Computer Vision: Eye Tracking By: Geraud Campion Michael O’Connor
Frame Work A brief overview of eye tracking, formulas for image dissection, and some current applications of eye tracking.
Eye Tracking Why Eyes? Failures of facial recognition due to poor alignment Eye and eye movement are important to human interaction
Eye Tracking Current Approaches Visible Spectrum Cameras Near-Infra-Red cameras (NIRs) Work well in optimal conditions: fast and accurate Not so good otherwise: a lot of false positives Not a great help in the field of psychology or neurology
Another Technique Reflected Light from the Eye:
Reflected Light
Eye Tracking Search Methods Probabilistic Methods Bayesian Inference Model Key to this is that an image is cut into a collage of rectangles of arbitrary size
Eye Tracking Y is a random matrix y is a specific point A = {a 1, a 2, … a n } where a i is a rectangle in Y H = {H 1, H 2, … H n } is a random vector assigning each patch H i a value: 1 object of interest, -1 background, 0 not rendered
Eye Tracking All this leads to this formula P(H = 1 | y) = Σ [P(H=1) p(h | H i = 1)p(y | h i H i = 1) ]/ p(y) which is the probability that a portion y holds our object of interest:
Eye Tracking Situation Based Reference Make a hierarchy of “context dependent experts” Each expert uses probabilistic methods Then we use this formula: p(o|y) = ∫p(s|y) p(o|sy) dh Y – an observed image S – contextual situation O – location of left eye of the face on image
Applications Camera Mouse Eyebrow/blink patterns for clicking Driver Fatigue Detection (750 deaths, 20,000 injuries / yr from commercial vehicles) Detecting Amblyopia in Children Toys
Some Results One Person Two People Eyebrow Clicker VTOY