Machine Learning Inspired Efficient Audio Drone Detection using Acoustic Features

摘要

With the recent proliferation of drones in the consumer market, drone detection has become critical to address the security and privacy issues raised by drone technology. This paper presents an efficient method for drone detection based on the drone's acoustic signature. Five different features are analyzed and compared to determine the best audio descriptor for drone detection. The selected features include Mel-frequency cepstral coefficients, Gammatone cepstral coefficients, linear prediction coefficients, spectral roll-off, and zero-crossing rate. Different support vector machine (SVM) classifier models are trained and tested on a large diverse database using 10-fold and 20% data holdout cross-validation schemes to evaluate the individual feature performance for efficient drone detection. Experimental results indicate that Gammatone cepstral coefficients are the most efficient features for audio drone detection. Further, the medium Gaussian SVM trained on all the investigated features achieves the classification accuracy of 99.9% with 99.8% recall and 100% precision, outperforming the compared existing state-of-the-art audio drone detection methods.