Online Proctoring Systems - A Hardware Implementation using Enhanced Eye Detection and Gaze Tracking

Online Proctoring Systems(OPS) have been on the rise due to the increasing dependence on video communication. With an important application in exam monitoring, these systems ensure maintenance of academic integrity by utilizing Artificial Intelligence and Computer Vision to detect suspicious behavior and automate the process of proctoring. Each OPS consists of 4 processes in the following order - Feature Extraction, Gaze Tracking and Estimation. Environment modeling and finally Inferential/ Prediction models for behavioral analysis. In this paper, we discuss the two most integral parts of any OPS, namely, Feature Extraction and Gaze Tracking. We compare some state-of-the-art methods available in both aspects of an OPS and attempt to clone them in a hardware setting. We indicate how deep learning-based methods offer more robust detection and tracking compared to template-matching-based classification. We show that CNN-based Deep Learning methods tend to exhibit higher accuracy and robustness to noise in detecting eye-features such as the pupil and iris as compared to Haar-based Cascading classifiers, which are a decent alternative for lightweight real-time applications in a controlled setting with minimal noise. For the second process in an OPS, we describe that Geometric and Camera Calibration is unmatched for gaze-tracking in controlled and constrained environment, while Optical Flow provides an effective and adaptable way to track gaze.