A Proposal of an End-to-End DoA Estimation System Aided by Deep Learning

Direction of arrival (DoA) estimation is a technique used in, for instance, radar systems, source localization, and wireless channel estimation, and improving its accuracy is becoming more important as its applications expand. In previous works, we verified that deep neural networks (DNN) trained offline are a viable tool for achieving great on-grid DoA estimation performance, even compared to the traditional root multiple signal classification (root-MUSIC) algorithm. Moreover, we separately proposed the signal-to-noise ratio (SNR)-based DNN selection method in order to fully leverage the DNN learning potential at specific SNRs; and the superposition of two DNNs (staggered DNNs) in order to overcome the estimation failure caused by incident radio waves at the grid border. In this paper, we propose an end-to-end DoA estimation system, which consists of three modules: DoA, SNR and source number estimators. Here, we expand the performance of the DoA estimator module by presenting a new algorithm more efficient in detecting DoA from the DNN output, and by using the combination of the two previously mentioned methods: SNR-based DNN selection and staggered DNNs. Also, an SNR estimation scheme with better precision is presented.