PaperlayerView exact revision on arXiv

Source-linked AI summary

Multistatic ISAC with Deep Learning-Based Clutter Suppression for 5G and Beyond

Yalin E. Sagduyu, Tugba Erpek, Sastry Kompella

arXiv:2610.40586v1cs.NIcs.AIcs.LGeess.SP

TL;DR

Dense, dynamic clutter can overlap target returns in multistatic ISAC and degrade range-Doppler estimation, localization, and velocity estimation. The paper models varied clutter conditions and introduces residual and attention-aware deep-learning suppression methods. These methods outperform conventional approaches across clutter conditions and improve sensing estimates.

  • Problem

    In multistatic ISAC, dense and dynamic clutter can overlap target returns in bistatic range-Doppler space, degrading target localization and velocity estimation.

    Source: §I · p. 1

  • Method

    The paper trains residual U-Net and attention-gated anisotropic deep-learning methods to enhance target responses from clutter-contaminated matched-filter observations.

    Source: §V · p. 3, A. Residual U-Net-based DLCS · p. 3, B. Attention-Gated Anisotropic DLCS · p. 4

  • Results

    Across position-error thresholds and clutter settings, DLCS methods outperform NCS and RMHP, with DLCS-2 best and DLCS-1 close behind.

    Source: §VI · p. 5

  • Takeaways & Limitations

    The proposed methods substantially improve clutter suppression and reduce sensing errors relative to conventional approaches across diverse clutter conditions in multistatic ISAC.

    Source: §VII · p. 5

Abstract

from arXiv · show

Integrated sensing and communications (ISAC) systems using 5G and beyond signals of opportunity enable joint wireless sensing and communication using shared infrastructure and spectrum resources. While multistatic ISAC can increase sensing coverage with spatially distributed receivers, dense and dynamic clutter can significantly degrade bistatic range and Doppler estimation, leading to errors in target localization and velocity estimation. This paper presents a multistatic ISAC framework with clutter suppression for high-fidelity estimation of target position and velocity. Distributed clutter is modeled with randomly located scatterers with varying Doppler spreads and radar cross sections to capture dense and dynamic clutter effects. A broad range of conventional signal processing baselines are evaluated for clutter suppression, showing limited robustness under strong clutter conditions. To address this limitation, deep learning-based clutter suppression methods are introduced to improve robustness against dense and dynamic clutter by employing residual and attention-enhanced deep neural network architectures that directly enhance target responses from clutter-contaminated observations. These methods substantially outperform conventional approaches in target position and velocity estimation across diverse clutter conditions in multistatic ISAC.

I. INTRODUCTION

The paper studies 5G NR PRS-based multistatic sensing, where spatially distributed receivers broaden coverage but clutter can overlap target returns and degrade localization and velocity estimates. It introduces a randomized clutter model and deep-learning suppression methods that improve target-response estimation over conventional approaches.

I. INTRODUCTION

  • 5G NR positioning reference signals can support passive sensing without dedicated radar transmitters and persistent, low-cost sensing over large areas.Source: §I · p. 1
  • Spatially distributed receivers provide multistatic diversity and wider coverage, but receiver-dependent clutter can overlap target returns in bistatic range-Doppler space and degrade localization and velocity estimation..Source: §I · p. 1

II. MULTISTATIC ISAC SIGNAL AND CLUTTER MODEL

  • The framework uses 5G NR PRS, cross-ambiguity-function processing, and nonlinear multilateration to estimate bistatic range-Doppler, target position, and velocity.Source: §I · p. 1
  • Clutter is modeled with randomly located scatterers whose radar cross sections and Doppler spreads vary, including conditions with substantial Doppler overlap with target returns.Larger clutter density and Doppler spread represent denser and more dynamic clutter, respectively.Source: §I · p. 1, §II · p. 2

I. INTRODUCTION

  • Residual and attention-aware deep-learning methods substantially outperform conventional suppression across clutter conditions in bistatic range-Doppler estimation, localization, and velocity estimation.The methods directly enhance target responses from clutter-contaminated observations.Source: §I · p. 1

III. TARGET LOCALIZATION AND VELOCITY ESTIMATION

The estimation pipeline correlates received samples with the known PRS, forms a slow-time delay response, and applies CAF processing to obtain bistatic range-Doppler maps. It then refines range and radial velocity, jointly estimates position from bistatic ranges, and estimates Cartesian velocity from radial velocities.

