Research On Crowd Analysis Method Based On Deep Learning

In recent decades,China’s economy has developed rapidly,and at the same time,local governments have been continuously promoting urbanization construction.A large number of people who used to live in rural areas have entered cities to work,study,and live.This has led to a sharp increase in both the area and population of the entire city.In this context,the population density in the core areas of cities and many public places is relatively high,making it very easy for large-scale crowds to gather.If an abnormal event occurs in a crowded crowd,the crowd can disperse and easily cause a stampede accident.At a mild level,it can cause property damage,while at a severe level,it may lead to casualties,which is regrettable.Therefore,it is necessary to monitor and analyze these places in order to provide timely warnings for events that have already occurred or are about to occur.In addition,compared to the method of using manpower to monitor and analyze these places,a video based crowd analysis system that utilizes image processing,deep learning,and other technologies in a reasonable manner is clearly more efficient for dynamic monitoring and analysis.The topic of this article is the research on crowd analysis methods based on deep learning,which mainly includes the research on crowd counting methods and crowd behavior analysis methods.The specific content is as follows:(1)In the study of population counting,the use of CNN and fusion of multi-scale information can effectively count sparse areas,but cannot accurately estimate dense areas.At the same time,Transformers using self attention mechanisms can obtain the connections between inputs from a global perspective,thus effectively estimating dense scenes.However,due to the lack of location information,its counting effect in sparse regions is limited.Therefore,this article proposes a crowd counting method based on CNN and Transformer.The model structure adopts a dual branch form of CNN and Transformer,where each branch generates a density estimation map.At the same time,an adaptive selection module is used to select appropriate branch results for each region based on local density information to obtain a higher quality density estimation map and improve the counting effect.The crowd counting algorithm proposed in this article was trained and tested on the Shanghai Tech and UCF-QNRF datasets,respectively.Compared with the classical algorithm CSRNet,the model in this paper shows a significant decrease in MAE and MSE.On the Shanghai Tech dataset part A,it decreased by 19.94% and 22.52%,on the Shanghai Tech dataset part B,it decreased by 34.91% and 31.25%,and on the UCF-QNRF dataset,it decreased by 21.95%and 42.30%,respectively.The experimental results demonstrate the improvement effect of Transformer on counting performance,as well as the effectiveness and superiority of the population counting method proposed in this article.(2)In the study of crowd behavior analysis,this article proposes an energy model based crowd behavior analysis method for the abnormal behavior type of crowd dispersion.This method uses a weighted cumulative average energy of video frames to determine whether there are abnormalities in the crowd in the video and detect abnormal video frames.The test results of the crowd behavior analysis method proposed in this article on the UMN dataset show that compared with some classic algorithms in crowd behavior analysis,the performance has been improved to a certain extent.On the UMN dataset,the AUC values reached 0.977,0.962,and 0.985,respectively.The above experimental results demonstrate the effectiveness of the energy model based crowd behavior analysis method proposed in this paper.