Global Point-To-Hyperplane Icp: Local And Global Pose Estimation By Fusing Color And Depth

RGB-D view registration has been widely studied by the robotics and computer vision community. The well known Iterative Closest Points (ICP) method and its variants prevail for estimating the relative pose between sensors. However, the optimization is performed locally and by consequence it can get trapped in local minima. Global registration methods have been introduced as an approach to solve the local minima problem by exploiting the geometric structure of SE(3), and accelerated with local approaches. In this paper, a local hybrid approach named Point-to-hyperplane ICP has been combined with a global Branch and Bound strategy in order to estimate the 6DOF (degrees of freedom) pose parameters. Registration is performed by considering color and geometry at both, the matching and the error minimization stages. Results in real and synthetic environments demonstrate that the proposed method can improve global registration under challenging conditions such as partial overlapping and noisy datasets.