PaperlayerView exact revision on arXiv

Source-linked AI summary

FreqPKC: Frequency-Aware PeakConv for Radar Semantic Segmentation

Yangyong Deng, Pengyu Zhang, Shaoqiu Song, Xiaotao Huang, Jian Wang

arXiv:2610.33504v1cs.CV

TL;DR

Radar segmentation needs to handle sparse, peak-dominated measurements, while existing PKC variants do not explicitly use the operator’s frequency response to guide bandwidth. FreqPKC estimates adaptive guard bandwidths from spectral information and integrates them into a multi-view segmentation framework. FreqPKC-Net consistently outperforms existing methods across the reported datasets and views.

  • Problem

    Standard convolutions mismatch sparse, peak-dominated radar data, and existing PKC methods use fixed or blindly learned guard bandwidths without explicit frequency-response guidance.

    Source: §I · pp. 1–2

  • Method

    FreqPKC decouples frequency bands, modulates features by band, and uses spectral energy to adaptively set guard bandwidths within a multi-view segmentation network.

    Source: §I · p. 2

  • Results

    FreqPKC-Net consistently outperforms SoTA methods on CARRADA, CARRADA-RAC, and RS-SAR; on CARRADA RD it exceeds AXFL-Net by 0.6 points in mIoU.

    Source: 1) Quantitative Results Analysis · p. 8, §V · p. 9

  • Takeaways & Limitations

    Evaluations across datasets confirm the framework’s generalization across different radar sensors.

    Source: §V · p. 9

Abstract

from arXiv · show

Automotive millimeter-wave (mmWave) radar has attracted increasing attention for semantic segmentation in autonomous driving, owing to its robustness under adverse weather and lighting conditions. Recent works have introduced radar-specific neural operators such as PeakConv (PKC), inspired by constant false alarm rate (CFAR) detection, to better model the peak-dominated characteristics of radar images. However, existing methods typically treat the guard bandwidth as learnable parameters, without explicitly characterizing the frequency response of the PKC operator. In this work, we revisit PKC from a frequency-domain perspective and propose Frequency-Aware PeakConv (FreqPKC), which explicitly models and exploits its underlying spectral characteristics. By analyzing the frequency response of PKC, FreqPKC enables the network to capture a broader range of frequency components and adaptively adjust its effective bandwidth. With the proposed Frequency-Aware (FA) module, the network can infer appropriate guard bandwidths to balance high- and low-frequency information: it enhances high-frequency components in target regions for more precise segmentation while suppressing noise through low-frequency filtering in background areas. Building on FreqPKC, we further develop FreqPKC-Net, a radar semantic segmentation (RSS) framework that integrates frequency-guided adaptive modulation with cross-view fusion for multi-view inputs, and can also be adapted to single-view data. Experiments on the public datasets CARRADA, CARRADA-RAC, and a self-built dataset named RS-SAR demonstrate that FreqPKC-Net consistently outperforms existing SoTA methods on both range-Doppler and range-angle views.

I. INTRODUCTION

Radar semantic segmentation seeks pixel-level scene understanding from a sensor robust to adverse weather and lighting, but standard image convolutions mismatch sparse, peak-dominated radar data. The paper links PKC guard bandwidth to frequency selectivity and introduces frequency-guided adaptive PRF modulation.

