Research On Gate Transit Control Based On Human Behavior Detection And Recognition

As one of the important interactivate devices in the operation and management of urban rail transit network,the performance of gate transit control is directly related to the travel experience of passengers and the ticketing revenue of the enterprise.With the urban rail transit entering the "network operation" period,urban rail transit is undertaking more and more heavy urban passenger transport work.The existing algorithm based on infrared sensor to identify the passenger behavior in the gate channel has been difficult to meet the detection requirements in complex scenes.How to efficiently extract the more complete posture characteristics of passengers when they pass,and identify passengers’ transit behavior from them,so as to avoid various problems and risks when the passenger flow is large,has important theoretical significance and engineering practical value.Considering that the end-to-end supervised learning algorithm is difficult to meet the application and migration needs of the actual scene,this paper adds the middle link of human skeleton extraction,and uses the deep learning model to detect it,so as to identify the passenger’s transit behavior.The main contents of this paper are as follows:(1)Considering the limitation of computation and storage resources when the model is actually deployed,this paper makes full use of the advantages of deepwise separable convolutional neural network to reduce the parameter scale and improve the calculation speed.The built high-performance real-time multi-person pose estimation and tracking model can stably extract the three-dimensional coordinates of the upper body joint points that are not easy to be blocked from the image,and constitute the output of the body pose flow.The added lightweight attention model named SEBlock realizes the adaptive gating of different feature channels and accelerates the feature learning progress of the model.(2)Considering the diversity of camera perspective in the actual deployment of the model,this paper combines the distance information extracted by the depth sensing camera to complete the 3D reconstruction of the feature space based on the chessboard calibration method.The standardization of attitude feature based on the coordinate system of gate channel can effectively improve the generalization ability of subsequent behavior recognition model.(3)Taking standardized pose flow as input features,this paper builds a human action recognition model with alternating cascades of temporal convolutional neural network and graph convolutional neural network which can extract the passenger’s transit behavior from the temporal and spatial motion information of each joint.In the model,graph convolutional neural network mainly completes the spatial information fusion of human skeleton topology map in non-Euclidean space,while temporal convolutional neural network mainly completes the information tracking of variable length time series.At the same time,the added methods which can solve over fitting problem,effectively guarantee the generalization ability of the model.(4)In order to verify the actual performance of the models,an experimental platform was built,and training and validation were performed on the "MS COCO 2017" dataset and the "NTU RGB+D" dataset,respectively.The results of comparison with foreign classic algorithms show that the human pose estimation and action recognition model constructed in this paper has the advantages of fast operation speed and little space occupation on the basis of maintaining high detection accuracy,which provides theoretical and algorithmic support for the improvement and perfection of the existing urban rail transit gate control.