RetinaMOT: rethinking anchor-free YOLOv5 for online multiple object tracking

Abstract

Abstract In recent years, YOLOv5 networks have become a research focus in many fields because they are capable of outperforming state-of-the-art (SOTA) approaches in different computer vision tasks. Nevertheless, there is still room for improvement in YOLOv5 in terms of target tracking. We modified YOLOv5 according to the anchor-free paradigm to be on par with other state-of-the-art tracking paradigms and modified the network backbone to design an efficient module, thus proposing the RetinaYOLO detector, which, after combining state-of-the-art tracking algorithms, achieves state-of-the-art performance: we call it RetinaMOT. To the best of our knowledge, RetinaMOT is the first such approach. The anchor-free paradigm SOTA method for the YOLOv5 architecture and RetinaYOLO outperforms all lightweight YOLO architecture methods on the MS COCO dataset. In this paper, we show the details of the RetinaYOLO backbone, embedding Kalman filtering and the Hungarian algorithm into the network, with one framework used to accomplish two tasks. Our RetinaMOT shows that MOTA metrics reach 74.8, 74.1, and 66.8 on MOT Challenge MOT16, 17, and 20 test datasets, and our method is at the top of the list when compared with state-of-the-art methods.

Keywords

• Multiple object tracking
• Attention mechanism
• YOLOv5
• Scene analysis
• Pooling pyramid