Research On Implementation Of Subway Passenger Flow Detection System Based On Deep Network

With the rapid development of urbanization in China,more and more people flow to the city,bringing strong pressure on urban traffic.In order to better alleviate traffic pressure,strengthening urban rail transit construction has become a major attempt to meet the development needs of the new era.As a new type of traffic that is more popular nowadays,the subway is being chosen by more and more people because of its punctuality and high security.In recent years,various provinces and cities have continuously promoted the construction of rail transit,and the subway network has been further expanded.The rapid increase in passenger traffic has brought enormous challenges to the daily work of subway stations.At present,China's urban rail transit passenger flow control is mainly artificial?Systematization,precision,and intelligence are seriously inadequate.The foreign technical equipment has defects such as poor applicability,high installation conditions,and high equipment costs.Existing related theories cannot meet the actual needs of large-scale subway network operations.Recently,video-based passenger flow analysis technology has developed rapidly,but most of the existing passenger flow detection relies on traditional image processing technology,based on manually selected features.Then appropriate classifiers are used to classify feature vectors.In practical applications,a subway station often has dozens to dozens of cameras installed,and the traditional image processing method cannot meet the requirements of real-time detection.This paper first introduces the traditional feature extraction method with good pedestrian target detection in the field of computer vision,Histogram of Oriented Gradients,HOG Meanwhile,I introduce the development and differences of the deep network which is two-step.That is,the region proposal is extracted first,then the region proposals are feed into CNN network to extract feature map.Then,according to the actual situation of the subway passenger flow,the head-shoulder parts of the subway passengers are marked,and a total of 47,000 samples are marked,which constitutes a unique subway passenger dataset for the subway.According to this dataset,the K-means algorithm is used to cluster the size of the target in the dataset.According to the clustering center,the YOLO-v3 network parameters are adjusted,so that the trained network model is more suitable for the dataset.The results show that the detection accuracy rate reached 93%.On the server equipped with four 1080Ti graphics cards,the real-time performance of the model multi-channel video processing is tested,and the strategy of detecting the video frame skipping frame is proposed.That is,using the Kalman filter and the Hungarian matching algorithm to track multiple targets in the video,some continuous frames are skipped during the tracking process,and at the same time,there is no large error in the tracking.Experiments show that the passenger can be counted when two frames are skipped.When the accuracy rate reaches 96%,the detection speed is greatly improved,and the effect of real-time detection of 20 videos is finally realized on the server.According to the trained model,a set of subway passenger flow detection system based on C/S architecture is developed.The system adopts sub-module design mode,which is divided into video decoding module,passenger detection and tracking module,passenger parameter calculation module and video forwarding.The module finally realizes real-time display of the video detection result and parameter calculation result of the server through the client.