RadarMOTR: Multi-Object Tracking with Transformers on Range-Doppler Maps

Tracking with radar sensors play a crucial role in the perception of many systems, including autonomous vehicles. They are particularly effective in various environmental conditions that require reliable performance. In this paper, we propose RadarMOTR, an adaptation of MOTRv2 [1], combining both traditional radar-based object detection and Transformer-based association on Range-Doppler data. Additionally, we introduce a new dataset, based on RADDet [2], that is suitable for Multi-Object Tracking on Range-Azimuth-Doppler and Carte-sian Bird’s-Eye-View data. To annotate the data automatically, we used the SORT [3] tracker on Cartesian data. Our model achieves 78.9 % HOTA and a 10.4 p.p. higher association accuracy than a baseline linear Kalman filter on Range-Doppler data.