<![CDATA[Eye gaze tracking serves as a pivotal instrument in realizing adaptive Human-Centric AI interactions within the Industry 5.0 era. However, current appearance-based gaze estimation methods primarily rely on large-scale end-to-end deep learning regression architectures characterized by high network complexity and computational overhead. This research proposes a hybrid framework integrating the YOLOv8-Pose computer vision model with the Random Forest Classifier machine learning algorithm for the spatial classification of gaze directions. Raw image pixel matrices are efficiently reduced by localizing 55 ocular keypoints, which are subsequently extracted into 113 tabular geometric feature attributes, encompassing Eye Aspect Ratio (EAR) parameters and the relative spatial shifts of the iris. The ensemble model was trained utilizing a synthetic dataset generated via the UnityEyes simulator. Experimental results demonstrate that the proposed hybrid system achieved a global accuracy of 82.75%. The highest feature space discrimination capabilities were recorded in the TopRight (F1-score 0.91) and MiddleRight (F1-score 0.85) spatial classes. This structured feature engineering approach is empirically proven to effectively mitigate anisotropic classification errors without necessitating high-end computational resources. For further development, future research is directed towards exploring aggressive hyperparameter tuning, meta-heuristic feature selection using Particle Swarm Optimization (PSO), and evaluating the model's robustness limit using real-world datasets in unconstrained open domains.]]>
Copyrights © 2026
