MixSORT: Adaptive Hybrid Real-time Tracking System

One of the challenges in multi-object tracking (MOT) is maintaining ID consistency under real-time constraints. In particular, existing trackers need to fuse IoU and appearance embeddings without incurring latency or ID switches in complex environments such as crowded scenes, long occlusions, and reappearances. To mitigate this, we propose MixSORT, a system that integrates the real-time efficiency of simple online and real-time tracking (SORT) with the ID stability of DeepSORT by actively adopting distance correlations for matching sequences. MixSORT consists of a score-aware, two-stage association pipeline in which high-confidence detections are matched by IoU, and the remaining candidates are refined using distance-correlation-based appearance matching. In experiments on the MOT17-11 validation sequence, MixSORT enhances ID stability while improving MOTP, IDF1, IDSW, and long-term tracking metrics compared to DeepSORT. Future research should consider developing approximate distance-correlation computations when the number of target identities increases due to computational intensity.