PaperlayerView exact revision on arXiv

Source-linked AI summary

Super-Resolution of Range-Doppler Maps: A Case Study with Chirp-Sequence Radar and Transformer

Sven Hinderer, Jonathan Riese, Zheming Yin, Bin Yang

arXiv:2608.27354v1eess.SP

TL;DR

The paper studies how to super-resolve range-Doppler maps when radar hardware limits range and Doppler resolution. It uses real-world Infineon radar data, an efficient DPSWIN transformer, and CFAR-guided training, reporting improved reconstruction and detection metrics with computationally efficient processing. The authors note that supervised training and the absence of downstream-task evaluation limit direct practical application.

  • Problem

    Hardware constraints limit radar range and Doppler resolution, while existing evidence is limited by simplified scenarios and unavailable benchmark datasets.

  • Method

    The paper trains an efficient DPSWIN transformer with 1D shifted-window and dual-path attention on paired real-world low- and high-resolution rD maps, using reconstruction and CFAR-based objectives.

  • Results

    The framework achieves efficient real-world rD-map super-resolution, with CFAR losses improving RMSE, SNR, and IoU and power-law compression giving overall best results.

  • Takeaways & Limitations

    Accurate and computationally efficient rD-map super-resolution is feasible in challenging real-world environments.

  • Takeaways & Limitations

    Supervised training has range-cell-migration and incomplete-chirp issues, and the study does not evaluate downstream tasks such as target localization.

Abstract

from arXiv · show

Range-Doppler (rD) maps produced by chirp- sequence (CS) radar systems are fundamentally limited in reso- lution by bandwidth, carrier frequency, and coherent processing interval constraints. Improving resolution through hardware is often impractical due to regulatory, cost, and real-time operation requirements. In this work, we investigate deep learning-based super- resolution of rD maps in both range and Doppler using a real- world dataset collected with an Infineon millimeter-wave CS radar. We propose a memory-efficient transformer architecture based on dual-path shifted window (DPSWIN) attention, which combines axial/dual-path attention with 1D shifted window (SWIN) attention for scalable processing of high-dimensional radar data. Our model is benchmarked against existing 2D SWIN attention-based super-resolution models, which we adapt to the rD-map super-resolution task. We prioritize reconstruction fidelity over perceptual quality and employ root mean squared error (RMSE)-based training objectives. We avoid adversarial or perceptual losses that may introduce visually plausible but physically incorrect structures. In addition, we incorporate CFAR-based target-detection losses to optimize downstream target detectability in the super-resolved rD maps. We further study the impact of signal processing and training design choices on the super-resolution, including magnitude compression, spatial upsampling strategies, and loss formulations. Experimental results demonstrate that the proposed framework achieves computationally efficient rD map super-resolution on previously unseen real-world environments.

I. INTRODUCTION

The paper addresses limited and underexplored radar rD-map super-resolution in real-world environments with an efficient transformer framework and domain-specific training choices.

  • Motivation: Hardware limits and violated assumptions constrain conventional radar resolution-enhancement methods in complex real-world environments.Classical methods also face computational costs that hinder real-time high-resolution rD-map processing.
  • Research gap: Prior radar super-resolution studies mainly use simplified scenarios and CNNs, while this work evaluates real indoor and outdoor pedestrian measurements.The authors position real-world radar measurements and radar-tailored transformers as departures from much of the existing literature.
  • Design studies: The study examines CFAR-based losses, magnitude compression, and FFT zero-padding as domain-specific design choices for rD-map super-resolution.These choices target reconstruction quality, detection, dynamic range, and spatial upsampling.
  • Contributions: The authors collect paired low- and high-resolution rD maps from an Infineon chirp-sequence radar in indoor and outdoor pedestrian environments.Controlled baseband sub-sampling produces the paired training data.
  • Contributions: DPSWIN combines dual-path attention with 1D shifted-window attention to reduce computational and memory complexity relative to conventional 2D SWIN transformers.The architecture is designed to preserve spectral structure while scaling to high-dimensional radar data.
  • Results: Power-law compression gives the overall best results, while FFT zero-padding outperforms pixel-shuffle upsampling for this task.The study concludes that accurate, computationally efficient rD-map super-resolution is feasible in challenging real-world environments.

A. CNNs and U-Nets for radar super-resolution

Earlier radar super-resolution work largely relies on CNNs and simplified data, whereas transformer approaches exploit efficient attention to process radar structure more effectively.

