SLAM 點雲圖鑑Point clouds for construction
M-01 · 2014

MRSMap

MRSMap 把每張 RGB-D 影像轉成八元樹多解析度面元地圖：各層節點都以單次掃描累加的充分統計量，保存點位置與 Lαβ 色彩的六維常態分布，並依最多六個觀測方向分開保存面元；最細解析度隨深度平方放寬，以反映 RGB-D 深度雜訊。配準時由最細解析度開始，在鄰域中尋找形狀與紋理描述子相符的面元配對，先以 Levenberg-Marquardt 初始化、再以牛頓法最大化配對似然，並以閉式近似估計位姿共變異。SLAM 以關鍵視角為節點，每一影格隨機抽選一個鄰近關鍵視角嘗試配準以發現迴圈，再以 g2o 最佳化位姿圖，全程在 CPU 上即時執行，同一方法也用於物件建模與追蹤。

具全域修正的完整 SLAM經典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
ScienceDirect HTML full text of the version of record (JVCIR 25(1), 2014) plus the corrigendum (JVCIR 26, 2015, p. 349)
全文取得途徑
國立臺灣大學機構訂閱
本頁內容

CPU RGB-D registration and key-view SLAM on octree multi-resolution surfel maps: each node stores a 6D position-colour Gaussian per view direction with depth-adapted finest resolution; maps are registered by descriptor-gated multi-resolution surfel association with LM then Newton optimization and closed-form pose covariance, and key views are linked by randomized loop-closure tests and g2o pose-graph optimization.

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 at VGA 640x480 and 30 Hz (TUM Freiburg benchmark sequences and the authors' object dataset; sensor model not named); QVGA used on the robot
原文測試平台RGB-D camera moved through the TUM Freiburg scenes and around objects for the authors' object dataset (carrying mode not stated, Secs. 5 and 6)、mobile manipulation robot Cosero (RoboCup@Home 2011 and 2012 demonstrations; locomotion not described, Sec. 6.4)
狀態估計Key-view SLAM: each frame is registered to the current reference key view by maximizing the surfel-match likelihood (approximate Levenberg-Marquardt initialization, then Newton's method with trilinear interpolation, typically 10 to 20 LM and 5 Newton iterations); key views linked by relative-pose constraints with closed-form covariance; pose graph solved by sparse Cholesky in g2o, one iteration per frame
資料關聯Multi-resolution surfel association starting at the finest resolution with a cubic volume query around the transformed surfel mean (side twice the node resolution), bootstrapped from previous associations via the 26-neighbourhood; accepted only if shape-texture descriptors (surfel-pair angle histograms and luminance and chrominance contrasts) differ by at most 0.1 and contour flags agree
時間表示discrete poses
去畸變不適用 (RGB-D input)
迴圈閉合Randomized hypothesis-and-test: per frame one key view, sampled with probability decreasing with distance and angle from the reference, is registered; the constraint is accepted if its bidirectional matching likelihood is at least a fraction of that of the key view's initial constraint
全域最佳化Pose graph over key views in g2o, iterated once per frame (median 0.79 ms, max 4.01 ms on freiburg2_desk with 64 key views and 138 edges); optimized key views fused into one multi-view map
地圖表示octree multi-resolution surfel map: every node stores sufficient statistics of a 6D Gaussian of position and L-alpha-beta colour, up to six surfels per node for orthogonal view directions; finest node size adapted to squared depth (0.0125 m limit); border and occluded-background surfels excluded
先驗資訊none
可輸出幾何multi-view multi-resolution surfel map of a scene or an object model (visualized by sampling the surfel distributions; object models about 54 MB for a chair and 19 MB for a humanoid)
計算需求CPU only (multi-core parallel association and derivatives); notebook Intel Core i7 3610QM 2.3 GHz; registration about 15 Hz at VGA (61 to 75 ms per frame); real-time SLAM drops frames and limits maximum resolution to 0.05 m

