OKR-Net: Overlapping Keypoints Registration Network for Large-Scale LiDAR Point Clouds

Abstract

Point cloud registration is a fundamental task in various intelligence applications, including simultaneous localization and mapping as well as scene reconstruction. However, in large-scale scenes, the majority of point clouds exhibit partial overlap, posing a significant challenge to the registration process. This study introduces a registration network, named OKR-Net, specifically designed to efficiently align partially overlapping point clouds. The OKR-Net comprises two innovative modules: a joint estimation module adept at identifying the keypoints within the overlapping region; and a coarse-to-fine registration module designed to aggregate the overlap and descriptor information, thereby reducing the outliers and yielding robust corresponding point pairs. In addition, an overlap labeling method for generated keypoints is introduced. The efficiency of the proposed registration network is validated utilizing two large-scale outdoor datasets: KITTI and NuScenes. The results demonstrate that the proposed method outperforms existing global registration methods, encompassing both classical and learning-based methods in real-world scenarios.