6D SLAM (Kurt3D, stop-scan-go ICP SLAM)
本文提出以三維雷射掃描為基礎的 6D SLAM（六自由度同時定位與建圖）：機器人以停下、掃描、再前進（stop-scan-go）的方式取得每一幅三維點雲，先把輪式里程計外推為六自由度初值，再以八元樹（octree）由粗到細搜尋初始對齊，之後用 ICP（Iterative Closest Point）逐幅配準（registration）。偵測到迴圈時，把閉合誤差依行經路徑長度比例分攤給迴圈內各幅掃描；資料收集完成後，再以同步配準（simultaneous matching）式的全域鬆弛（global relaxation）反覆將每幅掃描對其重疊鄰居重新配準。為壓低計算量，作者使用點數縮減、近似 k-d tree 與快取 k-d tree（cached k-d tree）搜尋。整套系統只維持單一位姿假設，未使用機率式不確定性表示。
本頁內容
A stop-scan-go 6-DoF SLAM system that registers 3D laser scans with ICP (octree-based initial guess, fast k-d tree variants), distributes loop-closing error along the path, and refines the whole map by offline neighbour-wise global relaxation without probabilistic uncertainty.
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 laser range finder built from a SICK 2D scanner on a servo-driven pitch mount (Sec. 5.1)、wheel odometry used only for initial pose extrapolation (Sec. 3.1) |
|---|---|
| 原文測試平台 | wheeled UGV (Kurt3D outdoor version, six-wheel skid steer; Sec. 5.1) |
| 狀態估計 | single-hypothesis deterministic pose estimation: odometry extrapolated to 6 DoF, octree-based coarse-to-fine initial alignment, then ICP with closed-form SVD solution per scan; explicitly no probabilistic filter or covariance (Sec. 3, Sec. 6) |
| 資料關聯 | point-to-point closest-point correspondences (ICP) using k-d tree, approximate k-d tree and cached k-d tree search; octree cube-overlap counting for the initial guess (Sec. 3.2-3.3, Sec. 4) |
| 時間表示 | discrete poses (one 6-DoF pose per stop-scan-go 3D scan) |
| 去畸變 | avoided by design: the robot stands still while each 3D scan is taken (stop-scan-go, Sec. 3 and Sec. 5.1); no in-motion distortion correction |
| 迴圈閉合 | loop hypothesis from maximum laser range and current pose, revised by octree matching; accepted when the number of closest-point pairs exceeds a threshold; closing error distributed over the loop's scans in proportion to travelled path length (translation linearly, rotation by quaternion interpolation) (Sec. 3.4) |
| 全域最佳化 | 'simultaneous matching' global relaxation inspired by Pulli: queue-based re-registration of each scan against the union of overlapping neighbours (overlap = more than 250 point pairs) until no scan moves more than 5 cm; the first scan is fixed; run offline after acquisition (Sec. 3.5, Sec. 4) |
| 地圖表示 | set of registered 3D point clouds (scans); octree used only for initial alignment; voxel view shown for a RoboCup arena map (Fig. 19) |
| 先驗資訊 | none (odometry only as initial guess) |
| 可輸出幾何 | globally registered 3D point cloud and 6-DoF scan poses; a continuous trajectory is reconstructed afterwards by distributing the gaps between trajectory patches (Sec. 5.4, Fig. 18); no covariance output |
| 計算需求 | for 77 scans: scan registration plus loop detection about 10 min, global relaxation about 2 h on a Pentium-IV 2.8 GHz (Sec. 5.4); octree heuristic up to about 2 s per scan pair if applied naively (Sec. 4); approximate k-d tree search cuts ICP time to roughly 75% (Sec. 4.3); on-board CPU Pentium Centrino 1.4 GHz (Sec. 5.1) |
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 |
|---|---|---|---|---|---|
| LiDAR | 3D laser range finder built from a SICK 2D laser range finder on a servo-driven pitch mount (SICK model not stated) | 方法輸入 | 未標示 | scans up to 180 deg (h) x 120 deg (v); horizontal resolutions 181, 361, 721 and vertical 128, 225, 420, 500; a 181-point plane takes 13 ms; a 181 x 256 scan takes 3.4 s; stop-scan-go acquisition | (Nüchter et al., 2007, Sec. 5.1, Fig. 8) |
