Face plays an important role in human-machine interaction like communication between humans for it can give many information on user. And due to this aspect many application fields of tracking face grow up over and over. But it is not easy to track face in real-time due to many problems such as variation of illumination, scale, partial occlusion and so on.
In this thesis, an algorithm for real-time face tracker is proposed. The proposed method utilizes color information to detect face region in clutter. Condensation algorithm using color as feature performs detecting skin-color region and this region considered as face candidate. To verify face region, three factors are considered: face aspect ratio, the location of eyes, nose, and mouth, and relative position of skin-color region in image. To remove the influence of variation of illumination on color information, HSV color space which separates pure color information from illumination is used. But big change of illumination causes a little variation of pure color information. To adapt this change, fuzzy inference engin to adjust hue element is introduced. Face aspect is also estimated to infer attention of human using the relative distance between the location of eyes, nose, and mouth and the center position of face region.
To implement the proposed algorithm and track face actively, a pan-tilt unit is devised. This active camera is developed for a vision system of human-friendly interactive robot. The proposed method is implemented using this active camera and shows real-time performance and robustness to variation of illumination, scale change and face aspect is also estimated successfully. Finally simple interaction method using face aspect and robot motion is developed for interaction between human user and interactive robot and this shows the usefulness of proposed algorithm.