Transfer Learning in Automotive Radar Using Simulated Training Data Sets

For a reliable detection and classification of road users in modern automotive radar systems, latest research introduces machine-learning (ML) based algorithms in competition to implementations based on the classical radar signal processing chain. Suitable training datasets for ML systems based on real-world radar measurements are however either rarely available or lack the specific radar raw data. A training approach based on transfer-learning methods from data generated by a simulation framework is presented for the range-Doppler-representation of radar measurement data. In particular, the impact of dataset size and sample quality in relation to the performance of the ML system in the radar domain is examined.