Design and Implementation of Vehicle Density Detection Method Based on Deep Learning

Abstract

With the development of society, more and more people rely on cars to travel and cause traffic congestion. Vehicle density detection can provide intelligent decision-making assistance for traffic video surveillance system and improve traffic efficiency. The camera view can cause the scale difference between vehicles in the same scene. To solve the multi-scale problem, a vehicle density detection method based on deep learning semantic segmentation is proposed. VGG-16 is used to extract vehicle features at the front end of the network, and the output feature image is 1/8 of the original image, which can improve the accuracy of the prediction density map. A dilated convolution module is designed at the back end to capture multi-scale features of the vehicle, enabling the network to capture more scale details and edge information. Finally, the network cascades the output with $1 \times 1$ convolution to get the prediction density map. And in the output characteristics of the last layer of the network, the quantity is predicted by adding the fully connected layer, and then the final predicted quantity is obtained by summing up the counting results of the network output density map. The performance of the network is tested on the TRANCOS dataset, and the test results show that the MAE of our network is 34.1% higher than that of the Hydra-3s network.