Synthetic FMCW Radar Range Azimuth Maps Augmentation with Generative Diffusion Model

TL;DR

This work tackles the data scarcity challenge in automotive radar perception by introducing a conditional diffusion model that synthesizes multi-class radar Range-Azimuth maps guided by Confidence Maps (ConfMaps). Geometry-aware conditioning and a Target-Consistency Regularization are incorporated to reflect radar physics and maintain semantically faithful target energy while preserving background diversity. On the ROD2021 dataset, the approach yields a 3.6 dB improvement in signal fidelity over baselines and a 4.15% boost in mean Average Precision when detectors are trained with the synthetic data. The proposed framework enables scalable generation of physically plausible radar signatures and enhances the generalization of downstream radar perception tasks.

Abstract

The scarcity and low diversity of well-annotated automotive radar datasets often limit the performance of deep-learning-based environmental perception. To overcome these challenges, we propose a conditional generative framework for synthesizing realistic Frequency-Modulated Continuous-Wave radar Range-Azimuth Maps. Our approach leverages a generative diffusion model to generate radar data for multiple object categories, including pedestrians, cars, and cyclists. Specifically, conditioning is achieved via Confidence Maps, where each channel represents a semantic class and encodes Gaussian-distributed annotations at target locations. To address radar-specific characteristics, we incorporate Geometry Aware Conditioning and Temporal Consistency Regularization into the generative process. Experiments on the ROD2021 dataset demonstrate that signal reconstruction quality improves by \SI{3.6}{dB} in Peak Signal-to-Noise Ratio over baseline methods, while training with a combination of real and synthetic datasets improves overall mean Average Precision by 4.15% compared with conventional image-processing-based augmentation. These results indicate that our generative framework not only produces physically plausible and diverse radar spectrum but also substantially improves model generalization in downstream tasks.

Figures (2)

• Figure 1: Diffusion Model for reconstruction conditioned by . A U-Net-based network $\mathbf{p_\theta(\cdot)}$ predicts noise involved in the noisy image $\mathbf{x_t}$ under the guidance of ConfMap $\mathbf{C}$ obtained from the approach detailed in Sec.\ref{['methodology']}, and timestamp $t$ in each denoising step to reconstruct the realistic image $\mathbf{x_0}$ iteratively. The RGB image is only used for scene visualization.
• Figure 2: Qualitative results of our generation. Each example is organized into four rows. The first row shows the original RGB images. The second row presents the ConfMaps for different object categories, where red indicates pedestrians, green indicates cyclists, and blue indicates cars. The third row shows the ground-truth , and the fourth row displays the generated by our model.