Noise-Resistant Sound Source Localization Based on Filter Banks in Disaster Scenarios
Jungyu Choi, Eojin Kim, Juseong Kim, Joonhwi Kim, Sungbin Im · 2025
With the increasing frequency of extreme weather and disasters, the number of casualties is also on the rise. Therefore, in order to minimize casualties, it is crucial to swiftly detect survivors immediately after a disaster and accurately estimate their location. In this study, we propose a method that combines filter banks, generalized cross-correlation phase transform (Gee-PRAT), and kernel density estimation (KDE) to achieve robust position estimation even in complex environments with noise and reverberation. First, the input signal is divided into multiple frequency bands using a filter bank, allowing for the analysis of signal characteristics in each band. Gee-PRAT is then applied to the signals in each frequency band to estimate the time difference of arrival (TDOA). Based on the estimated TDOA in each frequency band, KDE is performed, and the time with the highest probability is selected as the final TDOA. Using this final TDOA, we estimate the location of the sound source, improving the accuracy of position estimation even in noisy environments. The proposed method showed improved performance over existing techniques in the presence of noise and echo, and was experimentally validated as a robust solution for sound source localization in complex real-world environments.