Head Pose Estimation Using a Texture Model Based on Gabor Wavelets

Head pose estimation from a monocular camera or a simple image is a challenging topic. It is the process of inferring the orientation of a human head from digital imagery. Several processing steps are performed in order to transform a pixel-based representation of the head into a high-level concept of direction. The head pose is important in a lot of domains like human-computer interfaces, video conferencing or driver monitoring. Head pose estimation is often linked with visual gaze estimation (Lablack et al., 2009) which is the ability to characterize the direction and focus of attention of a person looking to a poster (Smith et al., 2008) or to another person during meeting scenarios (Voit & Stiefelhagen, 2008) for example. The head pose provides a coarse indication of the gaze that can be estimated in situations when the eyes of a person are not visible (like low-resolution imagery, or in the presence of eye-occluding objects like sunglasses). When the eyes are visible, head pose becomes a requirement to accurately predict gaze direction (Valenti et al., 2009). The aim of our work is to analyze the behaviour of the people passing in front of a target scene (Lablack & Djeraba, 2008) in order to extract the person's location of interest. The success of this kind of system highly depends upon a correct estimation of the head pose. In this paper, we present a template based approach which considers the head pose estimation as an image classification problem. Thus, the Pointing database (Gourier et al., 2004) has been used to build and test our head pose model. The feature vectors of different persons taken at the same pose will serve to learn a head pose classifier. The texture model is learned from feature vectors composed of the properties extracted from the real, imaginary and magnitude response of Gabor wavelets (due to the evolution of the head pose in orientation) and singular Value decomposition (SVD). The head pose estimation is then applied on the testing dataset. Finally, the classification accuracy is compared to the state of the art results that used the Pointing database. The paper is organized as follows. First, we highlight in Section 2 relevant works in head pose estimation. We then describe the method used for the head pose estimation and the database associated in Section 3. Sections 4 and 5 provide two representations of feature vectors extracted from SVD and the 3 different responses of Gabor wavelets. We apply on them two supervised learning SVM and KNN and the Frobenius distance. We discuss the 18