A Deep Learning Based Super Resolution DoA Estimator With Single Snapshot MIMO Radar Data

In this article, we study the radar angular resolution, the capability of distinguishing multiple targets in different directions of arrivals (DoA). We present a deep learning-based super resolution DoA estimator for multiple input multiple output (MIMO) radar with single snapshot data. The estimator consists of a deep learning DoA classifier (DLDC) in the central bearing angle zone, which can simultaneously detect up to 11 targets in
$[ { - 2^\circ,\ 2^\circ } ]$
with 32 virtual antenna elements, and a spatial filtering pre-rotator (SFPR) that makes the DLDC supporting any bearing angle location in a wider radar field of view (FoV). We present the DLDC-SFPR structure and provide detailed parameter setting. In the performance evaluation, the resolution of proposed method is first compared with the theoretical angular resolution limit (ARL). It is shown that the proposed estimator reaches the Chernoff ARL. The probability of resolution (PoR) is then investigated and compared with typical and state-of-the-art DoA estimators in various conditions. The numerical results show that the DLDC-SFPR achieves 0.4
^°
resolution with 90% PoR to distinguish 2 targets at signal to noise ratio (SNR) of 7 dB if the targets are on-gird. In the case of any-angle, the same resolution and probability are achieved at SNR = 12 dB. The proposed method is robust to multi-target concurrent cases, and outperforms the existing reported approaches. Two procedures, uniform scanning and progressive scanning, are proposed for wider FoV applications.