High-Precision 3D Coarse Registration Using RANSAC and Randomly-Picked Rejections

Abstract

A point cloud registration is an essential process of finding a spatial transformation between two point clouds in computer vision. The Iterative Closest Point (ICP) algorithm is one of the most widely used registration methods. Since the ICP algorithm is a locally optimal registration method, it is not guaranteed to converge to an exact solution because of local-minimum problem. In addition, the ICP algorithm is a time-consuming task. Because the ICP algorithm is performed repeatedly to find the best transformation, it tends to be slow. For those reasons, a coarse registration, which helps point clouds align fast and exactly, is needed before fine alignment. This paper provides a 3D coarse registration method to solve the local-minimum problem in the ICP algorithm. First of all, an initial matching is computed by performing feature extraction using Fast Point Feature Histogram (FPFH) feature which establishes good initial correspondences. Since these correspondences are not accurate yet, we need to reject outlier correspondences. Inlier correspondences are picked out through two rejection methods, RANSAC rejection and Randomly-picked rejection we propose. With these organized correspondences, a transformation matrix between point clouds is obtained. As a result, it is helpful to avoid the local-minimum problem in the ICP algorithm. Moreover, it is quite efficient to register point clouds with noise and large transformations.