Watchfull-Eyes: A Gaze Based Tracking System for Exam Integrity Using Computer Vision and Machine Learning
Nayana G B1, Prajwal K V2, Shashank S3, V Ronald Supreeth4, Mrs. Manjula H5
1Department of Computer Science and Engineering, PES Institute of Technology and Management
2Department of Computer Science and Engineering, PES Institute of Technology and Management
3Department of Computer Science and Engineering, PES Institute of Technology and Management
4Department of Computer Science and Engineering, PES Institute of Technology and Management
5Department of Computer Science and Engineering, PES Institute of Technology and Management
Abstract - Online education platforms face significant challenges in maintaining exam integrity and detecting fraudulent behavior. This work presents WatchFull Eyes, a comprehensive gaze-based exam proctoring system that leverages computer vision, deep learning, and real-time analytics to monitor student behavior during online examinations. The system integrates MediaPipe Face Mesh for accurate eye-gaze tracking with calibration-based coordinate mapping, COCO-SSD (MobileNetV2) object detection for violation identification, and advanced heatmap visualization for focus analysis. The full-stack application is built using React (frontend), Django REST Framework (backend), and SQLite (database), operating entirely offline to ensure data privacy and institutional compliance. Seven regression algorithms were evaluated for focus score prediction, with Ridge Regression achieving the highest accuracy (R2 = 0.87, RMSE = 4.23) on validation data. The system detects six violation categories: unauthorized devices, additional persons, tab switching, head deviation, eye closure, and excessive gaze deviation. Field testing across 50 student sessions demonstrated 94.2% precision in violation detection and 99.8% system uptime. This work contributes a practical, privacy-preserving framework for academic proctoring and demonstrates the effectiveness of browser-based AI integration for real-time behavioral monitoring without external dependencies.
Key Words: groundwater recharge, machine learning, neural networks, water management, hybrid approach, spatial analysis
