Study And Application Of Intensive People Flow Counting Based On Deep Learning

In recent years,the rapid development of artificial intelligence,computer vision and deep learning has greatly promoted the research of people flow counting based on surveillance video.The people flow counting based on surveillance video is the basis for the rational allocation of resources,the collection of commercial information and the intelligent management.It is also an important functional component of modern video intelligent monitoring,with relatively high research significance and application value.The content of this paper is intensive people flow counting based on deep learning.The YOLOv3 algorithm is used to realize pedestrian detection.The DeepSort tracking algorithm is used to realize local area tracking.The two-line method is used to realize the intensive people flow counting at the door.The number of people entering and exiting is stored in the database for storage and recording.Finally,an application example of intensive people flow counting system in the mall is designed and implemented.The main work of this paper includes the following aspects:(1)Pedestrian detectionAiming at the existence of mutual occlusion in intensive people flow,this paper designs a method based on the Smooth_L1 loss function to calculate the width and height loss of the bounding box and the human body detection scheme based on tilt imaging.Based on the YOL0v3 algorithm,according to the application scenario of this paper,combined with the influence of mutual occlusion on the number of people in the human body,this paper constructs the data set under the oblique angle,and forther divides the tilt angle according to the degree of occlusion between the targets at different angles.At 30 degrees,60 degrees and 90 degrees,three data sets were established,and the head and shoulder areas of the hunan body in the data were marked to overcome the problem of mutual occlusion between dense pedestrians.The data sets of the three angles were trained,and three models were compared for comparison.The actual detection results of the three angles were evaluated,and the optimal angle was selected.In the process of updating the parameters of YOLOv3 network back propagation,the L2 loss function is sensitive to the outliers and the model is unstable.In this paper,the loss of the bound:ing box is calculated by using the Smooth_L1 loss function.The results show that the loss function of Smooth_L1 is not sensitive to the outliers.Avoiding outliers with large parameter changes leads to model instability.(2)Pedestrian tracking and people flow countingIn this paper,a DeepSort algorithm using Top3 cosine distance mean is designed for human target tracking in intensive people flow.In this paper,the target tracking is performed by the DeepSort algorithm,and the target local area tracking is set,which greatly reduces the amount of calculation.In order to improve the tracking effect under crowded conditions,this paper compares the tracking matching results obtained by using the minimum cosine distance,the Top3 cosine distance mean and the Top5 cosine distance mean,and obtains the matching result using the Top3 cosine distance mean value to obtain fewer IDs.The number of switchings,so the Top3 cosine distance mean is placed in the cost matrix.In people flow counting,the two-line countlng method i5 used to track the human target and count the number of people to improve the counting accuracy.(3)Application of intensive people flow counting system in shopping mallsThis paper designs a people flow counting service system based on surveillance video,which pro-vides business decisions for shopping malls.In this paper,the data such as the number of people's statistics are stored and recorded using SQlite,and the application examples of the intensive people flow counting system in the shopping mall are designed and implemented.The demographic results provided by the system can provide business decisions for the mall.For example,controlling the opening of air eonditioners by eounting the total number of people in the mall i5 eonducive to energy conservation and emission reduction,and creating green shopping malls.The experimental results show tihat the proposed method is more accurate for the intensive people flow counting,and the average accuracy is 97.3%.