PaperlayerView exact revision on arXiv

Source-linked AI summary

StarBOA: Real-Time Mamba State-Space Unrolling for Sparse Radar Micro-Doppler in ISAC Networks

Mustafa Bora Çelik, Ceren Çelik, Orhan Gazi

arXiv:2609.33408v1cs.LGcs.ITcs.NImath.IT

TL;DR

Under ISAC chirp subsampling, high-fidelity micro-Doppler reconstruction is needed despite severe pulse dropouts. StarBOA replaces short-buffer attention with recurrent Mamba state and outperforms STAR across tested sparsity tiers, reaching a +0.2472 SSIM gain at 90% sparsity.

Abstract

from arXiv · show

In Integrated Sensing and Communications (ISAC), radar sensing must operate under chirp subsampling with up to 90\% missing data. An attention-based baseline, limited to a 52~ms buffer, collapses toward maximum uniform entropy ($H=2.584$ bits) as sparsity increases, failing to capture long-range gait-cycle context. We propose StarBOA, which replaces attention with a causal Mamba state-space model that updates incrementally on a per-window basis without re-scanning past reconstructions. By maintaining a persistent state, StarBOA integrates over $100\times$ more temporal history at no additional per-step computational cost. StarBOA outperforms the baseline's published results across all sparsity levels, with SSIM gains increasing from $+0.0379$ at 50\% missing data to $+0.2472$ at 90\%. Each window is processed in 1.53~ms with zero lookahead, demonstrating efficient causal reconstruction under extreme chirp subsampling.

I. INTRODUCTION

Under severe chirp subsampling, short-buffer attention can lose gait-cycle context; StarBOA replaces it with recurrent state-space modeling and reports stronger reconstruction across sparsity tiers.

I. INTRODUCTION

  • The introduction motivates high-fidelity micro-Doppler reconstruction because sparse chirps can produce velocity aliasing and noise that obscure activity-relevant kinematic signatures.In the described ISAC setting, radar sensing is confined to a 10% duty cycle.Source: §I · p. 1

B. Contributions: Continuous-Time State-Space Dynamics

  • StarBOA replaces STAR’s non-parametric attention with a selective continuous-time state space, carrying context across 128 windows with constant per-step update cost.STAR uses a fixed 6-window buffer; StarBOA’s context is extendable.Source: B. Contributions: Continuous-Time… · p. 1
  • At 90% chirp missingness, STAR’s attention reaches maximum uniform entropy, H = 2.5840 bits, and degrades to moving-average blur.The analysis attributes the collapse to noisy candidate spectra flattening attention weights across the historical buffer.Source: B. Contributions: Continuous-Time… · p. 1, A. Limitations of Attention Under… · p. 1
  • At 50%, 75%, and 90% missing chirps, StarBOA outperforms STAR-Attention in decoupled evaluation and keeps kinematic structure correlation above s > 0.888.The evaluation uses a frozen LIHT front-end on the public DISC 60 GHz benchmark.Source: B. Contributions: Continuous-Time… · p. 1
  • +0.0379 at 50% and +0.2472 at 90% sparsity are StarBOA’s SSIM gains over STAR’s officially reported results.The gain widens across the tested sparsity levels.Source: B. Contributions: Continuous-Time… · p. 1
  • StarBOA’s recurrent state updates without reprocessing history, supporting incremental streaming rather than repeated comparisons against a stored frame buffer.The contribution describes this as constant-time updating.Source: B. Contributions: Continuous-Time… · pp. 1–2

II. PRELIMINARIES: THE STAR FRAMEWORK

StarBOA builds on the unrolled architecture introduced by STAR, whose three cascaded blocks are summarized in this section.

II. PRELIMINARIES: THE STAR FRAMEWORK

  • StarBOA builds on STAR’s unrolled architecture, which the paper summarizes as three cascaded blocks.The passage frames this summary as background for the proposed method.Source: §II · p. 2

A. Channel Model and Incomplete Sampling

The channel model represents intermittently sampled slow-time CIR data as incomplete measurements of a sparse Doppler spectrum with noise. The STAR pipeline applies a single learned IHT step, then temporal attention and solution refinement; StarBOA retains Blocks A and C while replacing attention.

A. Channel Model and Incomplete Sampling

  • Only a subset of K chirps is gathered in each slow-time window, yielding an incomplete measurement vector h[t] for the sparse Doppler spectrum.The model specifies the sampled-index selector, inverse Fourier dictionary, unknown spectrum, and noise.Source: A. Channel Model and Incomplete Sampling · p. 2

B. The Three-Block STAR Pipeline

  • STAR uses a single learned IHT step to produce a coarse spectrum, then compares it with six past recovered spectra to extract temporal context.The context vector is used to refine the coarse spectrum through additive transformation and multiplicative gating.Source: B. The Three-Block STAR Pipeline · p. 2
  • StarBOA retains STAR’s physical LIHT front-end and refinement block, replacing only the non-parametric attention block with continuous-time selective state-space modeling.Its recurrent state accumulates context across windows rather than matching against a noisy historical buffer.Source: B. The Three-Block STAR Pipeline · p. 2, Figure 1 · p. 2

A. Sequential State-Space Context Modeling (Block B)

StarBOA replaces STAR’s limited-buffer attention with continuous-time selective state-space modeling that assimilates historical information into new slow-time windows. Its discretized recurrent update is causal for streaming operation, and the benchmark table evaluates the frozen-LIHT front end at three sparsity levels.