| 輪式或腿式里程計 | Kurt3D wheel odometer | 方法輸入 | 未標示 | planar odometry extrapolated to 6 DoF as initial guess only | (Nüchter et al., 2007, Sec. 3.1) |
| 載具平台 | Kurt3D (outdoor version) | 方法輸入 | 未標示 | 45 cm x 33 cm x 29 cm, 22.6 kg; two 90 W motors driving six skid-steered wheels; 16-bit CMOS microcontroller for motor control | (Nüchter et al., 2007, Sec. 5.1, Fig. 8) |
| 運算硬體 | Pentium-Centrino-1400 with 768 MB RAM, Linux (robot core computer) | 執行運算平台 | 未標示 | on-board computer of Kurt3D; no processing times are reported on it | (Nüchter et al., 2007, Sec. 5.1) |
| 運算硬體 | Pentium-IV-2800 MHz | 執行運算平台 | 未標示 | scan registration and loop detection for 77 scans about 10 min; global relaxation run for 2 h | (Nüchter et al., 2007, Sec. 5.4) |
| 其他 | Meter rule | 參考或真值量測 | 未標示 | reference distance of a closed-loop span (2080 cm) and pose shifts for matchability tests | (Nüchter et al., 2007, Sec. 4.3, Fig. 13) |
| 其他 | Uncalibrated mid-resolution aerial image | 參考或真值量測 | 未標示 | distance ratios between reference points A-D compared with the point model | (Nüchter et al., 2007, Sec. 5.4, Fig. 15, Table II) |
作者報告的優勢與限制
優勢
- Octree-based initial estimation lets ICP match scans from rudimentary odometry guesses; without it ICP would likely converge to a wrong minimum in the campus experiment (Sec. 3.2, Sec. 5.4)
- ICP tolerated initial errors of about 1 m in position and about 15 degrees in orientation in the Birlinghoven data set (Sec. 5.3)
- 3D scans taken at sparse locations were aligned where a 2D slice-based alignment of the same data showed noticeable errors (Sec. 5.2, Fig. 10)
- Length ratios measured in the final 77-scan campus map deviated 0.5-3.8% from ratios measured in an uncalibrated aerial image (Table II)
- A closed-loop distance of 2096 cm in the point cloud versus 2080 cm by meter rule (Fig. 13 caption)
- Cached k-d tree search gives exact correspondences faster than conventional k-d tree search after the first ICP iteration (Sec. 4.4, Fig. 7)
限制
- The approach concentrates on single loops and keeps one pose hypothesis; multi-hypothesis tracking was considered computationally infeasible (Sec. 3, Sec. 6)
- The octree initial-guess heuristic works best outdoors; symmetric indoor scenes such as corridors risk many plausible matches (Sec. 3.2)
- No outdoor ground truth; the aerial-image comparison is described by the authors as sketchy and a proper SLAM performance metric as missing (Sec. 5.6)
- Poses are corrected only at scan poses, leaving gaps in the trajectory that must be patched afterwards (Sec. 5.4, Fig. 18)
- No covariance from scan matching; explicit uncertainty representation left to future work (Sec. 6)
- (inference) Stop-scan-go acquisition (3.4 s for a 181x256 scan, Sec. 5.1) is incompatible with continuous handheld or walking capture typical of current construction mobile mapping. Bosse and Zlot 2009 (Sec. I, full text) cite stop-and-scan systems as refs. [3]-[5], but the reference list is not displayed on IEEE Xplore, so whether this paper is among them is unverified; Zlot and Bosse 2014 (Sec. 1) list the group's earlier mine-mapping system (Nüchter et al. 2004) among stop-and-scan solutions.
