Face tracking - the continuous monitoring of head position, orientation, and geometry - has numerous practical applications for human-computer interaction, such as a perceptual form of multi-modal input.
There are several non-invasive and computationally inexpensive techniques for face tracking that draw upon algorithms from computer vision. Of them, Bradski's CAMSHIFT algorithm is appealing because it requires minimal training. These techniques are particularly attractive in light of the growing installed base of fast desktop computers and cheap, low-end desktop digital video cameras. Low-end cameras, however, have characteristics that make them a poor fit for some such face tracking algorithms. In this paper, I introduce the problem of face tracking, provide an overview of the operation of CAMSHIFT as an example of a non-invasive vision-based face tracking algorithms, and describe my experiences attempting to employ video obtained from a low-end desktop digital video camera source in face tracking. I conclude this paper by offering conclusions and recommendations drawn upon my experiences.