SLAM 點雲圖鑑Point clouds for construction
M-01 · 2007

GMapping

GMapping 在 Rao-Blackwellized 粒子濾波（每個粒子攜帶一張佔據網格地圖）上提出兩項改良：以掃描匹配結果與里程計共同計算較準確的提議分布（proposal distribution），以及依有效樣本數選擇性重取樣以減少粒子耗盡。作者報告所需粒子數大約比先前方法少一個數量級。當掃描匹配失敗（如大空曠區域多為最大量程讀值）時，系統退回原始里程計運動模型。

具全域修正的完整 SLAM經典已讀全文同儕審查已出版機率基礎、綜述與 2D SLAM

閱讀版本
author-hosted manuscript of the T-RO paper (University of Freiburg, grisetti07tro.pdf, 12 pages, generated 2007-01-17); layout differs from the T-RO version of record 23(1):34-46, which was not compared line by line
全文取得途徑
作者自存版本
本頁內容

RBPF grid mapping with a scan-matching-informed proposal and adaptive resampling, reducing the particle count needed for correct maps.

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2D laser range finder、wheel odometry
原文測試平台wheeled UGV
狀態估計Rao-Blackwellized particle filter; per-particle Gaussian proposal fitted to K samples around the scan-matcher mode, weighted by observation likelihood and the odometry motion model; raw motion model used when scan matching fails; resampling only when Neff drops below N/2 (Sec. III-B to III-E)
資料關聯per-particle scan matching with the CARMEN 'vasco' matcher: gradient descent on the beam-endpoint likelihood of the current scan against the particle's own grid map, with the search bounded around the odometry-based initial guess (Sec. III-E, IV)
時間表示discrete poses; a filter update after each 0.5 m of travel or 25 deg of rotation (Sec. III-C, VI-F)
去畸變原文未報告
迴圈閉合implicit through particle filter (no explicit loop-closure module)
全域最佳化none
地圖表示2D occupancy grid per particle
先驗資訊none
可輸出幾何2D occupancy grid map
計算需求online on robots; per observation O(N) without resampling and O(NM) with resampling, M the grid size (Sec. V, Table I); Intel Lab log (45 min) corrected in under 30 min with 30 particles, 150 MB, 5 cm grid, 2.8 GHz PC; average execution times 1910 ms for proposal, weights and map update, 41 ms for the resampling test and 244 ms for resampling (Sec. VI-F, Table III)

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
LiDARSICK LMS方法輸入未標示原文未報告(Grisetti et al., 2007, Sec. VI)
LiDARSICK PLS方法輸入未標示原文未報告(Grisetti et al., 2007, Sec. VI)
載具平台ActivMedia Pioneer 2 AT方法輸入未標示equipped with SICK LMS or PLS laser range finders(Grisetti et al., 2007, Sec. VI, Fig. 3)
載具平台Pioneer 2 DX-8方法輸入未標示equipped with SICK LMS or PLS laser range finders(Grisetti et al., 2007, Sec. VI, Fig. 3)
載具平台iRobot B21r方法輸入未標示equipped with SICK LMS or PLS laser range finders(Grisetti et al., 2007, Sec. VI, Fig. 3)
載具平台Pioneer II robot資料集感測器Intel Research Labequipped with a SICK sensor(Grisetti et al., 2007, Sec. VI-A)
運算硬體standard PC with a 2.8 GHz processor執行運算平台未標示2.8 GHz(Grisetti et al., 2007, Sec. VI-F)

作者報告的優勢與限制

優勢

  • Particle count needed for a topologically correct map in at least 60% of runs was about one order of magnitude smaller than Hähnel et al.'s approach (Sec. VI.B).
  • Never more than 80 particles for datasets up to about 250 m x 250 m (Sec. VI).

限制

  • Scan matcher can fail in large open spaces (mostly maximum-range readings) or with poor overlap, then only odometry is used (Sec. III-C
  • Sec. VI-E) | Partly violates the planar-environment assumption outdoors (Sec. VI-A, Freiburg campus) | Only one scan-matcher mode is sampled, so in theory the filter may become overly confident in extremely cluttered environments with very noisy odometry
  • the authors never met this case with real robots (Sec. III-C) | With 60 particles the MIT Killian Court maps sometimes showed artificial double walls
  • 80 particles were needed for high quality (Sec. VI-A) | Each particle copies its full grid map at resampling, giving a worst-case O(NM) cost
  • adaptive resampling keeps such steps rare (Sec. V) | (inference) Evaluation targets topological correctness and particle counts rather than metric map error against an independent reference, so it cannot support 3D point-cloud accuracy claims

營建工程相關證據

未在工地測試。Freiburg 校園資料中，停車場因施工移走車輛導致掃描大多為最大量程讀值，掃描匹配失效一次（Sec. VI.E）；此為資料集中的偶發事件，並非營建場域驗證，但說明空曠、低特徵區域的風險。

原文驗證環境：公開基準、已完工建築、模擬

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 12 個比較組，合計 79 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 8 組列在最後，並連到性能比較頁。

Kümmerle et al., 2009 · Table 1 本方法 21 筆

表格設定（擷取紀錄原文）：Translational error of the relative-relation metric (Eq. 4) averaged over all manually verified relations for three mapping approaches; abs values in m, squared values in m^2 (Kümmerle et al., 2009, Table 1)

Translational error, Equation 4 using absolute errors (reported as mean ± std)，Aces · local relations provided by the authors

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 Kümmerle et al., 2009 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：indoor corridors, ACES building, University of Texas at Austin (Sec. 7)

