The Impact of the Number of k-Means Clusters on 3D Point Cloud Registration

Abstract

Point cloud registration plays a crucial role in many applications, from robotics and autonomous navigation to medical imaging and 3D scene reconstruction. While the Iterative Closest Point (ICP) algorithm is a well-known shape registration choice, its efficiency and accuracy can be affected by the vast search space for point correspondences. k-means clustering emerges as a promising solution for partitioning the search space into smaller clusters to reduce the computational complexity and increase the performance of the matching. However, the number and size of these clusters and how they affect the registration remains a critical and yet not fully explored factor. This paper delves into the relationship between the number of k-means clusters and point cloud registration accuracy. To determine the effect of the number of k-means clusters on registration accuracy and efficiency and to understand any emerging pattern, k-meansICP is developed to use the k-means algorithm to cluster the correspondence search space. Two sets of 3D molecular shapes with differing complexities are matched using initial rotation angles 15, 30, and 60 degrees with 2 to 10 k-means clusters. The results are then compared with the original ICP algorithm.