Research On Detection Algorithm Of Passing Passengers In Subway Turns Based On Deep Learning

With the rapid development of urban rail transit,subway travel has gradually become the first choice for urban public transportation,and the safety hazards and traffic efficiency problems of subway travel cannot be ignored.The running quality and efficiency of subway turnstiles directly affect the safety and experience of passengers,and also bring economic losses to the subway company.At present,the door-type gates widely used in China use infrared on-beam sensors for monitoring and passage recognition.This recognition mode cannot completely accurately identify various passage scenarios,and there is a certain probability of misjudgment,which causes the gates to be trapped.The occurrence of incidents such as people,evading fares,etc.,poses a safety hazard.In order to ensure the safety of passengers and improve the efficiency of travel,aiming at the shortcomings of existing recognition technology,this paper uses visual inspection technology to propose a deep learning-based subway gate passenger detection algorithm.The main research work is as follows:First of all,the data set is the basis of the research on the passenger detection and tracking algorithm in this paper.In order to ensure and improve the adaptability and generalization ability of the detection and tracking algorithm in the actual passing environment of subway turnstiles,a data collection platform was built at Shenzhen Shangsha Metro Station to collect data.According to the analysis of special traffic scenarios such as the occurrence of people caught in the gates,a dataset of passengers passing through the gates of the subway was established.The mainstream target detection and tracking algorithm in the vision field is researched and analyzed,and the evaluation index of the detection and tracking model used in this paper is determined.Secondly,this paper proposes a passenger detection algorithm based on YOLO-V3.Researched and analyzed the mainstream SSD and YOLO real-time detection algorithms,and selected the better performance YOLO-V3 algorithm as the basic detection algorithm of this article through comparative experiments.In order to further improve the detection performance of the network and ensure its detection speed,a feature extraction network that combines a cross-stage local fusion network and Dense Net is proposed,and then a feature extraction network based on the feature pyramid network and path enhancement network is proposed.The low-upward multi-layer feature fusion method enhances the model’s ability to learn features.Aiming at the problem of poor position prediction accuracy of the mean square error loss function,this paper uses the generalized intersection ratio loss to replace the original coordinate mean square error loss to further improve the performance of the model.Experiments show that the passenger detection algorithm based on YOLO-V3 proposed in this paper improves the detection accuracy by5.3% and the detection speed by 16.98% compared with the original algorithm.Thirdly,through the research and analysis of the tracking algorithm based on the twin network,aiming at the problem that the siam RPN needs to manually select the tracking target before tracking,a combination of the passenger detection algorithm and the siam RPN tracking algorithm is proposed,and the detection algorithm is used to automatically obtain the tracking target.Aiming at the lack of feature extraction capabilities of the twin network in siam RPN,a method of fusion residual connection is proposed to improve the tracking performance of the algorithm,and the Hungarian algorithm is introduced for data association to achieve real-time tracking of multiple passengers,and the tracking target update is set through Io U matching.And the tracking template update mechanism improves the effectiveness of tracking.Experiments show that the siam RPN passenger detection algorithm based on the connection of detection and fusion residuals proposed in this paper improves the tracking accuracy and tracking accuracy by 11.4% and 8.2%,respectively,and meets the real-time requirements.Finally,the reliability and feasibility of the passenger detection system for subway turnstiles are tested in the actual operating environment of the subway.The integration and platform construction of the passenger detection system have been completed,and the gate control logic has been designed according to the passage status of the passengers.The experimental results show that compared with the original sensor recognition mode,the deep learning-based method in this paper is more accurate and more reliable in special traffic scenarios,which proves the feasibility of the subway gate passing passenger detection system designed in this paper.