An improved performance of RetinaNet model for hand-gun detection in custom dataset and real time surveillance video

Abstract

The prevalence of armed robberies has become a significant concern in today’s world, necessitating the development of effective detection systems. While various detection devices exist in the market, they do not possess the capability to automatically detect and alarm the presence of guns during robbery activities. In order to address this issue, a deep learning-based approach using gun detection using RetinaNet model is proposed. The objective is to accurately detect guns and subsequently alert either the police station or the bank owner. RetinaNet, the core of the system, comprises three main components: the Residual Neural Network (ResNet), the Feature Pyramid Network (FPN), and the Fully Convolutional Networks (FCN). These components work together to enable real-time detection of guns without the need for human intervention. Proposed implementation uses a custom robbery detection dataset that consists of gun, no-gun and robbery activity classes. By evaluating the performance of the proposed model on our custom dataset, it is evident that the ResNet50 backbone architecture yields outperforms for the accuracy in robbery detection that reached in 0.92 of Mean Average Precision (mAP). The model effectiveness lies in its ability to accurately identify the presence of guns during robbery activities.

Keywords

• robbery activities
• retinanet
• resnet50
• fpn
• fcn