MSSNet: LiDAR-Camera Extrinsic Calibration Based on Multi-Modal Semantic Similarity

LiDAR-camera extrinsic calibration plays a crucial role in multi-modal fusion perception and autonomous driving. Existing methods relying on specific calibration targets are time-consuming and unsuitable for online calibration. A LiDAR-camera extrinsic calibration method utilizing multi-modal semantic similarity (MSSNet) from point clouds and images is proposed in this paper. In particular, correlation volume which is updated with an optical flow estimate module is used to compute the similarity between images and point clouds using the multi-modal semantic information as input. Next, the final updated correlation volume is used to regress the calibration parameters. We present qualitative and quantitative analysis of experimental results on the KITTI odometry dataset, demonstrating the effectiveness of our method for LiDAR-camera extrinsic calibration. Our code will be available at: https://github.com/BUCT-IUSRC/Research__MSSNet.