Eye Tracking and Gaze Estimation: A Systematic Review of Deep Learning Methods and Educational Applications

Abstract

Tracking of the eyes and head have occupied a crucial role in human computer interaction, virtual and augmented reality, accessibility systems and cognitive studies. This paper gives a critical review on gaze estimation and head pose tracking algorithms, data sets, and performance determinants. This article compares classical and deep learning-based methods, such as convolutional neural networks, transformer models, geometric, hybrid, and multi-task learning methods. The article also tests the robustness in the external conditions like blinking, occlusion, and changes in illumination, and cross-subject individual differences. The most important publicly available datasets are compared regarding the type of input, diversity of participants, the size of data, recording distance, and environmental conditions. This article also contrasts new webcam and specialized eye-tracking systems, and, in doing so, point out trade-offs between accuracy, cost and applicability in real-time. The review also stresses on the role of domain adaptation, contrastive learning, and multi-modal inputs in generalizing across domains. In general, the article contains an attempt to present a comprehensive picture of the existing situation to researchers and practitioners to assist them in designing and choosing gaze and head pose estimation systems to be used in various applications.