SelectMOT: Improving Data Association in Multiple Object Tracking via Quality-Aware Bounding Box Selection

Multiple object tracking (MOT) plays an important role in computer vision, powering real-world applications such as autonomous vehicles, robot navigation, and video surveillance. Most existing MOT methods use the tracking-by-detection (TBD) paradigm, where detections are associated into object tracklets using a data association algorithm. However, tracking performance remains limited due to confidence-biased two-stage matching (TSM) strategy, the impact of noisy detections during track updates, and difficulties in recovering lost targets. To address these challenges, we propose SelectMOT, a novel MOT framework that enhances tracking accuracy by selecting bounding boxes based on quality rather than confidence alone. Our two-stage selection matching (TSSM) strategy allows both high- and low-confidence detections to be matched based on spatial cost, refined by a detection selection module (DSM). In addition, we introduce a state selection module (SSM) to suppress noisy updates by dynamically choosing between detector outputs and Kalman predictions, as well as a recovery selection module (RSM) to enhance lost tracklet recovery using strict occlusion-aware constraints. Experiments on the MOT17, MOT20 and DanceTrack benchmarks show that SelectMOT outperforms state-of-the-art methods, achieving 65.5 HOTA, 78.6 IDF1, and 81.2 MOTA on MOT17; 63.7 HOTA, 75.1 IDF1, and 76.9 MOTA on MOT20; and 55.9 HOTA, 57.8 IDF1, and 91.2 MOTA on DanceTrack.