Automotive Radar Data Labeling Using Target Identifiers

Labeling of radar sensor point cloud data in automotive driving applications is a complex task. Current approaches involve collecting radar data, along with data from auxiliary cameras and LiDAR sensors, and applying postprocessing algorithms, possibly in combination with human annotations. We consider a system wherein an ego vehicle with an automotive radar collects point cloud data corresponding to targets in its environment. Alongside, object data related to the targets is collected that contains location and target class identifiers. We propose a method for offline labeling of radar point clouds using the target class and noisy locations. We validate the proposed labeling approach using simulated locations in the real-world RadarScenes public dataset as well as experimental datasets, achieving average ${F}_{{1}}$ scores of 0.96 and 0.94, respectively, thereby demonstrating the efficacy of the method.