Robust TDOA Estimation Using Kernel Density Estimation for Noisy Environments
Jungyu Choi, Joonhwi Kim, Sungbin Im · 2025
In various applications, the importance of localization has increased significantly. In this study, we propose a method combining a filter bank and kernel density estimation (KDE) for robust localization even in noisy disaster environments. The proposed technique applies a filter bank to the input signal to separate it into multiple frequency bands, allowing for the analysis of signal characteristics within each band. Subsequently, generalized cross correlation with phase transform (GCC-PHAT) is calculated for the signals in each separated frequency band. KDE is then applied to the computed GCC-PHAT to estimate the probability density function and determine the time difference of arrival (TDOA) values. Finally, the TDOA results are used to determine a final TDOA, and the coordinates where hyperbolas most closely intersect are estimated as the sound source's location. Experimental results demonstrate that the proposed method outperforms conventional techniques in noisy environments and is a robust solution for sound source localization in complex real-world scenarios.