A Preventive Framework for Physical Bullying Detection in Schools Using AI-Enabled CCTV and Surveillance Based Early Warning Systems

Abstract

Traditional school surveillance functions as a retrospective tool rather than a proactive deterrent. This paper proposes and theoretically validates an automated framework that transforms existing CCTV infrastructure into a real-time Early Warning System (EWS) for physical bullying in school environments. The framework employs a multi-layered deep learning architecture combining Convolutional Neural Networks (CNN) for spatial feature extraction and Long Short-Term Memory (LSTM) networks for temporal behaviour modelling. A Temporal Behavior Mapping module distinguishes between benign social interactions and aggressive actions including striking, shoving, and pursuit-based harassment. Benchmarking against comparable literature-reported implementations on public violence datasets (RWF-2000) indicates target detection performance of approximately 93% precision and 90% recall, with alert generation latency below 3 seconds in GPU-accelerated deployment. To address the intersection of school safety and civil liberties, the framework implements a Privacy-by-Design protocol incorporating skeleton-based pose estimation, eliminating the need for biometric facial recognition and ensuring student anonymity. A Human-in-the-Loop (HITL) verification workflow enables trained administrators to review AI-generated alerts before any intervention is enacted. The framework is designed for deployment scalability across varied school topographies and is evaluated for compliance with India’s Digital Personal Data Protection Act 2023 (DPDPA) and UNESCO AI Ethics recommendations. Results demonstrate that the proposed architecture offers a robust, ethically aligned, and technically feasible foundation for cultivating safer educational environments through predictive intelligence.

Introduction

Bullying in schools remains a major concern, negatively affecting students' academic performance, mental health, and long-term well-being. Traditional monitoring methods such as teacher supervision and passive CCTV systems often fail to detect physical bullying incidents promptly. To address this challenge, the proposed framework utilizes Artificial Intelligence (AI) with CNN-LSTM deep learning models to analyze real-time CCTV footage and identify aggressive behaviors such as pushing, hitting, and kicking. By combining spatial feature extraction through Convolutional Neural Networks (CNNs) with temporal behavior analysis using Long Short-Term Memory (LSTM) networks, the system distinguishes bullying from normal student interactions and generates timely alerts. The framework also introduces several novel contributions, including school-specific model adaptation, privacy-preserving skeleton-based pose estimation, a structured Human-in-the-Loop (HITL) decision process, compliance with India's Digital Personal Data Protection Act (DPDPA) 2023, and threshold sensitivity analysis for customizable deployment.

The proposed methodology follows a modular pipeline consisting of video acquisition, preprocessing, CNN-based spatial analysis, LSTM-based temporal modeling, aggression scoring, HITL verification, and real-time alert generation. Privacy is prioritized by analyzing only anonymized skeletal body keypoints rather than facial or biometric information. Experimental evaluation shows that the framework achieves approximately 93% precision, 90% recall, and an F1-score of 0.91, with alert generation in less than 3 seconds using a hybrid edge-cloud deployment. Compared to manual surveillance and conventional motion-based detection systems, the proposed approach offers significantly higher accuracy, faster response, and stronger privacy protection. Ethical safeguards, automated data retention policies, and human oversight ensure responsible AI deployment, making the framework an effective early warning system for improving student safety while maintaining legal and ethical compliance in school environments.

Conclusion

This paper has presented and theoretically validated an AI-enabled Early Warning System for real-time physical bullying detection in school environments. The proposed framework integrates CNN-based spatial feature extraction, LSTM-based temporal behaviour modelling, skeleton-based privacy-preserving pose estimation, and a formalised Human-in-the-Loop verification workflow within a legally compliant architecture grounded in India’s Digital Personal Data Protection Act 2023. Performance benchmarking against comparable literature-reported implementations indicates target detection performance of approximately 93% precision and 90% recall (F1 = 0.91 at ?* = 0.65), with alert latency below 3 seconds in hybrid edge-cloud deployment. The threshold sensitivity analysis provides practical calibration guidance for diverse school risk profiles, while the HITL decision matrix ensures that automated detection is always subject to trained human oversight before any intervention is enacted. The framework directly addresses three research gaps identified in existing literature: school-specific contextualisation of detection models, architecture-level privacy-by-design, and India DPDPA 2023 compliance mapping. These contributions provide a technically sound and ethically rigorous foundation for translating AI-enabled bullying detection from research into real-world Indian school deployments. Future work will focus on: (1) constructing a school-specific annotated dataset in collaboration with Indian educational institutions, (2) empirical evaluation of the complete pipeline under real deployment conditions, (3) multimodal extension integrating acoustic and wearable sensor streams, and (4) development of a lightweight MobileNet-based variant for edge deployment in legacy-infrastructure schools. Continued interdisciplinary collaboration between AI researchers, school administrators, child psychologists, legal experts, and policymakers will remain essential to ensuring that these systems serve their fundamental purpose: making every school a demonstrably safer place for every student.

