Crowd Counting Analysis Under Complicated Scenes

Population statistics,as an important part of group analysis,is one of the most important topics in the field of video surveillance.With the development of large-scale urbanization in China,the scale of urban expansion and population density have risen greatly.The urban public management in China is facing more challenges.Therefore,it is necessary to advance the crowd counting algorithm to meet the requirements of new era safety monitoring.Firstly,we look into the crowd counting algorithm based on CNN(convolutional neural network),which can solve the problems that traditional crowd counting method can't.This paper not only analyzes MCNN method(multi-convolutional neural network),but also research the working mechanism based on CNN algorithm through visualization technology.In order to make up the existing datasets that fail to meet the nowadays monitoring scenario test requirements,we collect a new dataset containing five scenes,and verify the practical value of the algorithm through experiments.Secondly,we found that the density distribution map obtained by CNN based algorithm has evident defect.Along with the idea of generating better density map,for the first time,we put forward novel crowd counting algorithm based on cGAN(conditional Generative Adversarial Networks).The algorithm trains the generator to generate the density distribution map and the discriminator identifies the realness of the pictures.This method can effectively improve the quality of the density distribution map generated by the generator.Experiments show that the accuracy of the proposed algorithm can exceed that of MCNN network.Meanwhile,the generated crowd density distribution map is of higher quality and the experiments is of higher training speed.To sum up,we concentration on explore the crowd counting algorithm based on CNN in this paper long with insightful analysis of the algorithm and experimental results.We not only proposed a totally new crowd counting algorithm based on cGAN,but also discuss the influencing factors of the two kinds of methods,the working mechanism and practical application.We conduct a relatively systematic study which is of certain theoretical innovation and application value.