M.S. Candidate: Orhun Olgun
Program: Multimedia Informatics
Date: 26.07.2019
Place: A-212
Direction-of-arrival (DOA) estimation of sound sources is a popular research topic and has several different applications including spatial audio. Recent advances in microphone arrays made more accurate sound field analysis possible. Spherical microphone arrays afford a trivial calculation of spherical harmonic decomposition of sound fields and can be employed in different DOA estimation methods in spherical harmonics domain.
This thesis proposes two extensions to the a novel DOA estimation method called Hierarchical Grid Refinement (HiGRID) for rigid spherical microphone arrays (RSMA). HiGRID is based on the calculation of the sector averaged directional response power of a steered beam over a sparse set of directions on the unit sphere. The selection of the direction for which response power is to be calculated is determined using spatial entropy as a criterion. This is followed by clustering of the resulting DOA map using a method based on connected components labelling is also proposed for counting sources and estimating their DOAs.
This thesis also investigates the extensions of several state-of-the-art DOA estimation techniques. These include the improvement of DOA estimation performance or computational efficiency of Eigenbeam Multiple Signal Classification (EB-MUSIC) and Direct Path Dominance (DPD) test. HiGRID is first used as source counting method prior to EB-MUSIC to decrease the computational cost of DOA estimation. HiGRID is then used as a DOA estimation method following the DPD test which increases the DOA estimation accuracy while reducing the total computational cost. A new data-driven statistical method for DPD test threshold selection is also proposed. This allows the an informed selection of DPD test threshold based on effective rank statistics of spatial correlation matrices obtained from RSMAs.
Comparison of HiGRID with previous DOA estimation methods with real and simulated recordings are presented. Evaluations of proposed algorithms for EB-MUSIC and DPD test are also presented in terms of DOA estimation errors using simulated recordings. HiGRID and its combinations with EB-MUSIC and DPD test performed favourably in comparison with other state-of-the-art DOA estimation methods indicating the utility of the proposed methods in multiple source DOA estimation.