Research On Population Density Estimation And Abnormal Aggregation Behavior Detection Based On Deep Learning

With the rapid development of economy and large-scale growth of population,the pace of urbanization is accelerating,and the density of urban population continues is increasing.Crowd gathering often occurs in urban centers and public areas.Once a person falls in a crowded crowd,it can cause traffic chaos and death.The number of cases of global stampede in recent years has caused serious loss of life and property.Therefore it is necessary to monitor the density of the crowd and issue an alert if abnormal gathering behavior may occur.Nowadays video surveillance has been widely used,which is one of the important means of security.It is widely used in crowded places such as the subway station,the railway station,and the shopping mall,etc.Compared with human surveillance,The crowd analysis system based on video surveillance has natural advantages.Using target detection and tracking,pedestrian recognition,and other technologies to estimate crowd density and count the number of people,and analyze crowd abnormal behavior,which has become a hot research topic in the intelligent video surveillance industry.Using deep learning,big data analysis,and image processing to analyze the crowd behavior in surveillance videos,we can discover the abnormal behavior of the crowd in time and provide decisions and basis for crowd safety management.In the paper,I am first to classify and analyze the traditional crowd counting algorithms,and then introduce the classic algorithm architecture of crowd counting based on convolutional neural networks in recent years.Compared with these architectures and then analyzed the design deficiencies,I propose own improved algorithm,and verify the feasibility and effectiveness of the algorithm in three public datasets ShanghaiTech,UCF-QNRF,UCF_CC₅0.The next thing to apply the improved network to detect abnormal behavior that based on video,realizing video-based crowd counting and abnormal aggregation early warning.The specific work of this paper is as follows:1.Based on the RestNet network,this paper presents a population density estimation algorithm based on global neural network.For tasks like crowd density estimation and crowd counting,the model requires a larger receptive field.The easiest way to obtain a large receptive field is to use a large convolution kernel,but the amount of parameters of the model will increase.Therefore,to realize a large receptive field and make amount of the model parameter not too large,and the model performance is not bad.In this paper,two sets of convolution kernels of 1 X 9 and 9 X 1 are superimposed on the residual network to obtain a broad receptive field.At the same time,the model can have better performance and reduce the parameters amount of the model.The improved network model is trained and tested on the three datasets ShanghaiTech,UCF-QNRF and UCF_CC₅0,and it achieves good results.Compared with the MCNN model,the values of MAE and MSE perform well on the model,which dropped by 33.3%and 26.2%on ShanghaiTech dataset PartA,and decreased by 70%and 69.9%on PartB.On UCF-QNRF Dataset,they decrease by 64.7%and 58%.On UCF_CC₅0,they fell by 32.4%and 18.8%.The experimental results show that the improved network model has greatly improved in all aspects of performance,the accuracy of the population estimation has been greatly improved,and the two evaluation indicators that average absolute error and mean square error have certain advantages under this network.It proves the feasibility,effectiveness and stability of the proposed network.2.In this paper,the video-based population statistics and the detection of crowd abnormal behavior is researched.The improved network model is applied to the detection of crowd abnormal aggregation behavior based on video.The method based on threshold is used to judge whether the video is abnormal.Firstly,useing the k-means method to complete the clustering.After the clustering is completed,the number of data points falling in different classes is counted to form two types of scales.Then,a threshold is defined to evaluate the difference between the two types of scales,to define abnormal aggregation.Realizing the detection of crowd counting and crowd abnormal behavior in the video.Finally,the test is conducted on datasets PETS2009 and Mall.Compared with the MCNN model,the accuracy of the number of abnormal frames detected by the method used in this paper is improved by 6.5%,17.2%,4.62%and 9.68%on the four video sequences on PETS2009.It's improved 14.25%on the Mall.The experimental results prove the feasibility of the improved network and the effectiveness of the method used.