Robust Point Cloud Registration Using Geometric Spatial Refinement

Abstract

Point cloud registration is a process used in computer vision and robotics for aligning different partial scans via rigid transform prediction. Recently, several existing traditional and learning-based point cloud registration methods have demonstrated significant progress. However, certain methods cannot handle outlier correspondence well and are highly dependent on the existence of a sufficient overlap between the input point clouds. In practical applications, the overlap between partial scans is typically very small, limiting the applicability of such methods. To overcome this problem, we propose the geometric spatial refinement network (GSRNet) for partial registration. In our method, we implement an attention module based on the geometric spatial feature differences for global feature refinement. This module uses low-dimensional robust geometric descriptors to refine high-dimensional features and then passes the obtained refined features through an overlapping prediction module to generate accurate overlapping point correspondences to achieve enhanced registration performance. Comparisons to the experimental results obtained by existing methods reveal that GSRNet achieves superior performance. In particular, excellent performance is obtained for low overlap rates.