PaperlayerView exact revision on arXiv

Source-linked AI summary

Target Localization Accuracy Gain in MIMO Radar Based Systems

Hana Godrich, Alexander M. Haimovich, Rick S. Blum

arXiv:0809.4058v1cs.IT

TL;DR

The paper studies how distributed MIMO radar configurations affect target-localization accuracy. It derives CRLBs for coherent and non-coherent processing, optimizes sensor placement, and develops BLUE and GDOP tools; coherent processing can provide a precision gain of order fc/β, while sensor placement yields an MN/2 variance reduction under optimal deployment.

  • Problem

    The paper addresses the need for analytical target-localization accuracy relations in distributed MIMO radar.

  • Method

    It derives closed-form CRLB expressions for coherent and non-coherent processing and develops a closed-form BLUE with GDOP-based deployment analysis.

  • Results

    Coherent processing provides a localization root-mean-square-error precision gain of order fc/β, while optimally placed M transmitters and N receivers reduce CRLB variance by MN/2.

  • Takeaways & Limitations

    Symmetric deployment around the target is optimal for minimizing the CRLB, and GDOP contours relate deployment to achievable accuracy across target locations.

  • Takeaways & Limitations

    The coherent analysis does not address phase-error sensitivity.

Abstract

from arXiv · show

This paper presents an analysis of target localization accuracy, attainable by the use of MIMO (Multiple-Input Multiple-Output) radar systems, configured with multiple transmit and receive sensors, widely distributed over a given area. The Cramer-Rao lower bound (CRLB) for target localization accuracy is developed for both coherent and non-coherent processing. Coherent processing requires a common phase reference for all transmit and receive sensors. The CRLB is shown to be inversely proportional to the signal effective bandwidth in the non-coherent case, but is approximately inversely proportional to the carrier frequency in the coherent case. We further prove that optimization over the sensors' positions lowers the CRLB by a factor equal to the product of the number of transmitting and receiving sensors. The best linear unbiased estimator (BLUE) is derived for the MIMO target localization problem. The BLUE's utility is in providing a closed form localization estimate that facilitates the analysis of the relations between sensors locations, target location, and localization accuracy. Geometric dilution of precision (GDOP) contours are used to map the relative performance accuracy for a given layout of radars over a given geographic area.

I. INTRODUCTION

The introduction situates distributed MIMO radar localization within broader radar and geolocation research. It motivates closed-form accuracy analysis by noting that prior work observed sensor-number and geometry effects without specifying exact relations.

  • MIMO radar localization: MIMO radar uses multiple antennas and waveforms, with distributed systems exploiting spatial diversity and multiple propagation paths for target localization.Localization can use direct multilateration from received echoes or indirect multilateration through estimated time-of-arrivals.
  • Processing architectures: Non-coherent processing requires time synchronization, whereas coherent processing requires both time and phase synchronization among radars.The paper’s terminology concerns localization processing; envelope-only processing can remain non-coherent despite in-phase and quadrature signal components.
  • Research gap: Prior analyses linked localization accuracy to carrier frequency, signal bandwidth, and relative sensor-target locations, but exact sensor-number relations remained unspecified.Earlier work also observed that increasing the number of sensors improves localization performance without developing an exact relation.
  • Research gap: The paper addresses these deficiencies by obtaining closed-form CRLB expressions for both coherent and non-coherent MIMO radar localization.The introduction also connects the problem to active range-based localization, whose accuracy is inversely proportional to effective signal bandwidth.
  • Geometric evaluation: Geolocation studies use sensor-location analysis and GDOP plots, motivating application of GDOP to evaluate MIMO radar localization performance.For optimally positioned GPS systems, the best achievable accuracy is inversely proportional to the square root of the number of participating GPS.

B. Main Contributions

The paper develops analytical localization bounds and evaluation tools for distributed MIMO radar. Its contributions cover coherent and non-coherent CRLBs, sensor-placement optimization, BLUE estimation, and GDOP-based deployment analysis.

  • CRLB development: The paper develops CRLB expressions for general MIMO radar with multiple transmitted waveforms, including orthogonal-waveform cases with non-coherent and coherent observations.The non-coherent case serves as a benchmark for evaluating coherent observations.
  • CRLB development: Both CRLB expressions factor into a bandwidth-and-SNR term and a term accounting for sensor-location effects.The factorization separates signal characteristics from the geometric contribution of the radar deployment.
  • Processing comparison: Non-coherent localization standard-deviation bounds are inversely proportional to the signals’ averaged effective bandwidth.The paper reports that substantially higher accuracy can be obtained from coherent observations.
  • Processing comparison: Coherent localization exploits phase information and yields a CRLB inversely proportional to carrier frequency.This improvement is called coherency gain.
  • Sensor-placement optimization: Optimally placed M transmitters and N receivers reduce the CRLB variance by a factor MN/2, called the MIMO radar gain.Symmetric deployment around the target is optimal for minimizing the CRLB.
  • Estimation and deployment analysis: A closed-form BLUE solution supports evaluation of estimator MSE and relates sensor locations, target location, and localization accuracy through GDOP.GDOP contour maps show achievable accuracy at various target locations for a given sensor deployment.

