Facial expression is the most natural and intuitive way between peoples’s non-verbal communication. Expression based nutural human-computer interaction will be one of the important development directions. This thesis focuses on the problems of face alignment and face tracking, and implements an expression cloning system. The main contributions of this thesis are as follows: 1. Studied the AAM-based face alignment problem and improved the texture representation and fitting algorithm. The AAM based on inverse compositional image alignment algorithm is very fast. However, it is sentitive to initial shape and has poor generalization performance. Moreover, it may generate an unallowable shape for the mouth. To address these problems, some improvements are proposed: (1) Proposed a two-stage fitting algorithm. At the first stage, the global affine transform parameters are updated to achieve a better overall initial shape. Then, at the second stage, all shape parameters are updated to obtain the final shape; (2) Proposed a multi-band texture representations: the intensity, x-direction gradient strength and y-direction strength, which are complementary and stable. The improved appearance model is more robust to illumination variations and has better generalization performance to unseen images. (3) Proposed a local shape constraint based on a penalty function to prevent the unlikely mouth shape which happens frequently in expression interactoin applications. 2. Studied the problem of 3D head pose and facial actions tracking and proposed several improvements for 2D+3D AAM according to face tracking applications. Firstly, the improved 2D AAM is extended to the 2D+3D AAM case and a deformable 3D face model-Candide-3 is adopted to extract head pose and facial action parameters. Secondly, several improvements are proposed in tracking environments: (1) Local matched feature points between successive frames are used to estimate the initial shape, and the resulting shape is closer to the ground-truth shape hence improves the stability in tracking fast face motions; (2) View-based AAM is adopted to handle large angles of head rotation; (3) A person-specific appearance model is incorporated into AAM fitting and improves the stability and accuracy. 3. Studied the problems of inter-frame information utlization and background interference and proposed AAM based face tracking with temporal matching and face segmentation constraints. By incorporating a temporal matching const...
修改评论