Maintaining the integrity of examinations is essential in educational settings. Conventionalsurveillanceistypicallybasedonmanualobservation,whichisinefficientanderror- prone. The project introduces a system based on deep learning intended to identify abnormal behaviorinexaminationroomsusingautomatic,real-timevideoanalysis.Byutilizingstate-of-the- artcomputervisiontechnology,thesystemanalyzescontinuousvideostreams,derivesbehavioral features, and labels activities as normal or abnormal. It covers behavior like unauthorized activities, communication, and object transfer that could imply academic dishonesty.
The system incorporates real-time alerting facilities—such as buzzers and immediate notifications—to facilitate prompt intervention by authorities.
This not only minimizes the requirement for constant human monitoring but also provides a more objective and scalable monitoring process. With flexibility and cross-platformcapabilities, the solution is accessible on low-cost hardware and can be utilized across different environments.
Aside from monitoring examinations, the technology can be modified for wider security use in public areas, business premises, and medical facilities. Future development, including multi- camera support, predictive analytics, and privacy-enhancing features, will further enhance its potential.Thisproject illustrateshowartificialintelligencecanrevolutionizesurveillancesystems to become proactive tools that improve fairness, security, and trust in high-stakes environments.
Introduction
Object detection in computer vision is crucial for applications like surveillance, autonomous vehicles, industrial automation, and healthcare. It involves accurately identifying and localizing multiple objects in images with class labels and bounding boxes. Early methods relied on handcrafted features (e.g., Haar cascades, HOG) and classical classifiers (SVM), which struggled with real-world complexities such as varying lighting, occlusion, and scale. The advent of deep learning, especially Convolutional Neural Networks (CNNs), revolutionized the field by enabling automatic hierarchical feature learning, greatly improving accuracy and efficiency.
Despite progress, challenges remain in balancing real-time performance with accuracy, particularly on limited hardware, and ensuring robustness across diverse environments and object classes. Current research focuses on developing efficient, flexible architectures for fast, precise detection suitable for real-time applications.
The literature survey highlights the rise of intelligent video surveillance using machine learning and deep learning for anomaly and suspicious activity detection, particularly in sensitive areas like exam halls. Methods include supervised, unsupervised, and semi-supervised learning, each with trade-offs in accuracy, data requirements, and false positives. Recent deep learning models show promise but often face computational constraints in real-time or low-resource environments. Related work in video steganography also offers insights on efficient video processing.
The problem addressed is the need for an automated, scalable, and efficient deep learning system to detect suspicious activities (e.g., cheating) in examination rooms using video surveillance. Conventional manual monitoring is insufficient due to its error-prone and limited scope. The system must work in real-time on affordable hardware while considering privacy and ethical concerns.
The proposed system combines high-resolution cameras with lightweight CNN models for accurate object detection and tracking. It processes video frames to detect multiple object classes and generates alerts based on suspicious behavior patterns. The system uses techniques like non-maximum suppression for precision and supports modular integration with IoT devices, utilizing GPU acceleration and model quantization for efficiency.
Existing systems rely heavily on manual monitoring and simple recording, lacking automation and real-time alerts. Modern systems use CNN-based feature extraction for activity classification and provide instant notifications but can be costly or complex.
The implemented system runs on mid-range hardware (Intel Core i3, 8GB RAM) with Python and open-source tools, designed for easy deployment in typical school environments. Experimental results on simulated exam videos show high accuracy, low false positives, and robust performance under varying lighting and crowd conditions. Metrics like precision, recall, and F1-score validate effectiveness.
A mathematical model outlines the system’s pipeline: video input → frame segmentation → background/foreground extraction → motion, head movement, and contact detection → activity classification → alerts and logs.
Training and validation losses decrease steadily over epochs, indicating effective learning without overfitting. Accuracy improves progressively, with validation accuracy surpassing 90%, demonstrating a strong, generalizable model.
Conclusion
The systemof suspicious activity detection designed for exam monitoring is a novel approach to enhancing academic integrity through real-time automated monitoring. Through the use of deep learning and computer vision methodologies, the system analyzes live video feeds, recognizes human activities, and labels them as normal or suspicious. Important processes include video segmentation, motion and head tracking, and contact detection to correctly recognize possible misconduct such as unauthorized communication or object transfer.
Thissmartsystemminimizesrelianceonmanualproctoringconsiderably,providingequalandfair monitoring throughout test sessions. It features automated alerting mechanisms—e.g., buzzers, emails, and SMS alerts—that quickly notify authorities if suspicious activity is identified. This enablespromptintervention, renderingtheexamprocessmoresecureandequitableforallpupils.
Withscalabilityinmind,thesystemcanbeinstallednotjust ineducationalinstitutionsbut alsoin other high-security zones such as government buildings, banks, and public places. Its modular design allows for the integration of upgrades such as predictive analytics and pose estimation, which can predict threats even before they completely emerge.
With advancements in artificial intelligence and deep learning technology, the system can be enhanced to function effectively in difficult conditions, including low lighting or heavy crowds. Integration with IoT devices, multi-camera systems, and privacy-protecting methods like anonymization are some of the future developments that can be envisioned.
Ingeneral, the project shows the real-world promise of AI-based surveillance inthe preservation of security and order. Through the automation of detection and alert mechanisms, it provides a foundation for wiser, more responsive monitoring solutions across various applications.
References
[1] Khan, S.,&Sultana,F. (2020).Suspicious ActivityDetectioninSurveillanceVideosusing DeepLearningApproaches.JournalofComputerScienceandTechnology,35(2),185-196.
[2] Ali,F.,&Mehmood, I. (2019).DeepLearningfor SuspiciousActivityDetectioninPublic Surveillance Videos. International Journal of Advanced Computer Science and Applications, 10(5), 249-255.
[3] Soleimani, M., &Ghafoorifard, H. (2018). An Intelligent Surveillance System for Exam HallCheatingDetection UsingDeep Learning. Proceedings of theInternationalConference on Machine Learning and Computer Vision (MLCV 2018).
[4] Zhao, R.,&Li, H. (2021). Real-timeDetection of Suspicious Behavior in ExamHalls using DeepLearningandSurveillanceSystems. JournalofArtificialIntelligenceResearch, 70,27- 41.
[5] Basharat, A., & Shah, M. (2020). Behavioral and Temporal Analysis for Suspicious Activity Detection in Surveillance Videos. IEEE Transactions on Image Processing, 29, 2227-2238.