On Performance Evaluation of Registration Algorithms for 3D Point Clouds

Abstract

3D point Geometric alignment is a challenging task encountered in many scientific applications related to different fields such as robotics and computer vision. For this reason, the well-known 3D registration problem has been extensively studied, and a lot of efficient 3D registration algorithms (RA) exist. Even though many surveys in the literature addressed RA's, none to our knowledge is especially interested in their use in robotic fields and more precisely in dimensional control of mechanical pieces. Our present work involving both a theoretical and an experimental study compares some local and non-rigid RAs, used to align large point clouds representing mechanical pieces. This paper is two-fold and permits first to uncover the similarities and differences between four known RAs which are ICP, NDT, Softassign and RANSAC and then to establish an inter RAs comparative performance evaluation based on accuracy, speed and other new specific metrics we have defined.