Group Behavior Recognition using Dynamic Bayesian Networks

In this PhD thesis we analyze the concepts involvedin the decision making of groups of agents and applythese concepts in creating a framework for performinggroup behavior recognition. We present an overview ofthe intention theory as studied by some greattheorists such as Searle Bratmann and Cohen andshow the link with more recent researches. We studythe advantages and drawbacks of some techniques inthe domain and create a new model for representingand detecting group behaviors the aim being tocreate a unified approach of the problem. Most ofthis thesis is consecrated in the detailedpresentation of the model as well as the algorithmresponsible for behavior recognition. Our model istested on two different applications involving humangesture analysis and multimodal fusion of audio andvideo data. By means of these applications weadvance the argument that multivariate sets ofcorrelated data can be efficiently analyzed under aunified framework of behavior recognition. We showthat the correlation between different sets of datacan be modeled as cooperation inside a team and thatbehavior recognition is a modern approach ofclassification and pattern recognition.