Robust techniques of gaze estimation, based on the eye centers localization and tracking and the exploitation of facial features
Loading...
Date
2023-11-01
Authors
Πουλόπουλος, Νικόλαος
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Nowadays, there is a growing interest in improving human-computer interaction (HCI) due to the penetration of computer systems in every aspect of everyday life. Eyes constitute the most characteristic features of the human face, as they provide valuable insights into human behavior and cognition. Eye gaze constitutes a revolutionary approach to interact without physical contact fostering a more natural and responsive interaction between humans and computers. The main objective of this thesis is to develop accurate and reliable methods for localizing the positions of the eye centers and estimating the eye gaze, using standard cameras under natural illumination, without the requirement of any dedicated equipment.
Despite the active research in this field, precise eye center localization and gaze estimation remain challenging problems due to many limitations related to the great variety in shape and color of human eyes, occlusions from hair or glasses, shadows and illumination variations, facial expression and head pose diversities, etc. and become even more challenging in applications where the need of real-time performance is crucial.
The methods developed for the localization of the eye centers aim to deal with the aforementioned degradations by introducing novel approaches and provide accurate results improving the state-of-the-art accuracy. A modified version of the Fast Radial Symmetry Transform, namely MFRST, that emphasizes on the shape of the iris and aims to find shapes with high radial symmetry, was introduced. Moreover, this shape-based transform was combined with a Convolutional Neural Network (CNN) in order to reduce the false detections of the MFRST and thus increase the overall performance. In the sequel, a novel end-to-end eye center localization architecture that performs image to heatmap regression between the eye regions and the corresponding heatmaps was introduced. Finally, a novel framework, dubbed PupilTAN, that tried to solve the eye localization problem in an unsupervised way was also presented, revealing a novel insight on solving the eye localization problem in the case of limited supervision.
Moreover, the gaze estimation problem is also extensively examined, providing solutions based on limited supervision by exploiting a large amount of unlabeled data. An unsupervised gaze representation learning approach based on gaze transfer that aims to disentangle the gaze feature, was introduced, and then, a gaze estimator was trained using few calibration samples.
All the proposed methods have been extensively evaluated in challenging datasets and compared to the state-of-the-art, revealing enhanced accuracy. This work has also led to novel insights on solving the eye localization and gaze estimation problems especially in the case of limited supervision, while the low complexity of the proposed techniques makes them a reliable solution for real-time applications.
Description
Keywords
Eye localization, Eye tracking, Gaze estimation, Facial features, Human-computer interaction, Deep neural networks