Using multicore processors to parallelize 3D point cloud registration with the Coarse Binary Cubes method

Abstract

This paper pursues speeding up 3D point cloud matching, which is crucial for mobile robotics. In previous work, we devised the Coarse Binary Cubes (CBC) method for fast and accurate registration of 3D scenes based on an integer objective function. Instead of point distance calculations, the method optimizes the number of coincident binary cubes between a pair of range images. In this paper, we propose taking advantage of widespread multicore and multithreaded processors to further speed-up CBC by parallel evaluation of prospective solutions in a globalized Nelder-Mead search. A performance analysis on two types of multicore processors is offered for indoor and outdoor scans from a 3D laser rangefinder. The proposed solution achieves a computational time gain close to the number of physical cores.