It must be possible somehow, remember the new sony feature, that automaticly waits for a smile until it's gonna take a photo.
It's easier to track the eyes. THe mouth, as it was said before, has less significance. Additionally it depends on:
is the mouth always closed or always open, or is the person talking? THis would make it even harder. Second: does the program know where the mouth is located in the beginning, or has it to recognize the mouth "from scratch"? Third: how is the error tolerance? I mean, is this thing guiding nuclear weapons, or doesn't it matter to much if it's gonna take a sandwich for a mouth sometimes? 5. Is it always the same mouth? If so, are the light conditions always the same? Does the program have a mouth reference picture for comparation?
That would be the easiest case. Simply compare the reference mouth with blocks onscreen. scale it, rotate it and distort it, until it fits the rference pic. Compare the pixels with some tolerance parameters. Working with down-scaled versions of both (blocks and reference) allows fast(er) shape detection. Noless it's gonna be dead slow, a bit like encoding an Mpeg2 movie. (That is infact about the same)
It's easier to track the eyes. THe mouth, as it was said before, has less significance. Additionally it depends on:
is the mouth always closed or always open, or is the person talking? THis would make it even harder. Second: does the program know where the mouth is located in the beginning, or has it to recognize the mouth "from scratch"? Third: how is the error tolerance? I mean, is this thing guiding nuclear weapons, or doesn't it matter to much if it's gonna take a sandwich for a mouth sometimes? 5. Is it always the same mouth? If so, are the light conditions always the same? Does the program have a mouth reference picture for comparation?
That would be the easiest case. Simply compare the reference mouth with blocks onscreen. scale it, rotate it and distort it, until it fits the rference pic. Compare the pixels with some tolerance parameters. Working with down-scaled versions of both (blocks and reference) allows fast(er) shape detection. Noless it's gonna be dead slow, a bit like encoding an Mpeg2 movie. (That is infact about the same)