Vehicle-Infrastructure Cooperative 3D Detection via Point Cloud Filtering

Abstract

Restricted by the vehicle's field of view and location, there are problems such as a lack of global perspective and limited long-distance sensing capability in the process of autonomous driving, so the assistance of road-side information is needed. At present, vehicle-infrastructure cooperation methods mainly suffer from the problems of vehicle-infrastructure sensor heterogeneity, spatial and temporal matching of vehicle-infrastructure sensors, high transmission cost, and significant computation on the vehicle-side. This paper proposes an early fusion with the point cloud filtering method for 3D object detection to reduce the transmission cost and alleviate the computational effort at the vehicle-side, which divides the whole 3D object detection process into three stages: First, a 3D object detection performed at the road-side, and the output prediction boxes are used as the road-side point cloud filter boxes, then the road-side point cloud is filtered according to the filter boxes, and only the point clouds within a specific range of the filter boxes are retained and transmitted to the vehicle-side. Finally, the vehicle-side fuses the point clouds from the road-side for vehicle 3D object detection and outputs the final detection results. The experimental results show that compared with the benchmark of the DAIR-V2X dataset, the transmission cost of this method is reduced by 92.07%. The detection accuracy is improved by 1.37% and 12.88% on average compared with the early fusion and late fusion methods in the benchmark, maintaining high accuracy while significantly reducing the communication load and saving the consumption of computational resources on the vehicle-side.