PaperlayerView exact revision on arXiv

Source-linked AI summary

Generalized Labeled Multi-Bernoulli Approximation of Multi-Object Densities

Francesco Papi, Ba-Ngu Vo, Ba-Tuong Vo, Claudio Fantacci, Michael Beard

arXiv:1412.5294v3stat.OT

TL;DR

Exact multi-object densities are generally intractable, and tractable implementations often impose independence assumptions that cannot represent object dependence or generic measurement models. The paper derives a tractable GLMB approximation matching cardinality and first-moment information while minimizing KLD over a special GLMB class, then applies it to tracking. The resulting approach is demonstrated in low-SNR radar TBD scenarios with closely spaced targets.

  • Problem

    Exact multi-object density computation is generally intractable, while common tractable methods often assume statistical independence and GLMB conjugacy can fail for generic measurement models.

  • Method

    The paper derives a tractable δ-GLMB approximation for arbitrary labeled multi-object densities using cardinality and PHD matching and KLD minimization, then applies it to generic-measurement tracking.

  • Results

    The approximation preserves cardinality and PHD, minimizes KLD over the specified δ-GLMB class, and yields effective tracking performance in challenging low-SNR radar TBD scenarios.

  • Takeaways & Limitations

    The result supports tractable recursive multi-object filters with formal track estimates for a wide range of non-standard measurement models.

Abstract

from arXiv · show

In multi-object inference, the multi-object probability density captures the uncertainty in the number and the states of the objects as well as the statistical dependence between the objects. Exact computation of the multi-object density is generally intractable and tractable implementations usually require statistical independence assumptions between objects. In this paper we propose a tractable multi-object density approximation that can capture statistical dependence between objects. In particular, we derive a tractable Generalized Labeled Multi-Bernoulli (GLMB) density that matches the cardinality distribution and the first moment of the labeled multi-object distribution of interest. It is also shown that the proposed approximation minimizes the Kullback-Leibler divergence over a special tractable class of GLMB densities. Based on the proposed GLMB approximation we further demonstrate a tractable multi-object tracking algorithm for generic measurement models. Simulation results for a multi-object Track-Before-Detect example using radar measurements in low signal-to-noise ratio (SNR) scenarios verify the applicability of the proposed approach.

I. INTRODUCTION

Multi-object inference estimates an unknown number of objects and their states from noisy observations, but exact multi-object densities are generally intractable. The paper motivates tractable GLMB approximations for dependent objects and generic measurement models.

  • Motivation: Multi-object densities represent uncertainty in object number, object states, and statistical dependence between objects.Such dependence can arise from posterior data association or object interactions.
  • Motivation: Exact computation is generally intractable, while common tractable methods often assume statistical independence between objects.PHD, CPHD, and multi-Bernoulli filters are cited as examples based on independent-object densities.
  • Motivation: GLMB densities can capture statistical dependence and are conjugate under the standard measurement likelihood.MHT can model dependence without a multi-object density, whereas JPDA has one only for a known object count.
  • Motivation: For generic measurement models, GLMB conjugacy can fail, making the resulting multi-object density numerically intractable.The paper identifies Track-Before-Detect, superpositional, merged, and video measurements as examples.
  • Contribution: The paper proposes a GLMB approximation that preserves cardinality and first-moment information, minimizes KLD over a tractable class, and supports generic-measurement tracking.A radar TBD example targets low-SNR, closely spaced objects.

B. Generalized Labeled Multi-Bernoulli

GLMBs form an important labeled-RFS family for analytic Bayes multi-object filtering. Their special LMB case represents independent tracks, while GLMB terms combine label-dependent weights with label- and state-dependent densities.

  • GLMB family: GLMBs provide the basis for an analytic Bayes multi-object filter and are conjugate under the standard multi-object likelihood.They are also closed under Chapman-Kolmogorov prediction.
  • GLMB structure: A GLMB is a mixture whose weights depend on object labels and whose multi-object exponentials depend on labels and kinematic or feature states.This separates label-set weighting from state-density structure.
  • GLMB statistics: The GLMB cardinality distribution and PHD are defined as key statistics of the density.These statistics are later used for matching in the proposed approximation.
  • LMB special case: An LMB is a one-term GLMB with track existence probabilities and conditional kinematic-state densities for each label.Its density factors over objects and can therefore be interpreted as multiple independent tracks.

