Aerial Drones with Direction Sensitive DeepEars

Recent years have seen a huge increase in the use of small unmanned aircrafts, otherwise known as micro aerial vehicles (MAVs), in a variety of monitoring applications. With regard to applications in emergency response, acoustic sensing plays a key role in locating sound emitting targets (e.g., a person in distress); especially in visually occluded environment. Our endeavour, therefore, is to provision the MAV Ears; and as part of the initial prerequisite, our aim is to develop a robust acoustic direction finding system. In order to achieve autonomous MAV navigation using acoustic signal, a direction-of-arrival (DoA) mechanism that is robust in low signal-to-noise ratio (SNR) conditions becomes a necessity. In this paper, we propose Drone-DeepEars: a new DoA estimation framework based on deep learning and optimized for sensor arrays with fewer sensing elements. We show that its DoA estimation accuracy is comparatively better than state-of-the-art techniques (such as MUSIC and ESPIRIT) at high noise levels, but at a relatively lower computational footprint.