  • CNNs and U-Nets: CNNs and U-Nets dominate prior radar super-resolution studies, including work on simulated point targets and selected radar representations.Applications span rD, range-azimuth, SAR, weather radar, and cross-modal radar-to-LiDAR upsampling.
  • CNNs and U-Nets: Diffusion-based Doppler upsampling may achieve strong offline performance but is less suitable for real-time processing because inference is iterative.The comparison highlights a deployment trade-off between potential performance and inference efficiency.
  • CNNs and U-Nets: The authors report that simple CNNs and U-Nets performed worse than transformer networks with comparable parameter counts in preliminary experiments.They therefore exclude U-Nets from further investigation.
  • Transformers: Transformers offer parallelized computation and scalable attention, but global attention has quadratic complexity with input length.Local attention is used to reduce this complexity.
  • Transformers: DPSWIN extends dual-path attention with efficient 1D axial shifted windows applied along both range and Doppler axes.This design reduces dual-path attention complexity while targeting radar’s two-dimensional spectral structure.
  • Transformers: Prior transformer radar super-resolution work includes 1D SWIN for SAR range profiles, but it uses simulated data and addresses only one axis.This work applies 1D SWIN to both range and Doppler.

IV. RANGE-DOPPLER SUPER-RESOLUTION PROCESSING

The processing pipeline converts temporal low-resolution radar frames into compressed rD-map stacks, transforms their features, and reconstructs super-resolved rD magnitudes.

  • Input sequence: The network receives the current and three preceding low-resolution frames, allowing temporal cues to support rD super-resolution.Using the current frame as input performed better, and the four-frame stack is used in all experiments.
  • Scope: The study restricts processing to rD-bin magnitudes for target detection, leaving phase-based extensions such as direction-of-arrival estimation outside scope.Phase information is identified as a possible future extension.
  • Pre-processing: Pre-processing forms rD maps from low-resolution time-domain frames, optionally zero-pads FFT inputs, drops the Nyquist range bin, and compresses magnitudes.The resulting compressed stack has four temporal channels.
  • Trainable pipeline: A convolutional feature extractor expands the input channels, after which the feature transformer produces an improved representation for reconstruction.The reconstruction block maps transformed features to a compressed super-resolved rD map approximating the high-resolution target.
  • Feature transformer: The architecture study is based on SwinIR and specifically investigates low-complexity 1D shifted windows instead of 2D windows.The choice is intended to exploit spectral rD-map structure.

A. Range-Doppler processing

The range-Doppler processing stage compares FFT zero-padding with learned spatial extension and shapes the resulting spectra for subsequent super-resolution.

  • Zero-padding: With zero-padding, fast-time samples are extended from Nr to 2Nr and slow-time samples from ND/2 to ND before the respective FFTs.The resulting spectrum retains ND Doppler bins and produces Nr + 1 non-negative range bins before shaping.
  • Zero-padding: Zero-padding expands low-resolution FFT inputs to produce the desired rD-map dimensions without adding information or reducing target main-lobe width.It oversamples the spectrum and can reduce target-peak frequency offsets from FFT-bin discretization.
  • Spatial extension: Without zero-padding, the reconstruction block instead extends the rD map’s range and Doppler dimensions using pixel-shuffle.The paper compares this alternative against the zero-padding version.
  • Dimension shaping: The maximum range bin is dropped so the range dimension becomes a power of two, simplifying upsampling and enabling possible future SWIN U-Net extensions.The authors assume the Nyquist bin contains no relevant information.
  • Dimension shaping: The shaped low-resolution rD stack is converted to Nr × ND dimensions before compression and feature extraction.Zero-padding allows the low-resolution map shape to match the high-resolution map shape before later processing.

C. Compression

The paper studies magnitude compression and transformer feature processing for rD-map super-resolution, emphasizing representations that handle radar’s large dynamic range while controlling computational cost.

  • Compression: Target returns span tens of decibels because received power depends on RCS, antenna gain, and inverse range to the fourth power.This makes direct linear-scale inputs difficult to use reliably.
  • Compression: The method tests linear, logarithmic, and power-law compression of non-negative frequency-domain magnitudes before network processing.A lower bound of 10^-12 prevents numerical problems during logarithmic compression.
  • Feature transformer: The feature extractor applies a stride-one 2D convolution to compressed rD frames, producing high-dimensional features for the transformer.The feature transformer then processes these features through residual Swin transformer blocks and convolutional layers.
  • Feature transformer: DPSWIN combines axial dual-path attention with 1D shifted windows along range and Doppler, reducing attention complexity from quadratic to linear in window size.The design uses shifted 1D windows to enlarge receptive fields while scaling better than 2D windows for high-dimensional inputs.

F. High-resolution range-Doppler map reconstruction

The reconstruction module extends rD features spatially when needed, then estimates a single super-resolution rD map. Training uses RMSE in compressed-magnitude space to emphasize reconstruction fidelity and potentially improve weaker-signal recovery.

