Toward a Social Attentive Machine

In this paper, we discuss the design of a new “intelligent” system capable of selecting the most “outstanding” user from a group of people in a scene. This ability to select a user to interact with is very important in natural interfaces and in emergency-related applications where several people can ask to communicate simultaneousely. The system uses both static and dynamic features such as speed, height and social features (interpersonal distances) which are all acquired using a RGB-Depth camera (Kinect). Those features are combined and a contrast-based approach is able to focus the system’ attention on a specific user without complex rules. People position with respect to the Kinect sensor and learning of the previous people behavior are also used in a top-down way to influence the decision on the most interesting people. This application is represented by a wall of HAL9000's eyes that search in the scene who is the most different person then track and focus at him until someone more "different'' shows up.