Drone Detection Using Deep Learning: A Benchmark Study

Abstract

Since Unmanned Aerial Vehicles (UAVs) became available to the civilian public, it has witnessed dramatic spread and exponential popularity. This escalation gave rise to privacy and security concerns, both on the recreational and institutional levels. Although it is mainly used for leisure and productivity activities, it is evident that UAVs can also be used for malicious purposes. Today, as legislation and law enforcement federations can hardly control every incident, many institutions resort to surveillance systems to prevent hostile drone intrusion. Although drone detection can be carried out using different technologies, such as radar or ultra-sonic, visual detection is arguably the most efficient method. Other than being cheap and readily available, cameras are typically a part of any surveillance system. Moreover, the rise of deep learning and neural network models rendered visual recognition very reliable [9, 21]. In this work, three state-of-the-art object detectors, namely YOLOv4, SSD-MobileNetv1 and SSD-VGG16, are tested and compared to find the best performing detector on our drone data-set of 23,863 collected and annotated images. The main work covers detailed reportage of the results of each model, as well as a comprehensive comparison between them. In terms of accuracy and real-time capability, the best performance was achieved by the SSD-VGG16 model, which scored average precision (AP50) of 90.4%, average recall (AR) of 72.7% and inference speed of 58 frames per second on the NVIDIA Jetson Xavier kit.