DOA Estimation by Particle Filtering with Frequency Selectivity

Abstract

This paper proposes a robust DOA finder to achieve accurate and stable estimation under noisy environments. The dominant frequency of a target signal, e.g. formant frequency of speech, is extracted in the same way to the spectral subtraction. To obtain smooth estimates in time, the proposed method models a DOA trajectory on the spatial state space, which is defined by time lag in cross-correlation, and estimates DOA through the state estimation procedure using particle filters. The subband cross-correlation function at the dominant frequency is selectively used for updating the particle weights to estimate posterior distribution. The feasibilities of both frequency selectivity and particle filtering have been confirmed in several noise conditions.