Generation of Synthetic Micro-Doppler Radar Data Using Generative Diffusion Models

This study investigates the generation of micro-Doppler radar data of cyclists using advanced generative models, namely a Denoising Diffusion Probabilistic Model (DDPM) and a Generative Adversarial Network (GAN). The main objective is to synthesize realistic raw radar data that can effectively mimic real-world scenarios including noise and address the significant challenge of dataset availability in radar technology due to high costs and difficulties in data collection. Radar data collected with a 77GHz automotive radar along the track of cyclists is used to evaluate and compare the performance of a DDPM and a GAN in generating complex-valued raw radar data. Both generative models demonstrated outstanding performance in generating realistic radar data. The DDPM shows a smoother convergence, less variation in loss, and better alignment with the statistical features of the original data compared to the GAN model.