Research Of Face Tracking And Expression Analysis In Real-Time Video Sequences

Human facial expression is a part of body language, which plays a crucial role in our daily exchanges. Facial expression analysis is to classify the specific people's facial expression state by computer and then to ascertain the subject's inner emotional or ideological activities, achieving smarter and more natural inter-action between human beings and computers. It is an important part of computer affective computing and has great value in video surveillance and other commercial applications, such as video phone, video conference, etc.The main content of this dissertation is to build a facial expression analysis system in accordance with the process as face detection, face tracking and expression analysis, and do research on the key issues of each step. Main work and innovation are as follows:(1) On the research of face detection and tracking, the dissertation achieves face detection by skin-color model and facial features validation, and focuses on the multiple-face matching in the process of tracking people. An algebra feature matching algorithm based on 2DPCA has been proposed, so as to solve the problems such as occlusion. Experiment shows that the method can achieve good results in speed and accuracy.(2) On the research of facial expression analysis, in the view of real-time video circumstance, this dissertation adopts DCT transform to reduce the whole face image dimension and obtains effective feature, which greatly reduces the dimension of the observation value sequence and lowers the complexity of training and recognition; Later, this dissertation designs a facial expression analysis algorithm based on Hidden Markov Model. We select the differences of 2D-DCT coefficients between expressive and neutral face images, which generate observation value sequence by ZIG-ZAG scanning and K-means clustering; ultimately the expressions are classified as six kinds: happy, anger, surprise, disgust, fear and sadness. Experiment shows that the method proposed by this dissertation achieves higher recognition rate than the PCA method in both single person and multiple person experiment.(3) On the research of the characteristics and the software development of TI TMS320C6000 series DSP, this dissertation transplant and optimize the code in DSP, and project is executed successfully on CCS simulator.