A. Sequential State-Space Context Modeling (Block B)

  • StarBOA replaces STAR’s dot-product attention with a continuous-time selective state-space model that uses historical information to reconstruct missing components in newly arriving windows.The model is intended to avoid affinity computations over a limited buffer of noisy frames.Source: A. Sequential State-Space Context… · p. 2
  • The coarse normalized spectrum from Block A is projected into state-space and gating branches, then processed with a causal convolution and input-dependent timescale and selective vectors.The convolution buffers the prior three steps; the state-space dimensions are D = 64 and N = 16.Source: A. Sequential State-Space Context… · p. 2
  • In online edge streaming, the recurrent state update is strictly causal with 0.00 ms lookahead; the benchmark evaluates a frozen LIHT front end at 50%, 75%, and 90% sparsity on DISC.Source: A. Sequential State-Space Context… · p. 3, Table I · p. 3

B. Residual Solution Refinement (Block C)

StarBOA applies solution refinement to Mamba’s gated context using a zero-initialized residual MLP. This makes the initial mapping an identity over the LIHT reconstruction while allowing the model to learn trajectory corrections.

B. Residual Solution Refinement (Block C)

C. Multi-Objective Loss Function

The network is trained end-to-end with a composite reconstruction objective. Its SSIM term uses an accelerated pooling operation during backpropagation, while testing computes canonical SSIM.

C. Multi-Objective Loss Function

F +0.5LSSIM(ˆS, SGT), (11)

IV. EXPERIMENTAL RESULTS AND DISCUSSION

StarBOA improves reconstruction fidelity across tested sparsity levels and processes each window with low latency and zero lookahead. Evaluation uses the DISC dataset and includes a controlled comparison with a frozen LIHT front-end.

A. Experimental Setup

  • The DISC dataset covers WALKING, RUNNING, SITTING, and HANDS, with 50%, 75%, and 90% missing chirps tested using identical data splits and 20 training epochs.Frames contain K = 64 slow-time samples, shifted by δ = 32.Source: A. Experimental Setup · p. 3

B. Decoupled Front-End Benchmark Comparison

  • +2.64%, +4.05%, and +9.77% SSIM improvement at 50%, 75%, and 90% sparsity, respectively, with kinematic structure correlation above s > 0.888.The decoupled comparison freezes the LIHT front-end so both refinement architectures receive identical coarse inputs.Source: B. Decoupled Front-End Benchmark… · p. 3

C. Real-Time Streaming and Recurrent Memory Horizon

  • 1.5255 ms per radar window with 0.00 ms lookahead; inference is over 11× faster than the physical window duration and over 5× faster than the arrival stride.Latency comprises 730.6 µs for LIHT, 670.0 µs for the Mamba step, and 124.9 µs for refinement.Source: C. Real-Time Streaming and Recurrent… · p. 3
  • In causal online streaming, recovery precision ascends monotonically as memory horizon expands, smoothly approaching steady-state performance at H = 128, spanning 1.11 s (128×8.64 ms).The evaluation varied H over {2, . . . , 128} past windows.Source: C. Real-Time Streaming and Recurrent… · p. 3

D. Root-Cause Analysis: Attention Entropy Collapse vs. Recurrent State Memory

Under 90% missingness, STAR's attention weights flatten across its six-window history, while StarBOA maintains recurrent state that integrates coherent trajectories across many windows.

D. Root-Cause Analysis: Attention Entropy Collapse vs. Recurrent State Memory

  • At 90% missingness, STAR attention entropy reaches 2.5840 bits, near the 2.5850-bit uniform maximum, as weights flatten across six historical windows.This collapse compounds STAR's 52 ms buffer limitation and is associated with moving-average blur.Source: D. Root-Cause Analysis: Attention… · p. 3
  • At 90% sparsity, StarBOA reconstructs a sharp torso line and continuous limb oscillation arcs (SSIM = 0.7832), compared with STAR-Attention's blurred limb arcs (SSIM = 0.536).The figure presents the ground truth, raw LIHT input, STAR-Attention, and StarBOA spectrograms over v ∈ [−4.5, 4.5] m/s and t ∈ [0, 4] s.Source: Figure 2 · p. 4
  • StarBOA's recurrent hidden state uses continuous recurrence and selective input projections to suppress instantaneous noise bursts and integrate coherent trajectories across H ≥ 128 windows.Unlike attention over an external buffer of corrupted past estimates, its state does not depend on static window averaging.Source: D. Root-Cause Analysis: Attention… · pp. 3–4

E. Visual Spectrogram Inspection

Visual inspection shows sharper gait trajectories with StarBOA than with STAR-Attention under extreme sparsity. In the full end-to-end evaluation, StarBOA's SSIM gains widen as missingness increases, and the conclusion reports 1.53 ms inference with zero lookahead.

E. Visual Spectrogram Inspection

  • Figure 2 visually corroborates Table I: StarBOA reconstructs a continuous torso trajectory and limb-oscillation arcs, whereas STAR-Attention exhibits moving-average smearing and fragmented footfalls.Source: E. Visual Spectrogram Inspection · p. 4

F. Full End-to-End Pipeline Evaluation

  • +0.0379 SSIM at 50% missing chirps, widening to +0.2472 at 90% sparsity against STAR's officially reported results.The authors relate the widening gain to the compounding buffer-span and entropy-collapse mechanisms.Source: F. Full End-to-End Pipeline Evaluation · p. 4

V. CONCLUSION

  • StarBOA consistently outperforms STAR across all sparsity tiers on DISC, with SSIM gains widening to +0.2472 at 90% missing chirps and execution in 1.53 ms with zero lookahead.The conclusion states that this allows base stations to allocate 90% of resources to communications while sustaining high-fidelity micro-Doppler sensing.Source: §V · p. 4

Related papers

Loading 2609.33408v1…