I. INTRODUCTION

  • Standard convolutions can waste computation on empty radar regions and smooth sharp target peaks because they are designed for dense RGB images.Radar measurements are sparse and dominated by localized peak responses.Source: §I · p. 1
  • Targets retain high-frequency spectral energy while background clutter is predominantly low-pass, and PKC guard bandwidth tunes its passband between these signatures.Smaller guard bandwidth shifts the passband higher; larger bandwidth shifts it lower.Source: §I · p. 2
  • FreqPKC uses frequency-domain information to adaptively modulate the peak receptive field through frequency decoupling, band-level feature modulation, and frequency-aware guard bandwidth estimation.The cascade translates global spectral context into local sampling strategies.Source: §I · p. 2
  • FreqPKC-Net takes range-Doppler and range-angle tensors, with angle-Doppler as an auxiliary input, and produces pixel-wise segmentations for both representations.The framework integrates cross-view fusion.Source: §I · p. 2
  • The paper constructs RS-SAR, a vehicle-mounted rotary-scan SAR dataset with finer than 0.8° angular resolution and full 360° panoramic coverage.Its coverage complements forward-facing FMCW radar datasets.Source: §I · p. 2

II. RELATED WORK

  • The related-work review covers automotive radar perception, radar semantic segmentation, and peak receptive fields, identifying limitations motivating frequency-aware design.Source: §II · p. 2

A. Automotive Radar Perception

Automotive radar perception uses point clouds or structured heatmaps, while radar segmentation methods have added multi-view fusion and temporal context. The section traces the peak receptive field to PKC and identifies the limits of spatial-only bandwidth adaptation.

A. Automotive Radar Perception

  • Point-cloud radar methods face information loss during conversion from raw signals, while sparse point representations constrain spatial resolution and localization accuracy.Source: A. Automotive Radar Perception · p. 2
  • Heatmap methods preserve radar tensors’ spectral content and fine-grained scattering structure, motivating task-specific operators for their sparse peak responses.Source: A. Automotive Radar Perception · p. 3

B. Radar Semantic Segmentation

  • Radar segmentation methods progressed from vision architectures to multi-view processing and alignment, but complex visual modules can retain computational redundancy by treating radar as generic imagery.AXFL addresses cross-view misalignment through axial-prior-guided alignment and adaptive fusion; transformer-based temporal methods face high computational costs.Source: B. Radar Semantic Segmentation · p. 3

C. Peak Receptive Field

  • PKC uses a learned peak receptive field to capture radar echo peaks, with reference cells outside the guard region used to estimate background interference.Source: C. Peak Receptive Field · p. 3
  • Existing adaptive PRF methods rely on spatial cues, while their bandwidth estimates may lack physical interpretability or generalize poorly in complex electromagnetic environments.The methods overlook the distinct spectral signatures of target structures and background clutter.Source: C. Peak Receptive Field · p. 3

III. METHODOLOGY

The methodology analyzes PKC as a frequency-selective operator, then relates its guard bandwidth to the position of its passband. The analysis predicts that smaller guards favor higher frequencies, while larger guards favor lower frequencies.

III. METHODOLOGY

  • The methodology first analyzes PKC’s frequency response, then introduces FreqPKC and the overall network.The analysis models PKC as an isotropic kernel with uniform weights and derives its response using a Hankel transform.Source: §III · p. 4, 1) Frequency Response of the PKC… · p. 4

1) Frequency Response of the PKC Operator:

  • PKC has zero DC gain, suppressing spatially uniform background when the center and surroundings share the same intensity.This is described as the frequency-domain manifestation of the CFAR principle.Source: 1) Frequency Response of the PKC… · p. 4
  • For bR ≫ bG, a smaller guard bandwidth shifts PKC’s passband toward higher frequencies, while a larger one shifts it lower.The paper reports that trained models assign smaller effective guards to target regions and larger ones to background regions.Source: 1) Frequency Response of the PKC… · p. 4
  • White-noise tests with five fixed guard bandwidths show the predicted bandpass response, with peak frequency decreasing as bG increases and all spectra reaching zero at DC.The passage reports empirical validation of the closed-form response.Source: 1) Frequency Response of the PKC… · p. 4

2) Frequency-Domain Estimation of the Guard Bandwidth:

Spatial features alone can leave guard-bandwidth estimation ambiguous and physically unconstrained. The proposed approach uses radar spectral signatures to guide guard-bandwidth decisions.

