An Automatic LiDAR-Camera Extrinsic Calibration Method for Sparse Point Clouds Using Boundary Features
Tiancheng Gu, Minqian Wang, Libo Weng, Yanjing Lei, Fei Gao
AI summary
Problem
Existing target-based calibration methods fail under sparse point cloud conditions due to insufficient 3D features, while targetless approaches struggle with initialization and environmental generalization.
Approach
The method automatically extracts complete checkerboard boundaries from images and sparse scans, aligns them using a dimension-reduced global-search and local-refinement strategy, and refines extrinsics via joint optimization of reprojection and normal consistency errors.
Key results
- Automatic complete checkerboard boundary extraction via line-segment clustering
- Dimension-reduced global-search and local-refinement boundary alignment
- Joint optimization of reprojection and normal consistency errors
- Sub-0.015 m translation and 0.3° rotation errors on simulated data, plus 90.9% mIoU on real sparse datasets
Why it matters
Enables reliable, low-cost sensor fusion for autonomous systems using affordable low-resolution or long-range LiDARs that produce sparse data.
Abstract
Extrinsic calibration for LiDAR and camera using sparse point clouds can significantly reduce cost and improve efficiency. However, most current target-based methods are designed for dense point clouds and are less effective in sparse scenarios, while targetless methods primarily rely on environ- mental features. To address this limitation, a LiDAR–camera extrinsic calibration method for sparse point clouds is proposed in this paper. First, the proposed method extracts the complete checkerboard image via line-segment direction clustering and midpoint-to-normal projection. Second, a constructed theo- retical checkerboard boundary point cloud is aligned to the scanned boundary point cloud using a proposed dimension- reduced, global-search and local-refinement (DGL) method. Third, coarse calibration is derived from the centroids of the checkerboard in images and aligned point clouds, followed by refinement through joint optimization of reprojection error and normal consistency error. Finally, experiments on simulated data achieve translation and rotation errors below 0.015 m and 0.3◦, respectively. On a self-collected dataset, the method attains 90.9% mIoU between reprojected checkerboard regions and images, outperforming state-of-the-art methods under sparse point cloud conditions.