Convolutional Neural Network Based Sound Recognition Methods for Detecting Presence of Amateur Drones in Unauthorized Zones

Abstract

Unmanned aerial vehicles (UAVs), or drones, have become an integral part of diverse civil and commercial applications. But illegal operations of UAVs pose serious risks to public safety, privacy and national security. Thus, the detection of these vehicles has become vital to identify and track amateur drones in unauthorized zones or restricted areas. This paper presents three sound event recognition (SER) methods developed based on the Mel-frequency cepstrum coefficients (MFCCs) and spectrogram features combined with three machine learning classifiers such as multi-class SVM, one-dimensional convolutional neural networks (1DCNN) and two-dimensional CNN (2DCNN) for recognizing drone sounds. The SER methods are evaluated using a wide variety of sounds such as music, speech, wind, rain, and vehicle. Results showed that the MFCC-SVM based and MFCC-1DCNN based SER methods had average recall rate (RR) = 92.53%, precision rate (PR) = 93.21% and F1-score = 92.84% and RR = 94.28%, PR = 96.57% and F1-score = 95.02% for a segment length of 1 s, respectively. The spectrogram-2DCNN based method had average RR = 73.27%, PR = 74.55% and F1-score = 73.37% for audio segment length of 500 ms. Preliminary results demonstrate that the MFCC-1DCNN based SER method achieves better recognition rates as compared to that of the MFCC-SVM based and spectrogram-2DCNN based methods in recognizing drone sounds.