Crowd Density Estimation Methods In Intelligent Video Surveillance

Crowd density estimation (CDE) is one of the most important content of intelligent video surveillance, which plays an important role in public security, management control, and business decision and so on. This thesis mainly studies the methods of crowd density estimation and proposes two real-time and effective CDE methods for different scenes, which have been applied to practical engineering projects.For low and median crowd density scenes, we propose a pixel-based method, which rely on very local features. The basic idea is that we assumed linear models to map foreground area, foreground edge length and foreground edge gradient direction histogram to the number of people. Therefore, we can first segment out crowd flows and extract crowd features, then use multiple linear regression to predict crowd density. To improve the estimation precision, geometric correction through proper weight assignments should be performed for perspective effect. For this, we propose a perspective effect correction algorithm based on piecewise linear interpolation. Experimental results show that the algorithm we propose is efficiency, the accuracy (ACC) is 92.78%, and the miss rate (MR) and the false alarm rate (FAR) of each class are all very low.For median and high crowd density scenes, we propose a Gabor filter and SVM based method. The basic idea is that we treat it as a classification problem, and extract crowd features mainly based on texture analysis. In the framework of supervised learning, we experiment and analysis lots of feature combinations. At last, we get a result that the features obtained by using the Gabor filters whose direction number is 6, and value of scale is 2 and 3 have a good ability to identify. Experimental results show that the algorithm we propose is efficiency, the ACC is 95.33%, the MR and the FAR of each class are all very low, and the misclassification cases are concentrated in the near classes.