A GMM based uncertainty model for point clouds registration

The existing methods for the registration of point clouds acquired by laser scanners have some limitations. Firstly, as some samples of surface, a point cloud acquired by the laser scanner, which normally works in a spherical fashion, has very limited density when the surface is far away from the laser scanner and the density varies a lot at different ranges. Current registration methods cannot accurately model the surface uncertainty for such kind of point clouds of limited and large varying density. Secondly, when the point cloud is acquired while the platform is simultaneously moving, the estimation error of the platform motion makes the acquired point cloud distorted. To deal with these problems, in this paper, we propose an uncertainty model based on the Gaussian Mixture Model (GMM) to represent the point cloud. Specifically, we construct the GMM piece-wisely on the underlying surface of point cloud, which will accurately model the surface uncertainty. Also a hierarchical structure is employed to increase the robustness of the registration. Furthermore, by assigning each Gaussian component with a pose, a probabilistic graph can be constructed to tackle the problem of registration when the platform is moving while scanning. In this way the distorted point cloud, caused by the estimation error of the platform's motion, can be corrected by performing graph optimization. Simulation and real world experimental results show that our method leads to better convergence than the state-of-the-art methods due to the accurate modeling of the surface uncertainty and the hierarchical structure, and it also enables us to correct the distorted point clouds.