VIOLENCE DETECTION IN VIDEOS USING DEEP LEARNING

Abstract

Video violence detection is a critical tool for enhancing public safety in smart cities, particularly in managing the rising tide of political violence. Current datasets and conventional methods often fail to address the intricate nature of political violence, resulting in models that lack robustness and precision. To overcome these limitations, this study introduces a novel approach by creating a specialized dataset tailored to political violence, consisting of 480 meticulously labeled video clips spanning four distinct categories: Normal, Clash, Fire, and Shooting. This dataset offers a comprehensive resource for understanding politi cal violence dynamics and serves as a vital benchmark for advancing research in this domain. We fine-tuned the Mobile Video Network (MoViNet-A0), achiev ing an impressive accuracy of 92.86% with a lightweight architecture of 1.904 million parameters, making it suitable for real-time applications. Additionally, we developed a custom keyframe extraction algorithm that integrates temporal and pixel-level features, enhancing stability and accuracy in identifying critical frames. Our methodology outperforms state-of-the-art techniques across multiple benchmark datasets, including Hockey Fight (98%), RLVS (98.3%), RWF-2000 (88.12%), and Surveillance Camera Fight (84%). These advancements highlight the system’s potential to strengthen security measures, inform policy decisions, and provide a scalable solution for detecting and classifying political violence in video surveillance, ultimately contributing to safer urban environments.