Once a face is selected, the face tracker starts the tracking process. During this process, the skin color model is used to find the face within the search window. The motion estimation and prediction are then based on the search result. The pan, tilt, and zoom of the camera are adjusted if needed. The skin-color model is updated in real-time based on the new
estimated parameters. If the tracking fails to find the face, the search window size will increase until the face is found
again. The face tracker can continuously track a person while he/she is moving freely (e.g., sitting, rising, walking).
The system has been running in our lab for about a year with continuous improvements in performance. The current tracking speed using an HP-9000 workstation is shown in table 2. The table suggests that the tracking speed greatly depends on the search window size. For example, when the face is closer to the camera, the face image is relatively bigger and so is the search window size.