Visual Semantic Representation Of Crowd Behavior Analysis And Application

Crowd behavior analysis and understanding, as a complex, various andchallenging topic, has gained more and more attention by many research agencies.Crowd behavior analysis is one of the most important parts in the domain of videomonitoring, human computer interaction, multimedia content understanding. It has agood effect on the corresponding human behavior detection, tracking and the scenemodeling. It is a foundational and meaningful work. Therefore, as the foundation ofcrowd behavior analysis and understanding, visual reasonable and effective semanticrepresentation of the crowd behavior seems especially important.This paper aims at analyzing and studying effective crowd behavior semanticrepresentation and semantic pattern mining strategy for the surveillance video, whichis from the viewpoint of the hierarchical crowd behavior representation analysis, andtargeting at the application of crowd abnormal detection and crowd motion patternmining. Thus, we propose the theoretical representation of semantic crowd behaviorand the framework based on the representation. Besides, several experiments on thebenchmark dataset validate the effectiveness and correctness of our framework.This paper firstly analyzes and summarizes the low-level features and thecorresponding semantic gap between low-level feature and high-level semantics andaddress the affect of social behavior model on interaction of crowd behavior.Compared with the traditional social force model and method, we propose the socialattributes hypothesis which is accorded with the social behavior model. The socialattribute assumption provide basis for the studies on the creating the social attributeforce model as mid-level semantic representation which produce the scene scaleestimation and the semantic context attributes. The representation effectivelycaptures instinct structure of crowd behavior the contextual interaction semantics viasocial attribute force model. It is also extended to global and local abnormaldetection application via supervised and unsupervised learning method. Based onfoundation of this work, we propose reasonable crowd behavior affective disciplinesby extension of the crowd social attributes hypothesis. Our affective modelingapproach based on Arousal-Valence model is proposed to map the motion features tothe corresponding affective states, which is aiming at make a high-level representation for crowd behavior feature. The model is applied to global abnormaldetection and crowd pattern matching by intuitively tracking the affective states ofthe crowd dynamics. Experiments on the benchmark datasets demonstrate theeffectiveness of our proposed semantic representation, which provides strongevidence for the correctness of the theoretical framework.The whole research process starts from the low-level feature of crowd behaviorand proposes hierarchical crowd behavior semantic representation from the bottomup and solves the theoretical application framework of semantic representation,which is believed to be valuable for the further study.