PL-ICP and CSM (laser_scan_matcher)
PL-ICP 是採用點到線(point-to-line)度量的 2D ICP 變體:參考掃描以相鄰點連成折線,目前掃描的每個點對應到最近兩點形成的線段,並以作者推導的精確閉式解最小化點到線距離。作者引用 Pottmann 等的結果說明,點到線度量在零殘差且初值良好時具二次收斂;並證明當參考曲面為折線時,演算法會在有限步內收斂到固定點或循環。附錄另提出利用射線角度排序、提前停止與跳躍表的快速對應搜尋。作者的 CSM 函式庫就是 ROS laser_scan_matcher 增量式 2D 雷射里程計所用的匹配核心。
本頁內容
2D ICP variant with a point-to-line metric solved in exact closed form each iteration, giving quadratic convergence near the solution and finite-step termination, with a fast correspondence search; implemented in the CSM library used by ROS laser_scan_matcher.
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 (SICK, 360 rays over 180 deg in the test log) |
|---|---|
| 原文測試平台 | wheeled UGV (robotic wheelchair; log from Minguez et al.) |
| 狀態估計 | iterative point-to-segment matching; each iteration minimizes the point-to-line error in exact closed form via Lagrange multipliers (a fourth-order polynomial in lambda); trimming rejects outliers (Sec. II; App. I) |
| 資料關聯 | each transformed point is matched to the segment between its two closest points in the reference scan, which is treated as a polyline; fast search exploits radial ordering, early stopping and precomputed jump tables (Sec. II; App. II) |
| 時間表示 | 不適用 (pairwise scan matching) |
| 去畸變 | 不適用 |
| 迴圈閉合 | 不適用 |
| 全域最佳化 | 不適用 |
| 地圖表示 | 不適用 (reference scan as polyline) |
| 先驗資訊 | initial guess required (odometry); less robust to large rotational initial errors |
| 可輸出幾何 | 2D rigid transform (t, theta) between two scans |
| 計算需求 | Pentium IV 1.8 GHz: on average 7.2 iterations and 0.0018 s (539 Hz) per matching for 360-ray scans (Sec. V.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 |
|---|---|---|---|---|---|
| LiDAR | Sick range-sensor | 資料集感測器 | Minguez et al. (2006) scan log | 360 rays over a 180 deg field of view | (Censi, 2008, Sec. V) |
| 載具平台 | robotic wheel-chair | 資料集感測器 | Minguez et al. (2006) scan log | scans taken about every 0.3 m with considerable odometry slip | (Censi, 2008, Sec. V) |
| 運算硬體 | Pentium IV 1.8GhZ | 執行運算平台 | 未標示 | 原文未報告 | (Censi, 2008, Sec. I; Sec. V.B) |
作者報告的優勢與限制
優勢
- In the reproduced artificial-error experiment PLICP put 99.85% of trials in the < 0.001 bucket for Experiment 1, versus 81.27% for MBICP, 83.31% for IDC and 57.78% for ICP (Fig. 3).
- On average 7.2 iterations and 0.0018 s per matching, versus 31.2 iterations and 0.076 s for MBICP (Sec. V.B).
- GPM coarse alignment followed by PLICP gave the best accuracy for 99.79% of trials in Experiment 6 (Sec. V.A).
- Needs fewer parameters than vanilla ICP: no convergence thresholds and no search bounds (Sec. VI).
限制
- Less robust to large rotational initial errors: in Experiment 6 (up to 45 deg) 24.81% of trials ended with errors above 0.05 (Fig. 3; Sec. V.A).
- In the test each scan is matched against itself, which the author calls unrealistic because real scans overlap only partially (Sec. V.A).
- Absolute timing comparisons depend heavily on the implementation (Sec. V.B).
- The theoretical convergence results hold for idealized algorithms without outlier rejection (Sec. III).
