PaperlayerView exact revision on arXiv

Source-linked AI summary

Sub-Nyquist Radar via Doppler Focusing

Omer Bar-Ilan, Yonina C. Eldar

arXiv:1211.0722v3cs.IT

TL;DR

The paper addresses sub-Nyquist detection and parameter estimation for sparse pulse-Doppler radar scenes, where prior CS methods had sampling, transmitter, dictionary-size, or noise limitations. It introduces Doppler focusing with Xampling and CS to recover targets from low-rate samples, achieving linear-in-P SNR scaling and near-full-Nyquist performance at one tenth the Nyquist rate for SNR above -25dB.

  • Problem

    Sub-Nyquist radar needs low-rate target recovery, but prior CS approaches may retain Nyquist sampling, constrain transmitters, require large dictionaries, or handle noise poorly.

  • Method

    Doppler focusing combines Xampling and CS to recover sparse delay-Doppler parameters from low-rate samples without transmitter restrictions or Nyquist-rate digital processing.

  • Results

    Doppler focusing provides SNR scaling linear in P and, at one tenth the Nyquist rate, achieves results almost equal to classic Nyquist-rate recovery for SNR above -25dB.

  • Takeaways & Limitations

    The approach supports low-rate ADC and DSP with a CS dictionary whose size does not grow with the number of pulses P.

Abstract

from arXiv · show

We investigate the problem of a monostatic pulse-Doppler radar transceiver trying to detect targets, sparsely populated in the radar's unambiguous time-frequency region. Several past works employ compressed sensing (CS) algorithms to this type of problem, but either do not address sample rate reduction, impose constraints on the radar transmitter, propose CS recovery methods with prohibitive dictionary size, or perform poorly in noisy conditions. Here we describe a sub-Nyquist sampling and recovery approach called Doppler focusing which addresses all of these problems: it performs low rate sampling and digital processing, imposes no restrictions on the transmitter, and uses a CS dictionary with size which does not increase with increasing number of pulses P. Furthermore, in the presence of noise, Doppler focusing enjoys an SNR increase which scales linearly with P, obtaining good detection performance even at SNRs as low as -25dB. The recovery is based on the Xampling framework, which allows reducing the number of samples needed to accurately represent the signal, directly in the analog-to-digital conversion process. After sampling, the entire digital recovery process is performed on the low rate samples without having to return to the Nyquist rate. Finally, our approach can be implemented in hardware using a previously suggested Xampling prototype.

I. INTRODUCTION

The paper develops Doppler focusing for sub-Nyquist pulse-Doppler radar, targeting sparse scenes while retaining low-rate acquisition and processing. It combines Xampling, FRI modeling, and coherent pulse processing to improve noisy recovery without restricting the transmitter.

  • Motivation: The method estimates sparse targets’ delays and Doppler frequencies using linear, non-adaptive sampling below the radar signal’s Nyquist rate.The setting is a monostatic narrowband radar with L sparse, non-fluctuating point targets in the unambiguous time-frequency region.
  • Contributions: Xampling acquires informative sub-Nyquist samples and supports complete digital recovery without returning to the Nyquist rate.The approach uses analog prefiltering before pointwise sampling, producing compressed samples for CS-based parameter recovery.
  • Model: The received signal has 3L degrees of freedom, comprising delay, Doppler frequency, and amplitude for each of L targets.This finite-rate-of-innovation structure motivates recovering target parameters directly from low-rate samples.
  • Doppler focusing: Doppler focusing coherently combines pulses whose Doppler frequencies match a selected value, reducing recovery to a one-dimensional delay problem.The coherent combination improves SNR and Doppler resolution while processing each selected Doppler frequency with a simple CS problem.
  • Results: For P pulses, Doppler focusing provides a factor P SNR improvement and a focus width of 2π/Pτ, matching optimal matched-filter SNR scaling.Targets separated in Doppler by more than 2π/Pτ create almost no interference.
  • Results: At one tenth the Nyquist rate, Doppler focusing outperforms classic and two-stage recovery, matching full-Nyquist classic performance at -25dB SNR.The reported comparison comes from simulations.

