Deep Neural Networks for Direction of Arrival Estimation of Multiple Targets With Sparse Prior for Line-of-Sight Scenarios

Received signal Direction of Arrival (DOA) estimation represents a significant problem with multiple applications, ranging from wireless communications to radars. This problem presents significant challenges, mainly given by a large number of closely located transmitters being difficultly separable. Currently available state of the art approaches fail in providing sufficient resolution to separate and recognize the DOA of closely located transmitters, unless using a large number of antennas and hence increasing the deployment and operation costs. In this paper, we present a deep learning framework for DOA estimation under Line-of-Sight scenarios, which able to distinguish a number of closely located sources higher than the number of receivers' antennas. We first propose a formulation that maps the received signal to a higher dimensional space that allows for better identification of signal sources. Secondly, we introduce a Deep Neural Network that learns the mapping from the receiver antenna space to the extended space to avoid relying on specific receiver antenna array structures. Thanks to our approach, we reduce the hardware complexity compared to state of the art solutions and allow reconfigurability of the receiver channels. Via extensive numerical simulations, we demonstrate the superiority of our proposed method compared to state-of-the-art deep learning-based DOA estimation methods, especially in demanding scenarios with low Signal-to-Noise Ratio and limited number of snapshots.

Automated summary

Received signal Direction of Arrival (DOA) estimation is a significant problem with wide-ranging applications. Current approaches struggle to separate closely located transmitters without using a large number of antennas, resulting in higher costs. In this paper, we propose a deep learning framework that can estimate DOA under Line-of-Sight scenarios, distinguishing more closely located sources than the number of receiver's antennas. Our approach reduces hardware complexity compared to state-of-the-art solutions and performs well in demanding scenarios with low SNR and limited snapshots.