Face Tracking Using Motion-Guided Dynamic Template Matching

Combining two sophisticated techniques of motion detection and template matching, this paper proposes a simple but effective algorithm for detection and tracking of human faces. First, we use a statistical model of skin color and shape information to detect face in the first frame, and initialize it as an appearance-based intensity template for subsequent tracking. Second, incorporating background subtraction, projection histograms of moving silhouette and geometric constraints of body parts, we can quickly determine a good approximation of the search region corresponding to head location. Finally, a correlation-based template matching procedure is applied to further localize human face accurately, and current template can be dynamically updated in size and content to adapt temporal changes of the tracked face’s scale and orientation. Moreover, a confidence measure representing the template’s reliability is presented to guide possible template re-initialization for continuous face tracking. Experimental results demonstrate the validity of our proposed method.