Abstract
The study aims to transform Convolutional Neural Networks (CNNs) for eye gaze estimation and prediction, providing relevant data on the limitations of traditional gaze tracking systems, which are often constrained by limited environments and expensive equipment. The authors propose a dual-task approach, where gaze estimation and gaze prediction are separated to enable more granular improvements in each process. Using the MPII Gaze dataset, collected under real-life conditions, various CNN architectures such as YOLOv3, SSD, and Mask R-CNN are evaluated and compared based on accuracy, precision, recall, and F1-measure. Each unique spatiotemporal sequence of eye images is utilized to enhance the predictive power of individual frames, allowing the model to identify temporal patterns and improve estimation through gaze continuity. Additional measures to increase model robustness and responsiveness include image normalization, region-of-interest extraction during preprocessing, and a geometric features-based blink detection mechanism. The results demonstrate that deep learning models can effectively improve gaze estimation accuracy under varying lighting conditions, head movements, and user diversity. This makes the technology applicable in fields such as education, medicine, automotive safety, adaptation, assistive technologies, and human-computer interaction. Overall, this research contributes to the development of scalable, adaptable, and precise gaze-tracking algorithms utilizing state-of-the-art automated learning methods, offering valuable insights for researchers in the field.