Head pose estimation and its application in TV viewers' behavior analysis

Head pose implies a person's visual attention and interest. It plays an important role in many applications. Existing head pose estimation methods work in the original head pose space. However, the large number of head pose candidates in the space makes the estimation task quite challenging. In this paper, we propose a coarse-to-fine head pose estimation method by decomposing the original pose space into a hierarchical structure. The estimation begins by detecting the region of interest (ROI) within a face image via measuring the importance scores of key image points. After that, a coarse head pose estimation step is applied to identify a subset of head pose candidates, based on Gabor filter and random forest. A fine estimation is then employed within the subset, using histogram of oriented gradient (HOG) and support vector machine (SVM). Finally, we apply the proposed method to TV viewers' behavior analysis by determining whether a viewer is focused or unfocused, which can be useful for marketing research.