Development of an Appearance-based Eye Tracking System with Convolutional Neural Network Integrated in a Learning Management Application

• Main Authors: Ching, Jason Samuel, Cresencio, Alec Franz, Dee, Issey Gabriel, Domingo, Florenz Howard, Cabatuan, Melvin, Gutierrez, Alma Maria Jennifer
• Format: Conference Proceeding
• Language: English
• Subjects: Accuracy; Calibration; CNN (Convolutional Neural Networks); deep learning; Error analysis; Euclidean distance; eye tracking; gaze map; Gaze tracking; IoT (Internet of Things); Predictive models; User interfaces
• ISSN: 2770-0682
• DOI: 10.1109/HNICEM60674.2023.10589191

This study introduces an appearance-based eye tracking system, integrated into a Learning Management System (LMS) with IoT implementation, for monitoring students during online quizzes. Utilizing DenseNet121 as the base Convolutional Neural Network (CNN) model, the system employs a 13-point calibration process to refine gaze predictions according to user position. The LMS is designed to record the user's face images and generate gaze and fixation maps during quizzes, which professors can review after the quiz. The output also includes a collated view of students' gaze data. Extensive testing, including face detection model comparisons, led to the selection of DenseNet121 for the model, which demonstrated high accuracy with an average Euclidean distance of 21.86 mm and an angular error of 2.50°. In the ideal setup, the system achieved a notable average Euclidean distance of 10.67 mm and a gaze angle error of 1.22°. The Multi-Layer Perceptron (MLP) calibration model further enhanced accuracy, achieving as low as 5.52 mm in average Euclidean distance. Controlled experiments with 33 participants indicated robust performance, particularly under direct lighting and without eyeglasses, with an average angular error of 4°, surpassing the targeted 5° error rate. The system's user interface likewise received positive feedback, reflected in the System Usability Scale (SUS) scores of 78.86 (Good) and 89.58 (Excellent) from the students and professors, respectively.