Point set registration via rigid transformation consensus

Abstract

Point set registration is a fundamental problem in many domains. This paper proposes a novel pair-wise registration algorithm based on the rigid transformation consensus. It starts by building a point correspondence set, which contains both inliers and outliers. Due to non-overlapping regions, it associates each point correspondence with a latent variable and formulates pair-wise registration as a maximum likelihood estimation problem, which is optimized by the expectation-maximum algorithm. Since all inliers follows the consensus of one similar rigid transformation, each correspondence is assigned a posterior probability to indicate whether it is inlier or outlier. To obtain the desired result, it requires to alternatively implement the establishment of point correspondence and maximum likelihood estimation. Given initial rigid transformation, the proposed algorithm is able to obtain a desired registration result for the pair-wise registration. Experiments tested on public available data sets illustrate its superior performance on accuracy and efficiency over previous algorithms.