X-ICP
X-ICP 針對 LiDAR 在幾何資訊不足環境(隧道、開放平面、狹窄走廊)中 ICP 沿弱約束方向發散的問題,先利用掃描與地圖的對應,分析各最佳化主方向的對齊強度,細緻判定可定位性(localizability)。再將此分析整合進掃描對地圖的點對平面 ICP,以約束最佳化控制或凍結退化方向的位姿更新。作者以 ANYmal 足式機器人在地下礦坑、營建工地與城市公園實測,並以 Leica RTC360 地面掃描作為礦坑與公園的參考地圖。
本頁內容
X-ICP detects fine-grained LiDAR localizability from scan-to-map correspondences and constrains point-to-plane ICP updates along degenerate directions; tested on a legged robot in a mine, a construction site and a park.
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 LiDAR (Velodyne VLP-16; Ouster OS0-128)、IMU、leg joint encoders (leg odometry prior) |
|---|---|
| 原文測試平台 | legged、simulation |
| 狀態估計 | scan-to-map point-to-plane ICP with localizability-driven constrained optimization (Lagrange multipliers) that fixes or limits updates along degenerate directions (abstract, Sec. VI) |
| 資料關聯 | scan-to-map correspondences analysed against principal optimization directions for fine-grained localizability (abstract) |
| 時間表示 | discrete poses |
| 去畸變 | point cloud motion compensation done at the LiDAR driver level with the leg-odometry pose estimates in the transformation tree; the pose prior is used to transform and undistort the input cloud |
| 迴圈閉合 | none |
| 全域最佳化 | none |
| 地圖表示 | point cloud map in a libpointmatcher-based registration framework (Sec. VII-A) |
| 先驗資訊 | leg odometry prior (IMU + joint encoders) as initial guess (Sec. VII-A) |
| 可輸出幾何 | pose updates; resulting point cloud maps compared with TLS reference maps (Sec. VII) |
| 計算需求 | mapping pipeline runs at 5 Hz (Sec. IV); single-threaded scan-to-map registration on the Rumlang data takes 32.19 (10.7) ms with X-ICP on an Intel i7-9750H laptop, stated to be equivalent to the robot computer, and 12.65 (3.51) ms on an Intel i9-13900K, vs 20.05 (2.21) ms and 11.1 (1.45) ms for the baseline without localizability awareness, whose statistics cover only the first 150 s of the run (Table V); localizability overhead per ICP iteration shown only as a plot (Fig. 20) |
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 | Velodyne VLP-16 | 方法輸入 | 未標示 | 16-beam (sparse) LiDAR on ANYmal-C; used in all field experiments | (Tuna et al., 2024, Sec. VII-A, VII-G3) |
| LiDAR | Ouster OS0-128 | 方法輸入 | Seemuhle mine repeat run | much higher point density and larger FoV than VLP-16; 128-beam (dense) LiDAR | (Tuna et al., 2024, Sec. VII-D2, VII-G3) |
| 地面雷射掃描儀(TLS) | Leica RTC 360歸入:Leica RTC360 | 參考或真值量測 | Seemuhle mine and Opfikon City Park ground-truth maps | 原文未報告 | (Tuna et al., 2024, Sec. VII-D, VII-F, Fig. 6-c) |
| 慣性量測單元(IMU) | inertial measurement unit (IMU) | 方法輸入 | 未標示 | 原文未報告 | (Tuna et al., 2024, Sec. VII-A) |
| 輪式或腿式里程計 | joint encoders | 方法輸入 | 未標示 | used with the IMU by ANYmal's leg odometry module (ref. [69]) to provide the registration prior | (Tuna et al., 2024, Sec. VII-A) |
| 載具平台 | ANYmal-C | 方法輸入 | 未標示 | legged robot | (Tuna et al., 2024, Sec. VII-A, Fig. 6) |
| 運算硬體 | Intel i7-9750H | 執行運算平台 | 未標示 | laptop CPU equivalent to the one on the robot; single-threaded timing | (Tuna et al., 2024, Sec. VII-A, Table V) |
| 運算硬體 | Intel i9-13900K | 執行運算平台 | 未標示 | desktop-class CPU; single-threaded timing | (Tuna et al., 2024, Sec. VII-G3, Table V) |
作者報告的優勢與限制
優勢
- accurate localizability detection and robust pose estimation without environment-specific parameter tuning, with the same kappa_1 = 250, kappa_2 = 180, kappa_3 = 35 for all environments and sensors (abstract, Sec. V-C)
- Seemuhle VLP-16 run: APE translation 2.05 (1.23) m with first-15 m alignment vs 3.36 (1.74) m for Zhang et al. and 5.79 (5.26) m for Hinduja et al., last-position error 0.27 m vs 6.37 m and 24.17 m (Table I)
- RPE per 10 m 0.17 m vs 0.20 m and 0.26 m (Table II)
- partial localizability improves translation over the binary variant Xs-ICP, last-position error 0.27 m vs 5.34 m (Tables III to IV)
- consistent map at the Rumlang construction site where the eigenvalue-threshold baseline performed poorly (Sec. VII-E)
限制
- sensitive to initial-guess quality like related methods, and with a truly bad prior the registration cannot be solved reliably (Sec. VI-B, VIII)
- the filtering parameter kappa_f must be re-adjusted per LiDAR (cos 80 deg for VLP-16, cos 60 deg for OS0-128) (Sec. V-B), and the authors plan to improve the sensor-dependent selection of kappa_1 with point-wise confidence weights (Sec. VIII)
- no ground-truth map or quantitative error is reported for the Rumlang construction-site run, which is judged from map appearance (Sec. VII-E)
