Heterogeneous Acoustic Features Space for Automatic Classification of Drone Audio Signals

Abstract

The incremented use of unmanned aerial vehicles (UAV) in recent years, have leaded to security flaws that demand a solution oriented to UAV monitoring. An attractive solution to this problem is based on the analysis of UAV audio signals. Such approach aims to extract a set of acoustic features and to use them as inputs of machine learning algorithms. Current works on this topic are mainly focused in using a specific set of acoustic features, such as linear prediction and cepstral metrics. However, relevant UAV acoustic information may be missing by considering a single type of features. In this work, we propose a heterogenous acoustic features space for solving UAV automatic classification problems. Temporal, spectral and time-frequency analysis are implemented to extract features from UAV audio signals and thus building a high dimensional features space. By applying features selection techniques, the most relevant acoustic features are identified and they are used to train machine learning algorithms. Our results show that, the heterogeneous features space yields high performance in automatic UAV classification tasks of binary and multiclass type. The classification results outperform the overall classification performance of other studies using set of homogeneous features. Furthermore, the metrics extracted using the wavelet packet transform are the most prevalent in the features spaces that yield the best classification results for the binary and muticlass classification tasks.