LSReg-Net: An End-to-End Registration Network for Large-Scale LiDAR Point Cloud in Autonomous Driving

Abstract

Large-scale point cloud registration is a fundamental problem for autonomous driving. To achieve alignment, most existing methods focus on local point cloud features for matching. However, these approaches fail to account for the sparse distribution of features in large-scale driving scenarios, limiting their effectiveness. This article introduces an end-to-end registration network, named LSReg-Net, specifically designed to address this challenge, offering a more reliable solution for large-scale LiDAR point cloud. The LSReg-Net performs registration by following sequential modules. First, this study develops a sampling strategy, distance-weighted farthest point sampling (DW-FPS), which effectively enhances sparse feature robustness. Then, a scale attention fusion (SAF) network is proposed to capture both local and geometric features. By considering both spatial sparsity and geometric context of the keypoints, the LSReg-Net obtains accurate correspondences after matching. Finally, a coarse-to-fine registration module is conducted with multilevel correspondences to acquire a precise rigid transformation. Extensive experiments on two large-scale LiDAR datasets demonstrate that the proposed approach outperforms existing registration methods. In addition, real-world experiments based on an experimental platform further validate the high efficiency.