Quad-rotor UAV Audio Recognition Based on Mel Spectrum with Binaural Representation and CNN

Abstract

With the wide application of UAV in civil and military fields, more and more attention has been paid to the accurate detection for UAV. In this paper, the noise of UAV is taken as the research object, and the audio recognition technology based on Mel spectrum with binaural representation and CNN is used for multi-UAV recognition. Firstly, the noise of UAV, bird chirp, traffic noise and other environmental noises are collected, and the audio dataset of UAV audio is made after preprocessing. Secondly, the Mel spectrum with binaural representation is proposed for audio feature extracting, and one-dimensional audio is converted to four channel spectrum with rich feature. Finally, designed deep CNN is trained by spectrum map for UAV audio recognition. Experimental results show that the proposed method is superior and the average accuracy is 99.24%.