Traffic Sign Recognition Algorithm Based on Improved YOLOv5

Abstract

Due to the low detection accuracy of the existing YOLOv5s recognition algorithm for traffic sign recognition, it cannot meet the application requirements of actual scenarios. This paper makes some improvements on the basis of the original YOLOv5s model. First, a detection layer is added to the original three detection layers to obtain more location information of the target in the underlying feature map. After that, a convolution extraction module with CBAM attention mechanism is designed in the backbone of the model, so that the network pays more attention to the feature information of traffic signs. Then, in the Neck part of the model, a lightweight neural network C3Ghost is designed to replace the original convolutional network, which reduces the network model parameters, and finally integrates more semantic features by adding a cross-layer connection structure. The experiment uses the traffic sign dataset jointly released by Tsinghua and Tencent, and trains 45 types of traffic signs with a large number. The test results show that the optimized network model reduces parameters, and the detection accuracy is improved by 5.3% compared with the original network, which meets the needs of actual scene detection accuracy.