3‐DPoint Correspondence and Registration

Abstract

AbstractEstablishing point correspondences between overlapping data sets is a fundamental yet challenging problem in the machine vision and pattern recognition literature. In this article, the reason the traditional closest point criterion has become a standard approach for establishing possible point correspondences between overlapping range images and why three‐dimension point correspondence and registration is challenging are first discussed. Then the classic quaternion method with and without taking the reliabilities of given point correspondences into account for camera motion estimation is outlined. To establish the possible point correspondences between overlapping range images, the improved version, SoftICP, of the SoftAssign algorithm is outlined. To gain an idea of the performance of the SoftICP algorithm, some experimental results based on both synthetic data and real range images have been presented. Although the SoftICP algorithm focuses on pairwise overlapping range image registration, some techniques for efficient point correspondence establishment and multiview range image registration are also outlined. Finally, some concluding remarks are made.