An Anthropometric Shape Model for Estimating Head Orientation from a Monocular Image Sequence
|
An Anthropometric Shape Model for Estimating Head Orientation from a Monocular Image Sequence
|
We present an algorithm for estimating the orientation of a human face from a single monocular image. The algorithm takes advantage of the geometric symmetries of typical faces to compute the yaw and roll components of orientation, and anthropometric modeling to estimate the pitch component.
Estimating head orientation is central in vision-based animation, gaze estimation and as a component of inferring the intentions of agents from their actions. Vision-based animation involves reproducing head actions as accurately as possible relative to the observed individual. Accurate gaze estimation requires the determination of the relative pose of the iris with respect to the head orientation. To infer human intentions such as avoidance of surveillance subtle spatio-temporal head and face motions should be powerful cues. We seek an approach that requires no prior knowledge of the exact face structure of the individual being observed. The diversity of face and head appearances due to hair (head and facial) and eyeglasses in addition to articulated jaw motion and facial surface deformations leave very few features geometrically stable and predictable across individuals and head orientations. The nose is the only feature not subject to significant local deformations. In addition, the eyes are often visible (although occasionally covered by eye-glasses). For estimating head orientation, we assume that both eyes and the nose are visible, thus avoiding near-profile poses.
In this work, a new approach for head orientation estimation is proposed. The approach employs image-based parameterized tracking [3] for face and face features to locate the area in which a sub-pixel parameterized shape estimation of the eye boundaries would be performed. This results in tracking of five points, four at the eye corners and the fifth at the tip of the nose. Although five points are not sufficient for recovering orientation in the absence of structure, we describe an algorithm that combines projective invariance of cross ratios from typical face symmetry and statistical modeling for face structure from anthropometry to estimate the three rotation angles. This approach consists of the following stages:
Region tracking of the face and the face features based on parameterized motion models (see [3]).
Refining the feature position estimation by color correlation and co-linearity of eye corners constraint.
Computing 3D orientation from these five points.
|
