Robust and accurate pupil detection for head-mounted eye tracking

Abstract In head-mounted eye tracking, robust and accurate pupil detection in real and mobile environments is still challenging, as pupil images are contaminated by various interferences, such as eyelids, eyelashes, and illumination. Based on the characteristics of pupil images, i.e., the pupil–iris contrast and the pupil darkness, the pupil region is represented by a horizontal weighted Haar-like feature. Based on this feature, a method called Compact Pupil Region Detection (CPRD) is proposed. Moreover, by adding the initial rectangle of the pupil region as prior knowledge, the CPRD-init method further improves the performance. Within the detected region, ellipse fitting is applied to extract the pupil ellipse. Experimental results on the benchmark database LPW show that the detection success ratio and 5-pixels error ratio increase by 7.1% and 3.9%, respectively, compared with state-of-the-art methods.