2) Frequency-Domain Estimation of the Guard Bandwidth:

  • Spatial intensity alone cannot distinguish a target edge from an isolated clutter spike, and spatial-feature regression may yield bandwidths that violate the radar signal model.The passage identifies this ambiguity and lack of physical constraints in AdaPKCθ and AdaPKCξ.Source: 2) Frequency-Domain Estimation of the… · p. 4
  • Targets have spatially localized, spectrally broad scattering, while background clutter is spatially smooth and predominantly low-pass.The passage describes these spectral signatures as complementary to spatial intensity for resolving ambiguity.Source: 2) Frequency-Domain Estimation of the… · p. 4
  • The FA module decomposes features into non-overlapping sub-bands to support physically grounded guard-bandwidth decisions and frequency-guided modulation.The passage states that multi-dataset evaluations validate generalization across different radar sensors.Source: 2) Frequency-Domain Estimation of the… · p. 4

B. Frequency-Aware PeakConv

FreqPKC estimates directional guard-bandwidth offsets from frequency-domain information. Its FA and modulation stages process and combine sub-band features before controlling the PKC operation.

B. Frequency-Aware PeakConv

  • The FA module addresses spatial-only guard-bandwidth estimation by using frequency-domain information.The passage frames this as addressing the spatial-only estimation discussed earlier.Source: B. Frequency-Aware PeakConv · p. 4
  • FreqPKC frequency-decouples input features into sub-bands, weights and modulates each band, then fuses them before producing offsets that control PKC.The figure depicts the offsets as BG† ∈ R4×H×W.Source: Figure 3 · p. 5
  • The FA module analyzes spectral energy across frequency bands and maps those cues to directional guard-bandwidth offsets at each spatial location.The resulting offsets guide adaptive sampling by the peak convolution on the original feature map.Source: B. Frequency-Aware PeakConv · p. 5

1) Frequency Decoupling Process:

Frequency decoupling centralizes the spectrum and uses low-pass masks to separate frequency components. Cascaded residual subtraction yields non-overlapping sub-bands ordered from fine target structures to coarse background residuals.

1) Frequency Decoupling Process:

  • A 2D FFT and spectrum centralization move zero frequency to the center, enabling mask-based filtering of the input feature map.The centralization allows masks to retain low-frequency energy in the central region.Source: 1) Frequency Decoupling Process · p. 5
  • The process uses progressively ordered cutoff scales and centered rectangular low-pass masks to retain central regions of the spectrum.A larger scale factor corresponds to a narrower low-frequency passband and lower effective cutoff frequency.Source: 1) Frequency Decoupling Process · p. 5
  • Cascaded residual subtraction partitions the spectrum into non-overlapping sub-bands, from fine high-frequency target scattering to a coarse low-frequency background residual.The passage says residual decoupling avoids band overlap and supports numerical stability and smooth gradient propagation.Source: 1) Frequency Decoupling Process · p. 5

2) Band-level Learnable Feature Modulation:

The module reweights decoupled frequency bands with learned spatial attention, then uses the resulting feature map to infer directional guard-bandwidth offsets.

2) Band-level Learnable Feature Modulation:

  • Learned spatial attention reweights each decoupled band, while an additional attention map also modulates the low-frequency residual before all M + 1 bands are summed.The design aims to enhance components associated with target structures and suppress bands dominated by noise or clutter.Source: 2) Band-level Learnable Feature… · pp. 5–6
  • A learnable output head maps the reweighted feature map to guard-bandwidth offsets for inferring the effective peak receptive field.Its convolution and batch-normalization layers fuse multi-band energy into directional control parameters.Source: 2) Band-level Learnable Feature… · p. 6
  • The attention mechanism lets the network discern which frequency components contribute positively or negatively to receptive-field expansion.The frequency-band attention maps are generated by dedicated depthwise convolutions initialized with zero weights.Source: 2) Band-level Learnable Feature… · pp. 5–6

3) Frequency-Aware Modulated Guard Bandwidth:

Predicted directional guard-bandwidth offsets deform the reference-cell grid, and differentiable interpolation samples the resulting coordinates for end-to-end learning.

