A real-time method for facial detection, tracking, and recognition.

摘要

While there are many current approaches to solving the difficulties that come with detecting, tracking, and recognizing a given face in a video sequence, the difficulties arising when there are differences in pose, facial expression, orientation, lighting, scaling, and location remain an open research problem. In this thesis we perform the study and analysis of an approach for each of the three processes, namely a template face detection, tracking, and recognition. In the face detection approach we detect a given face by finding the pupils of the eyes within a given face. Then in the tracking approach, the face is tracked by searching within a region in the face for the eyes. Finally, in the recognition approach, the face is recognized by scaling and rotating the template image and comparing it to reference face images in a given database. The proposed algorithms are faster relatively to other existing iterative methods. Unlike such iterative methods, in our proposed method we do not estimate the face rotation angle or scaling factor by looking into all possible face rotations or scaling factors. In particular, in the proposed tracking/recognition method, we take a vector distance between the two eyes in a given face to estimate the face rotation and scaling factor relatively to the image coordinate system. This is done once and for each frame. The reference face images in the database are normalized with respect to face translation, rotation, and scaling. We show here how the proposed method to estimate a given face image template rotation and scaling factor leads to real-time template image rotation and scaling corrections. This allows the recognition algorithm to be less computationally complex than iterative methods.