livox_camera_calib
本法不用棋盤格，而以自然場景中的邊緣特徵對齊 LiDAR 與相機。作者依 LiDAR 量測原理分析：深度不連續邊緣受前景與背景混合影響不可靠，因此改以體素切分與平面擬合取得深度連續邊緣，並分析邊緣分布對校正精度的敏感度。報告在多種室內外場景達到像素級精度。
本頁內容
Targetless LiDAR-camera extrinsic calibration from plane-intersection edges, avoiding unreliable depth-discontinuity edges, reaching pixel-level accuracy.
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 | Livox Avia solid-state LiDAR (non-repetitive scanning, 20 s accumulation) with Intel RealSense D435i camera (main suite)、Ouster OS2-64 spinning LiDAR with MV-CA013-21UC industrial camera (Sec. IV-C; detailed results in supplementary material) |
|---|---|
| 原文測試平台 | static sensor suite fixed in a stable position、spinning LiDAR moved slightly and tracked with LiDAR-inertial odometry to densify scans |
| 狀態估計 | rough calibration by alternating grid search (0.5° rotation, 2 cm translation) maximising the percentage of edge correspondences, then iterative maximum-likelihood point-to-edge reprojection estimation on SE(3) weighting LiDAR range and bearing noise and 1.5-pixel image edge noise; calibration covariance from the inverse Hessian |
| 資料關聯 | depth-continuous LiDAR edges from voxel cutting (e.g., 1 m outdoor, 0.5 m indoor), repeated RANSAC plane fitting and intersection of connected plane pairs at 30° to 150°; points sampled on each edge projected and matched to the κ nearest Canny edge pixels in a 2-D k-d tree, with a direction-orthogonality check |
| 時間表示 | 不適用 (static scenes) |
| 去畸變 | 不適用 |
| 迴圈閉合 | 不適用 |
| 全域最佳化 | none |
| 地圖表示 | 不適用 |
| 先驗資訊 | initial extrinsic (e.g., from CAD) |
| 可輸出幾何 | LiDAR-camera extrinsic for point cloud colorization |
| 計算需求 | offline; the entire pipeline (extraction, matching, rough and fine calibration) takes less than 60 s; hardware not 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 | Livox Avia | 方法輸入 | 未標示 | solid-state, non-repetitive scanning; 20 s accumulation per scene; vertical beam divergence angle 0.28° cited in Sec. IV-B | (Yuan et al., 2021, Fig. 10; Sec. IV, IV-B) |
| LiDAR | Ouster OS2-64 | 方法輸入 | 未標示 | spinning multi-line LiDAR with lower resolution at stationary | (Yuan et al., 2021, Fig. 10; Sec. IV-C) |
| 相機 | Intel RealSense D435i歸入：Intel RealSense D435I | 方法輸入 | 未標示 | written 'Intel Realsense-D435i'; intrinsics and distortion calibrated beforehand | (Yuan et al., 2021, Fig. 10; Sec. IV) |
| 相機 | MV-CA013-21UC | 方法輸入 | 未標示 | industrial camera | (Yuan et al., 2021, Fig. 10; Sec. IV-C) |
作者報告的優勢與限制
優勢
- Accuracy on par with target-based methods in natural scenes; robust to initial values (Sec. I, Sec. V)
- Converged in all 7 scene settings from 20 random initial values each (±5°, ±10 cm) (Sec. IV-A1)
- About 50% of residuals, including mean and median, within one pixel in all 36 cross-validation cases (Sec. IV-A3)
- Accuracy similar to the checkerboard method of Zhou et al. while using one data pair instead of more than 36 board poses (Sec. IV-B)
- Estimated covariances are consistent across scenes (Sec. IV-A2, Fig. 15)
限制
- Designed for high-resolution LiDAR; sparse spinning LiDAR needs scans accumulated by slight motion and LiDAR-inertial odometry (Sec. IV-C)
- Degrades in scenes with only cylindrical objects, unevenly distributed edges or edges in one direction only (Sec. IV-A4, Fig. 17)
- Fine calibration needs a good initial estimate; a rough grid search is used to widen convergence (Sec. III-E)
- Needs about 20 s of stationary Avia data per scene; the 4 s ACSC recordings required three clouds (Sec. IV, IV-B)
營建工程相關證據
原文未報告。作者以稠密點雲建圖、點雲著色與自動化 3D 測量為動機（Sec. I），但未在工地測試。文中量得棋盤格因雷射束發散造成的前景膨脹約 1.4 cm（6 m 距離，約 0.13°）（Sec. IV-B），提示深度不連續邊緣附近的尺寸量測需留意；外參誤差也直接影響著色點雲與影像量測（推論）。
原文驗證環境：受控實驗
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
Yuan et al., 2021
（2021）Pixel-Level Extrinsic Self Calibration of High Resolution LiDAR and Camera in Targetless EnvironmentsIEEE Robotics and Automation Letters, 6(4), pp. 7517-7524
DOI 10.1109/lra.2021.3098923arXiv 2103.01627程式碼
同儕審查已出版已讀全文近十年
相關版本
- 預印本：Pixel-level Extrinsic Self Calibration of High Resolution LiDAR and Camera in Targetless Environments https://arxiv.org/abs/2103.01627
程式碼：https://github.com/hku-mars/livox_camera_calib（授權：GPL-2.0）。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。