Benchmark Analysis of YOLOv8 for Edge AI Video-Surveillance Applications

Video surveillance systems (VSSs) are crucial for identifying crimes involving handheld weapons, but their effectiveness is limited by high operational costs and the need for continuous human monitoring. To overcome these challenges, research has turned to deep learning (DL) to automate weapon detection. Early efforts focused on detecting weapons in images using resource-intensive frameworks – and, thus, unaffordable – and relying on datasets that poorly represent real-world settings. Given the state-of-the-art limitations, this work introduces an edge artificial intelligence (AI)-based approach which explores the YOLOv8 architecture across different variants (nano, small, medium, large, extra-large), evaluated on a custom built dataset (WeaponSenseV2). Using the NVIDIA Jetson Nano as an edge device, the proposed study evaluates performance of the architectures on the custom-built WeaponSenseV2 dataset, which is designed to more accurately reflect real-world scenarios than existing datasets. By focusing on the intersection of edge computing and DL, this research aims to reduce the operational costs and computational demands associated with traditional VSSs. The results enrich the ongoing discussions on leveraging technological advancements to enhance public safety, providing valuable perspectives on the real-world deployment of scalable and efficient systems for detecting weapons.