Research On Urban Rail Transit Congestion Detection Algorithm Based On Convolutional Neural Network

In recent years,urban rail transit,as an important part of public transportation system,has been developing vigorously,and the continuous growth of the number of passengers has put forward new requirements for urban rail transit traffic organization and passenger transportation organization.Rail transit stations and train cars are very likely to become a high-density gathering place for passengers,which can easily lead to large-scale mass emergencies and security incidents,causing adverse social impact.Therefore,a convenient and efficient way to obtain passenger congestion status in real time is needed to provide strong technical support for passenger transportation organization.The current congestion detection techniques mainly rely on crowd counting algorithms combined with certain crowding classification rules to determine the crowding level of passengers.After considering the deficiencies of existing crowd counting algorithms and the actual needs of video detection methods,this article put forward three key issues: lack of datasets for in-car passenger counts in the field of passenger flow analysis;scale variation of passenger characteristics;quality problem of passenger distribution density map.On this basis,this paper investigated four aspects: the creation of the crowd counting dataset;the structural design of the convolutional neural network;the feature map recovery process;the application of the algorithm.The work in this paper can be summarized as follows.(1)The dataset is irreplaceable for the training of convolutional neural networks,and the existing crowd counting dataset cannot satisfy the congestion detection of passengers in subway cars.Therefore,this paper collected the on-board surveillance videos of Zhengzhou Metro Line 2 and used the manual labeling method to produce a dataset containing a large number of samples,which is named Zhengzhou＿MT.(2)To address the difficulty of multi-scale feature extraction in the field of passenger congestion detection,this paper combined the characteristics of multi-column neural network and dilated convolution,and proposed a multi-column atrous convolution neural network for metro passenger counting(MPCNet).By applying multi-column dilated convolution at the back end of the deep network to extract features at different scales of the population,the impact of scale variation on the counting performance of the network can be reduced and better counting results can be achieved.Finally,the robustness of the proposed algorithm in the face of multi-scale features was demonstrated by qualitative and quantitative experiments.(3)In order to solve the problem of poor quality of population distribution density maps generated by existing models,this paper proposed a dilated-transposed fully convolutional neural network(DT-CNN).The feature representation capability of the model is enhanced by using the feature learning process to introduce potential information from the original image into the feature map,thus creating a clear feature mapping.Compared with several excellent crowd counting algorithms on publicly available datasets,the proposed method in this paper generates higher quality density maps.(4)In order to test the real-time performance and effectiveness of the proposed algorithm in the actual detection process,this work has built a fully functional automatic detection system for congestion.The system used PyQt5 for visual operation interface development,visualizing the indicators of the test results,while enabling batch calculations and obtaining historical data in the form of line graphs.The effectiveness,real-time and practicality of the proposed algorithm and the designed system are proved by testing on offline in-vehicle surveillance video.