Analysis According to the Algorithm of Pedestrian Vehicle Target Detection in Hazy Weather Based on Improved YOLOv8

Identification and detection of vehicles and pedestrians is an important element in the development of urban traffic safety, both in normal traffic conditions and in driverless environments. Weather change is an important factor affecting traffic accidents, and bad weather is the main causative factor for traffic accidents. At present, the target detection research in harsh environments at home and abroad have made some progress, but the quality of traffic images obtained by the camera is very susceptible to adverse weather conditions such as haze, still facing the problem of target detection in the case of insufficient light, bad weather, poor visibility, blurring, and so on, so that it is not possible to accurately detect the vehicles and pedestrians on the road, which is very easy to lead to urban road traffic accidents. In view of the above problems, this paper proposes an improved pedestrian and vehicle detection algorithm based on YOLOv8 in hazy weather. First, the CBAM (Convolutional Block Attention Module) attention mechanism is introduced in the Backbone layer to enhance the extraction ability of the model for pedestrian and vehicle features. Secondly, BiFPN (Bidirectional Feature Pyramid Network) is adopted to replace PAN in the fusion layer to strengthen the feature fusion capability. Finally, the Wise-IoU loss function is used to replace the original loss function to solve the problem of ambiguous definition of the original loss function in aspect ratio, to improve the accuracy of the bounding box regression and accelerate the convergence of the model. It improves the efficiency of the model without losing the accuracy of the model, which results in better performance of the model in practical applications. The experimental results show that compared with the original YOLOv8s, the improved YOLOv8 algorithm achieves a mAP value of 64.63% on the dataset, which is an improvement of 2.12 percentage points, and the detection model based on the improved YOLOv8s is validated.