Vehicle Target Detection Based on Improved YOLOv3 Algorithm in Urban Transportation

In the context of visual detection of vehicles in urban road scenes, the problem of low detection accuracy or missed detection occurs due to vehicle density and the small scale of distant vehicles. This article proposes an improved YoloV3-SPP algorithm with DIOU-NMS Loss as the bounding box loss function to enhance the network's expressive power. To improve the algorithm's ability to extract features of small and occluded targets, this article introduces a dilated convolution module to increase the receptive field of the target. Experimental results show that the improved algorithm proposed in this article improves mAP by 1.79% in vehicle target detection and effectively reduces missed detection in dense vehicle target detection.