  • Received samples are correlated with the known PRS reference sequence to produce delay-bin correlation outputs for each slow-time observation.Source: §III · p. 2
  • The slow-time delay-response matrix is clutter-suppressed before coherent integration over a Doppler grid to form the CAF response.Source: §III · p. 2
  • CA-CFAR detects targets on the CAF magnitude map, and parabolic interpolation refines bistatic range and radial velocity estimates.Source: §III · p. 2
  • Target position is jointly fitted from receiver bistatic ranges, each defining an ellipse with the transmitter and receiver as foci; a multistart trust-region solver selects the minimum-residual solution.Source: §III · pp. 2–3
  • Ridge-regularized least squares estimates Cartesian velocity from the radial-velocity relations; speed and heading are computed from its components.Source: §III · p. 3

IV. BASELINES FOR CLUTTER SUPPRESSION

The paper benchmarks conventional clutter suppression methods on the slow-time delay correlation matrix before computing the CAF and estimating bistatic range-Doppler.

  • Conventional baselines operate on the slow-time delay correlation matrix before CAF computation and bistatic range-Doppler estimation.Source: §IV · p. 3
  • The evaluated methods suppress clutter through temporal differencing, subspace projection, Doppler-domain filtering, adaptive prediction, and smooth-component estimation.Source: §IV · p. 3
  • After suppression, the resulting correlation matrices feed CAF computation for bistatic range-Doppler estimation.Source: §IV · p. 3

1) No Clutter Suppression (NCS):

The baseline section includes an unsuppressed reference and a coherent current-reference subtraction method. CUR aligns the current observation's clutter phase to a clutter-only reference before subtraction, then uses the result for CAF-based range and Doppler estimation.

1) No Clutter Suppression (NCS):

2) Current-Reference Subtraction (CUR) :

3) Space-Projection Adaptive Difference (SP-AD):

These conventional methods suppress clutter through slow-time differencing, subspace or low-Doppler projection, filtering, adaptive prediction, and smooth-component estimation.

3) Space-Projection Adaptive Difference (SP-AD):

4) Rank-One Slow-Time Projection (SP-R1):

5) DPSS Band-Stop Filtering (DPSS):

6) Clutter Map Temporal Doppler Filtering (CMTD):

7) Total Variation High-Pass Filtering (TVHP):

8) Rolling-Median High-Pass Filtering (RMHP):

9) Wavelet High-Pass Filtering (WHP):

10) Adaptive Least Mean Squares (LMS) Residual Filtering (ALMS):

11) Polynomial High-Pass Filtering (POLYHP):

The section moves from polynomial slow-time clutter removal to supervised methods that estimate target-only responses from clutter-contaminated correlation matrices.

11) Polynomial High-Pass Filtering (POLYHP):

V. DEEP LEARNING-BASED CLUTTER SUPPRESSION

  • DLCS methods estimate the target-only correlation matrix from clutter-contaminated matched-filter observations for subsequent range-Doppler estimation.Training realizations vary receiver geometry, target state, clutter density, RCS, Doppler spread, and noise.Source: §V · p. 3

A. Residual U-Net-based DLCS

  • DLCS-1 uses a supervised residual U-Net to estimate the target-only matched-filter response directly from clutter-corrupted observations.Source: A. Residual U-Net-based DLCS · p. 3

1) Input Representation:

  • The input representation uses in-phase and quadrature components of the correlation matrix, normalized by the input tensor's RMS magnitude.Source: 1) Input Representation · p. 3

2) Network Architecture:

DLCS-1 uses an encoder-decoder U-Net with skip connections, trains against target responses with reconstruction and sparsity terms, and applies AdamW optimization.

