Gun Detection: A Comparative Study of RetinaNet, EfficientDet and YOLOv8 on Custom Dataset
Pyone Pyone Khin, Nay Min Htaik
Abstract
The escalating threat of armed robberies has prompted the urgent need for advanced detection systems. Current solutions in the market lack the ability to autonomously identify and signal the presence of firearms during robbery incidents. To overcome this limitation, this research proposes a novel approach based on deep learning techniques, specifically employing the RetinaNet, EfficientDet, and YOLOv8 (You Only Look Once) models for gun detection. The objective of this study is to develop a sophisticated system that can automatically and accurately identify firearms in surveillance footage during robbery activities. The three selected models bring unique strengths to the system, contributing to its overall efficacy. The proposed system aims to revolutionize the way we address security challenges associated with armed robberies. By leveraging the capabilities of deep learning models, it promises to provide a more proactive and accurate means of gun detection, thus improving overall public safety. The implementation proposed in this study employs a custom robbery detection dataset specifically designed for robbery detection, encompassing classes such as gun, no-gun, knife, no robbery activity, and robbery activity. This study offers a comparison of gun detection algorithms, specifically focusing on different three models which are RetinaNet, EfficientDet and YOLOv8. Through a meticulous evaluation of the three models’ performance on this custom dataset, the good results are obtained according to mAP. Among them, it becomes apparent that the YOLOv8 architecture surpasses others, achieving a remarkable accuracy of 91.4% Mean Average Precision (mAP) in robbery detection.