Research On Abnormal Behavior Detection Of Human Movement

Along with the development of the information technology, especially thetechnology in computer vision field is developed very much, at the same time, peoplepay more and more attention to safety, so traditional video surveillance systems cannotsatisfy people’s requirement for intelligence. Computer vision community has paidmore and more attention to human behavior research, which promotes the research anddevelopment of abnormal behavior. More commonly, in intelligent video surveillancesystem the following steps are taken: capture real-time video frames, build backgroundmodel to extract foreground which contains the moving targets, if the area includesshadows and noise, it is necessary to take process to get rid of that. then based on therequirement suitable algorithm is selected to track moving targets, then exact behaviorfeather and expose the standard of abnormal behavior to analyze the detected behaviorand judge them. Finally behavior detection result is obtained.In this thesis, the image processing technology and computer vision relatedtechnologies were introduced first, mainly including the foreground extractiontechnology and the behavior definition and feature extraction technology. In the wholeprocess of this thesis, first the foreground extraction step was took, to satisfy theforeground objects extraction’s requirements for real-time and accuracy, comprehensiveconsidered the Gaussian model’s merits and demerits, the RGB color channel separationmethod was proposed to exact foreground target in time, this method took fulladvantage of image’s color space information. To enhance the background model’srefresh rate, this method took pixel and frame update. In the foreground, there may existshadows, based on their luminance in HSV color space was different from the objects,so in this way, the shadows can also be removed. Thus, the ideal objects area wasobtained.In this thesis, two kinds of violence detection were considered. In the fewerinteractive behaviors scene, first each moving target was tracked and got their positioninformation, and based on the traditional optical flow field feather, more feathers wereproposed to detect the abnormal behavior in a global energy way, global energy included the behaviors’ individual energy and also included the interactive energy. Bysetting energy threshold, the behavior energy types can be judged and detected theabnormal. To detect the abnormal behaviors in the crowd scene, the method ofmodeling whole group as particle group was proposed and the social interaction forcemodel is used to assess the group’s active. To increase the accuracy of detection, theassessment of the video frame sequence’s social interaction force was proposed. Testedthis two proposed abnormal behavior detection method on different kinds of actionvideos, the results in fewer people scene and in crowd scene both showed the idealeffects.