VisionGuard: AI Online Proctoring System Based on Computer Vision

Core Overview of VisionGuard

VisionGuard is an open-source AI online proctoring system released by developer SAKSHI-221 on GitHub (June 16, 2026). It uses technologies like MediaPipe Face Mesh, head pose estimation, and eye tracking to monitor examinees' suspicious behaviors in real time, addressing the integrity issues of remote exams. The system has advantages such as open-source transparency and local deployment for privacy protection, providing intelligent proctoring solutions for educational institutions and certification organizations.

Integrity Dilemma of Remote Exams

The popularization of online education has made remote exams an important assessment method, but examinee cheating (such as consulting materials, proxy testing, etc.) undermines fairness.

Limitations of Traditional Proctoring

Manual video proctoring has limited coverage, is prone to fatigue, and involves subjective judgment, making it difficult to meet the needs of large-scale exams. An intelligent automatic proctoring solution is urgently needed.

System Positioning

An open-source AI proctoring framework whose core concept is to identify cheating behaviors under privacy protection. Compared with commercial software, it has advantages like transparent code, customizability, and local deployment.

Core Technologies

• MediaPipe Face Mesh: Real-time detection of 468 3D facial key points;
• Head pose estimation: Calculate head pitch/yaw/roll angles and detect abnormal head turns;
• Eye tracking: Analyze eye position to estimate gaze direction and identify gaze deviation;
• Real-time anomaly detection: Identify gaze deviation, abnormal head rotation, face disappearance, multi-person detection, abnormal movements, etc.

Modular Architecture

1. Video Capture Module: Obtain real-time video streams and preprocess them;
2. Perception Engine: Run MediaPipe models to output facial key points;
3. Behavior Analysis Module: Calculate posture/eye movement trajectories and count abnormal patterns;
4. Alarm and Report Module: Trigger alarms (including screenshots) and generate post-exam reports for manual review.

Main Application Scenarios

• Higher Education: Integrate with LMS to monitor midterm/final/qualification exams;
• Professional Certification: Reduce labor costs and improve monitoring consistency;
• Corporate Training: Ensure employees complete tests seriously and maintain training effectiveness.

Practical Value

AI-assisted proctoring improves coverage and efficiency, reducing manual burden.

Privacy Protection Measures

• Local Processing Priority: Reduce cloud data uploads;
• Data Minimization: Only store screenshots of abnormal events and metadata;
• Transparent Notification: Clearly inform examinees of the type of monitoring;
• Manual Review: AI results are for reference, and final judgments are made by manual review. These measures balance exam integrity and examinee privacy.

Technical Implementation

Developed in Python, relying on OpenCV (video processing) and MediaPipe (facial detection), with installation and configuration guides provided.

Usage Methods

• Command-line Launch: Specify camera, output directory, and sensitivity;
• Operation Modes: Support real-time monitoring and offline video analysis;
• Secondary Development: Provide modular API interfaces for easy integration or function expansion.

Project Significance

The open-source nature provides a platform for joint improvement by academia and industry, promoting transparency and standardization of online proctoring technology.

Future Outlook

Technology needs to be combined with institutional design, clear rules, and humanized implementation, allowing educators to focus on teaching rather than monitoring. VisionGuard provides technical support for maintaining exam integrity.