An Approach for Improving Accuracy and Optimizing Resource Usage for Violence Detection in Surveillance Cameras in IoT systems

Abstract

Violence is a serious issue that can happen in many places, like streets, schools, or homes, where people hurt each other or damage things. To help prevent violence, some places use special cameras called surveillance cameras. These cameras watch over areas and look for signs of violence, like people fighting or breaking things. When they see something, they can send an alert to the police or others who can help. However, building models to detect violence in videos or surveillance cameras can be challenging. Current models may not be lightweight, fast, or use fewer resources, which means they may not work well on all devices or in all situations. So, there is a need for new models that are better at detecting violence while still being fast and using fewer resources. In this study, we focus on training multiple models to detect violence. Specifically, we fine-tune various models, including MobileNet, MobileNetV2, DenseNet121, and ResNet50V2, combined with Bidirectional Long Short-Term Memory (BiLSTM) networks. Among these models, MobileNetV2 for spatial feature extraction combined with BiLSTM for temporal feature extraction stands out for its compact size, quick processing time, and ability to achieve satisfactory results. This combination offers a lightweight solution that can efficiently detect violence in videos or surveillance footage while maintaining good performance levels in IoT systems.