A cosine similarity-based maximal clique point cloud registration algorithm

When dynamically collecting point cloud data with sensors for map reconstruction or autonomous navigation, complex terrains often cause excessive sensor rotation angles, leading to point clouds containing a large number of outliers and substantial initial angular differences. This results in poor registration performance, making it difficult to meet practical application requirements. To address this issue, this paper proposes a maximal clique point cloud registration algorithm based on cosine similarity (CS-MAC). First, cosine similarity is used to construct a compatibility graph that represents the similarity relationships between the initial feature point pairs. Second, the igraph algorithm is employed to search for and filter maximal cliques with strong similarity correlations in the graph. Third, singular value decomposition is applied to estimate transformation hypotheses for each maximal clique, and the best hypothesis is selected. Finally, the optimal hypothesis is used to perform point cloud registration. Experimental results on the 3DMatch and KITTI public datasets, as well as a self-constructed dataset, demonstrate that under point cloud registration scenarios with large initial angular differences, the registration accuracy of many classical point cloud registration algorithms significantly decreases. In contrast, the proposed CS-MAC algorithm maintains high registration performance and outperforms other tested registration algorithms. Furthermore, the proposed algorithm exhibits excellent registration performance across three different registration scenarios, demonstrating strong generalization capabilities.