Abstract
Eye tracking spreads through a vast area of applications from ophthalmology,assistive technologies to gaming and virtual reality. Precisely detecting thepupil's contour and center is the very first step in many of these tasks, henceneeds to be performed accurately. Although detection of pupil is a simpleproblem when it is entirely visible; occlusions and oblique view anglescomplicate the solution. In this study, we propose APPD, an adaptive andprecise pupil boundary detection method that is able to infer whether entirepupil is in clearly visible by a heuristic that estimates the shape of acontour in a computationally efficient way. Thus, a faster detection isperformed with the assumption of no occlusions. If the heuristic fails, a morecomprehensive search among extracted image features is executed to maintainaccuracy. Furthermore, the algorithm can find out if there is no pupil as anhelpful information for many applications. We provide a dataset containing 3904high resolution eye images collected from 12 subjects and perform an extensiveset of experiments to obtain quantitative results in terms of accuracy,localization and timing. The proposed method outperforms three other state ofthe art algorithms and has an average execution time $\sim$5 ms insingle-thread on a standard laptop computer for 720p images.