Model Order Estimation using a Multi-Layer Perceptron for Direction-of-Arrival Estimation in Automotive Radar Sensors


In this work, a machine-learning-based approach to decide whether one or two targets are present in the same range-velocity cell of a chirp-sequence FMCW radar system is evaluated. An experimental setup for generating sufficient large sets of training and testing data using real measurement data from automotive 77GHz radar sensors is presented. Using this data a multi-layer perceptron is trained to directly estimate the number of present targets from the re- ceived signals in order to determine if resolution in the spatial domain is necessary. Evaluations of the trained model show that the network is able to inherently learn the underlying signal model and reach super-resolution performance.

In Topical Conference on Wireless Sensors and Sensor Networks (WiSNet), IEEE.