Enhancing Fourier-based Doppler Resolution with Diffusion Models

TL;DR

We address Doppler-resolution limitations in radar by learning to convert a low-resolution RD map produced by a zero-padded FFT into a high-resolution map. The method uses SR3, a conditional diffusion probabilistic model, to model $p(\mathbf{x}_{HR}|\mathbf{y}_{SR})$, trained with a forward diffusion on $\mathbf{x}_{HR}$ and inputs $\mathbf{y}_{SR}$. Inference proceeds with iterative denoising to yield $\mathbf{x}_{SR}$ and is evaluated on simulated RD maps with downsampling factors $s \in \{2,4,8\}$, showing improved target separability and CFAR detections over FFT and MUSIC in several setups, though hallucinations can occur at aggressive downsampling. The results suggest that diffusion-based post-processing can boost Doppler resolution without longer observation times, promoting more reliable separation of slow targets from clutter in practical radar systems.

Abstract

In radar systems, high resolution in the Doppler dimension is important for detecting slow-moving targets as it allows for more distinct separation between these targets and clutter, or stationary objects. However, achieving sufficient resolution is constrained by hardware capabilities and physical factors, leading to the development of processing techniques to enhance the resolution after acquisition. In this work, we leverage artificial intelligence to increase the Doppler resolution in range-Doppler maps. Based on a zero-padded FFT, a refinement via the generative neural networks of diffusion models is achieved. We demonstrate that our method overcomes the limitations of traditional FFT, generating data where closely spaced targets are effectively separated.

Figures (6)

• Figure 1: Training and inference for SR3. During training, RD maps are infused with noise, concatenated with the corresponding zero-padded FFT maps and fed to the UNet, which learns to predict the noise added to the input samples. During inference, noise sampled from a normal distribution is concatenated with a chosen RD map and fed to UNet that iteratively removes the artifacts in order to obtain the clean RD map.
• Figure 2: HR map, LR map, SR map via zero-padded FFT and SR map via SR3 in linear scale given from left to right. RD maps are resolved from a LR version of downsampling factor of $2$.
• Figure 3: HR map, LR map, SR map via zero-padded FFT and SR map via SR3 in logarithmic scale given from left to right. RD maps are resolved from a LR version of downsampling factor of $4$.
• Figure 4: HR (left), SR FFT (middle), and SR3 (right) range-Doppler maps with downsampling factor of $8$ in logarithmic scale.
• Figure 5: HR map, and CFAR detections from the SR map via zero-padded FFT, the SR map via MUSIC and the SR map via SR3 given from left to right. RD maps are resolved from a LR version of downsampling factor of $4$.