III. LOCALIZATION CRLB

The CRLB analysis begins from the Fisher information for the received MIMO radar observations and uses a parameter transformation through propagation delays. The resulting geometry enters through derivatives of sensor-target delays.

  • CRLB framework: The CRLB lower-bounds the mean-squared error of unbiased estimators for unknown parameters.The Fisher information matrix is formed from the conditional joint probability density of the observations.
  • CRLB framework: A chain-rule formulation computes the FIM using time delays and complex amplitudes before transforming back to the target parameters.The delay vector is relevant because received signals depend on propagation delays and complex amplitudes.
  • Non-coherent processing: Non-coherent processing has no common phase reference, so sensor-specific complex amplitudes remain nuisance parameters.The localization parameters are the target coordinates, while amplitude components act as nuisance parameters.
  • Geometric dependence: The geometry matrix H contains derivatives of propagation delays with respect to x and y, expressed through sine and cosine functions of transmitter and receiver bearing angles.These angle terms connect the CRLB to sensor and target locations.
  • Closed-form specialization: For orthogonal waveforms, the FIM and CRLB expressions simplify, allowing the localization bound to be computed in closed form.The CRLB remains linked to sensor-target geometry through H and to waveform correlations through the auxiliary matrices.

1) Orthogonal Waveforms:

For orthogonal waveforms, the paper derives closed-form localization CRLB submatrices for non-coherent and coherent processing. The bounds separate waveform-dependent factors from geometry, with coherent processing gaining from phase information.

  • Non-coherent case: The non-coherent CRLB provides closed-form lower bounds for the x and y localization variances after treating complex amplitudes as nuisance parameters.The coordinate bounds are obtained from the upper-left 2×2 CRLB submatrix.
  • Non-coherent case: Non-coherent variance bounds are inversely proportional to the averaged effective bandwidth and depend on sensor geometry through angular sine and cosine terms.The waveform-dependent factor is related to single-antenna range-estimation CRLB, while remaining terms capture sensor effects.
  • Coherent case: Coherent processing exploits phase terms that depend on propagation delays and target coordinates, reducing the number of unknown phase-related parameters.The phase information is measured relative to a common phase reference.
  • Coherent case: Coherent target-location variance bounds are inversely proportional to the square of the carrier frequency.The coherent expressions include additional geometry-related elements compared with the non-coherent case.

C. Discussion

The discussion contrasts coherent and non-coherent localization bounds, relates accuracy to bandwidth, carrier frequency, and radar geometry, and outlines important CRLB limitations.

  • The non-coherent localization variance bound is inversely proportional to the averaged effective signal bandwidth.
  • The coherent CRLB is inversely proportional to carrier frequency and independent of individual signal effective bandwidth because phase information is used across paths.
  • Coherent processing provides a localization precision gain of order fc/β; choosing fc/β between 100 and 1000 yields substantial coherency gain.
  • CRLB terms depend strongly on the relative geographical spread of radar systems and the target.
  • Localization variances trade off across horizontal and vertical axes, so minimizing one direction can increase error in the other.
  • The analysis is moderated because CRLB bounds small errors and omits effects that can produce large errors; coherent systems also face sidelobes and phase-error sensitivity.
  • The CRLB is tight at high SNR but not at low SNR, where waveform ambiguities can cause erroneous time estimates.

IV. EFFECT OF SENSORS LOCATIONS

This section optimizes distributed transmitter and receiver locations to minimize the coherent localization CRLB, reformulating the nonconvex problem through convex optimization and deriving symmetric optimal layouts.

  • The coherent CRLB localization gain depends on transmitter and receiver positions relative to the target.
  • The placement objective minimizes the trace of the 2×2 coherent CRLB submatrix for joint x- and y-axis localization accuracy.
  • The original nonconvex placement problem is transformed into an equivalent convex optimization problem solvable by routine techniques.
  • KKT conditions and Lagrange duality are used to characterize the optimal solution of the convex formulation.
  • M ≥3 transmitting and N ≥3 receiving sensors symmetrically spaced on circles around the target minimize the localization CRLB.
  • The optimal angular spacings are 2π/M for transmitters and 2π/N for receivers, with superpositions of symmetric sets also admissible.
  • A SIMO configuration with one transmitter does not achieve optimality and increases estimation error by a factor of 2 relative to the M-transmitter, N-receiver case.

A. Discussion

The discussion identifies sensor geometries and deployment choices that minimize localization CRLB, while noting practical limits on controlling sensor positions. It also introduces GDOP contour mapping for evaluating radar layouts.

