Abstract
Nowadays, video surveillance has a crucial role. Analyzing surveillance videos is, however, a time consuming and tiresome procedure. In the last years, artificial intelligence paved the way for automatic and accurate surveillance-video analysis. In parallel to the development of artificial-intelligence methodologies, edge computing is becoming an active field of research with the final goal to provide cost-effective and real time deployment of the developed methodologies. In this work, we present an edge artificial intelligence application to video surveillance. Our approach relies on a set of four IP cameras, which acquire video frames that are processed on the edge using the NVIDIA® Jetson Nano. A state-of-the-art deep-learning model, called Single Shot multibox Detector (SSD) MobileNetV2 network, is used to perform object and people detection in real-time. The proposed infrastructure obtained an inference speed of ∼10.0 Frames per Second (FPS) for each parallel video stream. These results prompt the possibility of translating our work into a real word scenario. The integration of the presented application into a wider monitoring system with a central unit could bring benefits to the overall infrastructure. Indeed our application could send only video-related high-level information to the central unit, allowing it to combine information with data coming from other sensing devices without unuseful data overload. This would ensure a fast response in case of emergency or detected anomalies. We hope this work will contribute to stimulate the research in the field of edge artificial intelligence for video surveillance.