Research On Key Problems Of Crowd Counting In Subway Scene

The study of population counting has a long time.At the earliest,people used detection and regression methods to carry out population counting.With the development of deep learning,people also begin to use deep learning methods to get the number of people in a picture.In a subway scenario,implementing crowd counting makes a lot of sense.In this paper,the scene of the subway platform is divided into dense scenes according to the density of the crowd,and the scene of the station hall and the entrance and exit is divided into sparse scene and counted respectively.First,static crowd counting network in dense scene.Dense scenes are counted using GRCNET(self-adaptive far and near range division counting network).The model is divided into two phases,the first stage is the image division stage,the stage regression method was used for each group photo adaptively divided into far close shot pattern,and joined in feature extraction part SPPNet better context information of the fusion image,the second phase is counting stage,the stage respectively using far far close shot of the different network to close shot pictures to count,add sum again,it is concluded that the final number.Because the crowd mobility is not strong on the platform,static image counting can complete this part of the task.Second,dynamic target detection network in sparse scene.In sparse scenes,a single-stage target detection algorithm with strong real-time performance,YOLO V4,is used for crowd counting.In order to better training,is adopted in the experiment YOLO v4 algorithm on COCO data set of weights,and use the K-means clustering algorithm to cluster the size of the box,get nine box type,in order to keep the model more suitable for the underground scene,this paper will take video frames and the head and shoulders,made a subway real-world data sets.This data set is used to train the model,and the final number of passengers is obtained by detecting their head and shoulders.Due to the strong mobility of the crowd at the station hall and entrance,dynamic video counting was carried out in this part.Third,the number of people statistics system.After the complete algorithm part,this paper through the use of Python Django Web framework,set up a complete traffic statistics system,respectively,to the number of static and dynamic traffic statistics,in the static part to realize image upload function,the dynamic part of the video can be counted,and counted in and out of the population,to enhance the practicability of this article.The main innovations are: 1)Divide the scene into dense and sparse situations.2)An adaptive network is proposed for the crowd counting network in the dense scene.3)For the sparse part of the target detection network,the attention mechanism is integrated.