Traffic sign detection is an important link in the field of autonomous driving, and given the problems of missed detections, false detections, many model parameters, and common and complex representative real environment conditions, such as poor robustness in foggy days, an improved YOLOv5 micro-target traffic sign recognition algorithm was proposed. Firstly, the dataset was atomized to adapt to the accurate identification in the foggy weather, and the PC3 feature extraction module was constructed by using a lighter partial convolution (PConv), and then the Extended Feature Pyramid Network (EFPN) was proposed in the neck network Finally, Focal-EIOU is introduced to replace CIOU as the loss function to solve the problem of false detection and missed detection of micro targets, and the CBAM attention mechanism is embedded to realize the lightweight model and significantly improves the feature extraction ability of the network model. Compared with the original YOLOv5 algorithm, the improved model is increased by 8.9% and 4.4% respectively on P and mAP0.5, the number of parameters is reduced by 44.4%, and the FPS value on NVIDIA 3080 device is 151.5, which can meet the real-time detection of traffic signs in the real scenes.