Monocular head-pose estimation utilizing optical flow and tracking of stable features

Monocular appearance-based head pose estimation by means of optical flow computation suffers particularly under large out-of-plane head rotations from stability problems due to the so-called aperture problem. In this work an independent estimation of the head pose is performed by computation of a projective transformation based on the registration of stable features between consecutive frames to cope with this problem. The robustness of the estimation is enhanced by stabilizing the optical flow with a motion-dependent regularization term. The selection of suitable points to compute the transformation is discussed in detail as well as the influence of the motion-dependent regularization on the conditioning number of the Hessian matrix and the system's performance. Results of the proposed method are validated by means of representative reference data sets.