Traffic sign is an important guide for vehicles in the process of standardized driving. Traffic sign recognition is an essential and important content in the environmental perception of driverless vehicles. Based on YOLOX-S algorithm, this paper strengthens the features obtained from the feature extraction network by adding CBAM attention mechanism module at the end of the backbone network. Utilizes Focal Loss function to better eliminate the imbalance between positive and negative samples and focuses on samples difficult to classify. Using the GIOU Loss function, the problems of inconsistent optimization and scale sensitivity of the original loss function are solved, and the recognition accuracy of the model is further improved. In this paper, the proposed algorithm is tested based on TT100k data set, and the recognition effects are compared with which of several mainstream algorithms. Experimental results show that under the premise of high FPS, the detection accuracy of most target categories is improved. Compared with the YOLOX-S model, the coco accuracy evaluation index Map_50 of the proposed model increased by 1.9%, Map_50:95 increased by 2.1%, and FPS is 35.6. The effectiveness of the improvement is proved.