  • Spatial upsampling: Pixel-shuffle reshapes low-resolution features from (N_r/2 × N_D/2 × C) into super-resolution features of (N_r × N_D × C/4).Spatial extension occurs in feature space before reducing channels to one output map.
  • Spatial upsampling: With zero-padding, the model skips pixel-shuffle and applies the final 3 × 3 convolution directly to an N_r × N_D feature map.This variant has a few negligible extra parameters because its channel dimension remains C rather than C/4.
  • Output constraints: An optional softplus activation enforces nonnegative outputs for linear- or power-law-compressed magnitudes; alternatively, outputs are rounded to zero afterward.
  • Training objective: RMSE measures the difference between compressed high-resolution and super-resolution rD maps across all bins.The loss is defined in compressed rD-map space rather than directly on uncompressed magnitudes.
  • Training objective: The square root after aggregation increases the impact of weaker signals compared with standard MSE, potentially benefiting compressions with larger dynamic ranges.RMSE also matches the paper’s log-spectral-distance definition for the log-compressed baseline.

V. CFAR DETECTION AND CFAR LOSSES

The paper adapts differentiable CA-CFAR behavior to rD-map super-resolution by comparing detector masks from high-resolution and super-resolution maps. It accounts for non-stationary clutter while acknowledging that standard CA-CFAR assumptions are violated in real scenes.

  • CFAR loss formulation: Differentiable CFAR losses compare soft CFAR threshold masks computed from high-resolution and super-resolution rD maps.The losses are introduced to align downstream target-detection behavior.
  • CFAR loss formulation: CA-CFAR losses are studied as additions to RMSE through cell-averaging variants for rD-map super-resolution.
  • CFAR detection: CFAR adapts detection thresholds to distinguish interference-only bins from target-plus-interference bins under a specified false-alarm probability.In 2D rD maps, the detector uses local reference cells around each cell under test.
  • CA-CFAR detection: CA-CFAR estimates interference by averaging reference-cell powers while excluding guard cells that may contain neighboring target power.Guard and reference bands, along with the false-alarm probability, are typically tuned for the application.
  • Differentiability: The arithmetic-mean interference statistic is differentiable, leaving hard thresholding as the operation replaced by a differentiable approximation.Border handling uses mirrored rD bins to avoid incomplete bands.
  • Assumptions and scope: Real scenes violate CA-CFAR assumptions through non-Gaussian noise, nearby person reflections, and correlated non-uniform clutter.The loss therefore aims to reproduce real high-resolution detector behavior rather than establish practical CFAR validity.

B. CA-CFAR losses with differentiable interference statistic

The paper replaces hard CFAR thresholding with differentiable sigmoid masks and evaluates multiple loss interpretations and sigmoid arguments. It identifies t2 and t3 as power-scale invariant, while t3 is more stable and physically grounded.

  • Soft masks: A differentiable sigmoid replaces hard thresholding, with ξ controlling sigmoid steepness and producing a soft mask value for each rD bin.
  • Sigmoid variants: The paper evaluates three sigmoid arguments, t1, t2, and t3, and compares their resulting CFAR masks.The comparison is illustrated in Fig. 7.
  • Sigmoid variants: t2 can saturate at sigmoid(−ξ) for zero CUT power, while high ξ may create large gradients and earlier saturation.These effects can impair low mask values, training stability, or convergence.
  • Sigmoid variants: t3 uses a log-ratio that makes the mask symmetric with respect to the CUT-to-threshold power ratio.The paper describes t3 as numerically more stable and physically and probabilistically grounded.
  • Mask losses: CFAR masks are trained using bin-wise BCE, KLD between normalized mask distributions, or MSE regression formulations.
  • Mask losses: t2 and t3 are invariant to equal scaling of target and interference power, whereas t1 violates this CFAR power-invariance property.

VI. EXPERIMENTS

The experiments use a newly recorded dataset spanning diverse indoor and outdoor pedestrian environments, with independent environment splits for evaluation. The study also documents CFAR-mask settings, dataset size reporting, radar parameters, and visualization conventions.

  • Dataset: Because no benchmark dataset exists for deep-learning rD-map super-resolution, the authors recorded a dataset for this task and planned open release.
  • Dataset: The dataset contains hand-held radar measurements from pedestrian settings in Stuttgart, split roughly between varied outdoor public spaces and indoor or partly closed environments.Recordings include varying walking speeds.
  • Dataset split: Training, validation, and test sets use completely independent environments or substantially different zones within the same street.The test set is reserved for final evaluation.
  • Sample construction: Each sample contains four consecutive high-resolution time-domain frames separated by 100 ms, formed from eight-frame recording stacks.Breaks between stacks reduce redundant, highly correlated samples.
  • CFAR settings: Fig. 7 compares CFAR masks using ξ = 1 or ξ = 10, PF A = 10^-2, two-cell guard bands, and five-cell reference bands.
  • Data and visualization: Visualizations use the log-spectrum representation, while privacy laws prevented recording corresponding images.
  • Radar configuration: The radar parameters balance maximum range, maximum velocity, range resolution, and velocity resolution for pedestrian sensing.The asymmetric range and Doppler dimensions permit attention windows spanning the Doppler axis to remain shiftable in range.

