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 regis-tration 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.

OL-Reg: Registration of Image and Sparse LiDAR Point Cloud with Object-Level Dense Correspondences

OLFF-Net: Robust Registration of 3D Point Cloud based on Overlapped Local Feature Fusion

End-to-end 2D-3D Registration between Image and LiDAR Point Cloud for Vehicle Localization

CFI2P: Coarse-to-Fine Cross-Modal Correspondence Learning for Image-to-Point Cloud Registration

DeepICP: An End-to-End Deep Neural Network for 3D Point Cloud Registration

Point cloud registration for LiDAR and photogrammetric data: A critical synthesis and performance analysis on classic and deep learning algorithms

Efficient and Deterministic Search Strategy Based on Residual Projections for Point Cloud Registration with Correspondences

Learning Sufficient Correlations among Points for 3D Non-rigid Point Cloud Registration

L-PR: Exploiting LiDAR Fiducial Marker for Unordered Low Overlap Multiview Point Cloud Registration

HRegNet: A Hierarchical Network for Efficient and Accurate Outdoor LiDAR Point Cloud Registration

An Automated Multi-Constraint Joint Registration Method for Mobile LiDAR Point Cloud in Repeated Areas

HRegNet: A Hierarchical Network for Large-scale Outdoor LiDAR Point Cloud Registration

Global Localization in Large-scale Point Clouds via Roll-pitch-yaw Invariant Place Recognition and Low-overlap Global Registration

Robust Point Cloud Registration Framework Based on Deep Graph Matching

Robust Point Cloud Registration Network for Complex Conditions

Sparse and Low-Overlapping Point Cloud Registration Network for Indoor Building Environments

A new method for two-stage partial-to-partial 3D point cloud registration: multi-level interaction perception

Robust point cloud registration using Hough voting-based correspondence outlier rejection

Multi-Model 3D Registration: Finding Multiple Moving Objects in Cluttered Point Clouds

Point Cloud Registration using Representative Overlapping Points

2D3D-MATR: 2D-3D Matching Transformer for Detection-free Registration between Images and Point Clouds