Probabilistic Head Pose Tracking Evaluation in Single and Multiple Camera Setups

This paper presents our participation in the CLEAR 07 evaluation workshop head pose estimation tasks where two head pose estimation tasks were to be addressed. The first task estimates head poses with respect to (w.r.t.) a single camera capturing people seated in a meeting room scenario. The second task consisted of estimating the head pose of people moving in a room from four cameras w.r.t. a global room coordinate. To solve the first task, we used a probabilistic exemplar-based head pose tracking method using a mixed state particle filter based on a represention in a joint state space of head localization and pose variable. This state space representation allows the combined search for both the optimal head location and pose. To solve the second task, we first applied the same head tracking framework to estimate the head pose w.r.t each of the four camera. Then, using the camera calibration parameters, the head poses w.r.t. individual cameras were transformed into head poses w.r.t to the global room coordinates, and the measures obtained from the four cameras were fused using reliability measures based on skin detection. Good head pose tracking performances were obtained for both tasks.