III. CLASSIC PULSE-DOPPLER PROCESSING

Classic pulse-Doppler processing samples at the waveform bandwidth, then applies matched filtering, pulse-dimension Doppler processing, and peak detection. Its computational and sampling demands grow with bandwidth, motivating low-rate processing that preserves SNR scaling.

  • Processing pipeline: Classic processing samples each incoming pulse at the Nyquist rate Bh, requiring N = τBh samples per pulse.The ADC stage creates discrete samples before matched filtering and Doppler processing.
  • Processing pipeline: Matched filtering produces time resolution 1/Bh, while a P-point Doppler transform produces frequency resolution 2π/Pτ.The resulting delay-Doppler map has P Doppler bins and N delay samples.
  • Processing pipeline: Classic processing forms a delay-Doppler map and detects targets using peaks, with prior knowledge potentially guiding selection of the strongest L points.Peak detection is described as heuristic.
  • Performance: Doppler processing increases SNR by P relative to a single pulse, which is optimal for P pulses under matched-filter processing.The pulse-dimension operation acts as a matched filter for a constant radial-velocity target.
  • Limitations: Sampling and computation scale with bandwidth because the system requires P convolutions of length N = τBh and N FFTs of length P.Increasing bandwidth therefore increases both ADC and digital-processing demands.
  • Motivation: The paper combines FRI and Xampling to break the link between radar bandwidth and sampling rate while retaining the same SNR scaling.The approach targets low-rate sampling and processing regardless of signal bandwidth.

IV. DOPPLER FOCUSING

Doppler focusing coherently superimposes echoes from multiple pulses after selecting a Doppler frequency, producing a delay-focused signal. In-focus targets add coherently while out-of-focus targets are attenuated, enabling sub-Nyquist delay recovery.

  • Doppler focusing: Doppler focusing combines target echoes from different pulses into one superimposed pulse, improving SNR and implicitly estimating Doppler frequency.It corresponds to interchanging classic delay and Doppler processing stages, which are both linear time-invariant operations.
  • Focus zone: For a selected ν, targets within a Doppler band of width 2π/Pτ achieve coherent integration and an SNR boost relative to one pulse.The focus zone is centered on the selected Doppler frequency.
  • Focus zone: Targets outside the focus zone approximately cancel, with g(ν|νℓ) approximately zero when |ν−νℓ| > π/Pτ.For P = 200 and νℓ = 0, the figure marks the focus zone as |ν| < π/Pτ.
  • Recovery benefit: The method reduces joint delay-Doppler estimation to delay-only estimation over a small Doppler range with increased target amplitude.This extra separation dimension can distinguish closely delayed targets with different Doppler frequencies.
  • Sub-Nyquist implementation: Doppler focusing can be performed in the frequency domain on low-rate Xamples, enabling sub-Nyquist implementation.The paper proceeds by sampling in time while extracting frequency-domain information.
  • Recovery benefit: In a delay-Doppler map, focusing at one ν makes only targets in that Doppler focus zone visible while out-of-focus targets disappear.The schematic marks the out-of-focus region in red and the currently focused target region separately.

V. SUB-NYQUIST DELAY RECOVERY

This section shows how Xampling acquires the Fourier coefficients needed for delay recovery directly at sub-Nyquist rates. A sparse delay model is then recovered using spectral-analysis or compressed-sensing methods operating on those coefficients.

  • Xampling: Xampling uses analog prefiltering before low-rate sampling to extract the information required for recovery without simply discarding Nyquist-rate samples.The method can be interpreted as compressed sampling performed during analog-to-digital conversion.
  • Delay recovery: 2L Fourier coefficients suffice to recover the unknown target amplitudes and delays in the noiseless delay-only problem.The lower bound applies when noise is negligible; annihilating filter, matrix pencil, ESPRIT, and MUSIC are listed as recovery options.
  • Xampling: Direct multichannel sampling obtains arbitrary Fourier coefficients by mixing the input with harmonics, integrating over the PRI, and sampling each channel.The number of extracted coefficients per pulse controls the trade-off between sampling rate and recovery performance.
  • CS delay recovery: The delay vector is L-sparse, with each nonzero index representing a target at delay n∆τ, and recovery finds its support through a CS dictionary.The dictionary is built from Fourier-coefficient measurements and a discretized delay grid.
  • CS delay recovery: Randomly selecting coefficient indices can provide favorable CS recovery conditions in noise, with RIP guaranteed with large probability when |κ| ≥ cL(logNτ)^4.The stated condition assumes a positive constant c and uniformly random coefficient indices.