營建工程相關證據
未在營建場域驗證;資料為機器人輪椅搭載 SICK 的室內記錄。它是 2D 掃描匹配元件,透過 ROS laser_scan_matcher 可在缺少輪式里程計的平台上提供增量式雷射里程計;在長走廊等沿牆方向缺少約束的環境中,點到線度量對沿牆平移的約束很弱,這項風險論文未評估(推論)。
原文驗證環境: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 3 個比較組,合計 28 筆紀錄。
Censi, 2008 · Fig. 3 table 本方法 24 筆
表格設定(擷取紀錄原文):Minguez et al. (2006) artificial-error experiment: each of 778 scans matched against a copy of itself displaced by a uniform random error up to the listed bound, 100 trials per scan; errors bucketed by the maximum absolute component in m and rad; MBICP, IDC and ICP columns copied from Minguez et al. Column order follows the table header (MBICP, IDC, ICP); the text lists them as MBICP, ICP, IDC. Only the '< 0.001' and '> 0.05' buckets are extracted. (Censi, 2008, Fig. 3 table)
% of trials with error < 0.001 (m, rad),Minguez et al. (2006) wheelchair SICK log (778 scans) · Experiment 1 (0.05 m, 0.05 m, 2 deg)
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。
按 Tab 進入圖表後,用上下方向鍵逐一瀏覽各類別,Esc 關閉提示框;也可開啟表格檢視閱讀全部數值。
這些是 Censi, 2008 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。
資料來源作者報告值(Censi, 2008, Fig. 3 table)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| MBICP | 81.27% | (Censi, 2008, Fig. 3 (table)) |
| IDC | 83.31% | (Censi, 2008, Fig. 3 (table)) |
| ICP | 57.78% | (Censi, 2008, Fig. 3 (table)) |
| PLICP本方法原文提出 | 99.85% | (Censi, 2008, Fig. 3 (table)) |
| GPM | 1.86% | (Censi, 2008, Fig. 3 (table)) |
| GPM o PLICP本方法原文提出 | 99.98% | (Censi, 2008, Fig. 3 (table)) |
Censi, 2008 · Text App. II.B table 本方法 2 筆
資料集與序列Minguez et al. (2006) wheelchair SICK log · full log
表格設定(擷取紀錄原文):Correspondence search cost on the same log; naive search uses max |t| = 0.5 m and max |theta| = 25 deg (Censi, 2008, Text App. II.B table)
avg. comparisons per ray per iteration,Minguez et al. (2006) wheelchair SICK log · full log
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。
按 Tab 進入圖表後,用上下方向鍵逐一瀏覽各類別,Esc 關閉提示框;也可開啟表格檢視閱讀全部數值。
這些是 Censi, 2008 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。
資料來源作者報告值(Censi, 2008, Text App. II.B table)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| smart correspondence search本方法原文提出硬體:Pentium IV 1.8 GHz | 6 comparisons | (Censi, 2008, App. II.B) |
| naive correspondence search硬體:Pentium IV 1.8 GHz | 124.9 comparisons | (Censi, 2008, App. II.B) |
Censi, 2008 · Text Sec. V.B table 本方法 2 筆
資料集與序列Minguez et al. (2006) wheelchair SICK log · full log
表格設定(擷取紀錄原文):Average iterations and execution time per scan matching on the same log; MBICP, ICP and IDC values copied from Minguez et al.; the author cautions that absolute timing depends heavily on implementation (Censi, 2008, Text Sec. V.B table)
avg. iterations,Minguez et al. (2006) wheelchair SICK log · full log
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。
按 Tab 進入圖表後,用上下方向鍵逐一瀏覽各類別,Esc 關閉提示框;也可開啟表格檢視閱讀全部數值。
這些是 Censi, 2008 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。
資料來源作者報告值(Censi, 2008, Text Sec. V.B table)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| MBICP硬體:Pentium IV 1.8 GHz | 31.2 iterations | (Censi, 2008, Sec. V.B) |
| ICP硬體:Pentium IV 1.8 GHz | 34.7 iterations | (Censi, 2008, Sec. V.B) |
| IDC硬體:Pentium IV 1.8 GHz | 30.4 iterations | (Censi, 2008, Sec. V.B) |
| PLICP本方法原文提出硬體:Pentium IV 1.8 GHz | 7.2 iterations | (Censi, 2008, Sec. V.B) |
來源
Censi, 2008
(2008)An ICP variant using a point-to-line metric2008 IEEE International Conference on Robotics and Automation (ICRA), Pasadena, CA, pp. 19-25
DOI 10.1109/robot.2008.4543181程式碼
同儕審查已出版已讀全文經典
相關版本
- 程式碼釋出:CSM (C scan matcher) stated in the paper at purl.org/censi/2007/csm; GitHub AndreaCensi/csm https://github.com/AndreaCensi/csm
- software using this method:ROS laser_scan_matcher (CCNYRoboticsLab/scan_tools), incremental scan matcher built on CSM https://github.com/CCNYRoboticsLab/scan_tools
- repository record:Caltech authors repository record listed by OpenAlex (not read) https://authors.library.caltech.edu/18274/
程式碼:https://github.com/AndreaCensi/csm(授權:LGPL (csm package.xml); the laser_scan_matcher package.xml states CSM is LGPLv3 and the wrapper BSD)。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。