A Fast Multi-Source Sound DOA Estimator Considering Colored Noise in Circular Array

Abstract

Sound direction of arrival (DOA) estimation keeps finding new applications. For instance, real-time sound DOA estimation enabled by acoustic sensors on low-cost and limited-resource devices can enhance or create many consumer use cases. Indeed, tracking multiple moving sound sources in noisy or reverberated condition with a cost-effective platform is highly desirable. In this paper, we develop a computation-aware DOA estimator that uses weighting in frequency domain. It relies on single source zones (SSZs) and circular integrated cross spectrum (CICS) techniques and peak search on DOA histogram. The existing histogram-based multi-source DOA estimation is fundamentally limited in angular resolution, and the CICS process is very time consuming. To tackle these problems, we propose a weighting method that exploits the unevenness of intermediate DOA estimation errors over the frequencies of interest, and a simple yet effective strategy to reduce search range. Two underlying mechanisms, the impact of the colored noise of microphones and the frequency-dependent behavior (statistic model) of DOA estimation errors, are unveiled. We have developed an experimental platform using commercially off-the-shelf devices to verify design objectives, including feasibility and real-time requirement. The proposed techniques are validated on the developed platform and the system performance is evaluated experimentally in a real-world environment. It is observed the DOA estimation resolution is improved roughly from 30° to 15° and the overall computation is reduced by 68%.