Acoustic-based drone localization and DOA estimation for highly noisy environments

Abstract

ABSTRACT The constant development of unmanned aerial vehicles (UAVs) has drawn attention from sectors such as law enforcement, agriculture, emergency response services, and defense. This growing interest can be attributed to their simplicity and cost-effectiveness. However, the fact that these devices can be easily used for illegal activities becomes a concern, as they could be employed in actions such as terrorism, unauthorized surveillance, and multiple types of cyberattacks, including eavesdropping, jamming, and spoofing. Therefore, the ability to estimate the location and direction of arrival (DOA) of drones is necessary to ensure the security of critical infrastructure and to enable the deployment of countermeasures. The main goal of this thesis is to estimate the drone’s location and DOA, as these are the most important parameters for launching acoustic-based countermeasures. The contributions to drone DOA estimation include: (1) the recognition of the potential of the zero cyclic sum (ZCS) condition for evaluating the consistency of multiple time delay estimates by exploring various combinations of primary and secondary peaks in the cross-correlation function; (2) the mitigation of the computational complexity of exhaustive search through a heuristic strategy based on genetic algorithms, while preserving estimation accuracy; (3) the proposal of a refinement step based on a least squares (LS) solution, which complements the ZCS condition and increases robustness against spurious peaks; and (4) the development of a passive DOA estimator that is simple to configure and does not require prior training, making it suitable for noisy environments. The proposed methods achieved success rates of up to 94.0±3.1% in correctly identifying the DOA in experiments with real drones in reverberant settings. The contributions to drone localization based on TDOA measurements between microphones include: (1) the formulation of optimization methods that exploit both primary and secondary peaks of the cross-correlation functions; (2) the validation of these optimization methods through real-world experiments; (3) the comparison of these methods with a neural network-based approach, which confirmed the validity of the proposed LS-based estimators that achieved localization errors below 1 meter; and (4) the demonstration that the proposed methods are flexible and can be used for both DOA estimation and localization tasks. The lowest mean localization error obtained using the ZCS-LS method was 0.55±0.35 m, while the neural network approach achieved 1.02±0.02 m at its best position, confirming the effectiveness of the proposed techniques.