FreqPKC: Frequency-Aware PeakConv for Radar Semantic Segmentation

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

cs.CVOct 24, 2026

FreqPKC: Frequency-Aware PeakConv for Radar Semantic Segmentation

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

Organizations: College of Electronic Science and Technology, National University of Defense Technology, Changsha, Hunan 410073, China.

Abstract

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.

Figures & tables

Panel of Figure 1 — Fig. 1 : Data processing and spectral analysis visualization. (a) Optical camera image of a driving scene. (b) Corresponding Range-Angle (RA) radar map. (c) Class-wise normalized mean power spectra computed over the entire CARRADA-RAC dataset, showing the spectral separability between different semantic categories (Car, Pedestrian, Cyclist, and Background). The yellow band highlights the mid-frequency region $\rho\in[1/8,3/8]$ where target-background discrimination is most pronounced. (d) Measured radial power spectra of the PKC operator under white-noise excitation for varying guard bandwidths $b_{G}$ , validating the bandpass nature of PKC and the monotonic decrease of the passband center $\rho_{\rm peak}$ with $b_{G}$ .

Panel of Figure 2 — Fig. 2 : Schematic diagram of the Peak Receptive Field (PRF). For a center unit $x_{c}$ , the reference cells $\{x_{r}^{(i)}\}_{i=1}^{N_{r}}$ are sampled outside the guard region defined by guard bandwidth $\mathbf{b}_{G}$ and reference bandwidth $\mathbf{b}_{R}$ .

Fig. 3 : Overview of the Frequency-Aware PeakConv (FreqPKC). The input features are fed into the Frequency-Aware (FA) module for frequency decoupling, producing multiple sub-band features. Each sub-band is weighted and modulated by learnable weights (Multiply) and the modulated features are fused (Add); the Modulation Module then outputs the guard bandwidth offsets $B^{G\dagger}\in\mathbb{R}^{4\times H\times W}$ that control the PKC operation, with $b_{1}^{G\dagger}$ and $b_{2}^{G\dagger}$ representing the offsets at two spatial locations, each specifying bandwidth adjustments in four directions (up, down, left, right).

Fig. 4 : Schematic of the Frequency Decoupling process.

Fig. 5 : The Rotary Scan SAR imaging system, consisting of an array radar subsystem with MIMO antennas, a rotating arm subsystem, and an unmanned ground vehicle (UGV). The LiDAR provides auxiliary data to facilitate the annotation of the radar images.

Fig. 6 : Output power distribution of FreqPKC and standard convolution layers across three radar views (RD, RA, AD) under white-noise input. The power ratio in each normalized band

[f_{l},f_{h})

reflects the layer’s frequency selectivity. Bold lines: group averages;

\Delta

: FreqPKC average minus Conv average.

Model	RD View		RA View
Model	mIoU	mDice	mIoU	mDice
PKCIn-Net (w/o FA)	60.1	72.6	47.3	58.7
w/o frequency decomposition	64.1	76.0	49.2	60.9
w/o band-level modulation	62.9	74.7	49.7	61.4
w/o adaptive $B^{G\dagger}$	63.5	75.3	49.5	61.1
FreqPKC-Net (full FA)	65.6	77.2	52.3	64.2

TABLE I : Component-wise ablation of the FA module on the CARRADA-RAC dataset.

Frameworks	#Params @Frames	RD View		RA View
Frameworks	#Params @Frames	mIoU	mDice	mIoU	mDice
RAMP-CNN [ 45 ]	106.4M@9	56.6%	68.5%	27.9%	30.5%
MVA-Net [ 22 ]	4.8M@3	53.5%	65.3%	37.1%	44.8%
TMVA-Net [ 22 ]	5.6M@5	56.1%	68.0%	37.7%	46.2%
TransRadar [ 49 ]	4.9M@5	56.6%	68.9%	37.9%	46.8%
PKCOn-Net [ 26 ]	6.3M@1	59.5%	71.3%	39.5%	48.3%

TABLE II : Quantitative comparison with SoTA methods on the CARRADA dataset. The best results are highlighted in bold . The second best results are underlined .

Frameworks	#Params @Frames	RD View		RA View
Frameworks	#Params @Frames	mIoU	mDice	mIoU	mDice
RAMP-CNN [ 45 ]	106.4M@9	56.5%	67.5%	37.9%	40.5%
MVA-Net [ 22 ]	4.8M@3	53.6%	66.3%	40.1%	53.9%
TMVA-Net [ 22 ]	5.6M@5	56.3%	68.1%	46.6%	57.9%
PKCOn-Net [ 26 ]	6.3M@1	59.9%	71.4%	45.5%	57.3%
PKCIn-Net [ 26 ]	6.3M@5	60.1%	72.6%	47.3%	58.7%

TABLE III : Quantitative comparison with SoTA methods on the CARRADA-RAC dataset. The best results are highlighted in bold . The second best results are underlined .