References

[1] S. A. Putri, A. Rifai, and I. Nawawi, \"Development of an intelligent violence detection system for bullying monitoring using deep learning models,\" Journal of Scientific and Applied Informatics, vol. 7, no. 2, Jun. 2024. doi: 10.36085/jsai.v7i2.6451. [2] S. G. Satpute and G. T. Rajeshwar, \"Real-time AI solutions for monitoring and preventing bullying in educational institutions,\" International Journal of Advanced Scientific Research, vol. 10, no. 3, pp. 48–51, Aug. 2025. [3] TechTRP Editorial Board, \"AI can help combat disciplinary and bullying cases in schools,\" TechTRP, Feb. 2024. [4] \"Takeaways from our investigation on AI-powered school surveillance,\" AP News, Nov. 2024. [5] \"Can AI identify safety threats in schools? One district wants to try,\" The Washington Post, Jun. 2025. [6] A.Nurbek and A. Altayeva, \"Comparative evaluation of machine learning methods for bullying detection in surveillance footage,\" European Research Materials, no. 9, pp. 1–14, 2025. doi:10.36085/erm.v0i9.5777. [7] L. Siddique et al., \"Analysis of real-time hostile activity detection from spatiotemporal features using deep CNNs, RNNs and attention-based mechanisms,\" arXiv Preprint, 2023. [8] I.-A. Haiura and A. Iftene, \"Detecting violence in videos using convolutional neural networks,\" Procedia Computer Science, vol. 240, pp. 465–475, 2024. [9] Natha, S., Ahmed, F., Siraj, M., et al., \"Deep BiLSTM attention model for spatial and temporal anomaly detection in video surveillance,\" Sensors, vol. 25, no. 1, 251, 2025. doi:10.3390/s25010251. [10] Lee, S., et al., \"Unified video anomaly detection model for detecting different anomaly types,\" Proceedings of IEEE/CVF WACV 2026. [11] [MDPI, \"Spatially time-based robust tracking and re-identification of kindergarten students: A hybrid deep learning framework combining YOLOv8n and ViT,\" Journal of Imaging, vol. 12, no. 4, 150, 2026. doi:10.3390/jimaging12040150. [12] G.Orru et al., \"Development of technologies for the detection of (cyber)bullying: The BullyBuster project,\" Information, vol. 14, no. 8, 430, 2023. doi:10.3390/info14080430. [13] Mazhar AA, Zada I, et al., \"AI-powered detection of cyberbullying in short-form video content: A hybrid deep learning framework,\" PLoS One, vol. 21, no. 2, e0338799, 2026. [14] V.Gattulli et al., \"Human activity recognition for the identification of bullying and cyberbullying using smartphone sensors,\" Electronics, vol. 12, no. 2, 261, 2023. doi:10.3390/electronics12020261. [15] T.Hassner, Y. Itcher, and O. Kliper-Gross, \"Violent flows: Real-time detection of violent crowd behavior,\" in IEEE CVPRW 2012, pp. 1–6. [16] \"Violence detection in surveillance videos with deep network using transfer learning,\" IEEE Xplore, 2019. doi:10.1109/ICDM.2019.8910041. [17] A. Ullah et al., \"Action recognition in video sequences using deep bi-directional LSTM with CNN features,\" IEEE Access, vol. 6, pp. 1155–1166, 2018. [18] J. Li et al., \"Efficient violence detection using 3D convolutional neural networks,\" in IEEE AVSS 2019, pp. 1–8. [19] W.Sultani, C. Chen, and M. Shah, \"Real-world anomaly detection in surveillance videos,\" in IEEE/CVF CVPR 2018, pp. 6479–6488. [20] S. Sharma et al., \"A fully integrated violence detection system using CNN and LSTM,\" IJECE, vol. 11, no. 4, pp. 3374–3380, 2021. [21] M.Inayathulla and K. Rajasekhara Rao, \"Enhancing real-time violence detection in video surveillance using hybrid deep learning model,\" JOWUA, vol. 16, no. 1, pp. 344–361, 2025. [22] M. Cheng, K. Cai, and M. Li, \"RWF-2000: An open large scale video database for violence detection,\" arXiv:1911.05913, 2019. [23] \"Intelligent video surveillance violence detection model with MobileNet V2 and LSTM,\" PMC, 2025. [24] B.Zajime, \"Ethical AI in schools: Balancing automation, privacy, and human oversight,\" WJAETS, vol. 15, no. 1, pp. 924–934, 2025. [25] UNESCO, \"Recommendation on the Ethics of Artificial Intelligence,\" Paris, France, 2021. [26] UNESCO, \"Recommendation on the Ethics of Artificial Intelligence: Safeguarding privacy and personal rights,\" 2024. [27] \"AI adoption in education and associated policy frameworks,\" OECD Policy Survey on School Education in the Digital Age, 2025. [28] Y. Yan et al., \"A systematic review of AI ethics in education: privacy, fairness, transparency and governance,\" Educ. Inf. Technol., 2025. [29] M. Campbell et al., \"Investigation of the privacy concerns in AI systems for young digital citizens,\" arXiv:2501.13321, 2025. [30] S.Muigai et al., \"Enhancing public safety through advanced video analysis: A Conv-LSTM-SVM model for violence detection,\" EAJIT, vol. 7, no. 1, pp. 1–17, 2025. [31] M. Cheng, K. Cai, and M. Li, \"RWF-2000: An open large scale video database for violence detection,\" arXiv:1911.05913, 2019. [32] \"Efficient violence detection in surveillance,\" PubMed Central, 2025. [33] Altaf Hussain, \"Detection and recognition of real-time violence and human actions recognition using lightweight MobileNet model,\" JIAP, vol. 1, no. 3, pp. 125–146, 2025. [34] UNESCO, Recommendation on the Ethics of Artificial Intelligence, Paris, France, 2021. [35] \"AI in school surveillance systems and human rights,\" AI Values, 2025. [36] Ministry of Law and Justice, Government of India, \"The Digital Personal Data Protection Act, 2023,\" Gazette of India, No. 60, Aug. 2023. [Newly added]

Copyright

Copyright © 2026 Akhilaa Vellalar H, Anjit Raja R, Dr. R Manickam. 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.