DroneEARS: Robust Acoustic Source Localization with Aerial Drones

Abstract

Micro aerial vehicles (MAVs), an emerging class of aerial drones, are fast turning into high value mobile sensing assets. While MAVs have a large sensory gamut at their disposal; vision continues to dominate the external sensing scene, with limited usability in scenarios that offer acoustic clues. Therefore, we endeavor to provision a MAV auditory system (i.e., ears); and as part of this goal, our preliminary aim is to develop a robust acoustic localization system for detecting sound sources in the physical space-of-interest. However, devising this capability is extremely challenging due to strong ego-noise from the MAV propeller units, which is both wideband and non-stationary. It is well known that beamformers with large sensor arrays can overcome high noise levels; but in an attempt to cater to the platform (i.e., space, payload and computation) constraints of a MAV, we propose DroneEARS: a binaural sensing system for geo-locating sound sources. It combines the benefits of sparse (two elements) sensor array design (for meeting the platform constraints), and our proposed mobility-aided beamforming (for overcoming the severe ego-noise and its other complex characteristics) to significantly enhance the received signal-to-noise ratio (SNR). We demonstrate the efficacy of DroneEARS by empirical evaluations, and show that it provides a SNR improvement of 15–18 dB compared to many conventional and widely used techniques. This SNR gain translates to a source localization accuracy of approximately 40 cm within a scan region of 6 \mathbf{m}\times 3 \mathbf{m}6 \mathbf{m}\times 3 \mathbf{m}, that is, one order of magnitude better than competing methodologies.