Research on traffic sign detection based on improved YOLOv4

Abstract

Traffic signs contain important traffic information. The traditional traffic sign detection method cannot solve the problem of low detection accuracy caused by the small, occupied area of traffic sign images. Based on this, a traffic sign detection algorithm based on improved YOLOv4 is proposed. Firstly, the 13 × 13 large receptive field detection layer is removed on the YOLOv4 structure, and the 104 × 104 detection layer is added. It obtains more global feature information and improves detection accuracy. The attention mechanism is introduced into the algorithm, that is, the backbone network extracts three feature layers and then adds the scSE module. Make the network focus on the target area and improve the algorithm detection ability. Secondly, in order to speed up the convergence of the network and improve the detection accuracy, a dynamic residual connection is added to the backbone network. It promotes the spread of well-performing signals. And use the decoupled-head detection head to use different branches to calculate classification and positioning tasks. By evaluating the average accuracy of the detection effect of CCTSDB traffic sign data set, mAP reaches 97.68 %, which is 3.78 % higher than YOLOv4. Moreover, network convergence experiments have shown that the improved model converges faster. Compared with other models, the improved model has better detection performance for smaller traffic signs and can better meet the actual needs of high-precision detection.