Drone Detection Using Convolutional Neural Networks with Acoustic STFT Features

As the use of drones has raised in the city, the regulation of malicious usage of drones is an important issue. However, it is difficult to detect a drone due to its miniaturization and modification. In this paper, we investigate the performance of using the convolutional neural networks (CNN) for detecting drones with the acoustic signals received by a microphone. Since the harmonic characteristics of drones are different from those of the objects that produce similar noise including scooters and motorcycles, the two-dimensional feature employed in the study is made of the normalized short time Fourier transform (STFT) magnitude. The performance of the proposed approach is evaluated in terms of detection rate and false alarm rate under various environments. The dataset used in this study consists of the measurements through experiments. The experiments are carried out in the open space with a hovering drone, which is DJI Phantom 3 or Phantom 4. The dataset contains 68,931 frames of drone sound and 41,958 frames of non-drone sound. For the 100-epoch model, the detection rate is 98.97 % and the false alarm rate is 1.28, while the 10-epoch model demonstrates the detection rate of 98.77 % and the false alarm rate of 1.62 %.