III. MULTI-OBJECT ESTIMATION WITH GLMBS

The paper develops two GLMB-based approaches for multi-object estimation: a separable-likelihood approximation exploiting conjugacy and a more principled approximation matching PHD and cardinality.

  • Separable likelihood: The section first uses a separable multi-object likelihood to obtain a tractable approximation that exploits GLMB conjugacy.This approach assumes the likelihood factors through a non-negative function on the single-object state space.
  • Separable likelihood: Poisson, IID cluster, multi-Bernoulli, and GLMB densities are conjugate with respect to separable multi-object likelihoods.Under this assumption, a GLMB prior yields a GLMB posterior.
  • Assumption: The true multi-object likelihood is generally non-separable, although separability can be reasonable when objects do not overlap in measurement space.Thus the approximation is most directly supported in scenarios with limited measurement-space overlap.

B. Labeled RFS Density Approximation

The paper approximates arbitrary labeled multi-object densities with tractable δ-GLMBs by matching cardinality and first-moment information. The approximation minimizes Kullback-Leibler divergence over a specified δ-GLMB class and extends to broader GLMB forms through marginal factorization.

  • The proposed approximation uses a tractable δ-GLMB class to approximate arbitrary labeled multi-object densities.The class is numerically evaluated through explicit enumeration of label sets.
  • It seeks a density whose parameter set matches the target density’s PHD and cardinality distribution.The target representation uses joint label-existence probabilities and label-conditioned state densities.
  • The approach is intended for general multi-object estimation and provides a basis for tracking with non-standard measurement models.The paper notes that the result is not restricted to tracking applications and relates it to an efficient multi-object tracking filter.
  • Proposition 2 replaces label-conditioned joint densities with products of their marginals while preserving the target PHD and cardinality distribution.The resulting δ-GLMB is defined for any labeled multi-object density in the specified class.
  • The resulting δ-GLMB minimizes Kullback-Leibler divergence from the target density over the class defined by the approximation.The proof uses equality of the label-set weights and marginal-density KL-minimization.
  • The matching strategy extends to more general GLMB densities by approximating each component’s joint density with a product of its marginals.This broader approximation preserves cardinality and PHD, but KL-divergence results are difficult to establish for the more general class.

IV. APPLICATION TO MULTI-TARGET TRACKING

The proposed GLMB approximation is applied to construct a multi-target tracking filter for generic measurement models.

  • The paper proposes a multi-target tracking filter for generic measurement models by applying the GLMB approximation result of Proposition 2.

A. Multi-target Filtering

The filtering formulation represents labeled multi-target states as finite sets and propagates their density through a multi-target Bayes filter.

  • Multi-target state and labels: Each target receives a distinct ordered-pair label encoding birth time and a unique within-birth-time index.
  • Multi-target state and labels: The label space at time k combines labels of previously existing targets with those assigned to targets born at k.
  • Multi-target state and labels: A multi-target state is a finite subset of the labeled state space, with transition and likelihood functions defining the system model.
  • Multi-target state and labels: Including distinct labels in target states accommodates target trajectories or tracks in the random finite set formulation.
  • Filtering density: The filtering density is a marginal of the multi-target posterior that can be propagated recursively by the multi-target Bayes filter.
  • Filtering density: An analytic solution supports labeled-state filtering and track estimation, although the paper does not distinguish filtering density from multi-target posterior terminology.

B. Update

The update applies the proposed δ-GLMB approximation to tracking with generic measurement models, while recognizing that the exact posterior need not remain a GLMB.

  • Update: The approach assumes no particular structure for the multitarget likelihood, covering point detections, superpositional sensors, and imprecise measurements.
  • Update: After updating, each multi-object exponential is not necessarily a multi-object exponential, so the resulting density is generally not a GLMB.
  • Update: When targets are well separated in measurement space, the likelihood can be approximated by a separable likelihood.

