Automatic Extrinsic Calibration of Low-FOV LiDAR and Camera for Urban Rail Applications

In urban rail transit, integrating non-repetitive scanning solid-state LiDAR with cameras significantly enhances situational awareness for tasks like autonomous driving, train positioning, and obstacle detection. However, the limited field of view (FOV) of LiDAR results in sparse feature distributions, creating substantial calibration challenges. To address these, we propose an end-to-end extrinsic calibration method for RGB cameras and low-FOV LiDAR, specifically designed for urban rail environments. Our approach integrates a feature extraction module utilizing a pre-trained ResNet-18, multi-feature fusion of depth, intensity, and density to enrich feature representations, and a stereo matching module using multi-peak disparity matching and GRU-based iterative refinement. These components collectively improve alignment accuracy and calibration robustness under challenging conditions. Validation on urban rail and KITTI-odometry datasets shows our method achieves translation and rotation errors of 1.89 cm and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0.36}^{\circ}$</tex>, respectively, outperforming state-of-the-art methods and demonstrating its effectiveness for real-world urban rail applications.