Traffic sign detection algorithm based on improved YOLOv4

Abstract

Aiming at the needs of Intelligent Traffic System, this paper proposes an improved algorithm based on YOLOv4 to solve the problems of low accuracy, low real-time detection speed and missed detection in small objects of current traffic sign detection. Firstly, the conventional convolution in the convolution block of YOLOv4 neck part is replaced by the depth separable convolution to reduce the parameters and computation of the algorithm. Secondly, the PANet structure is replaced with a new structure named PANet-A which is proposed in this paper, which can better combine low-level semantic information with high-level semantic information to reduce the information loss of small objects and obtain richer feature information. Finally, the feature map obtained from the backbone is embedded into the CBAM attention module, so that the network can effectively suppress useless information while obtaining important feature information, and help the whole model to detect traffic signs more accurately and the computation amount is almost unchanged. The experimental results show that, on the CCTSDB traffic sign dataset, the number of parameters of the improved algorithm is reduced by 33% compared with the original algorithm, the mAP is increased by 1.02%, and the FPS decreases only in a small range, which basically achieves a balance between recognition accuracy and speed. The experimental results prove the effectiveness of the improved network in small object detection and overall detection, which can meet the requirements of traffic sign detection in real scenes.