Eye Blink Detection Using Variance of Motion Vectors

A new eye blink detection algorithm is proposed. It is based on analyzing the variance of the vertical component of motion vectors in the eye region. Face and eyes are detected with Viola – Jones type algorithm. Next, a grid of points is placed over the eye regions and tracked with a KLT tracker. Eye regions are divided into 3×3 cells. For each cell an average motion vector is estimated from motion vectors of the individual tracked points. Simple state machines are setup to analyse these variances for each eye. The solution is this way more robust and with a lower false positive rate compared to other methods based on tracking. We achieve the best results on the Talking face dataset (mean accuracy 99%) and state-of-the-art results on the ZJU dataset.