Head Pose Estimation using Adaptively Scaled Template Matching

This paper proposes a real-time head pose estimation system usinga new image matchingtechnique. The system consists ofa trainingstage,in which subspace dictionaries for classifyinghead poses are computed using template matchingand the factorization method, and a recognition stage,in which head poses are estimated usingthe subspace method. The system uses the method ofadaptivelyscaled template matching,in which the expansion ratio ofthe template is adapted to the size ofthe tracked object and a search distribution with high center densityis used for position and expansion ratio search.It is eãective for accuratelytrackingregions and feature points on a face. The new method also increases the number ofsuccessful detections ofthe object because it rarelymisses the point ofmaximum similarity. A head-centered coordinate system (H-coordinates) is also proposed for representinghead poses independently ofcamera position. UsingH-coordinates,we can classifyhead poses bytwo variables (horizontal and vertical angles)independentlyofinclininghead motion (tilt angle),which does not change the gaze. Numerical experiments show the eãectiveness ofthe proposed method.