Violence Detection Using MobilenetV2

Abstract

In today\'s world, ensuring public safety through technology has become increasingly critical. This project presents a Violence Detection using deep learning and computer vision techniques. The system is designed to monitor video feeds and detect violent activities in real-time, enabling swift responses to potentially dangerous situations. By integrating models trained on violence-related datasets with optimized video processing pipelines, the application identifies violent behavior using frame-by-frame analysis. The backend is served using a lightweight API framework, and the system supports live video input from cameras. This innovative solution has potential applications in surveillance, public transport security, and smart city infrastructure. The aim is to provide an intelligent, automated approach to enhance safety and security in real-world environments. The application supports voice input and output, ensuring a hands-free, realistic interaction. Leveraging speech-to-text and text-to-speech capabilities, it helps users improve both their communication and technical answering skills. The system architecture efficiently integrates Open AI APIs to generate intelligent questions and analyze user responses, offering constructive feedback for continuous improvement.

Introduction

Public safety in areas such as transportation hubs, schools, offices, and urban environments requires effective monitoring systems. Traditional surveillance relies heavily on human operators, which can lead to fatigue, errors, and delayed responses. To address these challenges, the proposed Real-Time Violence Detection System uses Artificial Intelligence (AI) and computer vision to automatically detect violent activities from live video streams.

The system analyzes video frames using deep learning models trained on violent and non-violent actions. Key features include real-time video capture, frame preprocessing, feature extraction, violence classification, and an alert mechanism that notifies authorities when suspicious behavior is detected. Built with FastAPI, the system ensures low-latency and efficient real-time performance.

The literature review highlights several approaches to violence detection, including CNN-based frame classification, CNN-LSTM architectures for temporal analysis, anomaly detection using autoencoders, optical flow combined with CNNs, and advanced 3D models such as I3D. These studies demonstrate the importance of capturing both spatial and temporal information for accurate violence recognition.

The proposed methodology consists of five main modules: video acquisition, preprocessing, feature extraction, classification, and alert generation. Features such as HOG, HOF, and optical flow are extracted, while CNNs, LSTMs, and SVMs are used for classification. The system processes video frames continuously and generates instant alerts when violent behavior is detected.

For evaluation, the model is trained on labeled datasets containing violent and non-violent video clips. Performance is measured using metrics such as Accuracy, Precision, Recall, F1-Score, ROC-AUC, and Mean Average Precision (mAP). Experimental results compare the proposed approach with CNN-SVM, 3D-CNN, CNN-LSTM, and optical flow-based methods, demonstrating its effectiveness in accurately detecting violence while supporting real-time surveillance applications.

Conclusion

In this project, various violence detection techniques were explored using both spatial and temporal features extracted from video data. Experimental results demonstrate that incorporating temporal context through advanced models, such as attention-based transformers, significantly improves detection performance compared to traditional frame-based CNNs and optical flow-based approaches. The proposed method achieved superior results across key evaluation metrics, including accuracy, precision, recall, and F1-score, while maintaining efficient inference time suitable for real-time applications. These findings emphasize the critical role of temporal dynamics in accurately identifying violent activities. Overall, the system provides a robust and reliable solution for intelligent surveillance and public safety monitoring. Future work can focus on enhancing dataset diversity and integrating multimodal inputs, such as audio and contextual data, to further improve detection accuracy and system performance.

References

[1] Hassner, T., Itcher, Y., & Kliper-Gross, O. (2012). Violence Detection in Video Using Subclasses. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 1-8. https://doi.org/10.1109/CVPR.2012.6247951 [2] Tran, D., Bourdev, L., Fergus, R., Torresani, L., & Paluri, M. (2015). Learning Spatiotemporal Features with 3D Convolutional Networks. Proceedings of the IEEE International Conference on Computer Vision (ICCV), 4489–4497. https://doi.org/10.1109/ICCV.2015.510 [3] Simonyan, K., & Zisserman, A. (2014). Two-Stream Convolutional Networks for Action Recognition in Videos. Advances in Neural Information Processing Systems (NIPS), 568-576. https://arxiv.org/abs/1406.2199 [4] Ji, S., Xu, W., Yang, M., & Yu, K. (2013). 3D Convolutional Neural Networks for Human Action Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(1), 221–231. https://doi.org/10.1109/TPAMI.2012.59 [5] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, ?., & Polosukhin, I. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems (NeurIPS), 5998–6008. https://arxiv.org/abs/1706.03762 [6] Mo, S., & Bui, T. D. (2020). Violence Detection in Surveillance Videos Using CNN and LSTM Networks. IEEE Access, 8, 185153–185163. https://doi.org/10.1109/ACCESS.2020.3026754 [7] Google Dataset Search (if you used any public dataset): [Dataset Name], accessed May 2025, https://datasetsearch.research.google.com/

Copyright

Copyright © 2026 Dr. Prathibha G, Indraja K H, Abhishek S Nayak, Keerthana B C, Sanjay K R. 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.