Subspace-based DOA with linear phase approximation and frequency bin selection preprocessing for interactive robots in noisy environments

Abstract

This work develops a method of estimating subspace-based direction of arrival (DOA) that uses two proposed preprocesses. The method can be used in applications that involve interactive robots to calculate the direction to a noise-contaminated signal in noisy environments. The proposed method can be divided into two parts, which are linear phase approximation and frequency bin selection. Linear phase approximation rectifies the phases of the two-channel signals that are affected by noise, and reconstructs the covariance matrix of the received signals according to the compensative phases using phase line regression. To increase the accuracy of DOA result, a method of frequency bin selection that is based on eigenvalue decomposition (EVD) is utilized to detect and filter out the noisy frequency bins of the microphone signals. The proposed techniques are adopted in a method of subspace-based DOA estimation that is called multiple signal classification (MUSIC). Experimental results reveal that the mean estimation error obtained using proposed method can be reduced by 7.61° from the conventional MUSIC method. The proposed method is compared with the covariance-based DOA method that is called the minimum variance distortionless response (MVDR). The DOA improves the mean estimation accuracy by 4.98° relative to the conventional MVDR method. The experimental results demonstrate that both subspace-based and covariance-based DOA algorithms with the proposed preprocessing method outperform the DOA estimation in detecting the direction of signal in a noisy environment.