A Two-Stage Multi-Layer Perceptron for High-Resolution DOA Estimation

Nowadays, with the development of automotive and traffic radars, a higher angle resolution is required for an increasing demand of four-dimensional (4D) imaging radar. In this paper, the high-precision direction-of-arrival (DOA) estimation problem is solved by using a deep learning (DL) framework. Most existing on-grid DL-based methods have an upper limit of one-degree resolution. The DOA estimation performance under such resolution is still far behind conventional methods, and is not accurate enough in practical applications. Hence, we introduce a two-stage multi-layer perceptron (TS-MLP) framework to achieve higher resolution of DOA estimation with low complexity by dividing the problem into two main parts. The first MLP is used to determine the coarse grid point nearest to the source angle, and the second MLP fine tunes the estimate within the coarse grids. In addition, we propose a solution to the source association problem between the two stages when handling multiple targets. Our scheme shows much higher accuracy compared with existing DL-based methods, and has comparable performance with traditional high-resolution methods. Moreover, it performs quite robust in the presence of array imperfections.