PaperlayerView exact revision on arXiv

Source-linked AI summary

Sensor management for multi-target tracking via Multi-Bernoulli filtering

Hung Gia Hoang, Ba Tuong Vo

arXiv:1312.6215v2eess.SYcs.SY

TL;DR

Multi-target sensor management is difficult because object counts vary and observations include missed detections and false alarms, while standard optimal-control techniques do not directly apply. The paper formulates the problem as a POMDP and uses the CB-MeMBer filter with R´enyi-divergence and cardinality-variance objectives. In the reported range-only scenario, the R´enyi-divergence strategy performs better, though neither strategy achieves sufficiently low OSPA error.

  • Problem

    Multi-target sensor control must handle stochastic finite-set states and observations, including varying object counts, missed detections, and false alarms, for which standard optimal-control techniques are not directly applicable.

    Source: §1 · p. 1

  • Method

    The paper formulates sensor control as a POMDP and uses the CB-MeMBer filter to propagate an approximate multi-object posterior for R´enyi-divergence or cardinality-variance control.

    Source: §1 · pp. 1–2, §3.1 · p. 4

  • Results

    In the range-only scenario, the R´enyi-divergence strategy performs better, though neither strategy achieves sufficiently low OSPA error; elsewhere, CB-MeMBer strategies outperform PHD counterparts by approximately 10m in steady-state miss distance.

    Source: §5 · p. 8

  • Takeaways & Limitations

    The particle CB-MeMBer filter supports efficient calculation of the R´enyi-divergence reward and enables cardinality variance as a control objective for direct control of cardinality estimation error.

    Source: §6 · p. 9

    Target observability affects estimation performance across control policies, and further work is needed to investigate multi-target observability and suitable policies for different scenarios.

    Source: §5 · p. 9

Abstract

from arXiv · show

In multi-object stochastic systems, the issue of sensor management is a theoretically and computationally challenging problem. In this paper, we present a novel random finite set (RFS) approach to the multi-target sensor management problem within the partially observed Markov decision process (POMDP) framework. The multi-target state is modelled as a multi-Bernoulli RFS, and the multi-Bernoulli filter is used in conjunction with two different control objectives: maximizing the expected R\'enyi divergence between the predicted and updated densities, and minimizing the expected posterior cardinality variance. Numerical studies are presented in two scenarios where a mobile sensor tracks five moving targets with different levels of observability.

1 Introduction

Multi-target sensor management must handle finite-set-valued states and observations, random object counts, missed detections, and false alarms, making standard optimal-control techniques not directly applicable. The paper formulates sensor control as a POMDP and uses the CB-MeMBer filter for a computationally efficient algorithm with two control objectives.

1 Introduction

  • Multi-target states and observations are finite-set-valued, with random object counts, missed detections, and false alarms, so standard optimal-control techniques are not directly applicable.The paper casts sensor scheduling as a POMDP with finite-set-valued states and observations.Source: §1 · p. 1
  • The CB-MeMBer filter propagates a parametrized approximation to the multi-object posterior, which is cheaper than propagating the full posterior and eliminates clustering for state estimation.The approach is applicable to general nonlinear, non-Gaussian models, and its particle implementation further reduces computational load.Source: §1 · p. 2
  • The paper develops both expected R´enyi divergence and posterior cardinality variance as control objectives, with cardinality variance enabling direct control of cardinality estimation error.The R´enyi divergence objective was previously used for multi-target sensor control in.Source: §1 · p. 2
  • The paper introduces a computationally efficient multi-target sensor-control algorithm using the CB-MeMBer filter and assesses two control objectives.The paper also extends its preliminary conference work with full algorithm details, a cheaper alternative objective, and more complete numerical studies.Source: §1 · p. 2

2 Cardinality Balanced MeMBer filter

The section describes multi-object states and observations as finite sets and outlines Bayesian filtering before presenting the CB-MeMBer filter as a tractable approximation to the full posterior. Its recursion represents the posterior with multi-Bernoulli parameters.

2.1 General system model

  • Multi-object states and observations are random finite sets, and the system model captures object birth, death, transition, detection, and false alarms.Bayesian filtering predicts the multi-object density and updates it when new observations arrive, but the set-integral recursion has no general analytic solution and full-posterior SMC is computationally prohibitive for more than a small number of targets.Source: §2.1 · p. 2

2 Cardinality Balanced MeMBer filter

  • A Bernoulli RFS has an existence probability and a conditional state density; a multi-Bernoulli RFS is a union of a fixed number of independent Bernoulli RFSs.Its realizations are finite sets with cardinality no greater than M.Source: §2.2 · p. 3

2.2 CB-MeMBer Recursion

  • The CB-MeMBer filter approximates the multi-object posterior with a multi-Bernoulli RFS and recursively propagates its parameter pairs instead of the full posterior.The posterior approximation has unbiased cardinality; its update uses the measurement set, single-object likelihood, detection probability, and clutter intensity.Source: §2.2 · p. 3

3 Objective functions

The paper compares expected R´enyi divergence and MAP cardinality variance as sensor-control objectives. Both depend on future measurements, but the objectives differ in what they evaluate and in their practical limitations.