營建工程相關證據
作者在引言列舉建築（architecture）、隧道施工與維護、工廠設計、設施管理、都市與區域規劃等可能應用，並提到同組先前以 6D SLAM 測繪廢棄地下礦坑的工作（Sec. 1，Nüchter et al. 2004）。本文實驗則在 Schloss Dagstuhl 會議中心（84 幅掃描、240 m 閉合迴圈，Sec. 5.2）、Birlinghoven 研究室建築室內外（32 幅掃描，含 1.05 m 高差坡道，Sec. 5.3）、Birlinghoven 園區戶外（77 幅掃描，Sec. 5.4）與 RoboCup Rescue 競賽場地（Sec. 5.5）進行，另提及 ELROB 試驗；均不是營建工地、隧道或基礎設施的驗證。精度只以未校正航照的距離比例（Table II）與一段捲尺量測（Fig. 13）比對，作者自稱為粗略（sketchy）比較（Sec. 5.6），參考資料偏弱。
原文驗證環境：已完工建築、受控實驗、跨場域、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 5 個比較組，合計 10 筆紀錄。以下列出本方法紀錄最多的 4 組，其餘 1 組列在最後，並連到性能比較頁。
Nüchter et al., 2007 · Table II 本方法 4 筆
指標deviation between length ratio in aerial view and in point model
表格設定（擷取紀錄原文）：Length ratios measured in an uncalibrated aerial photo compared with ratios in the final 77-scan point model of the Schloss Birlinghoven campus (Nüchter et al., 2007, Table II)
deviation between length ratio in aerial view and in point model，author-collected Kurt3D data (Schloss Birlinghoven campus) · AB/BC (aerial 0.683, model 0.662)
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 Nüchter et al., 2007 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。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| 6D SLAM point model (ICP, loop closing, global relaxation)本方法原文提出 | 3.1% | (Nüchter et al., 2007, Table II) |
Nüchter et al., 2007 · Text Sec. 5.3 本方法 2 筆
資料集與序列author-collected Kurt3D data (Birlinghoven robotic lab) · 32 scans
表格設定（擷取紀錄原文）：Error tolerance of the initial estimate for successful ICP registration in the Birlinghoven data set (Nüchter et al., 2007, Text Sec. 5.3)
tolerated initial position error in x, y, z (about)，author-collected Kurt3D data (Birlinghoven robotic lab) · 32 scans
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 Nüchter et al., 2007 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。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| ICP scan matching本方法原文提出 | 1 m | (Nüchter et al., 2007, Sec. 5.3) |
Nüchter et al., 2007 · Text Sec. 5.4 本方法 2 筆
資料集與序列author-collected Kurt3D data (Schloss Birlinghoven campus) · 77 scans
表格設定（擷取紀錄原文）：Processing times for the 77-scan campus data set (Nüchter et al., 2007, Text Sec. 5.4)
time for scan registration and closed loop detection (total)，author-collected Kurt3D data (Schloss Birlinghoven campus) · 77 scans
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 Nüchter et al., 2007 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。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| 6D SLAM本方法原文提出硬體：Pentium-IV-2800 MHz | 10 min | (Nüchter et al., 2007, Sec. 5.4) |
Nüchter et al., 2007 · Text Fig. 13 caption 本方法 1 筆
指標distance d measured in the point cloud model (meter rule: 2080 cm)
資料集與序列author-collected Kurt3D data (Birlinghoven robotic lab, 32 scans) · closed loop distance d
表格設定（擷取紀錄原文）：Closed-loop span measured in the registered point cloud versus meter rule (2080 cm) (Nüchter et al., 2007, Text Fig. 13 caption)
distance d measured in the point cloud model (meter rule: 2080 cm)，author-collected Kurt3D data (Birlinghoven robotic lab, 32 scans) · closed loop distance d
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 Nüchter et al., 2007 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。
| 方法（原文寫法） | 報告值 | 出處 |
|---|---|---|
| 6D SLAM point model本方法原文提出 | 2096 cm | (Nüchter et al., 2007, Fig. 13 caption) |
其他比較組
列出其餘 1 個比較組
來源
Nüchter et al., 2007
（2007）6D SLAM—3D mapping outdoor environmentsJournal of Field Robotics, 24(8-9): 699-722
同儕審查已出版已讀全文經典查證後修正
相關版本
- 程式碼釋出：3DTK - The 3D Toolkit (slam6d); the 3DTK publications page lists this paper among those the toolkit implements; the paper itself does not cite a code URL https://slam6d.sourceforge.io/
程式碼：https://slam6d.sourceforge.io/（授權：GPL-3.0 (3DTK LICENSING file and package.xml in github.com/JMUWRobotics/3DTK, main branch; bundled third-party libraries keep their own licences; the file also places 3DTK-generated images and videos under CC BY-SA 3.0 and asks that the software not be used in military contexts). Applies to the current toolkit, not to a frozen release matching the paper.）。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。