Aiming at the problems of low accuracy and slow detection speed in traffic sign detection in complex scenes, an improved S-YOLO traffic sign algorithm based on YOLOv3 is proposed. Firstly, the batch normalization layer is merged into the convolution layer to improve the forward reasoning speed of the model; Secondly, binary K-means clustering algorithm is used to determine the a priori frame suitable for traffic signs; Then the spatial pyramid pooling module is introduced to extract the depth features of the feature map; Finally, CIoU regression loss function is introduced to improve the detection accuracy of the model. The experimental results show that under the reproduced CTSDB traffic sign dataset, the mAP and FPS of the proposed algorithm are improved by 4.26% and 15.19% respectively compared with YOLOv3. At the same time, compared with YOLOv4 and other algorithms, the proposed algorithm has better accuracy and speed for traffic sign recognition, has good robustness, and meets the efficient real-time detection of complex scenes.