1) Separable Likelihood:

The separable-likelihood case uses an approximate GLMB posterior, while the general case directly approximates the posterior when targets are closely spaced.

  • 2) General Case:: For closely spaced targets, the separable likelihood assumption is violated and the multi-target posterior must be approximated directly.
  • 2) General Case:: The direct approximation produces a δ-GLMB matching the posterior’s cardinality and PHD while minimizing Kullback-Leibler divergence.
  • 2) General Case:: For each label set, the approximating single-target densities are formed from marginals of the corresponding labeled multi-target density.
  • 2) General Case:: The approximation retains the label-set weights from the true posterior.
  • Prediction model: The standard dynamic model allows targets to survive and evolve or die, while new targets are superposed through a birth density.
  • Prediction model: The birth model includes labeled Poisson, labeled IID cluster, and LMB models; the implementation uses an LMB birth model.
  • Prediction model: A δ-GLMB filtering density remains a δ-GLMB under the stated multi-target prediction step.
  • Prediction model: The prediction equations explicitly calculate the new density parameters from those of the previous multi-target density.

V. NUMERICAL RESULTS

This section describes a particle-filter implementation for recursive multi-target tracking with radar power measurements, including separable and non-separable likelihood treatments.

  • Tracking implementation: The tracker uses a particle-filter approximation of the GLMB density for recursive multi-target tracking with radar power measurements.The implementation follows the GLMB density described in.
  • Target model: Each target state contains planar position, 2D velocity, and the modulus of its complex amplitude.Target dynamics use a Nearly Constant Velocity model, with amplitude fluctuations represented by a zero-mean Gaussian random walk.
  • Measurement model: Radar measurements are vectors of cell power returns generated from target templates and complex target echoes.The radar is positioned at the Cartesian origin, and target templates can overlap across measurement cells.
  • Measurement model: Each cell measurement follows a non-central chi-squared distribution with 2 degrees of freedom, while the zero-signal case is central chi-squared.The cell likelihood ratio uses a modified Bessel function approximation.
  • Likelihood treatment: The exact power-return likelihood captures superposition from overlapping target templates; the separable approximation assumes at most one target contributes per cell.The separable form is obtained by simplifying the full multi-target likelihood under the non-overlap assumption.

C. Separable Likelihood Results

The separable-likelihood experiment evaluates radar TBD tracking when targets do not overlap in measurement space. At 7dB SNR, tracking performance is reported as satisfactory despite slight target-count discrepancies from closely spaced targets.

  • Scenario assumptions: The separable likelihood is valid when targets do not overlap at any time in measurement space.The experiment assumes an informative birth density to avoid target-count bias from birth hypotheses that violate separability.
  • Scenario: The scenario has a time-varying target count caused by births and deaths, reaching a maximum of 5 targets mid-scenario.Figure 1 depicts targets appearing from the top right and moving toward the radar at the Cartesian origin.
  • Results: At 7dB SNR, the average estimated target count slightly differs from the truth because of closely spaced targets, while overall performance remains satisfactory.Figure 2 shows estimated trajectories along the x and y coordinates.

D. Non-Separable Likelihood Results

The non-separable experiment addresses closely spaced radar targets for which the separable likelihood can bias target-count estimates. The paper concludes that the GLMB approximation supports tractable tracking with statistical dependence in challenging low-SNR TBD scenarios.

  • Scenario: The non-separable scenario contains births and deaths with a time-varying target count reaching a maximum of 7 targets mid-scenario.The experiment uses non-separable likelihood parameters reported in Table III.
  • Measurements: The radar returns are visualized through range-azimuth, range-Doppler, and azimuth-Doppler maps for the non-separable scenario.The maps show the same target group across their third-coordinate indexing.
  • Measurements: A reflection near (1500m, 0.8°, 18m/s) is produced by two targets occupying the same radar cell.This illustrates the overlapping-target condition that violates separable-likelihood assumptions.
  • Conclusion: The proposed GLMB approximation captures target dependence, preserves cardinality and the PHD, minimizes KLD, and yields tractable recursive filters for non-standard measurement models.A low-SNR radar TBD example with time-varying, closely spaced targets was used to verify the theoretical result.
Loading 1412.5294v3…