Violence Detection and Automated Police Reporting System

This paper presented a physics-augmented skeleton-based approach to video violence detection that fundamentally reframes the problem from visual appearance classification to biomechanical event recognition. By explicitly modeling velocity, acceleration, jerk, kinetic energy, and torso stability, we capture the underlying physical nature of violent actions. Our fusion architecture combining temporal pose embeddings with physics features achieved 92.03% accuracy and 92.37% F1-score, demonstrating that violence detection through biomechanics provides robust, interpretable, and generalizable performance.   The proposed method offers significant advantages over traditional RGB-based approaches: independence from generalization across diverse surveillance environments. This work establishes biomechanical modeling as a promising direction for violence detection and opens visual appearance factors, computational efficiency enabling real-time CPU deployment, interpretable predictions grounded in measurable physics, and strong pathways for broader applications in  human activity analysis.  Furthermore, the proposed framework demonstrates strong potential for deployment in real-world surveillance systems, particularly in resource-constrained environments where GPU-based solutions may not be feasible. Its ability to operate efficiently on CPU systems while maintaining high performance makes it suitable for scalable and cost-effective security infrastructures.   Despite its promising results, certain limitations remain. The accuracy of the system is dependent on the quality of pose estimation, which can be affected by occlusions, low-resolution footage, or crowded scenes. Additionally, while biomechanical features improve interpretability, extreme variations in human behaviour or unconventional motion patterns may still challenge the model's generalization capability.   Future work can focus on enhancing robustness by integrating advanced pose estimation techniques and incorporating temporal attention mechanisms to better capture long-range dependencies in motion. Expanding the dataset with more diverse and complex scenarios, as well as exploring multimodal fusion with audio or contextual cues, could further strengthen detection performance.   In conclusion, this work highlights the effectiveness of a physics-augmented perspective in violence detection and emphasizes the importance of moving beyond appearance-based methods toward more meaningful, interpretable representations of human activity. The proposed approach lays a strong foundation for developing intelligent, reliable, and real-time violence detection systems for practical applications.   Ultimately, this study underscores a conceptual shift in violence detection by demonstrating that meaningful understanding of human motion can outperform purely appearance-driven methods. By grounding the detection process in fundamental physical principles, the proposed approach not only enhances predictive performance but also provides a transparent and explainable framework that aligns with real-world dynamics. This paradigm reinforces the idea that future intelligent surveillance systems should prioritize interpretability alongside accuracy, ensuring both technical reliability and ethical accountability. As research in this direction progresses, physics-informed models have the potential to redefine how complex human behaviors are analyzed, paving the way for more trustworthy, adaptable, and context-aware automated monitoring solutions.

1)         K. B. Kwan-Loo, J. C. Ortíz-Bayliss, S. E. Conant-Pablos, and H. Terashima-Marín, "Detection of violent behavior using neural networks and pose estimation," IEEE Access, vol. 10, pp. 86339–86359, 2022.   2)         G. Garcia-Cobo, J. C. Ruiz-Garcia, J. C. SanMiguel, and J. M. Martinez, "Human skeletons and change detection for efficient violence detection in surveillance videos," Computer Vision and Image Understanding, vol. 233, Aug. 2023.   3)         A. Traoré and M. A. Akhloufi, "Violence detection in videos using deep recurrent and convolutional neural networks," arXiv preprint arXiv:2409.03031, Sept. 2024.   4)         P. Zhang, "RTVD-Net: A real-time violence detection method based on YOLO-Pose and CNN-LSTM," in Proc. ACM Int. Conf. Multimedia Technol. Commun., 2023, pp. 156–163.     5)         P. P. Negre, R. Hamon, J. C. Burguillo, and J. Rivas-Echeverría, "Literature review of deep-learning-based detection of physical aggression from video," Sensors, vol. 24, no. 10, p. 3186, May 2024.   6)         M. Johnson and S. Patel, "Real-time vision-based violent actions detection through CCTV cameras with pose estimation," Cranfield University Research Reports, Tech. Rep. CU-2024-037, 2024.     7)         R. Kumar, A. Singh, and M. Verma, "Real time detection of school violence using machine learning and computer vision," Int. J. Adv. Eng., vol. 12, no. 3, pp. 245–258, 2023.   8)         B. Omarov, A. Narynov, Z. Zhumanov, A. Gumar, and M. Khassanova, "A skeleton-based approach for campus violence detection," Computers, Materials & Continua, vol. 72, no. 2, pp. 315–331, 2022.   9)         S. Asad, S. Rehman, R. Fraz, and F. Azam, "State-of-the-art violence detection techniques in video surveillance security systems: A systematic review," PLoS ONE, vol. 17, no. 6, e0269938, June 2022. 10)     C. Lugaresi, J. Tang, H. Nash, C. McClanahan, E. Uboweja, M. Hays, F. Zhang, C. Chang, M. G. Yong, J. Lee, W. Chang, W. Hua, M. Georg, and M. Grundmann, "MediaPipe: A framework for building perception pipelines," arXiv preprint arXiv:1906.08172, June 2019.   11)    C. Feichtenhofer, H. Fan, J. Malik, and K. He, "Slow Fast networks for video recognition," in Proc. IEEE Int. Conf. Computer Vision (ICCV), 2019, pp. 6202–6211.   12)    D. Tran, H. Wang, L. Torresani, J. Ray, Y. LeCun, and M. Paluri, "A closer look at spatiotemporal convolutions for action recognition," in Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2018, pp. 6450–6459.     13)    J. Carreira and A. Zisserman, "Quo Vadis, action recognition? A new model and the Kinetics dataset," in Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2017, pp. 6299–6308.   14)    K. Simonyan and A. Zisserman, "Two-stream convolutional networks for action recognition in videos," in Advances in Neural Information Processing Systems (NeurIPS), 2014, pp. 568–576.     15)    Z. Cao, G. Hidalgo, T. Simon, S. Wei, and Y. Sheikh, "OpenPose: Realtime multi-person 2D pose estimation using part affinity fields," IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 43, no. 1, pp. 172–186, 2021.   16)    T. Pfister, J. Charles, and A. Zisserman, "Flowing convnets for human pose estimation in videos," in Proc. IEEE Int. Conf. Computer Vision (ICCV), 2015, pp. 1913–1921.     17)    S. Hochreiter and J. Schmidhuber, "Long short-term memory," Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997.   18)    A. Vaswani et al., "Attention is all you need," in Advances in Neural Information Processing Systems (NeurIPS), 2017, pp. 5998–6008.     19)    M. Tan and Q. Le, "EfficientNet: Rethinking model scaling for convolutional neural networks," in Proc. Int. Conf. Machine Learning (ICML), 2019, pp. 6105–6114.   20)    A. Dosovitskiy et al., "An image is worth 16x16 words: Transformers for image recognition at scale," in Proc. Int. Conf. Learning Representations (ICLR), 2021. 21)    W. Kay et al., "The Kinetics human action video dataset," arXiv preprint arXiv:1705.06950, 2017.   22)    R. Girdhar, J. Carreira, C. Doersch, and A. Zisserman, "Video action transformer network," in Proc. IEEE Conf. Computer Vision and Pattern Recognition (CVPR), 2019, pp. 244–253.