  • MN ≫ (M + N) makes the multiple-transmitter, multiple-receiver advantage over one transmitter and MN receivers apparent.
  • 4ηc/(MN) is the CRLB for one transmit antenna with MN receive antennas, twice the localization-CRLB performance of MIMO with M + N sensors.
  • The best accuracy occurs when transmitting and receiving radars lie on a target-centered virtual circle with uniform angular spacings 2π/M and 2π/N, respectively.The discussion also allows any superposition of these arrangements.
  • Because real-time control of sensor positions relative to a target is impractical, selecting sensors most symmetrically placed around the target may improve localization accuracy.
  • GDOP metrics and contour mapping tools are introduced to evaluate radar-layout performance over a geographic area.

V. METHODS FOR TARGET LOCALIZATION

This section presents maximum-likelihood and best linear unbiased estimators for MIMO target localization. The MLE is asymptotically motivated but generally requires numerical optimization rather than a closed-form solution.

  • The section presents MLE and BLUE estimators, motivating MLE by asymptotic optimality and BLUE by its closed-form expression.
  • Under mild conditions, the MLE is asymptotically unbiased and asymptotically attains the relevant lower bound.
  • MLE: For coherent MIMO radar, the MLE jointly estimates θ = [x, y, ζ]^T from the observation vector using the likelihood p(r|θ).
  • MLE: Since the MLE has no closed-form solution, grid search or iterative likelihood maximization is required to determine the target coordinates.The resulting search can involve significant computational effort, though it may be restricted around a coarse non-coherent estimate.

B. BLUE Target Localization

The BLUE approach linearizes time-delay observations around a nominal target estimate and derives a closed-form localization estimator. Its variance supports analysis of sensor deployment, while GDOP mapping addresses visualization and design.

  • The BLUE estimator and its variances are available in closed form, enabling analysis without extensive numerical computation.
  • The BLUE model treats estimated time delays as observations that are linearized in target coordinates around a nominal location.The observed delay is modeled as μℓk = τℓk + εℓk, with delay-estimation error represented by εℓk.
  • The BLUE combines the time-error covariance matrix and angle-dependent transformation matrix D to estimate target location coordinates.
  • Unlike the MLE, which uses signal observations nonlinear in x and y, the BLUE uses time-delay observations and therefore requires an intermediate delay-estimation step.
  • The BLUE variances have functional dependencies on carrier frequency and sensor deployment similar to the CRLB, with matrix terms relating layout to variance.
  • The variance expressions do not readily visualize sensor-layout effects, motivating a mapping method proposed as a MIMO radar design and decision-making tool.

D. GDOP

GDOP summarizes how sensor and target locations affect localization accuracy and maps this performance across a geographic area. Symmetric sensor deployments around the target produce the lowest GDOP, while accuracy degrades outside the system footprint or under restricted viewing angles.

  • GDOP reduces the combined effect of sensor locations to a single metric and converts mapped values into localization error using the time-delay error term.The actual localization error is obtained by multiplying GDOP by cσε.
  • Targets inside the virtual (N + M)-sided footprint have higher localization accuracy than targets outside it, with the best localization at the system center.GDOP increases slowly toward the footprint boundaries but rises rapidly outside the footprint.
  • Symmetrical deployment around the target yields the lowest GDOP values, whereas restricted viewing angles markedly degrade GDOP.Widely spread but nonsymmetric radars retain some accurate regions, though their coverage is smaller than under symmetric deployment.
  • Coherent processing reduces CRLB standard deviation by a factor of fc/β relative to non-coherent observations, while optimized MIMO geometry yields gain proportional to the transmitter–receiver product.
  • The smallest CRLB is achieved when transmitting and receiving sensors are arrayed symmetrically around the target or as superpositions of such sets.

APPENDIX I DERIVATION OF THE FIM IN (22)

The appendix derives the Fisher information matrix for the unknown non-coherent and coherent parameter vectors from their conditional probability densities. It uses waveform orthogonality, differentiation with respect to time delays, and matrix definitions to obtain the required FIM forms.

  • The non-coherent FIM is developed for the unknown parameter vector ψnc using the conditional probability density and waveform orthogonality.Orthogonality removes cross elements for distinct waveform and sensor indices.
  • Matrix notation, diagonal-matrix definitions, and indexed sensor-waveform terms organize the second-derivative expressions used in both derivations.
  • The coherent FIM is developed analogously for ψc, with time-delay derivatives and matrices defined for the coherent model.

APPENDIX III COMPUTATION OF (36)

The appendix computes the coherent CRLB submatrix and derives the statistics of time-delay estimation errors. The calculation uses matrix determinant manipulations and a narrowband Gaussian-noise model.

  • The coherent localization CRLB submatrix is obtained by evaluating matrix terms, determinants, and the relevant submatrices of J(θc).
  • The time-delay estimates are represented as a vector of received-waveform estimates, and their noise terms are characterized through their mean and covariance.The receiver noise is modeled as a zero-mean Gaussian random process.
  • The appendix derives the required expressions by differentiating the likelihood-related terms and invoking the narrowband assumption.
Loading 0809.4058v1…