Fast QuadTree-Based Pose Estimation for Security Applications Using Face Biometrics

Face represents a convenient contactless biometric descriptor, currently exploited in a wide range of security applications, though its performance may be considerably affected by subject's pose variations with respect to enrolment pose. This issue is particularly challenging whether the face image is acquired in uncontrolled conditions, or it is extracted from video sequence, the latter representing a more and more frequent case given the huge diffusion of audiovisual content on the internet. To this regard, in this paper, a pose estimation method aimed at rapidly evaluating face rotations is presented. The proposed approach exploits a novel adaptation of quad-tree data structure to achieve an approximate estimate of face's yaw/pitch angles, enabling to select the face image most compliant to the stored template. Preliminary results confirm the efficiency of the proposed method, that provides a more than halved computing time with respect to the state of the art with further improvement margins.