Point cloud registration, Wasserstein Distance, Pre-alignment, Region matching, Computer vision, Segmentation, Slices, Indirect rotation
Abstract
In the field of computer vision, point cloud registration imposes a significant challenge. This is caused mainly by the complexities involved in matching points and the geometric transformations required. Sharp misalignment between clouds can compromise the effectiveness of registration algorithms. In this work, we propose an approach based on the Wasserstein distance to find similar regions between two point clouds, and to encounter a suitable orientation for maximum congruence between the identified regions. We thus obtain a pre-alignment method that increases the robustness of registration algorithms, specially in scenarios of sharp misalignment. The experiments suggest this approach’s effectiveness, not only in pre-alignment, but also in correspondence searching between distinct clouds. Furthermore, results indicate that it is possible to determine matching regions between two clouds and peform accurate registrations, combining both global and local registration features, opening a promising path of investigations.
