DopeNet: Range–Doppler Radar-based UAV Detection Using Convolutional Neural Network

Abstract

The proliferation of commercial unmanned aerial vehicles (UAVs) of various sizes and shapes, equipped with cameras and even signal sabotage devices, has raised concerns regarding privacy and safety. Some websites even offer weapons that can be attached to drones, adding to the security threats. As a result, researchers have been motivated to develop an intelligent system that can be integrated into surveillance systems to classify unauthorized UAVs that are flying in restricted areas. In this paper, we propose a convolutional neural network (CNN) for UAV detection based on Radar image data. The Radar system, called Real Doppler RAD-DAR (Radar with Digital Array Receiver), is a range-doppler system developed by the Microwave and Radar Group. We construct and analyze the CNN by adjusting its hyper-parameters using the RAD-DAR dataset. Our simulation results show that setting the number of filters to 32 results in the best time-wise accuracy. The network achieved an accuracy of 97.63%, which is higher than other benchmark image classifiers. Additionally, we conducted an ablation study to investigate and validate the contribution of each part of the neural network.