A Correspondence Selection Method by Updating Compatibility Matrix for Point Cloud Registration

Point cloud registration is an active research topic in the fields of computer vision, remote sensing, and robotics. The descriptor-based registration has been studied for a long time, in which the correspondence selection is an important step. However, the existing correspondence selection methods get either low recall or low precision. This will reduce the success rate (SR) of registration. For this reason, a new correspondence selection method is developed in this article. First, the point pair feature (PPF) constraint, which contains geometric distance and angle constraints, is used to calculate two compatibility scores for any two correspondences. The smaller one is treated as the final compatibility score so as to make full use of the geometric distance and angle constraints. Then, the compatibility scores are used to construct the compatibility matrix. The sum of the elements in a row is the voting score of a correspondence. The correspondences are sorted according to the voting scores and the top-K correspondences are used to update the compatibility matrix. In addition, the annealing technique is used to determine the value of K. The update process makes the correct correspondences rank ahead and incorrect correspondences rank behind. Thus, our correspondence selection method has a very good performance. The experiments implemented on three publicly available datasets well demonstrate that our correspondence selection method has high recall and high precision. Our point cloud registration method also gets high accuracy and SR. The code will be available on https://github.com/taowuyong?tab=repositories after publication