Traffic sign detection based on improved YOLOv3

Abstract:

Aiming at the problems of traffic sign detection such as large number of small targets, difficult location and low detection accuracy, a traffic sign detection algorithm based on YOLOv3 is proposed. First, the spatial pyramid pooling module is introduced into the network structure to perform the block pooling operation on three prediction feature maps with different scales, and the output of the same dimension is extracted, so as to solve the problem of information loss and scale disunity that may occur in the multi-scale prediction. In order to improve the detection accuracy of small target, the FI module is added to carry out information fusion of the three scale feature maps, and the small target information contained in the shallow large feature map is added to the deep small feature map. According to the characteristics of traffic sign dataset, the improved TIoU based on GIoU is used as the boundary box loss function to replace MSE function, which makes the boundary box regression more accurate. Finally, k-means++ algorithm is used to perform clustering analysis on the TT100K traffic sign dataset, generating new anchors with smaller size. Experimental results show that the proposed algorithm improves mAP by 11.1% compared with the original YOLOv3 algorithm, and the detection time of each image only increases by 6.6ms, which still meets the real-time detection requirements. Compared with other advanced algorithms, the proposed algorithm has better detection accuracy and detection speed.