolo et al., 2019, Table II)

來源

  • Palazzolo et al., 2019

    Emanuele Palazzolo, Jens Behley, Philipp Lottes, Philippe Giguère, Cyrill Stachniss（2019）ReFusion: 3D Reconstruction in Dynamic Environments for RGB-D Cameras Exploiting Residuals2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 7855-7862

    DOI 10.1109/iros40897.2019.8967590arXiv 1905.02082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/PRBonn/refusion（授權：CC BY-NC-SA 3.0 Unported (LICENSE.txt checked; non-commercial)）。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