Recognition Of Human Behavior In Intelligent Video Surveillance System

It has very important practical significance for recognition of human behavior in intelligentvideo surveillance system, and it is widely applied in these fields such as security-surveillance,intelligent buildings, human-computer interaction and so on. The intelligent algorithm was usedin computer to make it complete the monitoring task which was in video surveillance sceneautomatically. Thus the burden on monitor staff was reduced and a lot of human and financialresources were saved. The main work of this paper was the recognition of human action, to thisend, this paper used the space-time interest point methods to describe the action informationfrom the video. The interest points were detected by some interest point detection algorithms.Then it extracted a block of the video around the interest points and described them by somecertain methods. The main work of this paper is as follows:(1) This paper analysed the statistics characteristics of the gradient features and optical flowfeatures of different behavior, and find it was different between each behavior. Then, this paperput forward a new method of recognition of human behavior by utilizing the statistical propertiesof the local space-time feature. Firstly, this paper extracted the AGGD parameters from the twofeatures to describe human action respectively. In addition, the research of human actionrecognition was doing based on the gradient statistical features, optical flow statistical featuresand their fusion features. Finally, the performance was investigated in the Weizmann actiondataset and KTH action dataset. And the fusion feature could effectively improve the behaviorrecognition rate.(2) This paper analysed the histogram distribution of the local space time features ofdifferent behavior. And it was obvious different in each behavior’s histogram distribution of thelocal space time features. Based on this, this paper proposed a method of recognition of humanbehavior by utilizing weighted local space-time feature. This paper decomposed the brightnessgradient descriptor into three directions(X, Y,Z) and the optical flow descriptor into two velocitycomponents as (u, v). Their directional standard visual vocabulary codebooks of differentbehavior could be obtained by constructing visual vocabulary directly, and this paper used themto describe behavior. Then, this paper obtained the standard visual vocabulary of the fivedirection to describe action. And the behavior of the test video might be recognized by using theweighted similarity measure between each behavior’s standard vocabulary distribution and thevocabulary distribution of the test video. Finally, the experiment showed that all of the three features can improve the recognition rate of behavior. And the action reco gnition rate of fusionfeature was the best.