For example, for near-field IR eye images, apart from declaring a blink when there was no pupil blob, Chen and Epps (2014) employed two ellipse fittings to the detected pupil blob. One was fitted to the whole pupil contour, and the other was fitted to the bottom half contour, so that blink was determined by the degree of pupil occlusion by the eyelid. However, these methods require prior knowledge of pupil or iris blob and are not suitable for cases when the bottom of the blob is occluded. The second is based on the eyelid distance with a pre-defined threshold the ratio between the eye width and its aperture. However, the decision for the threshold may be especially difficult for the fine-grained eyelid trajectory.
The third involves eye motion. Mohanakrishnan et al. (2013) proposed motion vector difference in the face region and eye region to detect blink, since the motion vector Spain phone number list in the eye region can be “random” during a blink, while it is typically similar to the face region when the eyelid is still relative to the face. Instead of determining a similarity threshold, Appel et al. (2016) extracted features from the intensity difference of two adjacent frames specifically for near-field IR images. However, these methods are probably not able to detect long blinks since during these, the eyelid also does not move.
In the latest work, Fogelton and Benesova (2018) used motion vectors in the eye region and learned a sequence model of blink using blink completeness and achieved similar or slightly better performance than other methods in four datasets. The last detects blink directly from the eye appearance in images. Yahyavi et al. (2016) employed PCA and artificial neural networks for open and closed eye images. Mohammadi et al. (2015) detected the open and closed eye states by finding an appropriate threshold for the intensity change. Bacivarov et al. (2008) used the AAM parameters' difference in open and closed eye images to identify blink.