Video-based pedestrian head pose estimation for risk assessment

Pedestrian detection and action recognition is a demanding field in video-based driver assistance systems. Future systems not only try to detect pedestrians but also aim to predict the pedestrian's intention in order to guaranty the best safety for him/her and other traffic participants. Our contribution to that tendency consists of a method to reliably estimate the pedestrians' head pose in low resolution video sequences taken from an on-board camera. Assuming a pre-detected pedestrian, the head pose is initialized using normalized confidence values from a set of head pose detectors. Integrating the head pose predictions over time using particle-filtering will further result in a higher robustness and efficiency. Experiments on public available datasets (CHIL/CLEAR2007, CAVIAR) and real world scenarios show a performance improvement compared to single image based approaches. The developed method can be integrated easily into an overall system and will serve for a better pedestrian path prediction and intension estimation within risk assessment.