Support Vector Regression based Direction of Arrival Estimation of an Acoustic Source

Abstract

The direction-of-arrival (DOA) estimation of an acoustic is instrumental in many applications such as surveillance, robotics, defense, etc. This paper proposes the DOA estimation technique using a support vector regression (SVR) machine-learning model trained on the signals acquired from the uniform linear array (ULA) of microphones. The SVR machine-learning model has been trained using the correlation coefficients of signals at different microphones as the features of the model. The root-mean-square angular error (RMSAE) parameter has been used for performance comparison of SVR with that of Delay-and-Sum (DAS) beamforming, multivariate linear regression (MLR), and multivariate-curvilinear regression (MCR). From the results, it has been observed that the SVR model outperforms the DAS beamforming method as well as other regression models viz. MLR and MCR.