Off-grid DOA estimation via a deep learning framework

Direction-of-arrival (DOA) estimation problem is one of the most important tasks for array signal processing. Conventional methods are limited by either the computational complexity or the resolution. In this paper, a novel deep learning (DL) framework for super-resolution DOA estimation is developed, where the grid mismatch problem is fully considered into the DL DOA estimation and the offset between the real DOA and the discrete sampling grid is thereby trained as a part of the network output. Specifically, we first model the DOA estimation problem as a multi-classification task and obtain the on-grid estimation result with a resolution of 1 degrees as the first output. Then we designed another regression module to estimate the offset between target angles and the corresponding grids. By combining both on-grid and offset estimation results, the accurate off-grid estimation is achieved. In addition, the method of fusing low-level features and high-level features by skip connection is adopted between the on-grid estimation module and the offset estimation module. The high-level features in the first part of the network are combined with the original input to improve the training speed and the estimation accuracy. Herein, our training dataset contains a large number of off-grid signal data, which is close to the practical application. Also, the utilization of the off-grid part information improves DOA estimation performance compared to existing DL-based methods, achieving more accurate DOA estimations. Numerous simulation results have demonstrated the superiority of the proposed method in DOA estimation precision, and a good performance advantage maintains even when using a limited number of snapshots. Further, we test the proposed method on the real millimeter-wave radar system, and it outperforms the other state-of-the-art methods, especially for two adjacent targets in the same range and Doppler bin.