Deep Learning Based DOA Estimation in Low SNR and Multipath Scenarios

Deep Learning (DL) has revolutionized Signal and Image processing in complex environments where it is difficult to accurately define the signal model. The convolutional neural network (CNN) is DL technique that learns high level features directly from the data. By using different DL approaches, various studies have been done on direction of arrival (DOA) estimation of acoustic sources. While high-resolution algorithms MUltiple SIgnal Classification (MUSIC) have been studied in literature for DOA estimation, the performance of these algorithms is limited by constraints such as low Signal to Noise Ratio (SNR), antenna element spacing and multipath scenarios. In this paper, a Convolutional Neural Network (CNN) based technique is proposed for DOA estimation in noisy and multipath environment. It is shown that the proposed technique is robust towards these typical constraints even in scenario of closely spaced sources. Simulation results show that the proposed CNN algorithm outperforms MUSIC algorithm in low SNR, multipath environment while being computationally less intensive.