This repository contains the code for our project: Automated Online Proctoring Using Gaze View Tracking and Custom Object Detection
Due to COVID-19, there’s been a burst in online examinations. The main integrity safeguard so far is human proctoring, which requires a trained supervisor to constantly monitor each test-taker's video and audio through a webcam. To overcome such costliness and ineffectiveness, we have designed an automated online proctoring system that is effective, non-intrusive, and adaptable to different testing scenarios. Our approach presents a novel combination of (a) a gaze view tracking module using a mathematical 3D gaze motion formula and (b) a configurable cheating classification using a custom-trained object detection model. Our gaze view tracking leverages two cameras (a webcam and a follower-cam), working in tandem, and facial landmark detection to follow the test-taker’s gaze in a real-time and non-intrusive manner. It then feeds to our AI-based cheating classifier, which leverages TensorFlow object detection algorithm to identify pre-configured cheating targets. As a continuing project, this year we have revamped the 2D gaze approximation algorithm used in last year’s gaze view tracking and derived a mathematical formula to capture 3D gaze motion, allowing us to greatly enhance the gaze view tracking capability without cumbersome calibration. We also completed the configurable cheating detection module and integrated it with the gaze view tracking module. By calculating accuracy and latency of our end-to-end prototype through various trials, our system shows effectiveness in tracking the test-taker’s gaze and auto-detecting cheating targets. It serves as a great complement to the current online proctor suite and will influence online learning even after pandemics by reducing human toil.