3) Frequency-Aware Modulated Guard Bandwidth:

  • Directional guard-bandwidth components deform the fixed reference grid around each convolution window to produce updated reference-cell coordinates.A mapping function projects the four directional bandwidth scalars onto coordinate offsets.Source: 3) Frequency-Aware Modulated Guard… · p. 6
  • Bilinear interpolation samples each floating-point reference coordinate from its four neighboring feature-grid values.The interpolation resolves sub-pixel coordinates and preserves differentiability with respect to the predicted offset.Source: 3) Frequency-Aware Modulated Guard… · p. 6
  • Differentiable interpolation lets gradients of the guard bandwidth flow back to the frequency-aware module during end-to-end training.The module adds marginal computational overhead through lightweight FFT-based decoupling and a single estimation-head convolution.Source: 3) Frequency-Aware Modulated Guard… · p. 6

C. The Overall Framework of FreqPKC-Net

FreqPKC-Net combines multi-view radar features for range-Doppler and range-angle segmentation, and is also evaluated with a single-view architecture on RS-SAR. Its frequency-selectivity analysis finds that FreqPKC shifts output energy toward mid-frequency bands compared with standard convolution.

C. The Overall Framework of FreqPKC-Net

  • FreqPKC-Net uses three view-specific encoders, latent-space fusion, and two decoders to produce range-Doppler and range-angle segmentation masks.The angle-Doppler branch is auxiliary; each encoder uses FreqPKC blocks and multi-scale context aggregation.Source: C. The Overall Framework of FreqPKC-Net · p. 6

A. Datasets and Training Setting

  • The experiments evaluate frequency-aware radar segmentation on CARRADA, CARRADA-RAC, and the self-built RS-SAR dataset.RS-SAR contains single-view range-angle images with small- and large-vehicle classes, so it uses single-view architectures; the public datasets provide multi-view radar data.Source: A. Datasets and Training Setting · pp. 6–7, §IV · p. 6
  • Models use five consecutive frames as input and are evaluated on test-set mIoU and mDice averaged over classes.Training uses weighted cross-entropy, Dice, and coherence losses on CARRADA and CARRADA-RAC, while RS-SAR uses weighted cross-entropy and Dice losses.Source: A. Datasets and Training Setting · p. 7

B. Frequency Selectivity Analysis

  • Over 65% of standard convolution output energy falls in [0, 1/8), while FreqPKC gains +0.232 in [1/8, 1/4) and +0.203 in [1/4, 3/8).The comparison feeds white noise into trained layers and measures output power ratios across normalized frequency bands; the reinforced mid-bands coincide with the target-background discriminative region.Source: Figure 6 · p. 7, B. Frequency Selectivity Analysis · p. 7

C. Ablation Studies

FreqPKC-Net outperforms competing methods across the reported CARRADA and CARRADA-RAC comparisons. Ablations favor a four-band decomposition and show contributions from all three frequency-aware submodules.

C. Ablation Studies

  • The four-band configuration K = {2, 4, 8} achieves the best accuracy in the CARRADA-RAC sweep and is used in subsequent experiments.Coarser decompositions lack resolution to separate target scales, while finer ones add redundant high-frequency bands whose attention maps overfit training spectra.Source: C. Ablation Studies · p. 7
  • Removing any FA submodule lowers all four metrics; band-level modulation and adaptive bandwidth have the largest drops, at −2.7 and −2.1 RD mIoU.Single-removal variants still exceed PKCIn-Net by 2.8–4.0 RD mIoU, which the authors attribute to cooperation among the three components rather than parameter growth.Source: C. Ablation Studies · p. 7

1) Quantitative Results Analysis:

