Investigating Social Interactions Using Multi-Modal Nonverbal Features

Every day, humans are involved in social situations and interplays, with the goal of sharing emotions and thoughts, establishing relationships with or acting on other human beings. These interactions are possible thanks to what is called social intelligence, which is the ability to express and recognize social signals produced during the interactions. These signals aid the information exchange and are expressed through verbal and non-verbal behavioral cues, such as facial expressions, gestures, body pose or prosody. Recently, many works have demonstrated that social signals can be captured and analyzed by automatic systems, giving birth to a relatively new research area called social signal processing, which aims at replicating human social intelligence with machines. In this thesis, we explore the use of behavioral cues and computational methods for modeling and understanding social interactions. Concretely, we focus on several behavioral cues in three specic contexts: rst, we analyze the relationship between gaze and leadership in small group interactions. Second, we expand our analysis to face and head gestures in the context of deception detection in dyadic interactions. Finally, we analyze the whole body for group detection in mingling scenarios.