Compute-efficient eye state detection: algorithm, dataset and evaluations

Eye state can be used as an important cue to monitor the wellness of a patient. In this paper, we propose a computationally efficient eye state detection technique in the context of patient monitoring. The proposed method uses weighted accumulations of intensity and gradients, along with a color thresholding on a reduced set of pixels to extract the various features of the eye, which in turn are used for inferring the eye state. Additionally, we present a dataset of 2500 images that was created for evaluating the proposed technique. The method was shown to effectively differentiate open, closed and half-closed eye states with an accuracy of 91.3% when evaluated on the dataset. The computational cost of the proposed technique is evaluated and is shown to achieve about 67% savings with respect to the state of art.