1. Introduction
Suppose that, at times
t1, …,
tk, a single sensor collects a time-series
z1:k:
z1, …,
zk of measurements from a target with dynamically evolving state
x. Then, the Bayes-optimal approach for tracking the target is the recursive Bayes filter:
where
fk|k(
x|
z1:k) is the probability distribution of the unknown state
x at time
tk; where:
and where
fk|k−1(
x|
x′) is the target’s Markov state-transition density and
fk(
z|
x) is the sensor’s measurement density. Suppose that these are linear-Gaussian:
Then the family of linear-Gaussian distributions solves the Bayes filter in exact closed form. That is, if the initial distribution is linear-Gaussian—i.e., if
—then [
1]:
where …→ (
xk−1|k−1,
Pk−1|k−1) → (
xk|k−1,
Pk|k−1) → (
xk|k,
Pk|k) →… is the Kalman filter. The family of Gaussian mixture distributions also solves the Bayes filter in exact closed form [
2].
An unexpected recent development has been the generalization of this approach to the multitarget case. Let:
be the multitarget recursive Bayes filter, where
fk|k(
X|
Z1:k) is the probability distribution of the unknown multitarget state-set
X and
Z1:k:
Z1, …,
Zk is the time-sequence of collected multitarget measurement-sets. Suppose that
F is a family of parametrized multitarget distributions
f(
X|
p) with parameter
p ∈ ℘. Then
F solves the multitarget Bayes filter in exact closed form if
fk|k−1(
X|
Z1:k−1) =
f(
X|
pk|k−1) and
fk|k(
X|
Z1:k) =
f(
X|
pk|k) with
pk|k-1,
pk|k∈ ℘. In this case, the multitarget Bayes filter can be replaced by an equivalent, but potentially computationally more tractable, filter: …→
pk−1|k-1→
pk|k−1→
pk|k→….
Remark 1. References [3,4,5,6,7,8,9] employ the terminology “conjugate filter” rather than “exact closed-form filter.” The latter usage is more accurate since “conjugate” refers specifically to exact algebraic closure of F with respect to fk(Zk|X) (and not fk|k−1(X|X′)). Five such filters have been proposed (where the earliest-published papers are indicated):
The purpose of this review paper is to provide an in-depth assessment of these five filters, especially in regard to the following questions: Is this filter theoretically rigorous? Is it a true multitarget tracker? Is it actually exact closed-form?
The basic issue distinguishing (3a, 3b, 3c) from (1, 2) is the form of the initial multitarget distribution and the target-birth model: Poisson or non-Poisson? In particular, the Poisson component of the PMBM distribution fk|k(X|Z1:k−1) is claimed to be a model of the “undetected targets” at time tk—i.e., those targets never detected at times t1, …, tk.
This physical interpretation forces us to address the following question: What is an “undetected target”? This, in turn, requires the formal statistical theory of “undetected targets” developed in
Section 5. This theory results in the following formulas for the probability generating functionals (PGFLs) of the measurement-updated random finite set (RFS)
k|k and its associated detected-target RFS
and undetected-target RFS
(
Section 5.7):
The major conclusions of the paper are as follows:
The GLMB, LMBM filters solve the labeled multitarget Bayes filter in exact closed form.
They are, therefore, true multitarget trackers.
The U-PMBM filter solves the unlabeled multitarget Bayes filter in exact closed form.
The “undetected-targets” interpretation of the U-PMBM filter appears to be valid.
It is theoretically impossible to prune U-PMBM distributions in a practical manner.
The U-PMBM, LA-PMBM and H-PMBM filters are not true multitarget trackers.
The LA-PMBM and H-PMBM filters are theoretically and physically questionable.
In particular, the H-PMBM filter does not solve the “hybrid” multitarget Bayes filter in exact closed form.
The paper is organized as follows: overview of FISST-based multitarget tracking (
Section 2); the GLMB and LMBM filters (
Section 3); the three versions of the PMBM filter (
Section 4); a theory of undetected targets (
Section 5); mathematical derivations (
Section 6); and conclusions (
Section 7).
2. Overview of FISST-Based Multitarget Tracking
This section summarizes those concepts necessary to understand the paper. Greater detail can be found in books [
13,
14,
15,
16], tutorials [
17,
18,
19,
20], and a short survey of advances ca. 2015 [
21]. Additionally, systematic investigations of FISST vs. “point processes” can be found in [
19,
22] and of FISST vs. measurement-to-track approaches in [
23,
24].
2.1. Random Finite Sets (RFSs)
Let ℑ be a single-target state-space (e.g., a region of a Euclidean space) with x, x′ ∈ ℑ and let ℵ be the sensor measurement-space with z ∈ ℵ. Then the state of a multitarget system is represented as a finite subset X = {x1, …, xn} ⊆ ℑ with X = for n = 0. The number of elements in X is denoted as |X|. In a Bayesian approach, unknown states are random variables. Thus, an unknown multitarget state is a random finite set (RFS) ⊆ ℑ.
Similarly, the “measurement” collected from the targets in X is a finite subset Z = {z1, …, zm} ⊆ ℵ with Z = for m = 0. Since measurement-sets are random, multitarget measurements will be represented as random finite measurement-sets Σ ⊆ ℵ.
