Multi-frequency sequential sparse Bayesian learning for DOA estimation of the moving wideband sound source

Abstract

To obtain the direction of arrival (DOA) of the moving sound source from the sequential measurements collected by the microphone array is the main task in acoustic tracking and detection. Thanks to the development of compressive sensing and sparse Bayesian learning (SBL), treating time-varying DOA estimation as time-varying sparse signal recovery is considered to be a promising idea. However, most methods have assumed that the source is narrowband and the DOA is on the predefined sparse grid at each estimation step. In fact, most sound sources in the air are wideband and the DOA varies continuously. Therefore, the multi-frequency sequential SBL is proposed for the DOA estimation of the moving wideband sound source in this paper. In this method, gamma hyperprior is used as sparsity-promoting prior for multi-frequency bins so that the multi-frequency measurements can be utilized simultaneously, and with an inexact dynamic model, the sparsity-dependent information from the multi-frequency sequential measurements can be propagated to the next estimation step to improve the performance. Besides, the off-grid refinement is incorporated into the framework to adapt to the continuously varying DOA. Simulation results demonstrate that the proposed method has better performances under low signal-to-noise conditions with higher estimation accuracy and less computation time compared to other state-of-the-art methods. The field experiments show that our proposed method can has a stronger ability to suppress grating lobes and spatial aliasing than conventional methods in the estimation for wideband DOA and adapt to the scenarios where the number of sources also changes.