Vehicle Intelligent Application System Based on Improved YOLOv5 and DeepSORT

Nowadays, the coverage of the expressway is constantly expanding, and the traffic volume on each highway is also rapidly expanding. This research suggests a vehicle analysis system based on YOLOv5+DeepSORT to overcome the poor real-time performance of conventional traffic statistics and low accuracy in detection for original algorithms. In this paper, we propose a method named YOLOv5-Car based on YOLOv5 model, which mAP is 1.6% higher than that of the YOLOv5 model. The YOLOv5 model is improved in 2 ways: (1) When the Coordinate Attention (CoordAtt) module is added to the YOLOv5 backbone network, channel attention is divided into two 1-dimensional feature coding processes and aggregated along two spatial directions, respectively. This method can enhance the position and depiction of interesting objects in the input feature map. (2) Replace the Non Maximum Suppression (NMS) based technique with the more accurate and completely parallelizable Confidence Propagation Cluster (CP-Cluster). The outcomes of the experiments show that the strategy suggested in this study can improve the precision of the vehicle analysis system.