Traffic Target Detection System Based on Improved YOLOv7

With the intelligent development of the times, traffic scenes are becoming increasingly complex. In order to solve this problem, this paper proposes a traffic target detection system based on an improved YOLOv7 network. Based on the existing YOLOv7 network, it has been improved and applied for target detection in complex traffic scenarios. Firstly, the Wise_IoU loss function has been improved to enhanced Box frame edge information reading capability; without completely replacing the original GIoU loss function. Secondly, the SE attention mechanism is added to enable the image data to focus on important information as it passes through the neural network. This reduces the amount of irrelevant computation and enhances the overall convergence speed. Finally, for fixed region detection, set the region of interest (ROI) to detect vehicles, pedestrians, and other targets moving within that area. The improved YOLOv7 model presented in this paper has made a significant advancement in the performance metrics of target detection in traffic scenes, including PR precision, mAP@0.5 precision, F1 value, Box loss accuracy, and more. Comparison with the original YOLOv7 model shows an improvement in the F1 recognition rate from 89% to 95%, and the overall mAP@0.5 has also increased to 80%.