A novel method for head pose estimation based on the “Vitruvian Man”

In video-surveillance and ambient intelligence applications, head-pose estimation is an important and challenging task. Basically, the problem lies in assessing the pose of the head according to three reference angles, that indicate the head rotation with respect to three orthogonal axes, and are named roll, yaw, and pitch angles. The problem becomes particularly difficult if only 2D video-sequences or still images are available, thus information about the depth of the scene is missing. This makes the computation of the pitch angle very difficult. State-of-the-art methods usually add the information on the pitch angle separately, and this makes them strongly dependent on the hardware used and the scene under surveillance. Moreover, some of them require large training sets with head poses data. Finally, the extraction of several features from the detected face is often necessary. Since head-pose estimation is only a (small) part of a video-surveillance system as a whole, it is necessary to find novel approaches which make the head-pose estimation as simple as possible, in order to allow their use in real-time. In this paper, a novel method for automatic head-pose estimation is presented. This is based on a geometrical model relying on the exploitation of the Vitruvian man’s proportions and the related “Golden Ratio”. Our approach reduces the number of features extracted, avoiding the need for a training set as well as information on the hardware used or the scene under control. Simple ratios among eyes and nose positions, according to the assumed “Golden Ratio”, are used to compute, in particular, the pitch angle. Proposed method performs competitively with respect to state-of-the-art approaches, without requiring their working constraints and assumptions.