Fachbereich Informatik

Fachbereich Informatik - Aktuell


Disputation Thiago Caberlon Santini

am Montag, 12. August 2019, um 13:00 Uhr in Raum A301, Sand 14


Towards Ubiquitous Wearable Eye Tracking

Berichterstatter 1: Prof. Dr. Enkelejda Kasneci
Berichterstatter 2: Prof. Dr. Wolfgang Rosenstiel

It is often said that the eyes are the window to the soul. While science might never be able to prove
or disprove the mere existence of a soul, it has nonetheless found strong evidence that gaze is the
window to the mind. Not only has eye tracking played a key part in these findings, it also holds
immeasurable potential to further advance our understanding of the human mind and revolutionize
the way we interact with our devices. However, to fully realize this potential, it is imperative to bring
eye tracking out of the laboratory and into the wild. Towards enabling ubiquitous wearable eye
tracking, this thesis deals with the challenges involved in this transference, focusing on headmounted
video-based devices. At the feature extraction level, this work introduces novel methods to
reliably and robustly detect and track the pupil in real-time, proposing advanced metrics for their
evaluation, as well as exploring the influence of such methods on the resulting estimated gaze
signal. At the gaze estimation level, the thesis deals with two of the main factors hindering a wider
eye tracking adoption: Calibration and Slippage. To improve calibration usability, a innovative
approach is proposed, which enables the collection of plentiful calibration points quickly and
unsupervisedly virtually anywhere. Slippage is tackled by a novel hybrid method combining a
geometrical slippage-robust feature and traditional regression-based gaze-mapping functions. At the
eye movement detection level, a probabilistic approach is proposed to identify fixations, saccades,
and smooth pursuits at lower sampling rates. The methods herein introduced form the basis of a
complete state-of-the-art eye-tracking platform competitive with expensive commercial eye-tracking
systems. Unlike these commercial systems, the platform is open-source, able to track participants
outdoors and with glasses, and requires no additional hardware such as stereo eye cameras or
multi-glint patterns. The effectiveness of the resulting system is shown through a large-scale,
pervasive, and unconstrained eye-tracking study.