Research On Crowd Density Estimation Algorithm In Intelligent Monitor

With the increase of population and diversity of human activities,the phenomenon of crowded scenes can be seen everywhere.The occurrence probability of crowd accidents will be greater when the crowd is more crowded.The early warning ability of existing video monitoring system is becoming less feasible for hidden danger,because they require human operators on duty.As a result,the demand for intelligent monitoring system is becoming more and more urgent.In this paper,we focus on crowd density estimation in intelligent monitoring system,of which the key techniques include crowd foreground detection,crowd feature extraction and crowd feature regression.The main contributions in the thesis are as follows:This paper firstly analyzes the advantages and disadvantages of the existing moving target detection algorithms via experiments,including frame difference,mean function,Gaussian Background Modeling,codebook model and ViBe(Visual Background Extractor).Then the ViBe algorithm with moderate time complexity and better foreground extraction was selected.Finally,we solved the problem that the ViBe algorithm cannot update the background model of the target region in the initial frame by extracting the region of interest based on the frame difference method.Moreover,a shadow suppression algorithm based on HSV color space was applied.The improved ViBe algorithm can detect a more complete crowd foreground.Aiming at the problem of perspective effect in video surveillance,this paper firstly uses linear interpolation to generate a perspective correction matrix,which is used to improve the representational ability of pixel-based,texture-based and corner-point-based crowd features.Finally,a model between crowd features and crowd size is established by applying regression analysis to the fused features.Through experiments and analysis,we find that there are linear relationships between most of crowd features and crowd size,and there are irrelevant and redundant features in the crowd feature vector.In the light of this kind of problem,we constrain optimization objective function of linear regression by L1 norm for automatic feature selection.The optimal shrinkage rate is selected by the cross-validation method,so that the regression model can rule out the irrelevant and redundant features as much as possible so as to improve the precision and generalization ability of our regression model.Last but not least,we designed a crowd density estimation system based on MFC and OpenCV 2.4.10,and we adopt PETS2009 benchmark dataset for a test.Experimental results show that the precision of the proposed algorithm is relatively high.