AAM-based Localization Of Facial Feature And Face Recognition

Active Appearance Model (AAM), as an important feature extraction algorithm, was proposed by F. T. Cootes et al. in 1998 and has been employed to localization of facial feature. Due to its flexible structure and excellent performance, AAM has been extensively applied to face image processing, such as face detection, face recognition, face tracking and face animation. The paper mainly establishes a systematic software platform, which is used to label facial feature based on AAM automatically. In addition, a series of applications on pattern recognition have been embedded into the system. Firstly, the paper studies a facial tracking and recognition algorithm based on AAM. Through combining Lucas-Kanade (LK) with AAM, the facial tracking and recognition algorithm are implemented. It takes advantage of the fast fitting efficiency of LK and model-building ability of AAM, which could label facial feature with high-speed. The experimental results illustrate that this algorithm could accurately fit and track the face in the video.Secondly, an AAM fitting algorithm is studied from the point of view of classification. Combining AdaBoost and AAM, this algorithm requires building up the Point Distribution Model (PDM) and Boosting Appearance Model (BAM) for training multi-weak classifiers firstly. The optimal parameters got from the final strong classifier are applied to AAM fitting. The experimental results illustrate that this algorithm converts AAM fitting into the question of classification and could fit the faces accurately and efficiently.Finally, the facial video tracking and recognition systematic platform based on AAM is developed with Matlab GUI. Two model fitting algorithms based on LK and AdaBost are embedded into the platform. A series of function modules are integrated in it, such as landmarks labeling, AAM building, AAM fitting, video tracking and fitting information saving, etc. Then, various tasks especially for feature fitting and facial tracking and recognition based on AAM could be implemented efficiently. In conclusion, this platform lays solid foundation for the further researches based on AAM.