Real-Time Safety Alert System: Violence and Fall Detection with YOLOv8

The rapid proliferation of surveillance cameras in urban environments has created a data surplus that human operators can no longer effectively manage. While video feeds are abundant, the ability to interpret them in real-time to prevent harm remains a significant bottle neck. This paper introduces a comprehensive”Real-time Safety Alert System” designed to autonomously detect and flag two specific, high-risk scenarios: interpersonal violence and accidental falls. Leveraging the advanced architecture of the You Only Look Once version 8 (YOLOv8) model, our system shifts the paradigm from passive recording to active intervention. Unlike traditional motion-based sensors which suffer from high false-positive rates, our deep learning approach understands the semantic context of human actions, distinguishing a sudden fall from a person simply sitting down, or a violent altercation from benign social interaction. We further enhance the model’s reliability through a custom post- processing logic layer that evaluates the temporal consistency of detected events, filtering out momentary glitches. The system is optimized for edge deployment, processing video streams locally to ensure immediate alert generation while respecting user privacy. Extensive experimental validation confirms the system’s robustness, demonstrating high precision and recall rates that make it a viable candidate for deployment in sensitive environments such as schools, hospitals, and elderly care facilities.