To classify each frame as the onset, a temporal model of HMM-SVM to determine blink event. They found that this temporal model and the intensity feature were significantly better than using multi-class SVM and HOG, LBP, Gabor and optic flow features to detect blink. The reported blink detection performance among previous studies is usually high, above 90% in accuracy. On one hand, most of these results were for low resolution non-IR eye images, and it is questionable whether close-up IR images containing a variety of detailed eye appearance changes can be classified simply into open and closed eye states.
On the other hand, most datasets for evaluation were collected during leisure scenarios where voluntary blink often occurred, evidenced by completely closed eyes. However, during Spain phone number list tasks, most blinks are partial and most full blinks have the lids approaching each other but not necessarily touching (Brosch et al., 2017). These involuntary blinks may have different duration, amplitude and speed ratio of eye closing and opening to voluntary blinks (Abe et al., 2014). Pupil Contour Estimation Since the pupil is only distinct under IR illumination and its size can be accurately measured only when the resolution of pupil is good enough, pupil detection is always conducted on IR eye images.
Estimated pupil contour can be fitted by an ellipse model to obtain the pupil center, pupil size, and other features of interest. One straightforward approach is to segment the pupil from the background through binarization, however, it is challenging to find an adaptive threshold for a variety of eye images with large variations. Chen and Epps (2014) proposed a self-tuning threshold method which requires minimum parameter to tune to handle these variations for near-filed IR images. Other algorithms operate on remote IR images, with the aim of pupil center detection.