Deep Beamforming for Joint Direction of Arrival Estimation and Source Detection

Direction of arrival (DoA) estimation is a well studied problem with several significant applications in radar, sonar, wireless communications, and audio signal processing. A majority of conventional algorithms for DoA estimation require prior knowledge of the number of transmitters and/or sufficient measurements for estimating the received signal covariance matrix. When these requirements are not satisfied, the performance of such algorithms degrades considerably. Recently, some deep learning-based approaches to direction of arrival estimation have been proposed. However, similar to conventional algorithms, most of these methods require the number of transmitters to be known a priori or require a large number of snapshots. We propose a new deep learning approach to DoA estimation. Our approach is inspired by conventional beamforming-based methods and identifies both the number of transmitting sources as well as their angular positions. We demonstrate empirically that our method outperforms conventional methods and a recently proposed deep learning approach in the low-SNR and low-snapshot regimes.