Real-time train passenger flow detection algorithm based on convolutional neural network

As the backbone of urban public transportation system, urban rail transit plays an important role in alleviating the pressure of traffic travel and promoting the urban development. Scientific and accurate calculation of real-time train passenger flow is of great significance to improve the passenger service level of urban rail transit and develop a low-carbon and environmental-friendly green travel mode. Therefore, a real-time train passenger flow detection algorithm for urban rail transit based on convolutional neural network was proposed to realize the real-time passenger number detection in the field of the train monitoring video. The train passenger flow dataset was established on the vehicle monitoring video of Chengdu Metro Line 1, with the VGG-16 network excluding the fully connected layer as the basic network framework to extract the shallow detail features, such as the edge and corner point of the input image. The multi-scale convolutional layer and expansion convolutional structure were integrated to build the passenger multi-scale feature perception module. While maintaining the image resolution, the extraction performance of the scale context information was enhanced through different receptive fields, so as to improve the robustness of the network to the passenger scale change. The feature fusion network was constructed to fuse the detailed features extracted by the shallow network with deep semantic features after upsampling to improve the counting accuracy of small-scale passenger target and the information richness of the feature map. The experiments results show that the detection accuracy of the proposed algorithm in the urban rail transit train scenario is significantly improved, and the average absolute error, mean square error and average absolute percentage error indicators reach 1.1%, 1.8% and 5.4% respectively. Meanwhile, the detection rate of a single image is less than 60 ms, which can meet the real-time requirements of train passenger flow detection, and its performance is also improved to varying degrees on the two standard population datasets Shanghai Tech and UCF-CC-50, with good generalization performance. © 2023, Central South University Press. All rights reserved.