Redefining Automotive Radar Imaging: A Domain-Informed 1D Deep Learning Approach for High-Resolution and Efficient Performance

TL;DR

The paper addresses the challenge of limited azimuth resolution in automotive mmWave radar by reframing super-resolution as a 1D spectra estimation problem and introducing SR-SPECNet, a compact 4-layer MLP trained with a frequency-domain normalization and an SNR-guided loss. Ground truth spectra are derived from IAA, enabling high-quality RA heatmaps with significantly fewer parameters and faster inference than conventional 2D/3D CNN baselines. Across real-world 86-element MIMO radar data and cross-dataset tests, SR-SPECNet and its SNR-guided variant outperform baselines in NMSE, SSIM, and PSNR, while maintaining efficiency suitable for on-vehicle deployment. The work provides a practical, scalable solution for high-resolution radar imaging in autonomous driving and offers public data and code to accelerate further research.

Abstract

Millimeter-wave (mmWave) radars are indispensable for perception tasks of autonomous vehicles, thanks to their resilience in challenging weather conditions. Yet, their deployment is often limited by insufficient spatial resolution for precise semantic scene interpretation. Classical super-resolution techniques adapted from optical imaging inadequately address the distinct characteristics of radar signal data. In response, our study redefines radar imaging super-resolution as a one-dimensional (1D) signal super-resolution spectra estimation problem by harnessing the radar signal processing domain knowledge, introducing innovative data normalization and a domain-informed signal-to-noise ratio (SNR)-guided loss function. Our tailored deep learning network for automotive radar imaging exhibits remarkable scalability, parameter efficiency and fast inference speed, alongside enhanced performance in terms of radar imaging quality and resolution. Extensive testing confirms that our SR-SPECNet sets a new benchmark in producing high-resolution radar range-azimuth images, outperforming existing methods across varied antenna configurations and dataset sizes. Source code and new radar dataset will be made publicly available online.

Figures (5)

• Figure 1: Impact of antenna aperture and super-resolution algorithms on RA heatmap quality: (a) shows an RGB bus image. RA heatmaps using FFT for (b) 10, (c) 40, and (d) 86 antennas contrast with (e) LiDAR BEV. Heatmaps with IAA for (f) 10, (g) 40, and (h) 86 antennas highlight the improved clarity from both antenna count and super-resolution algorithm.
• Figure 2: Left to right columns: RA maps in polar coordinates, RA maps in Cartesian coordinates, LiDAR point clouds in bird's-eye view, and camera image.
• Figure 3: Radar signal processing pipeline
• Figure 4: Scalability of SR-SPECNet across variable $N_{\text{Range}}$ and $N_{\text{Doppler}}$. The figure contrasts RA heatmap reconstructions for 10 and 40 antenna elements, respectively. In the first row, $N_{\text{Range}} = 100$, $N_{\text{Doppler}} = 64$; in the second row, $N_{\text{Range}} = 50$, $N_{\text{Doppler}} = 40$; and in the third row, $N_{\text{Range}} = 200$, $N_{\text{Doppler}} = 40$, showcasing the model's scalability.
• Figure 5: RA heatmap quality comparison. Heatmaps from the same radar frame are reconstructed by SR-SPECNet, SR-SPECNet+, and baseline 2D U-Net and 3D U-Net models, alongside the ground truth, for both 10-element and 40-element antenna arrays. Each row corresponds to heatmaps generated from the same radar frame data.