Automatic Multipath Annotation for Conventional Automotive Radar Datasets

Automotive radars operating in dense urban environments experience a multipath propagation phenomenon, degrading radar performance, and challenging conventional radar processing. The deep learning (DL)-based automotive radar processing that can address multipath-induced challenges requires the availability of large and accurately annotated datasets. Publically available automotive radar datasets are missing accurate labeling of the multipath-induced 'ghost' targets and, therefore, cannot be used for DL-based automotive radar processing. This work proposes the automated multipath annotation approach to convert conventional large databases into multipath-labeled datasets. The proposed approach provides the fine-granularity 'ghost' targets and reflector labels to distinguish between the actual target and the reflector, to identify the reflector type, and to estimate the multipath reflection order. The multipath annotation performance of the proposed approach is evaluated using datasets of radar measurements recorded in typical automotive scenarios and a synthetic dataset of complex multitarget scenarios. It is shown that the proposed annotation approach can effectively annotate multipath radar detections in practical automotive scenarios and thus enables DL-based automotive radar processing in multipath-dominated urban environments.