The growing demand to upscale surveillance over public places and in online media provided one of the significant reasons for the need for violence detection in surveillance systems and in the moderation of online content. This paper proposes a real-time violence detection system that uses ResNet50V2 for violent activity detection with high precision, along with timestamping of incidents in video streams. It still has a confidence threshold at 0.6 for reliable detection with fewer false alarms. Extensive testing reveals high performance, such as achieved accuracy at 88 %, precision of 90 % for violence, and an F1score of 88%. The system thus assures better public security by ensuring timely responses and a holistic analysis of violent events.
DOI: https://doi.org/10.1109/icrito66076.2025.11241501
Publish Year: 2025
