CNN-Based Traffic Volume Video Detection Method

Traffic volume detection is the base for traffic management and even smart traffic construction. This paper proposes a method based on convolutional neural networks (CNN). Considering the camera always being fixed during traffic volume detection, a shallow residual neural network (ResNet) model is proposed in this paper, which uses road video data to train model parameters and extract vehicles feature. After training, this paper uses the model to identify the vehicles and a core correlation filter is proposed to track the target. Finally, the traffic volume count method is determined by judging whether the target passes through the region of interest (ROI). Compared with other traffic volume detection methods, this method is more suitable for classifying and counting vehicles in free flow because of its reliability and light weight. The experiment shows that the model has the recognition accuracy of 95.83% and the effective count rate is 88.37%.