Multi-view temporal tracking under weak perspective in real-time
Alex Po Leung · 2005
Haar-like wavelets selected by AdaBoost (Viola and Jones, 2002) can model non-rigid objects under different lighting conditions. However, if multi-views are considered, a huge amount of data is needed for each view for the AdaBoost training (Jones and Viola, 2003). Such a huge dataset is impractical to create and it is also computationally expensive to train such a multi-view model. We use the projective warping of 2D wavelets to track 3D objects in non-frontal views in real time. Transformed 2D wavelets can approximate relatively flat object features such as the two eyes on a face. In this paper, approximations to the transformed wavelets using weak perspective projection are derived and errors introduced by the approximations are analyzed. Kalman filters are used to temporally confine the parameter space of the transform. As features in non-frontal views are computed on-the-fly by projective transforms under weak perspective projection, our framework requires only frontal-view training samples.