A Geometric Information Extraction Method of Road Signs in LiDAR Point Cloud Based on RPCA

The extraction of geometric parameters of road signs, such as position and sizes, is a crucial aspect of transportation asset management and autonomous driving applications. In vehicular LiDAR point clouds, road signs occupy a small proportion, and are subject to significant interference from surrounding trees, resulting in blurred edges and noise. To accurately extracting the geometric information of road signs, a two-stage pole-like object point cloud segmentation method is proposed. Subsequently, robust principal component analysis (RPCA) is employed to eliminate noise and extraneous points around the signs. The components of independent central poles and sign planes are obtained through the shape analysis of point cloud clusters. Finally, introduce the annular region growth to fit the central poles, and employ normal vector projective sampling and oriented bounding box (OBB) to approxi-mate the signs. Thus, accurate geometric information is obtained for both the central pole and the sign. Experiments are conducted using laser point cloud from 34 different intersections in the Hongshan, Gaoxin, and Wuchang dis-tricts of Wuhan, China. Training and validation using the KPConv segmentation network achieves an accuracy of 90.31%, a precision of 91.07%, and 92.74% recall rate. Additionally, the extraction of geometric information is con-ducted on 98 road signs from 20 intersections within the data above. This method achieves an effective extraction rate of 89.80%, a positional accuracy of 0.062 1 m, and 8.07% geometric error. The experiments demonstrate that this method effectively eliminates noise and extraneous point interference, and performs well on those signs with missing point clouds within 20%.