A Target Re-Identification Method Based on Shot Boundary Object Detection for Single Object Tracking

With the advantages of simple model structure and performance-speed balance, single object tracking (SOT) based on a Transformer has become hot topic in current field. However, errors caused by target leaving shot, namely out-of-view, are more likely to occur videos than we imagine. To address this issue, proposed re-identification method for SOT called TRTrack. First, built bipartite matching candidate tracklets neighbor optimized Hopcroft–Karp algorithm, which is used preliminary judging leaves shot. It achieves 76.3% mAO benchmark Generic Object Tracking-10k (GOT-10k). Then, introduced alpha-IoU loss function YOLOv5-DeepSORT detect shot boundary objects attained 38.62% mAP75:95 Microsoft Common Objects Context 2017 (MS COCO 2017). Eventually, designed backtracking identification module TRTrack re-identify target. Experimental results confirmed effectiveness our method, superior most state-of-the-art models.