An embedded deep learning framework for real-time violence detection and alert generation.

Violence in public and semi-public environments poses a significant threat to social safety and requires timely intervention to prevent escalation and harm. The growing deployment of surveillance cameras has increased the demand for automated systems capable of identifying violent behavior in real time. Intelligent, on-device analysis offers a practical means to enhance situational awareness while reducing reliance on continuous human monitoring. This study presents an optimized embedded implementation for real-time violence detection designed for resource-constrained edge devices, using a lightweight CNN for spatial feature extraction and a GRU for temporal modeling. The proposed model integrates spatial feature extraction through a lightweight convolutional backbone with temporal sequence modeling via Gated Recurrent Units (GRU), enabling efficient spatio-temporal representation of violent actions. To ensure deployment feasibility on edge devices such as the Raspberry Pi 5, multiple optimization strategies—including quantization, structured pruning, and TensorRT acceleration—were employed to achieve high inference speed and low energy consumption without compromising accuracy. The framework was evaluated on five benchmark datasets (Hockey Fight, RLVS, Violent Flows, ShanghaiTech, UCF-Crime) and a custom six-class dataset, collectively encompassing diverse real-world scenarios. Experimental results demonstrated an average accuracy of 96.9%, F1-score of 96.3%, and ROC–AUC of 0.972, outperforming state-of-the-art models such as I3D, LRCN, and ViT while maintaining superior efficiency. The optimized system achieved 26 FPS and 38.4 ms/frame latency under live video conditions, confirming its capability for continuous real-time surveillance. Comprehensive ablation and statistical analyses validated the contribution of each architectural component and confirmed the significance of performance gains (p < 0.05). Qualitative evaluations using Grad-CAM visualizations further confirmed the model’s interpretability by accurately localizing violent regions within frames. While the CNN–GRU formulation follows established spatiotemporal modeling practices, the primary contribution of this work lies in the system-level design and deployment pipeline, including model compression and embedded runtime optimization for achieving real-time throughput under practical hardware constraints. The proposed framework thus establishes a robust, energy-efficient, and explainable solution for intelligent surveillance applications, providing a scalable foundation for future extensions involving multimodal sensing, federated learning, and transformer-based hybrid architectures in embedded vision systems.