Agentic AI Smart Exam Surveillance and Physical Alert System

Abstract

Academic examination integrity is a critical concern in modern educational institutions. Conventional proctoring methods are labour- intensive, inconsistent, and fail to monitor all students simultaneously in large examination halls. This paper presents an Agentic AI Smart Exam Surveillance and Physical Alert System—a real-time, multi-agent framework that automates the detection of cheating behaviours using computer vision and IoT hardware integration. The system deploys YOLOv11 for object detection of prohibited items (mobile phones, calculators, chit papers) and ByteTrack for stable multi-object tracking across video frames. Head pose estimation via MediaPipe Face Mesh employs the Perspective-n-Point (PnP) algorithm to derive Yaw/Pitch/Roll angles, enabling detection of head-turn and copying behaviours. Temporal logic filters requiring N-frame persistence eliminate false positives from transient movements. Eleven specialised AI agents cooperate through a shared-memory pipeline to deliver end-to-end examination monitoring. Upon detecting suspicious activity beyond configurable risk thresholds, the system captures timestamped evidence and triggers physical alerts via an Arduino-controlled servo-mounted laser pointer and buzzer. A Flask-based web dashboard provides invigilators with real-time visibility of student risk scores, alert history, and captured evidence. Experimental evaluation demonstrates real-time inference exceeding 30 FPS with mAP@50 of 91.0% on the custom examination dataset.

Introduction

The text presents a real-time AI-based examination surveillance system designed to detect and prevent cheating in exam halls where human invigilators are often insufficient. It argues that traditional monitoring methods struggle to scale and miss subtle or fast cheating behaviours such as using mobile phones, looking at others’ papers, or exchanging signals.

The proposed solution combines deep learning, computer vision, and multi-agent AI. It uses YOLOv11 for real-time object detection (e.g., phones, chits, calculators), ByteTrack for tracking students across frames, and MediaPipe Face Mesh for 3D head pose estimation to detect suspicious gaze and head movements. These components work together within an eleven-agent architecture, where each agent handles a specific task such as detection, tracking, behaviour analysis, risk scoring, decision-making, and alert generation.

A key feature is a risk scoring system that continuously evaluates each student based on detected objects and behaviours. Actions like using a mobile phone or sustained head turning increase the score, triggering escalating alerts. To reduce false positives, the system uses temporal logic filtering, meaning suspicious behaviour must persist across multiple frames before an alert is issued.

The system also includes an IoT-based response mechanism, where alerts can activate physical devices like buzzers or lasers for immediate deterrence, along with a dashboard for real-time monitoring and evidence recording.

The literature review shows that earlier systems relied on rule-based or basic deep learning methods, but lacked integration of object detection, pose estimation, and real-time response. More recent models like YOLOv11 and MediaPipe improve accuracy and speed, while multi-agent systems enhance modularity and scalability.

The methodology involves training a custom dataset of exam-related objects, using a YOLOv11 nano model optimized for real-time performance. Evaluation shows strong accuracy (around 91% mAP@50), with results demonstrating fast detection and effective cheating classification.

Conclusion

This paper presented an Agentic AI Smart Exam Surveillance and Physical Alert System combining YOLOv11 object detection, ByteTrack multi-object tracking, MediaPipe head pose estimation, and an eleven-agent multi-agent architecture. The system achieved 91.0% mAP@50 on the custom examination dataset and sustains 30 FPS display throughput with sub-150 ms alert latency on GPU. Head pose estimation via PnP and Rodrigues formula enables precise Yaw/Pitch/Roll-based behaviour classification, while temporal logic N-frame filtering eliminates false positives. The Arduino-integrated hardware alert subsystem provides on-site deterrence without cloud connectivity. Future work will investigate lip-movement detection, federated learning for privacy-preserving model improvement, and adaptation to online examination environments.

References

[1] Anish, S. R, A. H. Malini and T. Archana, \"Enhancing Surveillance Systems with YOLO Algorithm for Real-Time Object Detection and Tracking,\" 2023 2nd International Conference on Automation, Computing and Renewable Systems (ICACRS), Pudukkottai, India, 2023, pp. 1254-1257, doi: 10.1109/ICACRS58579.2023.10404710. [2] X. Yu, T. W. Kuan, Y. Zhang and T. Yan, \"YOLO v5 for SDSB Distant Tiny Object Detection,\" 2022 10th International Conference on Orange Technology (ICOT), Shanghai, China, 2022, pp. 1-4, doi: 10.1109/ICOT56925.2022.10008164. [3] Z. Mao, B. Li, R. Zhang and Q. Fei, \"DE-YOLO: Detail-Enhanced Maritime Object Detection Algorithm Based on YOLOv8,\" 2025 37th Chinese Control and Decision Conference (CCDC), Xiamen, China, 2025, pp. 3605-3610, doi: 10.1109/CCDC65474.2025.11090244. [4] J. Guria et al., \"AI Based Exam Invigilation System,\" 2026 6th Biennial International Conference on Nascent Technologies in Engineering (ICNTE), Navi Mumbai, India, 2026, pp. 1-6, doi: 10.1109/ICNTE66387.2026.11437509. [5] S. V, S. Shukla, V. Raina, Puru and S. N. Singh, \"A Multimodal Surveillance System for Detecting Cheating Behaviour in Online Exams,\" 2025 9th International Conference on Computational System and Information Technology for Sustainable Solutions (CSITSS), Bangalore, India, 2025, pp. 1-6, doi: 10.1109/CSITSS67709.2025.11295658. [6] A. S. Adhatrao, M. B. Patil, S. G. Sanmukh, A. Shinde and S. Musale, \"AI-based Surveillance for Exam Integrity: Real-Time Detection of Abnormal Student Behaviour,\" 2025 4th International Conference on Sentiment Analysis and Deep Learning (ICSADL), Bhimdatta, Nepal, 2025, pp. 1236-1241, doi: 10.1109/ICSADL65848.2025.10933137. [7] S. Essahraui, M. A. El Mrabet, M. F. Bouami, K. E. Makkaoui and A. Faize, \"An Intelligent Anticheating Model in Education Exams,\" 2022 5th International Conference on Advanced Communication Technologies and Networking (CommNet), Marrakech, Morocco, 2022, pp. 1-6, doi: 10.1109/CommNet56067.2022.9993953. [8] A. K, P. P and J. Paulose, \"Human Body Pose Estimation and Applications,\" 2021 Innovations in Power and Advanced Computing Technologies (i-PACT), Kuala Lumpur, Malaysia, 2021, pp. 1-6, doi: 10.1109/i-PACT52855.2021.9696513. [9] V. Bhagwat, C. Shewale, S. Kadam, S. Shirke, V. Kadam and M. Namose, \"Smart AI Proctor: Automated Exam Monitoring System,\" 2025 IEEE Pune Section International Conference (PuneCon), Pune, India, 2025, pp. 1-5, doi: 10.1109/PuneCon67554.2025.11378675. [10] P. K. Sahoo, S. Majee, S. Nath, B. Bhattacharjee, S. Deb and S. Patra, \"AI Enabled Secure Examination Environment with Real Time Surveillance and Automation,\" 2025 IEEE 4th Conference on Applied Signal Processing (ASPCON), Kolkata, India, 2025, pp. 77-82, doi: 10.1109/ASPCON66877.2025.11389621.

Copyright

Copyright © 2026 Shravan Wargantiwar, Anushree Khadgi, Vanshika Khedkar, Dr. Reema C. Roychaudhary . This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.