Abstract
In the global village of information, some virtualized technologies are now in bud. A technology called virtual meeting is still being developed. People from every part of the world can participate in the meeting through network communication on his or her computer at office. It is a good news to our busy life. However, virtual meeting needs a 3D virtual environment and adopts 3D facial models for character representation. The head motions of each user will be mapped to a 3D constructed head model. Therefore, it is an important topic to control the 3D head model by extracting motion parameters from human facial action units in the video sequence. Owing to the popularization of multimedia applications, the ISO MPEG-4 Committee is carrying on the standardization for multimedia communications .One essential issue of the standard is the synthetic and natural hybrid coding(SNHC),which includes 2D images and 3D objects like human bodies. According to the SNHC, a face object is defined with two kinds of parameters: the Facial Definition Parameters (FDPs) and the Facial Animation Parameters (FAPs). The FDPs are used to customize the facial model to a particular face while the FAPs are used for representation of most natural facial expressions. Applying these parameters on a virtual meeting can effectively achieve high data compression and enable augmented reality. Before abstracting face expressions to FAP parameters, it is necessary to find out the positions of each facial feature point. Hence, to track each feature point, and estimate head orientation is important pre-processing. In this thesis we mainly discuss about these two problems. It exists a relation, however, between position of face feature points and head orientation. We combine these problems in this thesis and try to find a solution for it. In addition, the aid of Kalman filter offers large improvements of the ability in tracking and estimation.