| Passengers are the main participants of urban rail transit and the main target of urban rail transit operations management.A comprehensive grasp of passenger crowdedness degree is of great significance to the all aspects of urban rail transit operations,such as passenger flow organization,station administration,train scheduling and risk prevention.The surveillance video collected by video surveillance system of the urban rail transit can reflect the situation of the passenger crowdedness degree.However,the information in the surveillance video is difficult to detect and utilize by computer.Based on the existing video surveillance system,a deep learning method is introduced to construct a detection method of passenger crowdedness degree with video in this thesis,which is suitable for urban rail transit scenarios.This research enriches the source of passenger flow information and has important meaning for improving the intelligent level of rail transit operation.In the field of computer vision,the problem of detecting the number of pedestrian targets in a video or image is referred to as a crowd density estimation problem.This paper first analyzes the relevant research results in the field of crowd density estimation and deep learning.And then,a summary of the development history is made and method of crowd density estimation is classified.With the shortcomings of existing research and the actual needs of detection methods into consideration,two key technical issues which are respectively about motion characteristics extraction and deep learning network construction are proposed.On this basis,emotion estimation and convolutional neural network are studied.The basic definition and the process of motion estimation technology are studied.Meanwhile,the origin,model definition,construction method and training method of convolutional neural network are studied in detail.The research results provide a theoretical basis for the study of video detection methods.The research on the method of crowdedness degree detection with video in urban rail transit starts from the process of detection.A three-step detection method including preprocessing module,core detection module and output module is built.The preprocessing module includes two functions of motion feature extraction and detection area selection.The core detection module is based on convolutional neural network and is divided into three stages: feature extraction,feature concatenation and result output.The output module is designed to convert the test results into practical information.During the research process,key parameters such as network structure and convolution kernel configuration were determined.Network training uses a multi-target training method based on stochastic gradient descent.The mean absolute error and the mean square error are selected as the evaluation indicators of the network.The ablation test experiment is used to verify the rationality of the relevant improvement in the core detection study of network.The advanced nature of the detection method was verified by comparative experiments.Finally,the method is used to detect the real surveillance video,which proves the practical value of the method. |
