FB-WKDE: A Robust Sound Source Localization Approach for TDOA Estimation in Noisy Environments

The frequency of extreme weather events and natural disasters has recently been increasing, leading to greater human casualties. To minimize such losses, rapid survivor detection and accurate localization are essential. However, conventional sound source localization methods such as generalized cross-correlation phase transform (GCC-PHAT), steered response power with phase transform (SRP-PHAT), and multiple signal classification (MUSIC) suffer from performance degradation in noisy environments. In particular, GCC-PHAT and SRP-PHAT are highly vulnerable to interference, while both SRP-PHAT and MUSIC rely on grid search, which not only increases computational cost but also fails to provide an estimate when the actual source lies outside the predefined grid. To address these limitations, this paper proposes filter bank-weighted kernel density estimation (FB-WKDE), a robust localization method that combines a filter bank (FB) with weighted kernel density estimation (WKDE). FB-WKDE computes the time difference of arrival (TDOA) by applying GCC-PHAT to each frequency band followed by WKDE. The proposed method was validated using simulation data generated with the pyroomacoustics platform as well as data collected from a real disaster training facility. Evaluation results show that FB-WKDE significantly improves accuracy and reliability compared to GCC-PHAT, SRP-PHAT, and MUSIC, consistently maintaining stable performance under various acoustic conditions. In particular, FB-WKDE demonstrates strong robustness even in high-noise environments, confirming its practical applicability in disaster scenarios where reliable sound source localization is critical under adverse conditions.