VMB Module for Low-Overlap Point Cloud Registration

Rigid point cloud registration plays an important role in self-driving, drones, robotics, and many other fields. The sensor limitations and external environment interference in practical applications lead to low overlap in a pair of point clouds which have some problems such as missing points and extraneous points. However, most of the current registration methods rely on complete point clouds to compute the correspondences, showing an unsatisfactory performance in low-overlap point cloud registration. In this letter, we present a voxel prelocalization multiscale relocalization block feature-based (VMB) overlap region extraction module which enables us to solve this problem. In this module, the key points for point cloud segmentation are located using voxel downsampling and relocated based on a weighted sum of multiscale normal angles. Then, the high-overlap point cloud blocks are extracted by matching three-point-combination-based feature of each block. Experiments on low-overlap datasets demonstrate the effectiveness and robustness of our method comparing with the other competing methods.