Embedded Station Passenger Counting System Based On Convolutional Neural Network

The station passenger counting system is the application of crowd courting technology in the field of public transportation,it can use video surveillance to realize passenger counting,the estimation of crowd density and the early warning of crowding risk,which is of great significance to improve the intelligence level of public transport system,improve the quality of urban public transport service,and build intelligent transportation system.Due to deployment cost,portability and other considerations,the station passenger counting system generally requires the use of low-cost embedded systems.However,The limited computing resources of embedded platform,complex scenes and other factors make the implementation of the passenger counting system based on surveillance video still a challenging task.This paper carefully analyzes the difficulties of crowd counting based on embedded applications,researches related literature and selects a suitable embedded computing platform,and then we propose a crowd counting algorithm based on convolutional neural network,and build a corresponding embedded station passenger counting system based on this algorithm.The main work and achievements of the paper are as follows:(1)This paper proposed a deeply-recursive convolutional neural network to predict crowd density map,and the total number of people is obtained by summing density maps.Experimental results have demonstrated that proposed method outperforms most state-of-the-art methods with far less number of parameters and fewer computations,but the number of parameters in our model is more than 2000 times smaller than in state-of-the-art method,solving that most CNN-based methods can not cope with the problem of limited computing resources in practical applications.(2)The paper organized and constructed a dataset for training and testing crowd counting algorithms.The dataset includes 1280 typical station scenes.The crowd distribution in the dataset is characterized by large changes in the number of people,large scale changes and varying degrees of dilution.It is more suitable for evaluating the performance of the station passenger counting system than the general crowd counting datasets.(3)This paper builded a station passenger counting system based on Hi3516D embedded platform.We simplified the crowd counting model based on convolutional neural network with recursive structure and successfully transplanted it to the embedded platform Hi3516D.The system has been successfully deployed to several bus stations and has passed the on-site accuracy test.The average accuracy of the estimated number of people per station reaches more than 85%,and the algorithm processing speed is 2 frames per minute,which meets the actual application requirements.