Research On Crowd Density Estimation In Video Surveillance

With the rapid growth of economy, the population density of cities increasescontinuously and social activities develop gradually frequently. It is often seen that thepopulation congestion in public occasions results in accidents. Effective estimation of crowddensity in scene surveillance can be used to help the relevant departments to defend publicsafety. It also means a lot to site management and people scheduling in some work places. Sohow to estimate the crowd density of video surveillance effectively is of profoundsignificance and wide applications.This paper works on the intelligent crowd automatic estimation technique which is basedon the video and image process．Firstly we introduce the development and basic principles ofthe crowd density estimation detailedly.The analysis shows that the pixel-based crowdestimation is simple but only be applied to the scenes of low density wihout the crowdbarrier,and the texture-based crowd estimation performs better in dense population but ismore complex and estimate error increases in the scenes of low density.To solve the problemspresented above,the paper proposes to research two methods of crowd density estimationaccording to the different scenes．For low and median crowd density scenes,we use the multiple linear regression ofpixel-based method to evaluate crowd density. Firstly we apply the techniques of backgroundsubtraction and adaptive background updating to acquire the crowd foreground of the imageand we implement edge detection as well.There is a good linear relationship between numberof people and foreground area,foreground edge length,foreground edge gradient directionhistogram.Then we obtain multiple linear model between the number and the characteristicsthrough multiple linear regression and can use it to estimate the number.We also use aperspective effect correction algorithm based on piecewise linear interpolation to improve theestimation accuracy.For median and high crowd density scenes,we use the crowd density estimation methodbased on the gray level co-occurrence matrix and fractal dimension. Because the images ofmoderate-to-high density crowd have more texture characteristics,so we can use the methodbased on texture analysis to extract the characteristics of population density.We extract thestatistical features based on gray level co-occurrence matrix, such as entropy, contrast and energy,and take difference box dimension method to calculate fractal dimension of textureimage. On the other hand, linear relation of number of people and features is not very obviousso that we apply nonlinear classification method to evaluate crowd density. Whereas supportvector machine to a large extent overcome the problems of local optimum and nonlinearity,wechoose support vector machine as the basic classifier.The results of simulation experimentsshow that the proposed method is effective and feasible.