Automated Labeling of Automotive Radar Azimuth Multipath

Abstract

Automotive radars are the key component in the autonomous vehicle’s sensing suite. Their role is particularly crucial in dense urban environments characterized by multipath propagation conditions induced by reflections from flat surfaces. Multipath propagation phenomena may generate ’ghost’ targets that can degrade radar performance. Deep neural network (DNN) based radar signal processing can address the multipath-induce phenomena. However, it requires the availability of extensive and annotated databases. Publically available automotive radar datasets lack accurately labeled multipath-induced "ghost" targets. Therefore, they are inappropriate for DNN-based radar processing. This work introduces an automated multipath annotation approach to transform conventional datasets into multipath-labeled ones. The derived approach provides detailed "ghost" targets and reflector labels, distinguishing actual targets from reflectors, identifying reflector types, and estimating multipath reflection order. The performance of the proposed labeling approach is evaluated using a manually labeled real-world multipath dataset, demonstrating its effectiveness in annotating multipath radar detections and facilitating DNN-based automotive radar processing in multipath-dominated urban environments.