Non-rigid face tracking, which is the estimation of a quasi-dense set of facial features in each frame of a video stream, is a difficult problem for which modern approaches borrow ideas from a number of related fields, including computer vision, computational geometry, machine learning, and image processing. Non-rigidity here refers to the fact that relative distances between facial features vary between facial expression and across the population, and is distinct from face detection and tracking, which aims only to find the location of the face in each frame, rather than the configuration of facial features. Non-rigid face tracking is a popular research topic that has been pursued for over two decades, but it is only recently that various approaches have become robust enough, and processors fast enough, which makes the building of commercial applications possible.
Although commercial-grade face tracking can be highly sophisticated and pose a challenge even...