A Point Cloud Registration Method Based on Segmenting Sphere Region Feature Descriptor and Overlapping Region Matching Strategy

Abstract

Accurately registering point clouds is challenging due to three primary reasons: 1) it is difficult for point cloud feature descriptors to handle noise in complex scenes; 2) poorly descriptive features lead to incorrect sets of corresponding points; and 3) non-overlapping regions in the scene can adversely affect registration results. To address these issues, our approach consists of three key contributions. First, we propose a segmenting sphere region (SSR) feature descriptor that comprehensively preserves point cloud spatial coordinate information through the “sphere segmentation-furthest point preservation” operation, enabling robust registration in complex scenarios. Second, we design SSR-Net to improve the descriptiveness of SSR features, generating a soft matching matrix to estimate the correspondence between the improved features. Finally, we design an overlap region estimation module in SSR-Net, which employs attention to find the overlap region, thereby reducing the negative impact of non-overlapping regions in the soft matching matrix on registration results. We conducted comprehensive experiments on the B3R, ModelNet40, KITTI, unmanned aerial vehicle (UAV), and 3DMatch datasets, demonstrating the effectiveness of our proposed method.