The thing is that this is actually possible with the technology we have these days.

Look at some of the works done by Iain Matthews and Simon Baker at CMU here:

http://www.ri.cmu.edu/research_project_detail.html?project_i...

and you can see that they have already realized realtime face tracking with dense registration ( which allows expression recognition ) based on the LK-inverse composition algorithm applied to AAMs

Facinating stuff =)