An automated multi-constraint joint registration method for mobile LiDAR point cloud in repeated areas

Abstract

Urban digitalization, autonomous driving, and 3D map construction all benefit greatly from the ability of mobile LiDAR systems (MLS) to swiftly capture centimeter-level accuracy point cloud data of urban scenes. However, the positional accuracy of MLS point clouds is affected by factors such as GNSS signal loss and IMU inertial drift, which can lead to ghosting and layering of point clouds in repeated areas, reducing its usability. Point cloud registration which was developed to solve the MLS localization problem has been studied for a long time. Most current registration methods only consider the point clouds globally, so the registration of repeated areas has not received adequate attention; and the few methods that target repeated areas fail to balance registration accuracy and automation. To address the above-mentioned issues, we propose an automatic multi-constraint joint registration method for MLS point clouds in repeated areas. Firstly, based on the strict correspondence between trajectory and point clouds, this method identifies point cloud repeated areas by searching for repeated areas in the trajectory. Secondly, it adaptively segments the point clouds of repeated areas based on changes in trajectory angle and speed. Then, it uses a deep learning network to automatically extract the scene features of these segmented point clouds, which are essentially constrained by a variety of basic features such as points, lines, and surfaces. Finally, these multi-constrained scene features are used to perform joint registration of point clouds in repeated areas. We carried out repeated areas registration experiments in a campus scene using our self-developed MLS. The results show that the best registration combination can produce errors of 0.021 m and 0.033 m, respectively, compared with pre-registration errors of 0.094 m and 0.103 m. Our method specifically considers MLS point clouds in repeated areas and ensures satisfactory accuracy while achieving automatic point cloud registration. It is also expected that the proposed method can be extended to other LiDAR platforms that also have trajectories, such as Airborne Laser Scanning (ALS) and Simultaneous Localization and Mapping (SLAM), to realize high-precision automatic registration of point clouds in repeated areas, with little adaptation to the complex scenes.