Research on Traffic Flow Detection by Incorporating Improved Deep Learning Algorithms under Intelligent Transportation Construction

In this paper, in order to solve the problem of inaccurate vehicle target localization of intelligent vehicle detection algorithms in complex scenarios, the YOLOv5s-ours algorithm is proposed. In the YOLOv5s-ours algorithm, the Attention Mechanism Module and the BiFPN structure are added to the YOLOv5s network architecture, and the Detection head of the YOLOv5 algorithm is replaced by the Decoupled Detection Head. The reliability of the YOLOv5s-ours algorithm is explored using frame-skipping detection and training loss detection methods. Compare the accuracy rate of the algorithm before and after the improvement in scenarios such as dense vehicles, dense vehicles at night, and multi-lane roads. The experiment proves that YOLOv5s_Ours takes less time than YOLOv5s. Furthermore, the YOLOv5-Ours algorithm begins to stabilize at 70 Epochs when the loss starts. The accuracy of the system that relies on YOLOv5s_Ours is 6% greater than that of YOLOv5s in the common scenario. The accuracy of traffic flow detection is more than 95% in the dense vehicle scenario. In the nighttime scenario, the accuracy of traffic flow detection reaches more than 90%. In the five-lane detection results, YOLOv5s_Ours improves the accuracy of the YOLOv5s algorithm by 8.78% compared to the YOLOv5s algorithm.