Facial Gesture Recognition

Wei-Kai Liao


Abstract

I will present my work on facial gesture recognition. In facial gesture recognition problem, different head poses will increase difficulty in recognition. We use a head pose estimation method based on iterative reweighted least-square to deal with this issue. After the recovery of head pose, we extract the dense flows effected by the expression change. Besides, the human face is divided into 9 regions according to its characteristics. In each region, an affine motion model is fitted based on those flows and parameters of the affine motion model are used for classification. The classification framework consists of graphical models and SVM. Instead of using single parametric distribution, we apply the Gaussian mixture to model the interactions between different regions, and the EM algorithm is used for training.


Maintained by Philippos Mordohai