Point Cloud Registration based on Gaussian Mixtures and Pairwise Wasserstein Distances

Point cloud registration has plenty of applications in robotics, e.g., for matching two consecutive LiDAR scans in order to estimate the motion of a mobile robot. In particular for dense point clouds, it can be advantageous to work with accumulative features such as Gaussian distributions. In this context, we propose a novel iterative method that directly aligns two Gaussian mixtures. This is achieved using an efficient approximation of the Gaussian Wasserstein distance, which we find a suitable metric capturing the similarity between shape and position of two components of the mixtures. The method is first analyzed in a simulation study, and afterwards further evaluated on real-world data. We find it to be a promising approach for point cloud registration, which can directly be expanded to LiDAR odometry.