Crowd Density Surveillance Based On Video Analysis

Intelligent video surveillance based on video analysis, being an important research branch in the field of computer vision and pattern recognition, has become one of the key technologies paid much attention in recent years. With crowd flowing more and more frequently in various public places, to effectively management the crowd by means of crowd density surveillance has become one of the important tasks in intelligent video surveillance. This thesis focused on the research work for crowd density surveillance based on video analysis. The main contribution can be summarized as following:1. Crowd surveillance based on people countingFirstly some crowd features including the area of crowd foreground, GLCM(Grey Level Co-occurrence Matrix)-related features are extracted based on background subtraction. Then the counting model is established in which the people number is linear with the crowd features. By means of LSM (Least Squares Method), the counting model can be regressed. By estimating the number of people, the crowd congestion of ROI (Region of Interest) can be surveillance.In order to avoid perspective distortion, this thesis proposed an optimal method for tiepoints selection based on RANSAC, in which the perspective transformation parameters are firstly deduced based on RANSAC, then a "density map" is produced by applying perspective transformation. With the crowd feature normalization by "density map", the accuracy of the counting model can be efficiently improved.2. Crowd surveillance based on classificationIn this part, the estimation of crowd congestion is regarded as the problem of multi-class classification by SVM(Support Vector Machine).In this proposed multi-class classification algorithm based on C-SVM, crowd features for different congestion level are firstly extracted from some typical video samples, and then the normalized features are used for training classifier. In the course of SVM training, the weighting modification is used to avoid sample imbalance. Furthermore, the optimal selection for C-SVM and RBF kernel function parameters are conducted by combining n-fold cross-validation and coarse-to-fine grid searching. The performance for several typical classifiers are also analyzed and compared.