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
載具平台Cosero mobile manipulation robot方法輸入未標示used the tracking method at RoboCup@Home 2011 and 2012(Stückler & Behnke, 2014, Sec. 6.4)
運算硬體notebook PC with Intel Core i7 3610QM 2.3 GHz (max. 3.3 GHz) QuadCore CPU執行運算平台未標示timings of all methods at VGA resolution(Stückler & Behnke, 2014, Sec. 6)
運算硬體quadcore notebook with Intel i7-Q720 CPU (Cosero's main computer)執行運算平台未標示RGB-D images subsampled to QVGA for tracking during RoboCup@Home demonstrations(Stückler & Behnke, 2014, Sec. 6.4)
其他external optical motion capture systems (models not reported)參考或真值量測TUM RGB-D and authors' object tracking datasetground-truth camera poses of the TUM benchmark and the authors' object dataset(Stückler & Behnke, 2014, Sec. 6)

作者報告的優勢與限制

優勢

  • Lowest median incremental-registration RPE on most of 22 TUM sequences against warp, GICP, 3D-NDT and fovis (Table 1)
  • About 15 Hz at VGA on a notebook CPU: 75 ms versus 4015 ms for GICP and 415 ms for 3D-NDT on fr1 desk (Table 2)
  • Keeps ICP-like robustness for larger frame skips where warp fails (Fig. 6)
  • Object models with median ATE of about 1 to 2 cm, and graph optimization strongly reduces error (Table 4)
  • Used for real-time object tracking on the Cosero robot at RoboCup@Home 2011 and 2012 (Sec. 6.4)

限制

  • Did not succeed on freiburg1_floor (little texture) or freiburg2_large_loop, and loses track when much of the image has invalid or highly uncertain distant depth (Sec. 6.2)
  • Real-time SLAM must drop frames and limit the maximum resolution to 0.05 m (Sec. 6.2)
  • Loop detection relies on pose proximity; appearance-based loop detection for larger scenes is future work (Sec. 5.1.1, Sec. 7)
  • Fovis is much faster (13 to 16 ms) and more accurate on some textured sequences (Tables 1 and 2)
  • The SLAM RPE values of Table 3 were corrected by a 2015 corrigendum; with the corrected values the method beats RGB-D SLAM on 8 of 11 sequences rather than the nine stated in Sec. 6.2 (inference from the corrigendum)

營建工程相關證據

論文未在施工現場測試，評估為 TUM (Freiburg) 室內小場景與桌上物件資料集，並在 RoboCup@Home 競賽中讓服務機器人即時追蹤桌子與鍋具。在 CPU 上以多解析度面元配準 RGB-D 影像，適合運算資源有限的室內機器人；但論文 Table 3 的 SLAM 精度已由 2015 年更正啟事（JVCIR 26:349）修訂，引用時應採更正後數值。corpus 中 Dai et al., 2017a、Whelan et al., 2015a、Whelan et al., 2015b 與 Schöps et al., 2019 皆以 MRSMap 作為比較基準。

原文驗證環境：公開基準、受控實驗、任務層驗證

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 7 個比較組，合計 60 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 3 組列在最後，並連到性能比較頁。

Stückler & Behnke, 2014 · Table 1 本方法 22 筆

指標median relative pose error (RPE) in mm

表格設定（擷取紀錄原文）：Incremental (frame-to-frame) registration on TUM Freiburg sequences; median translational relative pose error in mm (maximum values in brackets in the table not extracted); warp is the OpenCV reimplementation (Stückler & Behnke, 2014, Table 1)

median relative pose error (RPE) in mm，TUM RGB-D (Freiburg) · fr1 360

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 Stückler & Behnke, 2014 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：中位數（median）；對齊方式：原文未報告；單位：mm；場景：indoor office and structure/texture test scenes, RGB-D camera (carrying mode not stated in the paper)

資料來源作者報告值(Stückler & Behnke, 2014, Table 1)

數值與出處
方法（原文寫法）報告值出處
Ours (MRSMap)本方法原文提出5.1 mm(Stückler & Behnke, 2014, Table 1)
Warp [17] (OpenCV)5.9 mm(Stückler & Behnke, 2014, Table 1)
GICP [5]18.8 mm(Stückler & Behnke, 2014, Table 1)
3D-NDT [7]7.8 mm(Stückler & Behnke, 2014, Table 1)
Fovis [12]7.1 mm(Stückler & Behnke, 2014, Table 1)

