Robust point cloud registration based on semantic iterative closest point algorithm

Point cloud registration is a fundamental problem in computer vision, which is extremely challenging for LiDAR point clouds with a lot of noise, outliers, and poor initial position. To deal with these difficulties, this paper proposes a semantic-based iterative closest point algorithm, which utilizes bidirectional distance and correntropy for robust point cloud registration. Firstly, we propose a semantic-guided correspondence establishment strategy that utilizes semantic information to narrow down the search range of correspondences and improve registration accuracy. Secondly, a bidirectional semantic search point matching strategy is introduced to the algorithm, which increases its error correction ability. Thirdly, the maximum correntropy criterion strategy is used to suppress the noise and outliers to further enhance the algorithm in robustness. Experimental results demonstrate the accuracy and robustness of our algorithm compared with other registration methods.