VI. DELAY-DOPPLER RECOVERY

The delay-Doppler recovery procedure first Xamples each pulse, then coherently combines pulse coefficients through Doppler focusing. For each Doppler frequency, it solves a delay-only recovery problem whose dictionary does not grow with the number of pulses.

  • Doppler focusing: Targets satisfying |ν − νℓ| < π/Pτ contribute to the corresponding focused delay-estimation problem.The focus width is tied to the number of pulses and the PRI.
  • Doppler focusing: Doppler focusing combines Xampled coefficients across pulses for a chosen Doppler frequency, producing focused coefficients that support delay estimation.The operation can be performed directly on low-rate sub-Nyquist samples.
  • CS recovery: Separating Doppler estimation from delay estimation keeps the CS dictionary independent of the number of pulses P.In contrast, simultaneous delay-Doppler CS methods require dictionaries that grow with P.
  • Doppler focusing: Doppler focusing has no inherent blind speeds because it is continuous in Doppler frequency up to the PRF.Recovery searches focused outputs for large amplitudes and assigns their delay and Doppler coordinates.
  • CS recovery: The delay-recovery method can use CS in low-SNR settings or alternative spectral methods when noise and grid errors have different priorities.The focusing operation is independent of the underlying delay-estimation method.

C. SNR Analysis

The SNR analysis models focused coefficients as signal and noise contributions and shows that coherent pulse combining improves the target SNR linearly with the number of pulses.

  • SNR derivation: The focused SNR expression scales the unfocused SNR by P for each target.The derivation defines target and focused SNRs from the Fourier coefficients and their noise variance.
  • SNR mechanism: Noise remains incoherent across pulses because the pulse-domain noise samples are independent, while correctly focused target components add coherently.The focused noise is formed by summing P independent random variables.
  • SNR improvement: P times greater focused SNR is obtained after Doppler focusing compared with the unfocused target SNR.The improvement is linear in P and matches the improvement of an optimal matched filter.

D. Noiseless Recovery

The noiseless analysis establishes lower bounds on the required Fourier coefficients and pulses, then characterizes Doppler focusing’s sample complexity under continuous and grid-aligned Doppler assumptions.

  • Continuous Doppler: 4L^2 samples are required at minimum for perfect recovery of L targets without noise, with |κ| and P each at least 2L.The coefficient and pulse dimensions impose separate requirements rather than only a product constraint.
  • Grid-aligned Doppler: 2L min(M, 2L) samples suffice for perfect recovery when target Doppler frequencies lie on the specified grid.The result assumes no restriction on target delays.
  • Doppler focusing guarantee: Doppler focusing requires |κ| ≥ 2L and P ≥ M, with a total of 2LM samples for perfect recovery under the theorem’s conditions.The construction samples 2L Fourier coefficients for P = M pulses and uses annihilating-filter delay recovery.
  • Comparison with lower bounds: When M = O(L), Doppler focusing uses a number of pulses within order of magnitude of the lower bound.Its |κ| requirement matches the general lower bound per pulse.

E. Practical Considerations

The practical analysis addresses Doppler-grid implementation, target dynamic range, windowing trade-offs, and extensions to micro-Doppler targets. Windowing suppresses out-of-focus targets but broadens the focus zone, while the sparsity assumption can remain useful for many micro-Doppler scenes.

