Fast non-rigid points registration with cluster correspondences projection

Abstract

In this paper, we propose a fast non-rigid points registration method based on cluster correspondences projection. Firstly, a distance based clustering method is applied on the large scale points clouds. Then, the cluster centers are extract to represent the shape of the template and target object. After that, the target and template data are efficiently registered by an EM-like non-rigid registration method with a clustering projection between the cluster centers and all the points in the clusters. To more improve the computation efficiency, a fast deformation estimation is applied in reproducing kernel Hilbert space. As the size of cluster centers is much less than the size of the original point cloud and the cluster correspondences projection well keeps the points deformed smoothly in the clusters, the proposed method is robust and extremely efficient for large scale points registration with rigid and non-rigid deformations without loss accuracy. Experimental results show that, the efficiency of our approach outperforms on both synthesized and real datasets than the state-of-the-art methods under various types of distortions.