資料來源作者報告值(Kümmerle et al., 2009, Table 1)

數值與出處
方法（原文寫法）報告值出處
Scan Matching0.173 m原文指標寫法：Translational error, Equation 4 using absolute errors (reported as mean ± std; std = 0.614)(Kümmerle et al., 2009, Table 1)
RBPF (50 part.)本方法0.06 m原文指標寫法：Translational error, Equation 4 using absolute errors (reported as mean ± std; std = 0.049)(Kümmerle et al., 2009, Table 1)
Graph Mapping0.044 m原文指標寫法：Translational error, Equation 4 using absolute errors (reported as mean ± std; std = 0.044)(Kümmerle et al., 2009, Table 1)

Kümmerle et al., 2009 · Table 2 本方法 21 筆

表格設定（擷取紀錄原文）：Rotational error of the relative-relation metric (Eq. 4) averaged over all manually verified relations for three mapping approaches; abs values in deg, squared values in deg^2 (Kümmerle et al., 2009, Table 2)

Rotational error, Equation 4 using absolute errors (reported as mean ± std)，Aces · local relations provided by the authors

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 Kümmerle et al., 2009 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：indoor corridors, ACES building, University of Texas at Austin (Sec. 7)

資料來源作者報告值(Kümmerle et al., 2009, Table 2)

數值與出處
方法（原文寫法）報告值出處
Scan Matching1.2 deg原文指標寫法：Rotational error, Equation 4 using absolute errors (reported as mean ± std; std = 1.5)(Kümmerle et al., 2009, Table 2)
RBPF (50 part.)本方法1.2 deg原文指標寫法：Rotational error, Equation 4 using absolute errors (reported as mean ± std; std = 1.3)(Kümmerle et al., 2009, Table 2)
Graph Mapping0.4 deg原文指標寫法：Rotational error, Equation 4 using absolute errors (reported as mean ± std; std = 0.4)(Kümmerle et al., 2009, Table 2)

Labbé & Michaud, 2019 · Table 9 本方法 8 筆

表格設定（擷取紀錄原文）：MIT Stata Center 2012-01-25 sequences; RTAB-Map WheelIMU→S2M versus other ROS 2D lidar SLAM run with default parameters; Cartographer, GMapping and Karto use WheelIMU odometry, Hector SLAM uses none; GMapping ATE computed on the current best particle path (Labbé & Michaud, 2019, Table 9)

ATEend，MIT Stata Center (PR2) · 2012-01-25-12-14-25

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 Labbé & Michaud, 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 building; Long-range lidar

資料來源作者報告值(Labbé & Michaud, 2019, Table 9)

數值與出處
方法（原文寫法）報告值出處
RTAB-Map (WheelIMU→S2M) [Long-range lidar]原文提出0.05 m(Labbé & Michaud, 2019, Table 9)
Cartographer (WheelIMU) [Long-range lidar]0.11 m(Labbé & Michaud, 2019, Table 9)
GMapping (WheelIMU) [Long-range lidar]本方法0.19 m(Labbé & Michaud, 2019, Table 9)
Karto SLAM (WheelIMU) [Long-range lidar]0.22 m(Labbé & Michaud, 2019, Table 9)
Hector SLAM (no odometry) [Long-range lidar]0.06 m(Labbé & Michaud, 2019, Table 9)
RTAB-Map (WheelIMU→S2M) [Short-range lidar]原文提出0.07 m(Labbé & Michaud, 2019, Table 9)
Cartographer (WheelIMU) [Short-range lidar]0.45 m(Labbé & Michaud, 2019, Table 9)
GMapping (WheelIMU) [Short-range lidar]本方法1.71 m(Labbé & Michaud, 2019, Table 9)
Karto SLAM (WheelIMU) [Short-range lidar]0.48 m(Labbé & Michaud, 2019, Table 9)
Hector SLAM (no odometry) [Short-range lidar]4.59 m(Labbé & Michaud, 2019, Table 9)

Guadagnino et al., 2025a · Table VII 本方法 5 筆

指標ATE translation RMS [cm] (mean over 10 runs)

表格設定（擷取紀錄原文）：2D Monte-Carlo localization (RVP-Loc, Clearpath Dingo with SICK TiM781S) on a 2D map sliced from the KISS-SLAM 3D occupancy grid versus a GMapping map; pose-tracking ATE translation RMS, mean of 10 runs; ground truth from ceiling AprilTags seen by an upward camera; success rate and convergence time columns omitted (Guadagnino et al., 2025a, Table VII)

ATE translation RMS [cm] (mean over 10 runs)，authors' office sequences · Static Sequence 1

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 Guadagnino 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：均方根誤差（RMSE）；對齊方式：原文未報告；單位：cm；場景：office (static and dynamic scenes)

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

數值與出處
方法（原文寫法）報告值出處
GMapping map本方法9.48 cm(Guadagnino et al., 2025a, Table VII)
Ours (KISS-SLAM map)原文提出9.71 cm(Guadagnino et al., 2025a, Table VII)

其他比較組

列出其餘 8 個比較組

來源

  • Grisetti et al., 2007

    Giorgio Grisetti, Cyrill Stachniss, Wolfram Burgard（2007）Improved Techniques for Grid Mapping With Rao-Blackwellized Particle FiltersIEEE Transactions on Robotics, 23(1):34-46

    DOI 10.1109/tro.2006.889486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/OpenSLAM-org/openslam_gmapping（授權：BSD-3-Clause (stated on the OpenSLAM.org GMapping page, which links 'Get the Source Code!' to this repository; no LICENSE file at the repository root)）。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