Two Dimensional Convolutional Neural Network Frameworks Using Acoustic Nodes for UAV Security Applications

Unmanned Aerial Vehicles (UAVs) have become increasingly popular to the public with a multitude of associated applications in real life. However, malicious applications of UAVs, such as terrorism and air transport disturbance may occur with devastating consequences. Therefore, counter-UAV (C-UAV) systems are required to efficiently address these undesired situations. Artificial intelligence approaches, such as deep neural networks (DNNs), are deemed crucial for high accuracy detection and real-time response of the next generation C-UAV systems. UAV detection based on acoustic sensors has received a great research interest over the past years. Ηowever, the performance of the proposed solutions may vary in different environments since sound signals from UAVs are multi-source and unstructured. In this work, three different 2D Convolutional Neural Networks (CNNs) using short-time Fourier transform spectrograms as input, for UAV binary classification in real-world settings, are compared. The results confirm that 2D CNNs, along with data augmentation techniques in the image domain, can capture the spatio-temporal information of the UAV sound signals, even in noisy environments, and obtain a 95.35% macro average F1-Score in a dataset collected from two different environments.