2) Network Architecture:

  • The encoder-decoder U-Net uses convolutional blocks and max-pooling to extract hierarchical features, then reconstructs a target-enhanced residual map through bilinear upsampling and skip-connected features.Skip connections preserve fine spatial structures corresponding to localized target peaks in the range-Doppler domain.Source: 2) Network Architecture · pp. 3–4

3) Training Objective:

  • The training objective combines mean-squared error for target reconstruction with sparsity regularization to suppress residual clutter artifacts.The ℓ1 term promotes sparse target responses, while the ℓ2 term encourages accurate reconstruction.Source: 3) Training Objective · p. 4

4) Optimization:

  • The network uses AdamW with learning rate 10^-3 and weight decay 10^-5, with mini-batches of random realizations and gradient clipping at threshold 5.After training, its output is processed in the CAF and range-Doppler estimation stages.Source: 4) Optimization · p. 4

B. Attention-Gated Anisotropic DLCS

DLCS-2 extends the residual U-Net with multiple input representations and directional convolutional processing for clutter-dominated range-Doppler maps.

B. Attention-Gated Anisotropic DLCS

1) Multi-Channel Input Representation:

  • Its four complementary input representations include magnitude and logarithmic channels, which capture target energy and compress dynamic range to represent weak targets in strong clutter.Source: 1) Multi-Channel Input Representation · p. 4

2) Anisotropic Convolutional Encoder:

  • Anisotropic 5 × 1 and 1 × 5 kernels model slow-time Doppler structures and delay-domain correlations separately, helping distinguish target signatures from differently shaped clutter.Source: 2) Anisotropic Convolutional Encoder · p. 4

3) Attention-Gated Skip Connections:

Attention-gated skip features suppress clutter-dominated activations while retaining target-relevant structures. Across clutter settings, the DL methods outperform conventional suppression, with DLCS-2 performing best when trained with more data and epochs.

3) Attention-Gated Skip Connections:

  • Attention-gated skip connections refine encoder features using decoder context, suppressing clutter-dominated activations while preserving target-dominant spatial structures.The attention mask gates the encoder skip feature map before feature fusion.Source: 3) Attention-Gated Skip Connections · p. 4

4) Decoder Structure:

  • The decoder uses 5 × 1 and 1 × 5 anisotropic kernels to model slow-time Doppler structures and delay-domain correlations, respectively.It reconstructs the target-enhanced residual map through bilinear upsampling and concatenation with attention-refined skip features.Source: 4) Decoder Structure · p. 4

5) Loss Function:

  • The network combines reconstruction and sparsity losses, with total-variation regularization promoting spatial smoothness while preserving localized target peaks.The sparsity and total-variation terms are controlled by λs and λTV, respectively.Source: 5) Loss Function · p. 4

6) Optimization and Training:

  • After AdamW training with gradient clipping and weight decay, the network output feeds CAF-based range-Doppler estimation and bistatic localization.Source: 6) Optimization and Training · p. 4

VI. PERFORMANCE EVALUATION

  • Under strong clutter, DLCS-1 estimates bistatic range and radial velocity for all receivers, with 4.462 m position error and 0.445 m/s speed and 0.045° angle errors.NCS and RMHP produce large localization and velocity errors in this test; RMHP fails to recover target responses for some receivers.Source: §VI · p. 5, Table I · p. 5, Table II · p. 5
  • With Ntrain = 5000 and Nepoch = 500, DLCS-2 approaches the reference method and consistently outperforms DLCS-1 across clutter settings.Source: §VI · p. 5
  • Across position-error thresholds and clutter settings, DLCS-2 performs best, followed closely by DLCS-1, and both outperform NCS and RMHP.This comparison uses models trained with Ntrain = 5000 and Nepoch = 500.Source: §VI · p. 5

VII. CONCLUSION

The paper studies clutter suppression for multistatic ISAC using 5G NR PRS and evaluates bistatic range-Doppler estimation, localization, and velocity estimation. Its DL-based methods improve clutter suppression and reduce sensing errors relative to conventional approaches across diverse clutter conditions.

  • The study finds that clutter can severely degrade multistatic ISAC sensing accuracy, while conventional suppression provides limited gains under challenging conditions.Source: §VII · p. 5
  • The proposed DL-based methods directly enhance target responses from clutter-corrupted observations.Source: §VII · p. 5
  • Across diverse clutter conditions, the DL methods substantially improve clutter suppression and reduce sensing errors relative to conventional approaches.Source: §VII · p. 5
Loading 2610.40586v1…