An Improved YOLOv5 Real-time Traffic Sign Detection Algorithm
Yang Xiao-qiang, Ye Liu · 2023
Focusing on the problems of complex detection network structure, multiple network parameters, and low detection accuracy in the process of traffic sign detection, a lightweight traffic sign detection algorithm MC-YOLOv5 is proposed. Firstly, the backbone feature extraction structure CSPDarkNet-53 in YOLOv5 is replaced by Mobilenetv2, and the depth separable convolution is introduced to reduce the network parameters and computation; Then, the K-means++ clustering algorithm is used to generate a priori anchor frame automatically; Finally, CBAM attention mechanism is introduced into the network to strengthen the representation ability of essential features, reduce feature loss, and improve the network feature extraction ability. The average accuracy of the MC-YOLOv5 algorithm is verified by experiments on TT100K traffic sign dataset [email protected] it reaches 90.0%, the Recall rate reaches 98%, the detection accuracy of MC-YOLOv5 model is greatly improved compared with the original YOLOv5 model, and the model volume is reduced by 67%. The experimental results show that the algorithm can better meet the requirements of accuracy and real-time traffic sign detection, and realize the model's lightweight.