Using Random Forests for the Estimation of Multiple Users' Visual Focus of Attention from Head Pose.

When interacting with a group of people, a robot requires the ability to compute people's visual focus of attention in order to regulate the turn-taking, to determine attended objects, as well as to estimate the degree of users' engagement. This work aims at evaluating the possibility of computing real-time multiple users' focus of attention by combining a random forest approach for head pose estimation with the user's head joint tracking. The system has been tested both on single users and on couples of users interacting with a simple scenario designed to guide the user attention towards a specific space region. The aim is to highlight the possible requirements and problems arising when dealing with the presence of multiple users. Results show that while the approach is promising, datasets that are different from the ones available in the literature are required in order to improve performance.