Automated Violence Detection in Surveillance Networks with Deep Learning

The prevalence of violent incidents around the world can be quite overwhelming, posing serious threats to personal safety and social stability. Violence manifests in various forms, including communal, caste, and political conflicts, affecting both rural and urban areas. Moreover, college youth are increasingly influenced by these activities. To combat violence, several strategies have been implemented, including the installation of surveillance systems. This paper proposes a deep learning-based approach for the automatic detection of violent activities using Convolutional Neural Networks (CNNs). Specifically, it utilizes MobileNetV2, which offers high accuracy and serves as a foundational model for our system. The surveillance network processes real-time video from CCTV cameras installed on college campuses, enabling prompt alerts to be sent to relevant authorities via a Telegram network. The proposed methodology aims to overcome limitations found in existing violence detection techniques by employing frame grouping and Temporal Squeeze and Excitation Blocks. This enhances the extraction of spatio-temporal features and improves accuracy and efficiency, especially in scenarios involving occlusions and crowded environments.