Panel of Figure 10 — Fig. 7 : Qualitative segmentation results comparing FreqPKC-Net with selected baseline methods on the CARRADA-RAC dataset. One representative frame is shown from the test split. The top row displays the camera image and RD view results, while the bottom row shows the RA view results. (a) Input RD/RA tensor, (b) Ground truth mask, (c) PKCIn-Net, (d) AdaPKC θ -Net, (e) AdaPKC ξ -Net FiT , (f) FreqPKC-Net (Ours). Different colors represent different object categories. Black: background, Red : car, Yellow : pedestrian, Cyan : cyclist.

Frameworks	#Params @Frames	mIoU	mDice
MVA-Net [ 22 ]	1.30M@3	39.72%	45.19%
TMVA-Net [ 22 ]	1.20M@5	39.44%	44.61%
PKCIn-Net [ 26 ]	1.25M@5	39.80%	45.13%
AdaPKC ξ -Net [ 31 ]	1.25M@5	40.23%	45.94%
AdaPKC θ -Net [ 31 ]	1.25M@5	40.96%	47.04%
FreqPKC-Net (Ours)	1.26M@5	42.16%	49.10%

TABLE IV : Quantitative comparison with SoTA methods on the RS-SAR dataset. The best results are highlighted in bold . The second best results are underlined .

Semantic Segmentation

Semantic segmentation assigns a class label to every image pixel, enabling detailed scene understanding across domains and sensors.

See all 414 papers

Real-Time Radar Semantic and Panoptic Scene Understanding
Oct 19, 2026Santosh Parit, Julia Hindel, Rohit Mohan +4Autonomous Driving Perception Panoptic Segmentation
University of Freiburg · University of Oxford
RAVEN: Radar Adaptive Vision Encoders for Efficient Chirp-wise Object Detection and Segmentation
Apr 6, 2026Anuvab Sen, Mir Sayeed Mohammad, Saibal MukhopadhyaymmWave Radar Sensing Efficient Object Detection
Georgia Institute of Technology, Atlanta, Georgia, USA
Learning Structurally Consistent Representations for Multi-View Radar Semantic Segmentation
Jun 30, 2026Ali Zia, Muhammad Umer Ramzan, Abdelwahed Khamis +2Multi-View Consistency Multi-View Learning
School of Computing, Engineering and Mathematical Sciences, La Trobe University, Melbourne, Australia · La Trobe Institute for Sustainable Agriculture & Food (LISAF), Australia · School of Engineering and Applied Sciences, GIFT University, Pakistan +1

Oct 19, 2026J/K move · Enter open · S save

Real-Time Radar Semantic and Panoptic Scene Understanding

Santosh Parit, Julia Hindel, Rohit Mohan, Georgi Pramatarov+3

University of Freiburg · University of Oxford

Autonomous Driving Perception Panoptic Segmentation Semantic Segmentation mmWave Radar Sensing

Open paper AskPDF arXiv

Robust scene understanding under varying environmental conditions is essential for safe autonomous driving. While cameras and LiDAR degrade in adverse weather and poor lighting, Frequency-Modulated Continuous-Wave (FMCW) spinning radar provides dependable 360° perception at extended ranges with minimal compute cost. However, real-time FMCW radar scene understanding is challenging due to complex signal returns, limited temporal resolution, and scarce annotated datasets. We introduce StreamRPS, a novel framework for streaming FMCW radar panoptic segmentation, comprising an automated label-generation pipeline via LiDAR scene completion and a real-time streaming network for semantic and panoptic segmentation. Our encoder-decoder streaming architecture comprises a radar-specific hierarchical Transformer encoder and task-specific decoders for semantic or panoptic prediction. Our encoder introduces a stem to capture range-dependent signatures, anisotropic patch embeddings to preserve angular resolution, and axial polar attention to separately model the periodic azimuth and bounded range axes. We further introduce azimuth roll rotation augmentation to simulate ego-vehicle rotation during training. Comprehensive experiments on the SAX and BOREAS datasets validate both our auto-labeling methodology and streaming segmentation network against established baselines. We outperform the strongest baseline by at least 1.4pp while processing sectors at over 400 FPS in a 4-sector streaming setting. We additionally provide manual annotations for 100 SAX radar scans, enabling independent evaluation of the entire framework.

Abstract

Semantic Segmentation

Explore similar work

Real-Time Radar Semantic and Panoptic Scene Understanding

RAVEN: Radar Adaptive Vision Encoders for Efficient Chirp-wise Object Detection and Segmentation

Learning Structurally Consistent Representations for Multi-View Radar Semantic Segmentation