Meticulous Acquisition System for Tracking User’s Natural Kinetics (MAS TUNK): An Approach in Eye Tracking Dataset Collection for Neural Network Training
Muhammad Arief Nugroho, Maman Abdurohman, Bayu Erfianto, Mahmud Dwi Sulistiyo · 2024
We propose MAS TUNK, a novel data collection framework for eye-tracking, designed to enhance the future development of automatic proctoring systems. Confronting the constraints of traditional eye-tracking methods reliant on specialized hardware, MAS TUNK employs a game-based strategy using standard laptop webcams to amass a more diverse and natural set of eye movement data across different user environments. Our collected dataset features full face frames, eye-cropping regions, iris Euclidean distances, headpose estimations, and mouse coordinates, all timestamped to ensure synchronized data capture. We describe the development and assessment of a Convolutional Neural Network (CNN) model trained on this dataset, which exhibits high precision in gaze estimation. Results show the model achieving training and testing accuracies of 91.5% and 94.5%, respectively, with RMSE values of 0.059 and 0.054. The losses were minimized to 0.003 for both datasets, signifying the model’s aptitude for accurate gaze prediction. While the present study focuses on dataset creation and model testing, the findings provide a solid foundation for the future application of such models in automatic proctoring and impersonation detection.