Drone Sound Source Localization Using Distributed Microphone Arrays

Abstract

Sound source localization (SSL) with microphone arrays remains challenging, especially for tracking aerial targets such as drones. Early single-array beamforming methods have limited spatial resolution and rely on manually selected narrow frequency bands. Existing multiarray fusion methods have limited robustness and high computational costs. In this work, a fast and powerful drone SSL framework is proposed to address the above issues. First, to avoid manual frequency selection, the dynamic harmonic extraction (DHE) method is used to automatically select the localization frequency from the harmonic components of the drone noise. Second, an improved fast sparse Bayesian learning (FSBL) algorithm is used to estimate the direction of arrival (DOA) of drones, which adopts a greedy update strategy to estimate the noise variance and hyperparameters, thereby enhancing iterative stability and shortening the running time. Third, a fast fuzzy fusion (FFF) method is used to improve the computational efficiency of grid-based fusion methods. In the FFF method, the position of the drone is rapidly estimated by quasi-Newton iteration, and it can still maintain high robustness when there is a deviation in the DOA. Field experiments on drone localization show that the average localization error of our framework is 4.71 m, which is reduced by approximately 50.42% compared with the 9.50-m error of the delay-and-sum beamforming (DSB) combined with the fuzzy fusion (FF) localization method. The proposed drone localization framework has strong robustness and adaptability to environmental noise, signal fading, and sudden maneuvers of drones. The drone localization results indicate that the combination of DHE, DOA estimation based on the improved FSBL algorithm, and FFF provides a practical solution for drone localization.