Acoustic detection of drone range and type using nonuniform band energy features

Abstract

With increased use of drones in a variety of situations, it is imperative that efficient means of detecting the type of unmanned aerial vehicle and/or its range are developed from security, privacy, and safety perspective. This paper describes the application of an acoustic feature set in a deep learning network for the estimation of the line-of-sight range of drones. The set of spectral energy values over nonuniform bands within the range of audio recordings of open space drone noise has been shown to predict the range to a reasonable degree of accuracy. The energy feature set, when augmented with low frequency spectral components, raised prediction accuracy to within 85 cm of mean error and a standard deviation of 12 m for test cases ranging from 10 m to 935 m. Additionally, the spectral band energy applied to classify the range quantized into 1 m intervals resulted in better than 87% accuracy with a fixed error of ±50 cm over the entire range. Adding low frequency spectral components to the band energy set raised the correct range classification to 97%. For classification of seven tethered drones, the band energy feature set resulted in 99.9% accuracy.