Multi-LiDAR Roadside Intelligent Perception Method Fusing High-Definition Map

In the research of vehicle-road collaborative roadside perception, challenges such as low detection efficiency, unstable target trajectories, and inaccurate tracking arise due to the sheer volume of point cloud data and the inevitable obstruction of targets. To tackle these issues, a method of intelligent roadside perception utilizing multi-LiDAR fused with High-Definition (HD) maps is proposed. The goal is to enhance the accuracy and reliability of perception outcomes by incorporating detailed map information. Leveraging the calibration results of multi-LiDAR, the extraction of the region of interest (ROI) within the three-dimensional point cloud is achieved through HD maps, effectively reducing the quantity of point clouds for processing and enhancing computational efficiency. Employing the polar-image Gaussian mixture model (P-GMM) for background modeling, moving targets are swiftly identified using polar-images to circumvent direct processing of extensive LiDAR point clouds, thereby boosting detection efficiency. By enforcing the alignment between vehicle heading and lane line direction, the lane orientation in the HD map is translated into a linear constraint of vehicle state within the Kalman filter framework, thereby enhancing the efficacy of vehicle detection and trajectory tracking. Experimental validation is conducted using simulated crossroads and real-world roads with double T-shaped intersections. Compared to other methods, the method proposed yielded a 55% reduction in data volume, a 12% increase in target detection accuracy, and a 56% decrease in processing time. The improvements in extreme error, mean error, and root mean square error are also achieved in target tracking. The experimental results show that the method proposed can fuse HD map information effectively, achieving rapid detection and tracking of road-moving targets in a wide range of road scenarios.