3.1 R´enyi divergence

  • Expected R´enyi divergence selects controls that maximize information gain, measured by dissimilarity between predicted and updated densities.The policy is myopic, averaging the objective over possible next measurements.Source: §3.1 · p. 4, §3 · p. 4
  • R´enyi divergence generally has no analytic solution, making computation potentially expensive, and its information gain may not directly translate to practical performance criteria.The cited passage names cardinality and localization errors as examples of such criteria.Source: §3.1 · p. 4

3.2 MAP cardinality variance

  • Minimizing estimated MAP cardinality variance directly targets estimated cardinality error and can be computed analytically with the CB-MeMBer filter.The objective favors sensor positions yielding better MAP estimates of target count.Source: §3.2 · pp. 4–5
  • The cardinality-variance objective still requires approximating an expectation over possible future measurements to choose a control.The variance can be computed analytically, but it depends on future measurements.Source: §3.2 · pp. 4–5
  • Cardinality variance does not account for localization error and is expected to perform well only when the sensor provides reasonable information about target states.The paper notes degraded performance in pathological cases with insufficient state information, including bearing- or range-only sensing.Source: §3.2 · p. 5

4 SMC implementation

The sensor-management algorithm uses sequential Monte Carlo approximations with the CB-MeMBer filter and computes control commands according to either the R´enyi-divergence or cardinality-variance objective. The implementation uses predicted states and future-measurement approximations to reduce computational workload.

4 SMC implementation

  • The general SMC implementation approximates the CB-MeMBer recursions and R´enyi-divergence computation; the cardinality-variance objective permits a fast implementation without sampling the multi-target state space.The control-command subroutines differ according to the objective function.Source: §4 · p. 5

4.1 SMC implementation for maximizing R´enyi divergence

  • For R´enyi-divergence control, the cheaper implementation generates one ideal future measurement from predicted estimated states, assuming zero clutter and unity detection rate.The authors report more efficient and reliable state estimates than the PHD-filter approach in.Source: §4.1 · p. 5

4.2 SMC implementation for minimizing cardinality variance

  • Cardinality-variance control still needs Monte Carlo approximation because future measurements enter its expectation, and the straightforward procedure is computationally expensive.The algorithm uses nested loops to sample cardinality variance for each control command.Source: §4.2 · pp. 5–6
  • The reduced-workload approach uses the estimated predicted state to generate one ideal future measurement for computing expected cardinality variance.The passage describes truncating the predicted density and using its estimated state.Source: §4.2 · p. 6
  • The proposed approach reduces computational workload versus the CB-MeMBer strategies by using minimal particles, while eliminating clustering improves tracking performance versus the PHD counterpart.Source: §4.2 · p. 6

5 Simulation results

In range-bearing tests, CB-MeMBer-based control improves on PHD comparators; range-only tests expose poor performance from both objectives as observability changes.

5 Simulation results

  • Both policies move toward and remain near targets; cardinality variance yields smoother trajectories, whereas Rényi allows impulsive jumps.Source: §5 · p. 7, Figure 1 · p. 8
  • The objectives react differently to target births and deaths: Rényi stays near confirmed targets, while cardinality variance moves toward the estimated-position middle.The cardinality-variance strategy does not account for target localization error.Source: §5 · p. 7
  • 0.61 seconds (non-sampling) versus 41.35 seconds (with sampling) for cardinality-variance control; the non-sampling strategy is fastest.Source: §5 · p. 8
  • 10 m lower steady-state miss distance: under the same Rényi objective, CB-MeMBer outperforms PHD-based counterparts.The comparison uses range-bearing measurements.Source: Figure 2 · p. 8, §5 · p. 8
  • Range-bearing OSPA falls faster and is marginally lower at steady state for cardinality-variance CB-MeMBer than its Rényi counterpart.Source: Figure 2 · p. 8, §5 · p. 8
  • In range-only tests, Rényi performs better, but neither objective achieves sufficiently low OSPA; low observability leaves full states difficult to estimate.Source: §5 · p. 8

6 Conclusions and future works

The paper combines POMDP theory with RFS/FISST modeling for computationally tractable multi-target sensor management, while showing that control performance depends on target observability.

6 Conclusions and future works

  • With reasonably informative measurements, both proposed control strategies work well; under low observability, cardinality-variance control degrades noticeably more than Rényi-divergence control.The paper identifies observability as a factor in control-policy performance across scenarios.Source: §6 · p. 9
  • The CB-MeMBer filter approximately propagates a parametrized multi-object posterior, enabling efficient reward calculation and use of estimated cardinality variance as an objective.For the cardinality-variance objective, the proposed fast control algorithm avoids state-space sampling.Source: §6 · p. 9
  • Further work is needed to investigate multi-target observability and which control policies suit different scenarios.Source: §6 · p. 9
Loading 1312.6215v2…