- baselines were re-implemented by the authors, and the Zhang et al. eigenvalue threshold was 120 in all experiments except Opfikon, where it was tuned to 200 (Sec. VII-B, VII-F)
- even X-ICP shows about 2 m mean APE over the 521.8 m mine run (Table I
- reviewer observation)
- code release not verified (project website only)
營建工程相關證據
作者實測:瑞士 Rumlang 大型營建工地 153 m 路線(開放平面、原地旋轉),以地圖一致性定性比較;Seemuhle 地下礦坑隧道段以 TLS 參考地圖量化。屬單一研究、單一工地,不可推廣為一般工地性能。
原文驗證環境:模擬、施工中工地、地下或隧道、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 12 個比較組,合計 65 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。
Tuna et al., 2024 · Table III 本方法 18 筆
資料集與序列Seemuhle underground mine (authors' data) · VLP-16 run; first 15 m alignment
表格設定(擷取紀錄原文,這些數值分屬表中不同部分):(Tuna et al., 2024, Table III)
- Ablation, Seemuhle VLP-16 APE (same protocol as Table I); Xs-ICP drops the partial-localizability category and categorizes only in the first ICP iteration
- Same setting as other Table III rows
APE Translation mu(sigma) [m],Seemuhle underground mine (authors' data) · VLP-16 run; first 15 m alignment
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 Tuna et al., 2024 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(Tuna et al., 2024, Table III)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| X-ICP (Proposed)本方法原文提出 | 2.05 m | (Tuna et al., 2024, Table III) |
| Xs-ICP (Proposed)本方法原文提出 | 2.29 m | (Tuna et al., 2024, Table III) |
Tuna et al., 2024 · Table I 本方法 9 筆
資料集與序列Seemuhle underground mine (authors' data) · VLP-16 run; first 15 m alignment
表格設定(擷取紀錄原文,這些數值分屬表中不同部分):(Tuna et al., 2024, Table I)
- Seemuhle underground mine, ANYmal with VLP-16, 521.8 m; APE via EVO against Leica RTC 360 ground truth, mu (sigma); 'first 15 m' = trajectory aligned on the first 15 m (about 200 poses), 'origin' = aligned at the first pose; plus last-position error
- Same setting as other Table I rows
APE Translation mu(sigma) [m],Seemuhle underground mine (authors' data) · VLP-16 run; first 15 m alignment
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 Tuna et al., 2024 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(Tuna et al., 2024, Table I)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| X-ICP (Proposed)本方法原文提出 | 2.05 m | (Tuna et al., 2024, Table I) |
| Zhang et al. [12] | 3.36 m | (Tuna et al., 2024, Table I) |
| Hinduja et al. [17] | 5.79 m | (Tuna et al., 2024, Table I) |
Tuna et al., 2024 · Table IV 本方法 8 筆
資料集與序列Seemuhle underground mine (authors' data) · VLP-16 run
表格設定(擷取紀錄原文,這些數值分屬表中不同部分):(Tuna et al., 2024, Table IV)
- Ablation, RPE per 10 m traversed distance, Seemuhle VLP-16, mu (sigma)
- Same setting as other Table IV rows
RPE Translation mu(sigma) [m] per 10 m,Seemuhle underground mine (authors' data) · VLP-16 run
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 Tuna et al., 2024 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(Tuna et al., 2024, Table IV)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| X-ICP (Proposed)本方法原文提出 | 0.17 m | (Tuna et al., 2024, Table IV) |
| Xs-ICP (Proposed)本方法原文提出 | 0.19 m | (Tuna et al., 2024, Table IV) |
Tuna et al., 2024 · Table V 本方法 8 筆
資料集與序列Rumlang construction site (authors' data) · 153 m run
表格設定(擷取紀錄原文,這些數值分屬表中不同部分):(Tuna et al., 2024, Table V)
- Scan-to-map registration time per scan on the Rumlang construction-site data, single-threaded, mu (sigma); baseline statistics computed only until 150 s into the run to avoid the degenerate part
- Same setting as other Table V rows
scan-to-map registration time mu (sigma) [ms],Rumlang construction site (authors' data) · 153 m run
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 Tuna et al., 2024 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(Tuna et al., 2024, Table V)
| 方法(原文寫法) | 報告值 | 出處 |
|---|---|---|
| X-ICP(Intel i9-13900K, single thread)本方法原文提出 | 12.65 ms | (Tuna et al., 2024, Table V) |
| X-ICP(Intel i7-9750H, single thread)本方法原文提出 | 32.19 ms | (Tuna et al., 2024, Table V) |
| Xs-ICP(Intel i9-13900K, single thread)本方法原文提出 | 11.14 ms | (Tuna et al., 2024, Table V) |
| Xs-ICP(Intel i7-9750H, single thread)本方法原文提出 | 29.42 ms | (Tuna et al., 2024, Table V) |
其他比較組
來源
Tuna et al., 2024
(2024)X-ICP: Localizability-Aware LiDAR Registration for Robust Localization in Extreme EnvironmentsIEEE Transactions on Robotics, 40:452-471
DOI 10.1109/tro.2023.3335691arXiv 2211.16335
同儕審查已出版已讀全文近十年
相關版本
- 預印本:arXiv:2211.16335 (v1 2022-11-29; v4 2024-02-18) https://arxiv.org/abs/2211.16335