Machine Learning-Driven Image Noise Robustness Enhancement for Reliable Gaze Estimation

This research paper presents the development of a novel Human-Computer Interaction (HCI) system that enables users to control a computer mouse through eye movements by utilizing machine learning techniques. Traditional input devices, such as a mouse or keyboard, often present accessibility challenges for individuals with physical disabilities or limited mobility. To address this problem, I proposed machine learning-driven gaze estimation system for human computer interaction for quadriplegia patients. The proposed system takes both eye images as input to predict gaze vector. The system interprets real-time eye movement data to execute cursor movements which offers an alternative and inclusive method of computer interaction. To improve the accuracy of the gaze estimation system, I introduced a random gaussian noise-based denoising autoencoder. Experimental results demonstrate that this approach significantly enhanced accuracy reducing the angular error by 4.7 degrees on a public gaze estimation dataset. Additionally, I conducted real-world experiments to evaluate the model's performance in tracking the gaze of actual users. The results indicate that the proposed model achieved an accuracy of 91%, representing an enhancement of up to 15% compared to state-of-the-art gaze estimation methods.