For driverless cars, the detection and identification of traffic signs is critical to the understanding of the environment around the vehicle. There are many smaller traffic signs in the pictures taken during driving, which are difficult to detect by existing object detection methods.In view of the real-time and accuracy requirements of the traffic sign recognition problem of driverless cars, this paper improves the VGG convolutional neural network, proposes the VGG-8 model, and improves the VGG-16 structure. The network is optimized through SGD and Nesterov Momentum. Training and testing in the GTSRB traffic dataset. By comparing a small-scale four-layer convolutional neural network VGG-6 with a 13-layer convolutional neural network VGG-16, a six-layer convolutional neural network is proposed. VGG-8. Inspecting the ten traffic signs in the video, VGG8 has higher accuracy and running speed, and has certain feasibility in practical applications.