  • Grid implementation: A uniform Doppler grid can be processed efficiently using an M-point DFT or FFT of a length-P sequence.The resulting grid supports Doppler focusing through the described algorithm.
  • Dynamic range: Strong target amplitudes require aggressive windowing so weaker targets are not masked during focusing.Out-of-focus targets are those whose Doppler frequencies differ from the focus frequency by more than π/Pτ.
  • Windowing: Windowing reduces out-of-focus target effects but increases the frequency focus zone, potentially including more targets in each delay-estimation problem.For P = 100 pulses, Fig. 4 compares windowed and constant weighting and shows that windowing changes the focus zone.
  • Micro-Doppler: Micro-Doppler introduces Doppler spread, but reported dominant frequencies from vibration, rotation, or resonance can preserve a sparse representation in many cases.The paper treats micro-Doppler only briefly and within the assumption of a small number of dominant frequencies.

F. Clutter

The section examines clutter as a strong deterministic interference source and explains why coherent integration cannot improve signal-to-clutter ratio. Doppler focusing provides frequency-selective isolation that may reject clutter concentrated near zero Doppler.

  • Doppler-selective rejection: Doppler focusing appears to offer inherent clutter rejection, suggesting that dedicated prefilters such as MTI may not be required.The text presents this as a suggestion rather than a demonstrated universal guarantee.
  • Clutter characteristics: Clutter echoes can be several orders of magnitude stronger than target echoes and, unlike random noise, their signal-to-clutter ratio does not improve with increasing CPI.Clutter is a deterministic scaled, shifted, and modulated replica of the transmitted signal, so it benefits from coherent integration like the target.
  • Clutter characteristics: For a stationary radar, mostly static clutter is concentrated at zero Doppler, motivating conventional notch-based anti-clutter methods.The section uses this concentration to connect clutter rejection with Doppler-selective processing.
  • Doppler-selective rejection: Doppler focusing acts like a filter bank with pass-band width 2π/Pτ, and windowing controls attenuation at the cost of wider pass-bands.This creates adjustable isolation between delay-estimation problems for targets separated in Doppler by more than the pass-band width.

VII. COMPARISON TO PREVIOUS APPROACHES

The comparison contrasts simultaneous delay-Doppler CS, which performs well in noise but scales poorly in dictionary size and computation, with Doppler focusing’s fixed-in-P dictionary and comparable SNR scaling.

  • Simultaneous recovery: Simultaneous delay-Doppler CS uses a two-dimensional grid and a dictionary whose size grows rapidly with problem dimensions.Its matching operation resembles a discrete matched filter over the entire pulse train.
  • Simultaneous recovery: 10^9 dictionary elements can require many gigabytes of memory and high-end processing capabilities.The stated example concerns even moderate-sized problems with roughly 10^3 delay and Doppler grid points.
  • Doppler focusing: Doppler focusing retains SNR improvement that scales linearly with P while its dictionary depends only on delay parameters and remains fixed as pulses increase.The authors could not compare directly because their 12GB-RAM computer could not store the competing dictionary.
  • Complexity: Doppler focusing requires |κ|P samples and solves M CS delay-recovery problems, whereas the compared method samples at the Nyquist rate and scales poorly with bandwidth.The bandwidth-dependent computational scaling is identified as critical for high-resolution radar.

B. Two-Stage Recovery

The section compares sequential recovery and dictionary-design choices for noisy delay estimation. Doppler focusing preserves coherent phase information across pulses, while coherence can be selected according to whether small support errors are acceptable.

  • Two-stage recovery: MMV-based two-stage recovery jointly processes pulse measurements using shared delay support, but its row-norm operation destroys phase information and mixes signal with noise.The paper characterizes this operation as a form of non-coherent integration.
  • Two-stage recovery: Doppler focusing compensates exact pulse-to-pulse phase differences and therefore provides an SNR increase linear with P, exceeding the P^β scaling reported for two-stage methods.The cited comparison states 0.5 < β < 0.833, with β decreasing toward 0.5 as P increases.
  • Performance metric: Reducing the delay-grid step lowers quantization error but makes CS dictionary columns more similar and increases coherence.This creates a tension between off-grid accuracy and the usual preference for low coherence.
  • Performance metric: When delay errors up to τmax are acceptable, correlating each dictionary column with its K nearest neighbors can improve noisy hit performance.Here K = ⌊τmax/∆τ⌋, and a neighboring column may still count as a successful hit when the exact column is missed.
  • Performance metric: Consecutive Fourier coefficients produce higher coherence and suit tolerant support-recovery criteria, whereas random coefficients better suit exact recovery.Fig. 5 reports coherence values of 0.9 for the consecutive set and 0.3 for the random set.

