Abstract: To address the limitations of the DCP (Deep Closest Point) point cloud registration algorithm, including its poor fine-grained feature extraction capability, low computational efficiency, and feature misalignment issues, a deep learning-based point cloud registration network that integrates Mamba and SoftBBS (Soft Best Buddies Similarity) is proposed. Firstly, high-dimensional features are extracted from raw point cloud data using a DGCNN(Dynamic Graph Convolutional Neural Network)and the Mamba network, enhancing local feature extraction and computational efficiency. Secondly, SoftBBS is employed to compute an optimal point-pair similarity matrix, reducing the impact of low-reliability matches on registration results and improving robustness. Finally, a LS(Least Squares Method) method is used to calculate the optimal rigid transformation matrix, enhancing registration accuracy. Experimental results show that, compared to DCP, the proposed registration algorithm improves accuracy by 65. 1% and also outperforms several recently popular deep-learning-based registration networks in robustness.

Key words: point cloud registration, best buddy similarity, state space model, deep learning