Poisson multi-Bernoulli conjugate prior for multiple extended object filtering

I. INTRODUCTION

• Extended targets can generate multiple spatially distributed measurements per time step, motivating filters beyond the point-target setting and an extended-target PMBM conjugate prior.The PMBM representation separates detected targets from targets not yet detected; prior extended-target conjugacy had been shown for δ-GLMB, but only point-target conjugacy for PMBM.Source: §I · pp. 1–2

II. PROBLEM FORMULATION

• The problem is to propagate a multi-target random-set density through prediction and Bayes update under standard transition and measurement models, then implement and evaluate the PMBM filter.The paper focuses on filtering current target sets, not forming target trajectories.Source: §II · p. 2

1) Poisson point process:

• A PPP has Poisson-distributed cardinality and iid elements, with intensity D(x) = μf(x) combining a Poisson rate and spatial distribution.The intensity integrated over a state-space region gives the expected number of set members there.Source: 1) Poisson point process · p. 2

III. MODELING

• The paper’s PPP model covers clutter measurements, extended-target measurements, target birth, and targets that have not yet been detected.The target and measurement sets are modeled as random finite sets with unknown, time-varying cardinality.Source: §II · p. 2, 1) Poisson point process · p. 3

2) Bernoulli process:

• A Bernoulli RFS represents a target’s uncertain existence and state; an MB unions independent Bernoullis, and an MBM mixes MB densities with weights for data-association sequences.A Bernoulli RFS is empty with probability 1 − r or contains one state with density f(x).Source: 2) Bernoulli process · p. 3

B. Standard extended target measurement model

• The measurement set combines independent clutter and target-generated sets; a detected target generates PPP measurements with state-dependent rate and spatial distribution.Effective detection is pD(x)(1 − e^-γ(x)), incorporating the probability of generating at least one measurement.Source: B. Standard extended target… · p. 3
• Because measurement origins are unknown, each association partitions measurements among clutter or new targets and existing targets, with at most one target index per cell.The update must account for uncertainty about which measurements come from targets and which target generated each one.Source: B. Standard extended target… · p. 3, §IV · p. 3

C. Standard dynamic model

• Existing targets survive independently with state-dependent probability and evolve by a Markov transition; new targets appear independently as a PPP, while spawning is omitted.Source: C. Standard dynamic model · p. 3

A. PMBM density

• The PPP models undetected targets, while the MBM models targets detected at least once; each MB component represents a global data-association hypothesis.The PMBM partitions targets into disjoint undetected and detected subsets.Source: A. PMBM density · p. 4

B. PMBM filter recursion

• Under the standard dynamic model, prediction maps a posterior PMBM density to a predicted PMBM density.The prediction recursion updates the parameters of the PMBM representation.Source: B. PMBM filter recursion · p. 4
• For the standard measurement model, the update preserves PMBM form, but each predicted MB can generate a component for every possible association.The number of mixture components therefore increases with the update.Source: B. PMBM filter recursion · p. 4

V. COMPLEXITY, DATA ASSOCIATION APPROXIMATION, AND APPROXIMATION ERROR

• Unknown data associations make the MBM and PMBM parameter count grow rapidly as observations accumulate, motivating complexity control.The paper next analyzes association complexity and approximation error.Source: §V · p. 4

1) Number of data associations:

• The update association space can grow from exponential O(2^(|Z|+|Ij|)) to factorial O((|Z|+|Ij|)!), depending on the setting.Each detection may be assigned to a previously detected target or to background.Source: 1) Number of data associations · p. 5

2) Number of components in MBM:

• With an empty initial MBM, the number of hypotheses is Bell number B(sum |Zt|); two sets of two detections yield 15 hypotheses, versus 4140 for two sets of four.The Bell-number order is the sum of measurement-set cardinalities over time.Source: 2) Number of components in MBM · p. 5
• The number of unique Bernoulli components is determined by possible subsets of the accumulated measurements, representing the prediction and update operations required.Both updated MB counts and unique Bernoulli counts contribute to computational complexity.Source: 2) Number of components in MBM · p. 5

3) Discussion:

• An MBM with uncertain target existence has fewer hypotheses than its certain-existence MBM01 representation, reducing the number of update weights to compute.The PMBM prediction can be implemented without approximation, unlike the certain-existence δ-GLMB prediction described here.Source: 3) Discussion · p. 6
• For extended targets, the simulation results support the same uncertain-existence trade-off previously reported for point targets: better balance of tracking performance and computational cost.The cited point-target studies report that this trade-off favors global hypotheses with uncertain existence.Source: 3) Discussion · p. 6