  • 65.5% mIoU and 77.1% mDice on CARRADA RD, and 49.1% mIoU and 61.1% mDice on RA, exceed AXFL-Net by 0.6/0.5 and 0.9/0.8 points, respectively.CARRADA-RAC shows a similar trend, with consistent margins over AXFL-Net.Source: Table II · p. 8, Table III · p. 8, 1) Quantitative Results Analysis · p. 8
  • FreqPKC-Net exceeds MARSS by 2.2/1.9 RD mIoU/mDice and 2.1/2.3 RA points on CARRADA, and by 4.2/3.8 RD and 3.3/3.3 RA points on CARRADA-RAC.The authors note that subband decoupling without adaptive receptive-field modulation leaves accuracy gains untapped.Source: 1) Quantitative Results Analysis · p. 8
  • On CARRADA, the RA gains over AXFL-Net are larger than the RD gains: +0.9/+0.8 versus +0.6/+0.5 mIoU/mDice.The paper relates this view-dependent pattern to greater angular ambiguity and denser RA clutter, and describes FA as complementary to spatial feature learning.Source: 1) Quantitative Results Analysis · p. 8
  • The RD mIoU margin over AXFL-Net grows from 0.6 on CARRADA to 1.5 on CARRADA-RAC; margins over AdaPKCξ-NetF iT reach 4.8 RD and 3.5 RA mIoU on CARRADA-RAC.Relative to PKCIn-Net, the FA module contributes 5.5 RD mIoU on both datasets and 7.1/5.0 RA mIoU on CARRADA/CARRADA-RAC.Source: 1) Quantitative Results Analysis · p. 8
  • On CARRADA RA, FreqPKC-Net reaches 49.1% mIoU, 4.8 points above AdaPKCξ-NetF iT, while both comparison methods show sharp drops from RD to RA.AXFL-Net falls from 64.9% to 48.2% mIoU and AdaPKCξ-NetF iT from 62.1% to 44.3%.Source: 1) Quantitative Results Analysis · p. 8
  • FreqPKC-Net adds 0.17M parameters (+2.7%) and 1.95 GFLOPs (+1.8%) over PKCIn-Net while delivering the best accuracy in Tables II and III.Inference efficiency was benchmarked on one RTX 3090 GPU at batch size 1 in FP32; FPS and latency are reported in the Supplementary Material.Source: 1) Quantitative Results Analysis · p. 8

2) Qualitative Results Analysis:

FreqPKC-Net produces more complete contours and fewer false positives in the representative CARRADA-RAC RA scene, and outperforms competitors on RS-SAR. Across the three datasets, the paper reports consistent gains across sensor configurations, while noting that substantially different sensors may require recalibrating K.

2) Qualitative Results Analysis:

  • In the representative CARRADA-RAC test scene, FreqPKC-Net produces more complete RA contours, better continuity, and fewer false positives than the comparison models.All methods perform comparably on RD in this scene; comparison models show target incompleteness on RA. AXFL-Net is excluded because its open-source implementation is unavailable.Source: 2) Qualitative Results Analysis · p. 8, Figure 7 · p. 9

E. RSS Performance on RS-SAR Dataset

  • For substantially different radar configurations, K may need recalibration; systematic cross-sensor transfer remains future work.The evaluated settings include forward-facing 77 GHz FMCW radar and a 360° rotary-scan SAR system.Source: E. RSS Performance on RS-SAR Dataset · p. 9

V. CONCLUSION

  • The conclusion presents FreqPKC as mapping multi-band spectral statistics to directional guard-bandwidth offsets, balancing high-frequency target detail against low-frequency clutter.FreqPKC-Net combines this operator with cross-view fusion, and the paper reports consistent gains across CARRADA, CARRADA-RAC, and RS-SAR.Source: §V · p. 9

E. RSS Performance on RS-SAR Dataset

Related papers

Loading 2610.33504v1…