An End-to-End Radar Multi-object Tracking Algorithm Based on Detection Transformer and Hybrid Data Association

Radar Multi Target Tracking (MOT) is widely used in applications such as aerial surveillance and autonomous driving perception Unlike visual target tracking, radar targets typically do not Having appearance texture information, but also facing difficulties such as clutter interference, timing cold start, and data correlation, makes tracking performance prone to degradation To this end, an end-to-end multi-target tracking algorithm based on Detection Transformer is proposed, and a hybrid data association mechanism is designed for it This algorithm uses a sliding window in the data preprocessing stage Strategy organization training and inference samples, the model is limited to completing the learning of target motion patterns and initial features within a short time span, thereby alleviating clutter sensitivity and cold start Dynamic issues; In the post-processing stage, the query slot semantic matching is combined with the Hungarian algorithm, and the weighted fusion of Mahalanobis distance motion cost and cosine distance semantic cost is used as the matching Guidelines, target identity assignment can be completed without relying on appearance features Simulation experiments show that the algorithm outperforms current representative methods in multiple evaluation metrics