C. Network architecture, training, and FFT window

The model uses a compact DPSWIN configuration with fixed attention and feature dimensions, trained using RMSE, CFAR-related objectives, and regularization. Its processing pipeline evaluates multiple rD-map metrics and CFAR settings.

  • Network configuration: The reference model R0 uses two RSTBs, four STLs each, four attention heads per MSA, feature dimension C = 96, and 856.065 parameters.R0 uses DPSWIN with window size W = 64.
  • Training: Training uses mini-batches of 32, 400 epochs, Adam optimization, gradient clipping, stochastic depth, and dropout probability 0.05.The learning rate is reduced to 10% after 150 epochs.
  • Losses: The standard objective is RMSE between log-compressed super-resolution and high-resolution rD maps, supplemented by weighted combinations of RMSE, CFAR, and wLSD losses.The loss weights default to λ_i = 1 when unspecified.
  • Metrics: Evaluation uses linear-scale RMSE, SNR, LSD, weighted LSD, and IoU to assess reconstruction quality and target-bin overlap.IoU is computed using target-plus-interference bins from the high-resolution rD map.
  • CFAR settings: CFAR losses use guard and reference cell bands of widths two and five, with PF A settings including 10^-2 and 10^-3.The same CFAR parameters are used for the relevant linear- and super-resolution maps.

VII. RESULTS

The results compare attention mechanisms, spatial extension methods, and CFAR-loss formulations. Zero-padding improves sharpness but increases computational cost, while CFAR-enhanced objectives improve several metrics with trade-offs across detection and reconstruction measures.

  • Spatial extension: Zero-padding models outperform pixel-shuffle versions, and the zero-padding output is notably sharper at comparable effective bandwidth.The sharper result comes with increased computation because the network processes more rD bins.
  • Spatial extension: The effect of zero-padding is larger than the choice of attention window mechanism, although zero-padding increases processing cost.The paper identifies lower-resolution processing after feature extraction as a possible future efficiency strategy.
  • Attention windows: DPSWIN LSD improves with larger windows, while stacked STLs with small windows also provide an apparently sufficient and cheaper receptive-field increase.The best LSD result uses DPSWIN with a large window size.
  • Attention windows: 2D SWIN is overall worse than DPSWIN in LSD, wLSD, and IoU, although its W = 4 configuration achieves the best RMSE and SNR.The authors consider LSD metrics more representative because RMSE and SNR emphasize the strongest bins.
  • CFAR losses: MSE-based CFAR training with a sigmoid generally outperforms R0, and λ_CFAR = 0.5 improves every objective for the reported model.Higher CFAR weighting can improve IoU, RMSE, and SNR while worsening LSD and wLSD.
  • CFAR losses: The authors did not observe a visible reduction in blurring despite metric improvements, suggesting that the visual change may be too small to detect.They propose spectral layers or spectrum-based losses as possible future approaches.

C. Weighted LSD loss

The study evaluates weighted LSD and magnitude-compression choices alongside the broader training design. Power-law compression produces the most balanced results for the reported real-world dataset, while weighted LSD shifts metric trade-offs rather than uniformly improving them.

  • Weighted LSD loss: Weighted LSD has a small effect on LSD, increasing with lower compression and larger β, which down-weights low-power bins more strongly.The hyperparameter β is set to 0.3 for evaluation.
  • Weighted LSD loss: CFAR losses are more effective for increasing IoU, whereas weighted LSD with higher β gives the best RMSE and SNR.The comparison suggests different objectives favor detection overlap versus strong-bin reconstruction.
  • Magnitude compression: Linear-scale inputs can improve RMSE and SNR but produce worse IoU and very poor LSD, while log-compressed R0 gives the best LSD and PF A = 10^-2 IoU.Metric behavior depends strongly on the representation and loss pairing.
  • Magnitude compression: Power-law compression with its corresponding loss yields similar IoU to R0, a slight LSD increase, and notable RMSE and SNR improvements.The reported power-law exponent is γ = 0.3.
  • Magnitude compression: Power-law compression produces slightly less smoothing than log compression, improving RMSE and SNR but worsening LSD because LSD benefits from oversmoothing noise.Linear representations struggle to reconstruct bins outside the main scatterer.
  • Magnitude compression: Overall, power-law compression gives the most well-rounded results, with a small advantage over log compression and a large advantage over no compression.This conclusion is reported for the paper’s dataset.
  • Scope: The supervised setup has practical limitations, including possible range-cell migration, incomplete low-resolution chirps, and no evaluation of downstream target localization.The authors identify additional benchmarks and downstream tasks as future work.
Loading 2608.27354v1…