Research On Density Adaptive Crowd Counting Based On Multi-Feature Fusion In Video Surveillance

With the increasing enrichment of modern human life and production activities,social activities such as concerts,fairs,and large-scale job fairs have become more and more popular.The crowd gathering activities occurred in public places such as city squares,schools and large supermarkets,which lead to the frequent occurrence of group abnormal events,which seriously threatens social public safety and stability.Therefore,crowd research based on video surveillance is of great significance.Pedestrian statistics,as one of the important topics in population research,is increasingly favored by researchers.However,monitoring scenes are inevitably affected by factors such as object shadows,lighting changes,and occlusions,making it extremely difficult to build an accurate,real-time crowd counting system.Researchers have proposed many algorithms.Generally speaking,the existing algorithms can achieve high counting accuracy for low-density population,but for high-density population with more research value,there is no algorithm has been proposed which takes both accuracy and real-time into account.To solve these problems,this paper will carry out the following related research and analysis:(1)This paper innovatively combines the compressive and viscous features in fluid mechanics,constructs a fluid mechanics model,and then fuses fluid features and pixel features to solve the problem that single features in high-density populations cannot effectively characterize population information.(2)In this paper,a density adaptive population statistics method based on multi-feature fusion is proposed.Firstly,the concept of adaptive scene density is introduced.After foreground segmentation of moving objects,the scene is divided into different densities.Different scenes need to construct different feature vectors.Different feature regression models are obtained through training,and then the number of people can be predicted.(3)In this paper,several popular foreground extraction algorithms are compared experimentally,and finally the best ViBe algorithm is adopted.In addition,in order to obtain more complete foreground motion information,this paper expounds the principle and calculation process of two commonly used optical flow algorithms,gives the experimental results separately,analyzes their advantages and disadvantages,and selects a relatively complete one.The HS optical flow method for characterizing moving particle features is used as a method for extracting microscopic particle motion information.