Research On Abnormal Behavior Real-time Detection In Video Surveillance

With the continuous development of computer and Internet technology and the growing demand of security in each areas of society, intelligent monitoring technology has achieved a great improvement. Intelligent monitoring technology has been applied in various areas such as security checks at airports and train stations, traffic control, national security. Traditional monitoring system which mainly rely on peoples’ eyes can not adapt to the current massive video data and guarantee efficiency and reliability. Therefore efficient and reliable intelligent monitoring system has come into being, which is possible to improve the intelligence of video surveillance system.In the detection method of abnormal behavior in crowd in this paper, firstly the video collection is preprocessed thus the video sequence can be processed directly, then the global target feature such as spatial and temporal characteristics, motion vectors, the motion vector histogram and flux grid scale is abstracted. The analysis of abnormal behavior in video sequences is based on these characteristics. The following three steps was implemented: first, analyze the abnormal behavior based on feature energy information; If abnormal behavior is not detected, fuses the feature extraction dispersion; If abnormal behavior not exists, do analysis according to Lagrangian particle dynamics between the extracted features; if abnormal behavior still not exists, locally abnormal is detected. In the detection process, grid scale features not only can be used to detect abnormal behavior globally, but also be able to detect locally abnormal behavior in video with no global abnormal behavior.Next, local abnormal behavior will be detected. First extract the image blocks of normal activities and model the key areas of the image blocks and calculate correlations between the regions to obtain the model of normal activities. Then implement the extraction and optimization of image blocks in test videos to determine whether there is a local abnormal behavior and labeled.Experiments are implemented on UMN, UCSD and Subway database. Experimental results show that our method is able to detect global and local abnormal behaviors in the presence of shadows, lighting changes interference and the method outperforms others.