CompSLAM
CompSLAM 的 ICUAS 版本以鬆耦合方式，把視覺慣性里程計（ROVIO）或熱影像慣性里程計（ROTIO，使用完整輻射溫度影像）接到 LOAM 式 LiDAR 里程計與建圖：相機里程計在新點雲到達時提供掃描對掃描配準的初始值，並以 J^T J 的特徵值判斷掃描對掃描與掃描對地圖配準是否退化；一旦退化，就改用相機里程計的相對位移延續 LiDAR 位姿，並把當前點雲寫入地圖。相機里程計本身以共變異數成長的 D-optimality 與運動界限做健康檢查。熱影像在黑暗與粉塵中仍能提供約束，是本文與僅用可見光相機融合的主要差別。
本頁內容
Loosely coupled, degeneracy-aware fusion in which visual-inertial or thermal-inertial odometry provides priors to LOAM-style LiDAR odometry and mapping and takes over pose propagation when the LiDAR problem becomes ill-conditioned, with a D-optimality health check on the camera odometry.
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 PuckLITE on the underpass UAV; Ouster OS1-64 in the mine deployment)、IMU (VectorNav VN-100)、visual camera (FLIR Blackfly with shutter-synchronized LEDs) for VIO、LWIR thermal camera (FLIR Tau2, full radiometric imagery) for TIO |
|---|---|
| 原文測試平台 | UAV (quadrotor based on DJI Matrice M100 in the underpass test; aerial robots in the mine and in the SubT Tunnel Circuit) |
| 狀態估計 | loosely coupled, degeneracy-aware: ROVIO-based visual-inertial odometry, or its thermal variant on full radiometric thermal images (ROTIO), supplies the relative motion between successive point clouds as the prior for LOAM scan-to-scan matching; degeneracy of scan-to-scan and scan-to-map matching is detected from the eigenvalues of J^T J, and in degenerate steps the previous LiDAR odometry or mapping estimate is propagated with the camera-odometry relative transform; camera odometry is health-checked by the relative growth of its covariance (D-optimality) and by motion bounds (Sec. III, Eqs. 1-4, Fig. 4) |
| 資料關聯 | LOAM point-to-line and point-to-plane correspondences for scan-to-scan and scan-to-map matching (Sec. III) |
| 時間表示 | discrete per-scan poses; the higher-rate camera odometry is used to compute the relative transform between successive point clouds (Sec. III) |
| 去畸變 | 原文未報告 in the ICUAS paper |
| 迴圈閉合 | none reported in the ICUAS paper |
| 全域最佳化 | none reported in the ICUAS paper |
| 地圖表示 | LOAM point-cloud map; in ill-conditioned scan-to-map steps the current cloud is inserted with the prior mapping estimate plus the camera-odometry relative transform (Sec. III) |
| 先驗資訊 | none |
| 可輸出幾何 | robot pose and LiDAR point-cloud map (Figs. 5-7) |
| 計算需求 | onboard Intel NUC-i7 (NUC7i7BNH) on the underpass UAV, real time fully onboard (Sec. IV-A); no timing figures reported |
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 PuckLITE | 方法輸入 | 未標示 | point clouds at 10 Hz | (Khattak et al., 2020, Sec. IV-A) |
| LiDAR | Ouster OS1-64 | 方法輸入 | 未標示 | 64 beams; used with LOAM in the TRJV mine deployment | (Khattak et al., 2020, Sec. IV-B) |
| 慣性量測單元（IMU） | VectorNav VN-100歸入：VectorNav VN100 | 方法輸入 | 未標示 | inertial measurements at 200 Hz | (Khattak et al., 2020, Sec. IV-A; Sec. IV-B) |
| 相機 | FLIR Blackfly (with shutter-synchronized LEDs) | 方法輸入 | 未標示 | images at 20 Hz | (Khattak et al., 2020, Sec. IV-A) |
| 熱像儀 | FLIR Tau2 | 方法輸入 | 未標示 | full radiometric thermal imagery for ROTIO | (Khattak et al., 2020, Sec. IV-B) |
| 載具平台 | quadrotor based on DJI Matrice M100 | 方法輸入 | 未標示 | aerial robot used in the self-similar underpass experiment | (Khattak et al., 2020, Sec. IV-A) |
| 運算硬體 | Intel NUC-i7 (NUC7i7BNH) | 執行運算平台 | 未標示 | onboard computer running pose estimation in real time | (Khattak et al., 2020, Sec. IV-A) |
作者報告的優勢與限制
優勢
- In a self-similar highway underpass the LiDAR-only map is wrong in size, while the proposed fusion builds a correct map (Sec. IV-A; Fig. 5)
- Autonomous flight of about 410 m in an active underground mine in darkness and heavy airborne dust that returned to its take-off position (Sec. IV-B)
- Used as the onboard localization in an autonomous aerial mission of about 190 m in the DARPA SubT Tunnel Circuit, returning to the take-off spot (Sec. IV-C)
- The 2025 successor preprint reports deployment on all aerial, legged and wheeled robots of Team Cerberus in their competition-winning DARPA SubT final run (arXiv 2505.06483 abstract)
限制
- Evaluation in the ICUAS paper is qualitative; drift is judged only by the autonomous return to the take-off position in the absence of external ground truth (Sec. IV-B, IV-C)
- Thermal imagery becomes scarce in thermally flat scenes with small temperature variations (Sec. I; Sec. II)
- Visible-light cameras degrade in poor illumination, low texture and obscurants; LiDAR degrades in self-similar geometry and in dust or fog (Sec. I)
- The 2025 system version does not filter dynamic objects (arXiv 2505.06483 Sec. V)
營建工程相關證據
ICUAS 論文的場域為高速公路地下道、營運中的地下礦坑與 DARPA SubT 隧道，均無外部真值，屬定性驗證，沒有施工現場點雲精度證據。2025 年系統預印本指出 CompSLAM 曾用於長期營建作業與工業開挖等後續專案，並以 MapEval 及 DARPA 真值點雲評估 SubT 決賽場地的地圖（arXiv 2505.06483 Tables II-IV）；同組織的營建機器人融合研究可見（Nubert et al., 2022a）。熱影像與 LiDAR 互補的設計可供粉塵或無照明的地下工程參考（推論）。
原文驗證環境：地下或隧道、任務層驗證
報告的性能數據
性能數據仍在分批查證，目前尚未收錄此方法的報告值。
來源
Khattak et al., 2020
（2020）Complementary Multi–Modal Sensor Fusion for Resilient Robot Pose Estimation in Subterranean Environments2020 International Conference on Unmanned Aircraft Systems (ICUAS), pp. 1024-1029
DOI 10.1109/icuas48674.2020.9213865程式碼
同儕審查已出版已讀全文近十年查證後修正
相關版本
- successor：CompSLAM: Complementary Hierarchical Multi-Modal Localization and Mapping for Robot Autonomy in Underground Environments (Khattak, Homberger, Bernreiter, Nubert, Andersson, Siegwart, Alexis, Hutter; arXiv 2505.06483v1, 2025) https://arxiv.org/abs/2505.06483
- 程式碼釋出：leggedrobotics/compslam_subt (released with the 2025 preprint) https://github.com/leggedrobotics/compslam_subt
程式碼：https://github.com/leggedrobotics/compslam_subt（授權：BSD-3-Clause (GitHub license metadata; code released in 2025 with the successor preprint, not with the ICUAS 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。