Facial action tracking using particle filters and active appearance models

Tracking a face and its facial features in a video sequence is a challenging problem in computer vision. In this view, we propose a stochastic tracking system based on a particle- filtering scheme. In this paradigm, the unobserved state includes global face pose and appearance parameters coding both shape and texture information of the face. The adopted observations distribution is derived from an Active Appearance Model (AAM). The transition distribution and the particles number are adaptive in the sense that they are guided by an AAM deterministic search. This optimization stage adjusts the explored area of the state space to the quality of the prediction and enables a substantial gain in computing time. The observation model uses a robust distance measure in order to account for occlusions. Experiments on real video show encouraging results.