Group Behavior Recognition Using Dynamic Bayesian Networks: Understanding Intentions, Goals and Actions That Take Place Inside Teams

In this PhD thesis we analyze the concepts involved in the decision making of groups of agents and apply these concepts in creating a framework for performing group behavior recognition. We present an overview of the intention theory, as studied by some great theorists such as Searle, Bratmann and Cohen, and show the link with more recent researches. We study the advantages and drawbacks of some techniques in the domain and create a new model for representing and detecting group behaviors, the aim being to create a unified approach of the problem. Most of this thesis is consecrated in the detailed presentation of the model as well as the algorithm responsible for behavior recognition. Our model is tested on two different applications involving human gesture analysis and multimodal fusion of audio and video data. By means of these applications, we advance the argument that multivariate sets of correlated data can be efficiently analyzed under a unified framework of behavior recognition. We show that the correlation between different sets of data can be modeled as cooperation inside a team and that behavior recognition is a modern approach of classification and pattern recognition.