Real time Gaze Tracking in Remote Proctoring - A Study of Appearance-based Gaze Estimation

Abstract

Our thesis aims to address the critical issue of academic dishonesty in online examinations by proposing a proctoring system that integrates eye gaze tracking technology for the detection of suspicious behavior. The study begins by discussing the existing challenges of current examination systems and identifying the problems that need to be addressed. It emphasizes the necessity for a more advanced proctoring system with gaze tracking capabilities to effectively deter attempts at academic dishonesty. The research is divided into two main parts: the selection of an appropriate model and the incorporation of proctoring functionalities. Two models were chosen for evaluation, namely iTracker, which was pre-trained on the GazeCapture dataset, and L2cs-net, which we trained on the MPIIFaceGaze dataset. The findings from these experiments indicate that L2cs-net outperforms iTracker in terms of accuracy, speed, and latency but only when supplied with the processing power of a GPU, without one iTracker is better. Regarding the proctoring system aspect, it is noted that most of the existing research is commercially driven, with limited academic contributions. To optimize the proctoring system for online exams, we recognize the significant value of examinees’ eye gaze and define important regions on and off the screen through calibration using “magic pixels”. Moreover, we attribute cheating criteria using a formulated equation that takes into account factors such as Count, Frequency, Duration, and Regression. Two potential approaches for the proctoring system, namely Thresholding and Machine Learning (ML), are considered. However, our focus lies on the development of a thresholding-based approach. Overall, this thesis presents a comprehensive exploration of academic dishonesty in online examinations, proposes a proctoring system using eye gaze tracking technology, and compares the performance of different models and methodologies. The findings contribute to the advancement of proctoring systems and provide insights for the development of more effective measures against academic dishonesty.