B. Numerical Results

Numerical experiments compare Doppler focusing with classic and other sub-Nyquist recovery methods across hit rate, estimation error, SNR, and target-resolution scenarios. Doppler focusing generally performs best under sub-Nyquist sampling, while waveform bandwidth affects close-target resolution.

  • Experimental setup: 200 Fourier coefficients per pulse versus 2000 Nyquist-rate samples demonstrated a 1:10 sampling-rate reduction.The experiments used L=5 targets and P=100 pulses.
  • Recovery performance: Doppler focusing was superior to other sub-Nyquist recovery techniques in hit rate and RMS error across tested SNR values.Consecutive coefficients performed better at lower SNR, while random coefficients improved performance as SNR increased.
  • Recovery performance: Doppler focusing degraded gracefully with sample-rate reduction, whereas classic processing suffered significantly below the Nyquist rate.The comparison included classic processing, two-stage CS recovery, and Doppler focusing.
  • Waveform design: At one tenth the Nyquist rate, concentrating waveform energy in sampled frequencies enabled Doppler focusing to surpass classic processing using ten times as many samples at lower SNR.The modified waveform preserved target SNR while transferring energy into sampled frequencies.
  • Target resolution: Wideband classic processing distinguished very closely spaced equal-Doppler targets better than narrowband Doppler focusing.The comparison concerns resolution rather than the general hit-rate advantage of Doppler focusing.
  • Low-SNR scene: At -28dB SNR, Doppler focusing alone distinguished two targets near 4.2 µsec with nearly identical delays but different Doppler frequencies.The figure reports the highest hit rate among the sub-Nyquist methods in this scene.
  • Dynamic range: For two adjacent targets with a 20dB power difference, Doppler focusing at one tenth the Nyquist rate recovered both while matched-filter processing recovered only one.The matched-filter comparison was performed at both Nyquist and one tenth the Nyquist rate.

C. Clutter

The clutter experiments examine Doppler focusing with many near-zero-Doppler scatterers and strong target-to-clutter imbalance, alongside hardware implementation evidence. Windowing and Doppler-bin exclusion are used to mitigate localized clutter interference.

  • Clutter scenario: 4000 Swerling-0 clutter scatterers were simulated near DC with SCR=-50dB, while targets had SNR=-25dB.The clutter occupied a single Nyquist bin around DC and was distributed across all delays.
  • Dynamic range: Doppler focusing recovered both closely spaced targets with a 20dB power difference, whereas matched-filter processing recovered only one.This dynamic-range result was obtained without windowing at both Nyquist and one tenth the Nyquist rate.
  • Clutter mitigation: Ignoring the cluttered DC bin and its two neighboring bins, plus applying a Taylor window with -50dB attenuation, improved Doppler-frequency isolation.The procedure addressed clutter sidelobes that still covered targets at very low SCR.
  • Hardware experiment: A custom sub-Nyquist radar receiver board implemented Xampling and Doppler-focusing digital recovery for a real hardware experiment.The prototype used synthesized analog input signals and a receiver board based on previously described RF hardware.
  • Clutter mitigation: With 40dB windowing, five of nine targets were hits; with 50dB windowing, the entire nine-target scene was detected correctly.Without windowing, clutter sidelobes caused misdetections across the nonzero-Doppler region.
  • Hardware experiment: The prototype demonstrated that the sampling and recovery methodology is feasible with standard RF hardware while operating at a low sampling rate.The claim concerns practical feasibility of both analog sampling and digital recovery.
  • Scope boundary: The method’s remaining scope includes unknown target counts and improved dynamic range for strong targets and their sidelobes.These issues are identified as ongoing work rather than resolved capabilities.
Loading 1211.0722v3…