The Research Of Activity Analysis Based On VStop Visual Words In Complex Dynamic Scenes

In complex scenes, activity analysis, especially rare activity analysis is the basictopics of computer vision. In the research, pattern recognition, artificial intelligence, imageprocessing and other areas of knowledge is necessary. The knowledge of how to effectivelyintegrate those method is worth to study for the activity analysis knowledge. The subject ofstrong applications (Intelligent Caring, Intelligent Security, Intelligent Transportation,Intelligent City, etc.) adds to its theoretical significance and application value. Paperfocuses on the fixed camera in aerial complex scenes, the scene changes, the target varietyof different forms of exercise lead to a difficult target tracking, behavioral modelingdifficulties, rare behavior is difficult to model, pause, velocity anomalies, abnormaldirection. Its contents are as follows: visual word generation algorithm, rare behavioranalysis and online rare activity detection, the main work of this paper is summarized asfollows:1)Traditional visual word usually using only the direction information and locationinformation, lacking of velocity information and context information. The lack of target’sspeed information, some behaviors can not be effectively modeled. For example, the twotarget moving on the same trajectory but the with different speed. So we proposing a newvisual word generation algorithm-VStop visual words generation algorithm. The VStopvisual words not only reserved directions information and space information, but alsoadopt a rich velocity information and target stop state information. For velocityinformation, firstly using the adaptive the speed quantization process to establish thevelocity information called velocity word (V-word), solving the problem of lack of velocityof the target. Followed by the introduction of stop words (Stop-word) the concept ofdescription the behavior of velocity is zero, effectively solved the problem of target suspend or block in the scene. Through the Hospedales ICCV09data sets experimentalresults show that the VStop visual word generation algorithm can effectively improve theperformance of video moving target analysis.2) In activity offline clustering, traditional clustering process which using Flow visualwords, is short of high-level semantics and the ability to discrimination two activity incommonly trajectory. And lacking the ability to describe the details of the behavioralevents and can’t modeling rare activity in supervise way. this paper proposes the offlineactivity clustering method based on VStop visual word. The experimental results show thaton the Hospedales ICCV09dataset and MIT dataset, The algorithm can effectivelymodeling the behavior details, originally indistinguishable category refinement. At thesame time for the traditional identification methods can not effectively model rare activity,this paper presents the framework of auto-discovery rare activity.3) In online rare activity detection process, traditional methods can not quicklydetection rare activity, especially in the case of multi-target with multi-speed, and onlineupdates difficulties. so we propose the rare acivity detection framework based on mainactivity update algorithm. The algorithm inspired by the foreground detection based ongaussian mixture background modeling algorithm. We analogy main activity to dynamictexture. And modeling the main activities based on VStop visual word. Foreground regiondetection algorithm introduced to detecting rare activity region based on main activities.The results of the UCSD data Experiment show that the algorithm can effectively detectrare behavior with lower complexity and higher real-time respond.