Learning Representations for Neural Networks Applied to Spectrum-Based Direction-of-Arrival Estimation for Automotive Radar

This paper proposes a new approach to Direction-of-Arrival Estimation using Artificial Neural Networks. It is capable of estimating both, model-order and azimuth DoA in a single step. In a hybrid approach, we train on synthetic data generated from a signal model and validate on data obtained through a measurement setup. We show a proof-of-concept for the hybrid approach, validated with measurement data. Advances on the exactness of the signal model enable the trained ANN to handle real-world data out-of-the-box. Our findings indicate super-resolution performance and the capability of estimating even high model-orders while significantly reducing computation time.