Research On Crowd Counting Method Based On Surveillance Video

In the age of big data,with the rapid growth of the world's population and technological progress,crowd counting methods based on surveillance video have a wide range of applications,including security monitoring in public places,intelligence gathering and analysis in shopping malls,and public space design.Existing detection,regression,and convolutional neural network algorithms exhibit excellent performance,but at the same time,the poor generality of algorithms based on detection and regression methods and the complex architecture problem of convolutional neural network algorithms still exist.In this context,this article will focus on the above issues,and conduct research on the population density statistics of medium-density dynamic scenes and high-density scenes.The specific research content is as follows:Firstly,aiming the medium-density dynamic scenes,this paper proposes a crowd counting method based on feature points,which combines the advantages of detection and regression methods.This article first analyzes that there are crowd density changes and a certain degree of occlusion characteristics in medium-density dynamic scenes.For this reason,the paper selects a low complexity and versatile detection method.Based on this,efficient clustering of Dirichlet model is selected for characteristic of the detector output is not precise enough.Then,the method based on the number of feature points is used for greatly improves the accuracy of the crowd counting.Secondly,based on the crowd counting methods of feature points,an algorithm for the correlation between the number of feature points and crowd density is proposed,and a data fusion method is designed for this purpose.This paper analyzes the problem of crowd detection and missed detection in the case of rapid changes in population density in medium-density dynamic scenarios.Aiming at the problem,the relationship between the number of feature points and the population density is designed,data fusion performed in the adjacent three frames of the medium-high density intervals,which improved the robustness.The experimental results show that the algorithm surpasses the best performance of the database in the medium-low,medium-high density areas.Thirdly,For high-density crowds,this paper presents a convolutional neural network model for crowd estimation.By learning the non-linear function,a corresponding population density map is obtained,thereby avoiding design of the complex regression method.In the fully convolutional neural network model,a deconvolutional structure is embedded and the perspective information is used jointly so that the output map and the input have the same size and retain more spatial information.Experiments show that the method achieves the best estimation accuracy in high-density crowds.In this paper,two sets of solutions with different complexity are proposed to solve the problems existing in the crowd counting methods,and coverage of full-density scenes.Experiments show that the algorithm designed in this project has good performance and high practical value.