Multiple source localization in the spherical harmonic domain using augmented intensity vectors based on grid search

Abstract

Multiple source localization is an important task in acoustic signal processing with applications including dereverberation, source separation, source tracking and environment mapping. When using spherical microphone arrays, it has been previously shown that Pseudo-intensity Vectors (PIV), and Augmented Intensity Vectors (AIV), are an effective approach for direction of arrival estimation of a sound source. In this paper, we evaluate AIV-based localization in acoustic scenarios involving multiple sound sources. Simulations are conducted where the number of sources, their angular separation and the reverberation time of the room are varied. The results indicate that AIV outperforms PIV and Steered Response Power (SRP) with an average accuracy between 5 and 10 degrees for sources with angular separation of 30 degrees or more. AIV also shows better robustness to reverberation time than PIV and SRP.

Augmented Intensity Vectors for Direction of Arrival Estimation in the Spherical Harmonic Domain

“…In [38] and [23], we have shown that the conventional wideband PWD-SRP fails in localization of multiple sources especially for adjacent sources as the summation of power spectra can result in merging of adjacent peaks associated to different sources resulting in an erroneous estimated DOA between the sources as seen in Fig. 1.…”

“…In [38], [23] we proposed Augmented Intensity Vectors (AIVs) which exploits eigenbeams of order ≥ 2 to form vectors with improved DOA accuracy compared to PIVs. These vectors are obtained using spatially constrained grid search to minimize a cost function with initialisation derived from PIVs.…”

“…(1) Steered Response Power (SRP) methods [21] in which a beamformer such as Plane-Wave Decomposition (PWD) [22], [14], [11] is used to steer a beam into multiple directions and the source direction is found as the direction with the highest power. Due to limitations on spatial resolution, SRP methods can fail to localize closely spaced sources [23]; (2) Sub-space methods [24], [25] such as MUSIC (MUltiple SIgnal Classification) [6], [26] employ the eigenvalue decomposition to decompose the noisy signals spatial covariance matrix into the signal and noise subspaces, which then are used to estimate the DOAs. Although they are somewhat robust to noise and reverberation, their performance normally depends on a user-defined threshold.…”

Multiple DOA estimation based on estimation consistency and spherical harmonic multiple signal classification

“…The MUSIC spectrum in (11) and (18) was calculated with 1°resolution across azimuth and inclination (360 × 181). Incoherent DPD-MUSIC was excluded from our evaluation since studies in [13] show that incoherent DPD-MUSIC fails in case of low angular separation of sources as the two peaks associated with two adjacent sources can be merged into one peak over summation of MUSIC spectra which causes the second highest peak to be detected far from the sources.…”

“…The initial DOA estimates (one per TF bin) can be obtained by any SS DOA estimator such as PIV [10], as used in this work, Augmented Intensity Vectors (AIVs) [12], [13], Steered Response Power-based methods [14] or SS MUSIC. The initial DOA estimates are weighted based on their consistency within a time interval and the ones with the strongest weights are selected as the most consistent DOAs using the assumption of stationary sources.…”