Analysis Of Crowd Density And Detection Of Sudden Abnormal Behavior In Intelligent Video Surveillance

In current social environment, crowd scene is growingly universal, diversified and complicated. Accordingly, problems that relate to public management and public safety of the crowd are becoming increasingly serious, which brings great challenges to the supervisory personnel. Crowd density estimation and sudden abnormal behavior detection are most important research content in the intelligent crowd surveillance system. Both domestic and foreign scholars have done a lot of work in these areas, but there are still some problems left. In the crowd density estimation algorithm, description of the texture features of crowd only considers the overall features of the image while ignoring the local details of the image, which makes the description of image texture become not comprehensive enough. At the same time, the crowd density classifier training process should also be optimized further. In the crowd’s sudden abnormal behavior detection algorithm, most detection model cannot describe crowd motion state of variation properly, for example, there is no timely tracking when the crowd burst abnormal behaviors, and abnormal feature description was not significant.In view of the above questions, the improved methods as follow: In order to estimate crowd texture, this paper presents the description operator of fusion of global and local gray level co-occurrence matrix (GLCM).In the training process of the classifier, the paper filters out abnormal sample based on Bayesian estimation, and improves iterative training algorithm based on K-means clustering. For the sudden abnormal behavior detection, this paper also establishes the anomaly detection model based on transient energy feature of the crowd. The specific work of this paper is as follows:1. The crowd density estimation algorithm which combined the foreground pixel features with texture features. To begin with, according to the perspective model, the Region of Interest in the foreground image should be divided into blocks in proportion, and then the foreground pixels proportion of image in block is statistically analyzed. A specific area proportion is selected as threshold. When a proportion is less than threshold, statistical algorithm based on foreground pixels should be used for regression, otherwise, texture feature based machine learning algorithm should be adopted. Description operator of the fusion of global and local gray level co-occurrence matrix (GLCM) is used to extract crowd texture in video pictures. In process of training support vector machine model, the paper adopts the estimation method based on Bayesian to filter abnormal sample, and then obtain crowd density classifier by improved iterative training algorithm based on the K-means clustering.2. Crowd abnormal detection based on transient energy feature. The paper states a method that is based on statistical analysis of the pixel, to extract motion feature points, the extracted feature points more representative. Video grid processing is the first step of this method. Then, video background model is established by utilizing the Gaussian mixture model, and then foreground motion region is obtained by using background subtraction method. Finally, the motion feature point is gained according to the proportion of foreground pixels within the grid. To structure the transient energy feature of the crowd, firstly, the motion vector of feature point in the square grid should be acquired by using the optical flow method, and next step is to calculate the basic energy feature of grid. Then, establish transient energy features to detect the sudden abnormal behavior.According to simulation experiments, the conclusion is that the average accuracy rate of crowd density classification reaches 96.2% by using texture feature which fuse local and global co-occurrence matrix features. After utilizing training set which has filtered out abnormal samples based on Bayesian estimation, to train the classifier, the rate of average correct classification increases to 98%. Besides, after adopting the improved iterative training method based K-means clustering, the training time declines to 2.4s, and the test time decline to 43.9ms. In the experiment of comparing with the model based on kinetic energy features and the model based on Social Force, the sudden abnormal behaviors detection algorithm based on transient energy shows the prominent superiority both on the response speed and the description of the change of crowd movements.