Real-time acoustic detection and identification of drones in operational conditions

Abstract

Unmanned Aerial Vehicles (UAV), commonly known as drones, can fly below the radar horizon, can have very small radar cross-sections, and can drop munitions, which makes them a serious potential threat to public facilities, such as airports and shopping centers. However, UAVs emanate unique acoustic signatures that can be used to identify them and track their locations using microphone arrays. This paper presents a deep-learning approach to performing noise-robust drone detection and identification in real-time. In particular, a Convolutional Neural Network (CNN) with densely connected layers at the output was employed. Detection was accomplished by including an additional output for when only background environmental noise is present. The inputs were log-magnitude FFTs extracted from 10 ms Hanning windowed frames with 50% overlap. As long as the training data include a higher or equal level of noise than the target data, accuracy at detecting and identifying seven drones was above 90% for environmental noise levels as low as 20 dB SNR. Current research is focused on using a Bayesian version of the network to enable the classifier to report confidence levels in its decisions as well as detect novel drones, i.e., when a detected craft is not represented in the training data.