Research On Lightweight Passenger Flow Statistics Algorithm Based On Video Image Processing

With the continuous increase of passenger flow in public places and the continuous development of big data technology,passenger flow statistics in public places plays an increasingly important role in people flow scheduling,urban design,shopping mall planning,etc.Due to the low cost and convenient installation of cameras in public places,and with the continuous development of in-depth learning in vision in the past two years,many domestic and foreign manufacturers and researchers are committed to using video images as the main solution for passenger flow statistics,which is embodied in the recognition and tracking of pedestrians in the cameras to get the track of each pedestrian Through trajectory analysis to distinguish pedestrian's access behavior.In this paper,the passenger flow statistical system is divided into three main steps: Object detection,Multi-Object Tracking and Trajectory Analysis system.By introducing the lightweight deep learning algorithm into the object detection and target tracking algorithm of the system,the problems of low accuracy and huge calculation amount of the current passenger flow statistical system are solved.At the same time,the polygonal region based trajectory analysis method is designed to solve the problem of narrow scope of application.First of all,for Object Detection,the problems of large computation,insufficient combination of context information and non correspondence of target feature receptive field existing in the current single step target detector structure are analyzed.A new lightweight object detection network,CEFA-YOLO,is proposed by using Shuffle Net V2+ as the backbone network,supported by CEM,Carafe and FAM,which solves the above problems.Comparing this algorithm with other advanced algorithms on the opened dataset VOC.It compare with the widely used lightweight algorithm Tiny-Yolo,it has 15.1 m AP improvement and 90% reduction in computation.Then,for Multi-Object Tracking,a multi-level matching based multi-target tracking method is designed.The main purpose of this paper is to design a lightweight pedestrian recognition network to obtain the apparent features of the target.At the same time,the Kalman filter is used to predict the motion features of the target.Finally,the cascaded measurement based on the superposition of the two features can better express the information of the target object,which makes up for the inaccurate tracking caused by the traditional IOU matching only based on the location information.Through the test on the opened dataset MTO15,a better tracking effect is obtained,and the applicable scenarios of the system are further analyzed.Finally,for Trajectory Analysis System,this paper proposes a trajectory analysis method based on polygonal region.The position angle is used to determine the entrance and exit position of the region,and then the entrance and exit angle of the target trajectory is analyzed to identify the behavior of the target,so as to solve the problem that the traditional trajectory analysis method is applicable to a single scene.In this paper,a lightweight local detection scheme for polygonal region is proposed,which can greatly improve the efficiency of the system without affecting the operation of the original system,and alleviate the low efficiency of the traditional trajectory analysis method.At the same time,a character target information acquisition scheme is added to the system,and a feature discrimination criterion based on the average cosine distance and variance is proposed,which can find the target representative frame quickly and efficiently,and can greatly improve the available value of the system without increasing the system load.Based on the above algorithm,this paper uses the canteen of Nanjing University of Information Science And Technology canteen as the application scenario,tests and analyzes the canteen in different places,different shooting angles and different time periods.The experimental results show that the system has good performance of accuracy in various scenes with different angles and time periods,and it has fast computing speed and low computing consumption,which conforms to the application background of high efficiency and light weight.