Research on automated annotation technology for radar range-angle-doppler data based on visual supervision mechanism

Abstract With the development of autonomous driving technology, vehicles integrate camera, lidar, and radar modules for environment perception. In mass-produced models, cameras and lidars mostly serve as primary sensors, while radar is mainly limited to distance detection and collision prevention. Given radar’s unique working principle, it potentially offers stronger sensing capabilities. However, its signals are hard to process and annotate, and relevant research lacks sufficient dataset support. This study presents a method for processing radar raw signals and proposes an image-supervised automatic annotation approach for radar tensor data in the Range-Angle-Doppler (RAD) format. Meanwhile, we introduce the RADaSet dataset containing synchronized camera, lidar, and radar data, with the radar’s RAD data annotated via the self-annotation method.