Violence detection from surveillance cameras using deep learning models

Abstract

This study uses deep learning as a framework to solve the critical demand for automated violence detection in video surveillance systems to improve public safety. The inefficiency of traditional manual surveillance camera monitoring which involves thousands of cameras highlights the need for smart solutions. This study focuses on developing a violence detection system that classifies violent activities into two categories: fight and no fight. The project workflow involves gathering and annotating a custom video dataset with fight and no fight labels and preprocessing the data to ensure uniformity. The dataset is then divided into training, testing, and validation sets to train and evaluate ConvLSTM and the long-term recurrent convolutional network (LRCN) models which effectively extract spatiotemporal features for accurate classification. During extensive experiments, the model with the highest accuracy is determined. The most accurate model is deployed locally as a user-friendly interface, enabling violence detection and visualizing the results. This study also reviews related works in video action recognition and violence detection providing insights into the field. The proposed system contributes to intelligent video surveillance by offering a scalable solution for information security such as criminal activity and security threats to public safety.