Radar Obstacle Detection Based on Pre-Built Maps in Rail Transit Scenarios

Long distances generally pose challenges in judging whether obstacles encroach upon the clearance in rail transit scenarios. This paper proposes a radar obstacle detection system based on pre-built maps to overcome this issue. The system incorporates a set of customized high-resolution, long-range millimeter-wave radars adaptable to rail transit scenarios as perception sensors and pre-built maps to filter out obstacles. The map building process involves two steps. The first is to construct point cloud maps based on a normal distributions transform (NDT) algorithm with vehicle-mounted LiDAR; the second is to construct semantic maps based on the point cloud maps to delineate the track area boundaries. Localization is achieved for moving trains through real-time matching between the laser point clouds and point cloud maps. This system also enables real-time heading angle corrections and obstacle filtering based on positioning information and semantic map data. On-board experiments conducted in a rail transit environment demonstrated the system's capabilities in stably detecting obstacles at distance above 800 m and building maps to cover the entire test line. Furthermore, this system effectively filtered out irrelevant obstacles located outside the track area boundaries based on the maps, thereby ensuring accurate warning performance for obstacles within the clearance.