Estimation of the Orientation of a Face in a Knowledge–based Analysis–synthesis Coder

An algorithm for estimation of the orientation of a person’s face in a knowledge–based analysis–synthesis coder (KBASC) assuming a talking person appeared in an image sequence is described. For estimation of the orientation of a person’s face using a face model, it is assumed that the eye and mouth view lines of a perspective camera are known and the mouth position locates at the right position with respect to its view line. Then, exploiting that eye and mouth positions of the face model along their view lines in the 3D space build an isosceles triangle, eye positions of the face model are iteratively estimated by rotating the eye–mouth plane. For this, the mean square error between the projection of the face model and the real image is minimized using Gauss–Newton–Method. The presented algorithm has been implemented in a knowledge–based analysis–synthesis coder (KBASC) and is applied to improve the tracking accuracy in the KBASC.