Comparative study of coarse head pose estimation

Abstract

For many practical applications, it is sufficient to estimate a coarse head pose to infer gaze direction. Indeed, for any application in which the camera is situated unobtrusively in an overhead corner, the only possible inference is coarse pose because of the limitations of the quality and resolution of the incoming data. However, the vast majority of research in head pose estimation deals with tracking full rigid body motion (6 degrees of freedom) for a limited range of motion (typically ±45 degrees out-of-plane) and relatively high resolution data (usually 64x64 or more). We review the smaller body of research on coarse pose estimation. This work involves image-based learning, estimation of a wide range of pose, and is capable of real-time performance for low-resolution imagery. We evaluate two coarse pose estimation schemes, based on (1) a probabilistic model approach and (2) a neural network approach. We compare the results of the two techniques for varying resolution, head localization accuracy and required pose accuracy. We conclude with details for the implementation specifications for resolution and localization accuracy depending on system accuracy requirements.