Stückler & Behnke, 2014 · Table 3 (corrigendum) 本方法 22 筆

指標RMSE RPE in m

表格設定（擷取紀錄原文）：SLAM on TUM Freiburg sequences: RMSE of relative pose error averaged over all frame differences, in m; values from the 2015 corrigendum that replaces the originally printed Table 3 (Stückler & Behnke, 2014, Table 3 (corrigendum))

RMSE RPE in m，TUM RGB-D (Freiburg) · freiburg1_360

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 Stückler & Behnke, 2014 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 and structure/texture test scenes, RGB-D camera (carrying mode not stated in the paper)

資料來源作者報告值(Stückler & Behnke, 2014, Table 3 (corrigendum))

數值與出處
方法（原文寫法）報告值出處
Ours all frames (MRSMap SLAM)本方法原文提出0.123 m(Stückler & Behnke, 2014, Corrigendum Table 3 (replaces article Table 3))
Ours real-time (MRSMap SLAM, frames dropped, 0.05 m max. resolution)本方法原文提出0.126 m(Stückler & Behnke, 2014, Corrigendum Table 3 (replaces article Table 3))
RGB-D SLAM [24], [34] (Endres et al.)0.103 m(Stückler & Behnke, 2014, Corrigendum Table 3 (replaces article Table 3))

Yan et al., 2017 · Table 2 本方法 4 筆

指標ATE [m]

表格設定（擷取紀錄原文）：Complete SLAM absolute trajectory error on TUM RGB-D; '-' = no result given (Yan et al., 2017, Table 2)

ATE [m]，TUM RGB-D · fr1/desk

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 Yan et al., 2017 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 scenes, RGB-D camera (carrying mode not stated in the paper)

資料來源作者報告值(Yan et al., 2017, Table 2)

數值與出處
方法（原文寫法）報告值出處
RGB-D SLAM [4] (Endres et al.)0.023 m(Yan et al., 2017, Table 2)
Kintinuous [33]0.037 m(Yan et al., 2017, Table 2)
MRSMap [28]本方法0.043 m(Yan et al., 2017, Table 2)
ElasticFusion [34]0.02 m(Yan et al., 2017, Table 2)
DVO SLAM [14]0.021 m(Yan et al., 2017, Table 2)
sigma-DVO SLAM0.019 m(Yan et al., 2017, Table 2)
PSM SLAM原文提出0.016 m(Yan et al., 2017, Table 2)

Yan et al., 2017 · Table 4 本方法 4 筆

指標mean distance from points to nearest ground-truth surface (m)

表格設定（擷取紀錄原文）：Surface reconstruction accuracy on ICL-NUIM living room with noise: mean distance from reconstructed points to the nearest ground-truth surface (m) (Yan et al., 2017, Table 4)

mean distance from points to nearest ground-truth surface (m)，ICL-NUIM · lr kt0

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 Yan et al., 2017 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：原文未報告；單位：m；場景：synthetic living room

資料來源作者報告值(Yan et al., 2017, Table 4)

數值與出處
方法（原文寫法）報告值出處
RGB-D SLAM [4] (Endres et al.)0.044 m(Yan et al., 2017, Table 4)
Kintinuous [33]0.011 m(Yan et al., 2017, Table 4)
MRSMap [28]本方法0.061 m(Yan et al., 2017, Table 4)
DVO SLAM [14]0.032 m(Yan et al., 2017, Table 4)
ElasticFusion [34]0.007 m(Yan et al., 2017, Table 4)
PSM SLAM原文提出0.006 m(Yan et al., 2017, Table 4)

其他比較組

列出其餘 3 個比較組

來源

  • Stückler & Behnke, 2014

    Jörg Stückler, Sven Behnke（2014）Multi-resolution surfel maps for efficient dense 3D modeling and trackingJournal of Visual Communication and Image Representation, 25(1):137-147

    DOI 10.1016/j.jvcir.2013.02.008程式碼

    同儕審查已出版已讀全文經典查證後修正

程式碼：https://github.com/jstueckler/mrsmap（授權：BSD licence (LICENSE file checked, University of Bonn)）。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