Synthetic FMCW Radar Range–Azimuth Maps Augmentation with Generative Diffusion Model

Zhaoze Wang1,3, Changxu Zhang1,3, Tai Fei2, Christopher Grimm1, Yi Jin1, Claas Tebruegge1, Ernst Warsitz1, Markus Gardill3
1 HELLA GmbH & Co. KGaA, Lippstadt, Germany
2 Dortmund University of Applied Sciences and Arts, Dortmund, Germany
3 Brandenburg University of Technology, Cottbus, Germany
IEEE Radar Conference (RadarConf) 2026
Paper PDF Code arXiv
Conditional diffusion pipeline for FMCW radar RAMap synthesis

Diffusion model for RAMap reconstruction conditioned by semantic Confidence Maps. A U-Net predicts the noise in each denoising step to iteratively reconstruct a realistic FMCW radar range–azimuth map; the RGB image is shown only for scene visualization.

Abstract

The scarcity and low diversity of well-annotated automotive radar datasets often limit the performance of deep-learning-based environmental perception. We propose a conditional generative framework for synthesizing realistic FMCW radar Range–Azimuth Maps (RAMaps) for pedestrians, cars, and cyclists. Conditioning is achieved through class-wise Confidence Maps (ConfMaps), while radar-specific Geometry-Aware Conditioning (GAC) and Target-Consistency Regularization (TCR) adapt diffusion generation to radar signal characteristics. Experiments on ROD2021 demonstrate a 3.6 dB PSNR improvement over the baseline, and training with a mixture of real and synthetic data improves downstream radar object-detection mAP by 4.15% compared with conventional image-processing-based augmentation.

Method

Conditional RAMap Generation

Semantic ConfMaps encode object class, range and azimuth as spatially aligned Gaussian targets and guide a conditional diffusion model to synthesize multi-class FMCW radar spectra.

Geometry-Aware Conditioning

GAC incorporates range-dependent attenuation, antenna gain and inter-object occlusion into the conditioning signal, explicitly injecting radar geometry and propagation priors.

Target-Consistency Regularization

TCR introduces a differentiable CFAR-inspired focal objective that emphasizes sparse target energy while still allowing realistic stochastic variation in radar background noise.

Qualitative Results

Qualitative FMCW radar RAMap generation results

Qualitative RAMap generation across diverse driving scenes. From top to bottom: RGB images, class-wise ConfMaps, ground-truth RAMaps, and generated RAMaps. Red, green, and blue ConfMap responses denote pedestrians, cyclists, and cars, respectively.

Quantitative Results

+3.6 dB
PSNR improvement over the baseline signal-generation method
23.7 dB
Best RAMap reconstruction PSNR using both GAC and TCR
+4.15%
Overall downstream mAP gain from synthetic radar augmentation

Signal-Level Evaluation

MethodConditionGACTCRPSNR ↑
de Oliveira et al.BBX Mask××20.1
OursConfMap××22.1
OursConfMap✓×23.3
OursConfMap✓✓23.7

Downstream Radar Object Detection

MethodParking Lot mAPCampus Road mAPCity Street mAPHighway mAPOverall mAP ↑AP@0.5 ↑
RODNet58.4134.2717.7130.2431.3035.22
RAMP-CNN57.9535.1018.2530.9031.5035.10
Ours57.3536.9521.0133.8235.6539.75

BibTeX

@inproceedings{wang2026synthetic,
  title     = {Synthetic FMCW Radar Range--Azimuth Maps Augmentation with Generative Diffusion Model},
  author    = {Wang, Zhaoze and Zhang, Changxu and Fei, Tai and Grimm, Christopher and Jin, Yi and Tebruegge, Claas and Warsitz, Ernst and Gardill, Markus},
  booktitle = {2026 IEEE Radar Conference (RadarConf)},
  year      = {2026}
}