Vehicle Tracking Method Based on Attention-YOLOv5 and Optimized DeepSort Models

Zhuang Li,Xincheng Tian,Yan Liu,Xiaorui Shi

DOI: https://doi.org/10.1109/ddcls55054.2022.9858395

2022-01-01

Abstract:Vehicle tracking plays a vital role in traffic management and autonomous driving. In order to further improve the accuracy of vehicle tracking, reduce the number of ID-Switch and enhance the anti-interference ability to the outside world, we propose a vehicle detection and tracking model based on Attention-YOLOv5 and optimized DeepSort. Firstly, YOLOv5 with attention mechanism was used as the front-end target detector, that is, add CBAM and ECA attention mechanisms to the backbone network of the benchmark YOLOv5 separately to improve the feature extraction ability. Then use the transfer learning idea to train it on the BDD dataset, it is concluded that ECA-YOLOv5 has the highest detection accuracy. Further we optimize the DeepSort algorithm, in order to overcome the disadvantages of the pre-training model not targeting traffic vehicles, we improve the re-recognition network structure and optimize the VeRi re-identification dataset, next perform re-identification pre-training on the Improved-VeRi dataset based on transfer learning. In order to make the vehicle features evenly distributed around the center of the class and improve feature resolution, a loss function ${L_{vehicle\_CL}}$ combining the Softmax loss function and Center loss function is proposed. In the experimental part, collecting some highway traffic videos of different scenes. At first, conduct visual analysis, then perform ablation experiments, and evaluate by using MOTA, MOTP and ID-Switch. Compared with the benchmark algorithm, the tracking performance is significantly improved.