Robust Tracking Based on Improved YOLOv5s and Dynamic Neighborhood Target Association

In order to enhance the video single-target tracking task's ability to cope with complex scenarios such as, rotation, rapid scale change, background interference, and occlusion, this paper proposes a tracking method based on improved YOLO detection and dynamic neighboring region target association. First, the YOLOv5s model is improved by introducing different attention mechanisms such as CBAM, CA, and PA, and retraining to select the optimal model so as to enhance the target detection effect. Second, adaptive tracking search region adjustment is performed within the neighborhood of the tracked target, and an IOU-based similarity measure and filtering mechanism is used to reduce the number of detected target boxes, thus reducing the number of false matches and computation. Finally, the appearance features of the retained target boxes are extracted by a reidentification network, and the final tracking target box is determined by performing cosine matching followed by IOU matching to enhance the tracking stability in complex scenes. The experimental results show that the algorithm of this paper achieves an average success rate of 0.816 and an average accuracy rate of 0.946 on 30 vehicle video sequences, and has good robustness in a variety of complex scenarios.