Traffic vehicle detection algorithm based on YOLOv3

Vehicle detection based on machine vision is an important part of urban intelligent transportation, and vehicle detection technology combined with deep learning is the mainstream method. In order to overcome the low detection accuracy of traditional YOLOv3 algorithm for small vehicle targets. In this paper, we add a larger convolution layer on the basis of the traditional three convolution layers of YOLOv3, and use k-means++ clustering algorithm to get 12 anchor frames again. In this method, the newly added 104*104 feature scale is more suitable for small target detection than the original feature scale of YOLOv3, and it is easier to get the global optimal anchor point by using k-means++ algorithm, so as to improve the detection accuracy. The final experimental results show that the mean average precision (mAP) of the improved YOLOv3 is 91.01%, which is 2.93% higher than that of the traditional YOLOv3.