Research On Subway Inbound Passenger Flow Detection Algorithms Based On Deep Learning

As an important node in urban rail transit,the normal operation of subway station is of great significance to the operation safety and transportation efficiency of the whole subway network.Nowadays,large passenger flow has become the normal state of urban rail transit,accurately grasping the status of inbound passenger flow is the basis of scientific management and control passenger flow.In this paper,the passenger flow into the subway station is detected,the passenger face is taken as the detection target,and an improved lightweight passenger flow detection algorithm and a passenger flow detection algorithm under complex conditions are proposed for different passenger density scenarios.Combined with multi-face tracking algorithm,the key status information of inbound passenger flow can be acquired in real time and accurately.The main work includes the following aspects:(1)Based on the original detection network of Tiny YOL02 algorithm,an improved lightweight face detection algorithm Deep Tiny YOLO2 is proposed by using feature graph fusion and key convolution layer replacement.The two algorithms before and after the improvement were trained and verified on the same data set of WiderFace.The experimental results show that the average detection accuracy and speed of Deep Tiny YOL2O algorithm are improved by 7.3%and 11.9%compared with those before the improvement,and the generalization of face detection is stronger.The final training network model has simple structure and few training parameters,which is suitable for transplanting to embedded devices.(2)In view of the severe face occlusion and multi-scale face and other complex situations in high-density inbound passenger flow,this paper combines the deep residual network and the feature pyramid network to improve the YOLO2 network structure.Combining the dimensional clustering of data set and the improved non maximum suppression algorithm,the improved Deep YOLO2 face detection algorithm is proposed.This algorithm,together with YOLO2,Faster R-CNN and SSD deep learning algorithm,was verified on WiderFace verification set and self-built subway face data set,respectively.Experimental data show that the Deep YOLO2 algorithm improves the detection accuracy to varying degrees,and the detection speed meets the requirements of real-time detection.(3)A fast multi-face tracking model is proposed based on Kalman filter for inter-frame face location prediction and Hungarian matching data association.Combining the time and space information of the video data in the experiment,this paper calculates the real-time inbound flow,passenger flow density and inbound speed of passengers outside the station.Passenger flow statistics experiments were carried out on three video data of different time lengths and different scenes,the average accuracy of statistics was over 90%.