| Intelligent video surveillance is a hot issue in the machine vision field, having significant practical value and broad application prospects. Intelligent video surveillance technology mainly includes some operation on the video sequence, such as object detection, tracing, recognition and behavior understanding. In recent years, the increasing occurrence of the mass incidents has aroused the concern of managers and researchers, and has become a research focus of the intelligent video surveillance. This paper studied the anomaly identification method of the crowd by using the data mining techniques. The research includes crowd density and quantity estimation, crowd formation and evacuation identification and abnormal behavior identification about low-density crowd. Image texture analysis is used to achieve crowd density and quantity estimation in this paper. First, the foreground image of the object is gotten by the object detection, second, the GLCM of crowd image is calculated. The GLCM contrast, GLCM entropy, GLCM energy and GLCM uniformity are used as the characteristic value and the support vector machines are used to estimate the crowd density. Simultaneously the crowd quantity evaluation is achieved through the linear regression. By using background subtraction method to get the texture feature of the object image, the impact of background on the estimation results is avoided and the accuracy of estimation is improved.For the high-density crowd, an improved recognition method calculating crowd formation and evacuation is proposed based on sports corner. First, this method extracts the image corner. Second, extracts foreground object using the background subtraction, establishes the mask template and gets corner by using this templates. At last, by analyzing the object corner covariance matrix variation of the determinant, the crowd formation and evacuation identification are achieved by using the threshold method and the support vector machine method respectively. The experimental results show that the improved method can overcome the influence of the corner of stationary object after the crowd for the background corner.For low-density crowd abnormal behavior analysis, the low-density crowd identification is achieved by the object tracking method .First, the target is tracked through the Meanshift algorithm. Second, the target velocity, velocity amplitude and amplitude trajectory features are extracted. Third, decision tree is used to achieve object run, buildup, wandering and fighting etc. In order to overcome the camera shooting angle of the impact of behavior recognition, this paper use the camera calibration to get the parameters and use the turn image coordinates to world coordinates. According to the world coordinate of the object center of gravity, extract the motion characteristic value of object, and then go no the next identify work. Thus, the accuracy of behavior recognition is improved. |
