The Research Of Abnormal Behavior Analysis In IVS

In recent years, along with the rapid development of the global economy, socialprogress and the rapid rising of comprehensive national power, social securityproblem has become increasingly serious, ranging from banks, shopping malls,subways, railway stations, factories, small residential areas, schools, dormitoriesand others,as such public places’ needs for safety protection, real-time videorecording, and alarming are constantly increasing, also the requirements forintelligent degree are increasingly stringent. Tranditional monitoring mode has faraway to meet such demand, and intellectualization video surveillance system fittingvarious monitor scene is the hot trend of current development.Intelligent video surveillance is defined as the absence of any humanintervention, the full use of various methods of computer vision, digital imageprocessing, pattern recognition, to dispose and analyze the real-time video sequencesautomatically. Foreground detection and extraction to interested area amongmonitoring area and pre-defined behavioral event (or targets), then marking andtracing targets, finally analyze the human behavioral, note to video, meanwhile,automatic alarm once appeared abnormal behavior in the video.This paper mainly research on abnormal behavior analysis in-depth for videosurveillance, to make use of Chinese Academy of Sciences Institute of AutomationCASIA video database as a data source analysis. CASIA video database includessingle behavior, such as walking, running, jumping, faint, wandering etc., doubleinteractive behaviors such as fighting, overtake, robbing, meet apart, follow always,etc., extract key features to classify and identification the behavior for a variety ofacts. Main content of this paper contain following several aspects:1. First step is moving target detection, this paper studies the optical flowmethod, frame difference method and Gaussian background modeling method, do allthe simulation experiments to foreground extracting. After compared the effect of thethree methods, the extracting effect of mixed Gaussian background modeling is thebest, choose to use the improved Gaussian background modeling method as thesystem the final method. Besides of, choose kalman filter to complete the moving target tracking in several conditions such as target fusion tracking, target disappeartracking. Do experiment to simulate and realize.2. Second step is feature extraction, extract a set of key characteristics as a basisfor subsequent classification. We choose to select a set of characters which regardlessof the angle of shoot view, or the target size, shape, posture, such as moving velocity,target height-width ratio, field entropy of optical flow motion area, and main directionangle of target based on the amplitude weighted. Comparing all frames’ feature vectorof different behaviors verify the efficient discrimination.3. Finally, behavior classify, used SVM to achieve a multi-classification forvarious behaviors. To samples nearby the support vectors, choose KNN methodinstead, which will improve the classify accuracy. Totally, choose SVM-KNNclassifier to finish the classification on this paper. Besides, develop a experimentsystem to real-time show video and popup classification results.