Non-rigid point set registration based on local neighborhood information support

Non-rigid point set registration is a crucial task and an unsolved problem in the field of computer vi-sion. One commonly used method for solving the problem is based on the Gaussian mixture model (GMM). In this method, the point set registration is formalized as a probability density estimation prob-lem. Most GMM-based methods achieve registration by maintaining global and local structures of points. However, the previous methods did not filter the neighborhood information in the local structure, and the quality of local neighborhood information directly affects the accuracy of registration. Therefore, extracting effective local neighborhood information is still a challenge. We propose a novel point set registration method based on GMM by extracting local neighborhood information. The two point sets X and Y are regarded as the centroids of GMM and data points produced by GMM, respectively. Our method computes initial correspondences by comparing the feature descriptors of point sets, and the ini-tial correspondences are updated by considering the neighborhood information. Our method then uses the Expectation-Maximization method to solve the GMM. In the experimental results, the efficiency and advantages of our method relative to the current methods are verified by applying five commonly used datasets.(c) 2022 Elsevier Ltd. All rights reserved.