B. Approximations of the data association problem

• The implementation combines association reduction with post-update pruning, recycling, and merging to keep PMBM parameters tractable.These operations reduce associations before the update and parameters after it.Source: B. Approximations of the data… · p. 6

1) Reducing the number of associations:

• Gating, grouping, measurement partitioning, and Murty’s M-best assignments form a subset of associations that typically reduces update associations by several orders of magnitude.The gates account for target position, extent, and state uncertainty.Source: 1) Reducing the number of associations · p. 6

2) Reducing the number of parameters:

• Low-weight MBM components are pruned; low-existence Bernoullis are recycled into the undetected PPP, while similar PPP mixture components may be merged.A threshold controls recycling, and PPP components with low weights can also be pruned.Source: 2) Reducing the number of parameters · p. 6

C. Multi-Bernoulli mixture merging

• Rather than merging the whole PMBM into one PMB, the method merges similar MB densities using an assignment minimizing a KL-divergence upper bound below a threshold.The paper notes that merging densities with significantly different extent estimates can distort extent estimates.Source: C. Multi-Bernoulli mixture merging · p. 6

D. Approximation error

• Pruning low-weight components has L1 error bounded by the discarded weights, and retaining more components can make the approximation arbitrarily accurate.For a normalized original density and an unnormalized approximation, the normalized approximation's L1 error is less than twice the sum of truncated weights.Source: D. Approximation error · p. 7
• Recycling a Bernoulli as a PPP minimizes KL divergence; the paper also states that low recycling and merging thresholds yield small approximation errors.Pruning or merging similar PPP intensity components is also described as incurring small error.Source: D. Approximation error · p. 7

VI. GGIW IMPLEMENTATION

• The random-matrix implementation approximates target shape by an ellipse and supports radar, lidar, camera, and noisy nonlinear measurement models.The model was selected in part because it is used in other multiple extended-target filters, facilitating comparison.Source: §VI · p. 7

A. Single target models

• GGIW combines a gamma model for the Poisson measurement rate with Gaussian-inverse Wishart models for detections with unknown mean and covariance.The target state comprises kinematics, extent, and measurement rate.Source: A. Single target models · p. 7

B. Pseudo code for the update and the prediction

• The single-target GGIW parameters are updated from detections and predicted through the motion model, with the corresponding parameters given in Tables II and III.The PMBM recursion assumes GGIW-mixture birth and initial undetected-target intensities; its detection and survival assumptions hold trivially for constant probabilities and are expected to hold for sufficiently smooth functions within the estimate's uncertainty area.Source: A. Single target models · p. 7, B. Pseudo code for the update and the… · p. 7, Table II · p. 8
• The PMBM update applies missed-detection and detection updates to PPP and Bernoulli components, while new unassociated measurement clusters initiate Bernoullis.First-detection PPP densities can be multimodal and are reduced to a single GGIW; missed-detection Bernoullis have two modes that differ only in gamma parameters, and low-existence Bernoullis are pruned.Source: B. Pseudo code for the update and the… · pp. 8–9, Table VI · p. 8

VII. RESULTS

• The simulation study compares PMBM with δ-GLMB and adaptive-birth LMB, and the section also presents laser-range experiments.Birth processes for PMBM and δ-GLMB were tuned to the simulated scenarios.Source: §VII · p. 9
• The simulated target state includes position and velocity with a two-dimensional extent, and uses constant-velocity motion with an identity extent transformation.Source: §VII · p. 9
• The study reports GOSPA, average localization error, missed and false targets, cardinality errors, and sequence-processing time.GOSPA is presented as a unified metric, while the other measures capture properties important in multiple-target tracking.Source: §VII · p. 10
• The three simulated scenarios vary target births: separated targets later approach and split, births at a shared location occur at different times, or four birth locations are used.Source: Figure 1 · p. 10

