Extending tracking with deep learning on radar detections

Abstract

Abstract Recent advancements in machine learning suggest that adopting an end-to-end deep learning approach for multi object tracking in radar applications could be advantageous compared to the current methods based on Bayesian statistics. Utilizing and building upon the Detection Transformers architecture, developed for images, the MultiTarget Tracking Transformer v2 (MT3v2) was developed to handle point object tracking radar data in a model-based environment with promising potential for extended object tracking tasks. This work sets out to test previous proof-of-concept and expand into the domain of reality with the MT3v2 network utilized as a base architecture. A semi-model-free environment was used, with radar data generated from the CARLA traffic simulator to test various modes of tracking. Altering the architecture of the MT3v2 network to enable handling of extended objects, making it compatible with a multi extended object tracking scenario. To evaluate the deep learning approach to see if it is comparable with state of the art tracking method, a generalized optimal subpattern assignment metric is used for extended objects to grade the different trackers. During training and evaluation, the results shows that the MET3v2 network is able to learn from simulated radar data to enhance tracking performance over time. The results also suggest that a model-free approach when working with multiple extended object tracking problems for radar detections could be used to yield improved tracking performance.