Three-dimensional point clouds of urban areas are easily acquired using a mobile laser scanning (MLS) system. Due to the comprehensive influence of Global Navigation Satellite System (GNSS) positioning error, an inertial measurement unit (IMU) attitude fixing the error, scanner angle measurement and ranging error, and the location accuracy of the MLS point cloud varies greatly in the same direction and in different periods, which need to be registered to complete the correction of data position deviation. However, the difference between MLS point clouds from various sections is a nonrigid change, thereby making follow-up high-precision registration correction processing challenges. The field of view of the vehicle-borne platform scanner is limited, and there is mutual occlusion between ground objects and motion elements involving cars and pedestrians. This will result in the incompleteness of scene data. A terrestrial laser scanning (TLS) system is needed to make up for the deficiency of MLS system data coverage and scene expression by resurveying key areas and data holes. The MLS point cloud is the global coordinate system, and the TLS point cloud is the station’s local coordinate system, both of which must be registered to complete the unification of the coordinate datum. To address those issues, a registration method of road scene laser point cloud based on geographical object features was proposed.

Firstly, the data characteristics of the laser point cloud in the road scene were analyzed, and the MLS point cloud was segmented by combining the elevation error trend and the distribution characteristics of ground objects in the road scene. A continuous small-range segmented point cloud was created from the MLS point cloud with a continuous distribution in the strip. Secondly, in the registration data, the height of low-lying vegetation was determined. The elevation of low-lying plants was used as the threshold value, and pass-through filtering was used to separate the near-surface point cloud from the non-near-surface point cloud. For the near-surface point cloud, curbstone was extracted using the moving window discriminant method, and the road guardrail and isolation belt were extracted interactively. The rod-shaped objects were extracted from non-near-ground point clouds based on the rod-arc feature. To minimize scene complexity, the registration primitives were derived from the artificial geographic entity target characteristics. Then, the multi-scale key points of the registration primitives were retrieved using a combination of eigenvalues and Local Surface Patches (LSP). Finally, under the constraint of key points, 4-points congruent sets (4PCS) and the improved iterative closest point (ICP) algorithm were employed to finish the registration of multi-stage MLS point clouds and the registration of TLS and MLS point clouds.

For multi-stage MLS point cloud, we first created wheel track vector lines along the direction of the lane wheel track and then extracted one road point every one meter along the vector lines to analyze the relationship between the maximum deviation of elevation direction error and road length (Fig. 5). Based on the analysis findings and the features of the road scene, segmentation was performed using suitable thresholds (Table 3) to keep the nonrigid variation degree within a particular range, thereby facilitating data processing and reducing the influence of nonrigid error on registration. For road scene geographic entity objectives with different structures, a multi-scale keypoint detection method combining eigenvalue and LSP was proposed. Multi-scale key point descriptors were constructed via the analysis of octree voxel index, weighted covariance, curvature, and other attributes of different neighborhoods (Fig. 2) and compared with ISS, 3D-SIFT, and other feature point extraction methods. The findings suggest that the strategy used to extract key elements in this study is appropriate for road scenes (Fig. 7, Table 2). To complete the registration of multi-stage road vehicle-mounted laser point cloud and the registration of fixed stations and vehicle-mounted point cloud, a progressive registration mechanism integrating 4PCS and ICP algorithm was examined (Fig. 3). The registration accuracy of the multi-stage vehicle point cloud is within the range of 5 cm, and the maximum registration accuracy of fixed stations and the vehicle-mounted point cloud can reach 4.2 cm.

This paper takes multi-stage and multi-platform urban road point clouds as the research object, analyzes the characteristics of road scene laser point cloud, and proposes a road scene laser point cloud registration method based on the characteristics of geographical entities. The vehicle point cloud was divided on the basis of association between the difference in elevation deviation of multi-stage vehicle point cloud data and road length. To increase the accuracy and efficiency of the registration method, the geographical entity of the road was used as the registration primitive, and key locations with uniform distribution without losing a representative were extracted by integrating the eigenvalue and shape index. The registration of laser point cloud data of road scenes was completed. The multi-stage vehicle point cloud registration accuracy was within 5 cm, and the highest registration accuracy of fixed stations and vehicle-mounted point cloud was up to 4.2 cm.