A. Simulation study

• The study averaged results over 100 Monte Carlo runs for each of three scenarios, two drawn from prior work and one constructed for this paper.The scenarios cover close-target association, staggered target births and deaths, and higher target and clutter counts.Source: A. Simulation study · pp. 10–11
• PMBM had smaller GOSPA errors than both δ-GLMB and LMB in scenario 1, and its localization error did not increase when targets were close.The comparison is shown in Fig. 2 and summarized in Table XI.Source: A. Simulation study · p. 10, Figure 2 · p. 11, Table XI · p. 11
• For most time steps in scenario 2, PMBM had lower GOSPA error, lower NLE, and lower MT, but removed disappeared targets more slowly.The comparison is shown in Fig. 3 and summarized in Table XII.Source: A. Simulation study · p. 10, Figure 3 · p. 11, Table XII · p. 11
• In scenario 3, LMB had larger GOSPA error, NLE, and MT, while δ-GLMB’s GOSPA rose above PMBM’s after about time step 70.PMBM was slower to remove disappeared targets; the scenario results are shown in Fig. 4 and summarized in Table XIII.Source: A. Simulation study · p. 11
• PMBM’s cost was significantly below δ-GLMB’s but above LMB’s; the authors attribute its speed over δ-GLMB to uncertain existence and merging similar unlabelled MB densities.LMB maintains a single MB density, while PMBM maintains a mixture of MB densities.Source: A. Simulation study · p. 11
• Across the compared scenarios, PMBM offers an appealing compromise between computational cost and tracking performance.Its computational cost is lower than δ-GLMB’s but higher than LMB’s.Source: A. Simulation study · p. 11

B. Experiment

• The experiment emphasizes undetected-target intensity because lidar occlusions, not measurement noise or clutter, challenge this data.Detected-target tracking results were essentially identical to prior GGIW-PHD results on the data.Source: B. Experiment · p. 11
• A heterogeneous, time-varying detection probability was used to handle occlusions, allowing the PPP intensity to represent possible undetected targets in occluded regions.This approach had previously been used in.Source: B. Experiment · p. 12
• Behind a stationary pedestrian, the undetected-target PPP intensity increased in the occluded area, while unoccluded areas had very low intensity.Figure 6 visualizes the position component of that intensity; red dots mark lidar detections and the color scale runs from low blue through intermediate green to high yellow.Source: B. Experiment · p. 12

VIII. CONCLUSIONS

• The conclusion states that PMBM compares well with the extended-target δ-GLMB and LMB filters in simulation.Source: §VIII · p. 12
• The Lidar experiment illustrates that approximating detection probability lets the filter track detected targets during occlusions and represent areas where undetected targets may be located.The conclusion notes applications where detection varies in time and space, including scanning radars and airborne optical sensors.Source: §VIII · p. 12
• Using PMBM output to form target trajectories is left for future work.The paper distinguishes filtering current target sets from trajectory formation.Source: §VIII · p. 12

APPENDIX

• The appendix proves Theorem 2 using the probability generating transform, also called the probability generating functional.Source: APPENDIX · p. 12, A. PGFL background · p. 12

A. PGFL background

• A PGFL is a multitarget integral transform of a multi-object density, and the appendix gives the PPP, Bernoulli, and multi-Bernoulli PGFLs.The definitions are attributed to.Source: A. PGFL background · p. 12
• The PMBM PGFL follows from the MBM PGFL under the standard independence assumption.Source: A. PGFL background · p. 13

B. PGFL form of Bayes update

• The Bayes update is formulated in PGFL form for a PPP-clutter measurement model with PPP target measurements, then applied to a PMBM prior to obtain the updated PGFL.The passage identifies the update as corresponding to the Bayes-updated multi-target density.Source: B. PGFL form of Bayes update · p. 13

C. Preliminaries to proof of Theorem 2

• The proof derives functional-derivative rules for the clutter and target terms, including differentiation with respect to a single measurement or a measurement set.The preliminary results support the differentiation needed in the proof of Theorem 2.Source: C. Preliminaries to proof of Theorem 2 · pp. 13–14
• For a single measurement, the partition with that measurement alone represents clutter or a new target; placing it with an existing target represents an association to that target.Source: C. Preliminaries to proof of Theorem 2 · p. 14
• The induction step forms new association partitions either by adding a measurement index to an existing cell or by creating a new cell.Together, these cases enumerate the partitions of the measurement indices and existing-target indices.Source: C. Preliminaries to proof of Theorem 2 · p. 14
• The appendix also states a scalar identity that rewrites a sum involving an inner product in a normalized Bernoulli-like form.The identity defines L, r, and f from a, b, and c.Source: C. Preliminaries to proof of Theorem 2 · p. 14

• Under prior PMBM pgfl (43) and measurement model (45a), the Bayes-updated pgfl has the PMBM form stated in Theorem 2.Source: D. Proof of Theorem 2 · pp. 14–15
• The derivation differentiates the prior pgfl, applies Lemma 5, then takes a ratio to obtain the updated density's pgfl.Source: D. Proof of Theorem 2 · pp. 14–15
• The resulting pgfl factors into a PPP pgfl with intensity (68b) and an MBM pgfl with parameters (65).Source: D. Proof of Theorem 2 · p. 15