Remark 2. It is sometimes claimed that multitarget states can be rigorously modeled as variable-length concatenated vectors (
x1, …,
xn) ∈ ∪
n ≥ 0 ℑ
n. This is not the case—see Section 2.4 of [19]. 2.2. Multitarget Calculus
A
multitarget density function is a function
f(
X) ≥ 0 of the finite-set variable
X ⊆ ℑ such that the units of measurement of
f(
X) are
ι −|X| where
ι is the unit of measurement of ℑ. The
set integral of
f(
X) is:
where
fn(
x1, …,
xn) =
f({
x1, …,
xn})/
n! for distinct
x1, …,
xn [
14] (p. 361). Every random finite state-set
has a multitarget probability distribution
(
X): ∫
(
X)δX = 1. The
cardinality distribution of
is:
The
probability generating functional (PGFL) of
is, for “test functions” 0 ≤
h(
x) ≤ 1:
where
hX = 1 if
X =
and
hX = Π
x∈Xh(
x) otherwise. The simplest nontrivial PGFLs are:
where
s(
x) ≥ 0 is a density function on ℑ. The power functional
hX satisfies the generalized binomial theorem [
13] (Equation (3.6)):
The intuitive definition of the
functional derivative of
[
h] is:
where
δx(
y) is the Dirac delta function concentrated at
x. (For a rigorous definition see [
13,
18].) If
X = {
x1, …,
xn} with |
X|=
n then the
general functional derivative is
[
h] if
X = and, otherwise:
The PGFL and multitarget distribution of an RFS are related by:
The probability hypothesis density (PHD) of
is:
FISST includes an extensive “toolbox” of “turn-the-crank” rules for set integrals and functional derivatives—see [
14] (pp. 383–389) or [
13] (pp. 69–80).
2.3. Important RFSs
Various RFSs of importance to this paper are most easily described using their PGFLs:
Poisson RFS: [h] = eD[h−1] where D[h] = ∫h(x)⋅D(x)dx and where D(x) ≥ 0 is a PHD—i.e., a density function on x∈ ℑ.
Bernoulli RFS: [h] = 1 – q + q⋅s[h] where 0 ≤ q ≤ 1 and probability density s(x) are, respectively, the existence probability and spatial distribution of a single target.
Multi-Bernoulli (MB) RFS: where 0 ≤ qi ≤ 1 and probability density si(x) are, respectively, the existence probability and spatial distribution of the i-th of N targets.
Multi-Bernoulli Mixture (MBM) RFS: .
Poisson Multi-Bernoulli (PMB) RFS: .
Poisson Multi-Bernoulli Mixture (PMBM) RFS: .
2.4. Multitarget Recursive Bayes Filter
As noted earlier, this is:
where:
and where
fk|k−1(
X|
X′,
Z1:k−1) is the multitarget Markov state-transition density and
fk(
Z|
X,
Z1:k−1) is the sensor’s multitarget measurement density. It is usually assumed that
fk|k−1(
X|
X′,
Z1:k−1) =
fk|k−1(
X|
X′) and
fk(
Z|
X,
Z1:k−1) =
fk(
Z|
X); but the original forms allow (for example) the target-birth process and the clutter process, respectively, to be estimated from the measurements in
Z1:k−1.
In this paper we will be concerned with
fk(
Z|
X) for only the “standard” multitarget measurement model, which has PGFL:
Here, at time tk, is the sensor probability of detection, is the sensor measurement density, kk(z) is the intensity function of a Poisson clutter process, and . For notational simplicity we will usually suppress the time-index k—e.g., etc.
Likewise, we will be concerned with
fk|k−1(
X|
X′) for only the “standard” multitarget motion model, which has PGFL:
Here, at time tk, is the target probability of survival, is the target Markov density, is the PGFL of a multitarget birth RFS, and . For notational simplicity we will usually suppress the time-index k—e.g., , , etc.
2.5. PGFL Form of the Multitarget Bayes Recursive Filter
The PGFL form of Equation (16) for the standard motion model is [
14] (Equation (14.273)), [
13] (Equation (5.94)):
The PGFL form of Equation (17) is [
14] (Equation (14.280)), [
13] (Equation (5.58)):
where, for the standard measurement model [
14] (Equation (14.290)), [
13] (Equation (5.104)):
In what follows, we will notationally suppress the dependence of these PGFLs on Z1:k−1.
3. The GLMB and LMBM Filters
3.1. Labeled Random Finite Sets (LRFSs)
Track labeling (or, more generally, target identity) in an RFS context was first addressed in 1997 in [
25] (pp. 135, 196–197) and in 2007 in [
14] (pp. 505–508). However, the first implementations of RFS filters did not take track labels into account because of computational concerns. Later implementations, such as the Gaussian mixture cardinalized probability hypothesis density (GM-CPHD) filter, addressed labeling heuristically [
13] (pp. 244–250). The labeling issue was not addressed in a theoretically rigorous and systematic fashion until 2011 in the
labeled RFS (LRFS) papers of Vo and Vo [
7,
8].
In LRFS theory, single-target states are assumed to have the form
x = (
u,ℓ) ∈ ℑ ×
L0 where
u ∈ ℑ is a kinematic target state-vector and ℓ is an element of a countable set
L0 of target labels. The integral on ℑ ×
L0 is defined by:
where, by assumption, ∫
f(
u,ℓ)
du = 0 for all but a finite number of ℓ. The corresponding set integral is:
Let X = {(u1,ℓ1), …, (un,ℓn)} ⊆ ℑ × L0. Then the set of labels of the targets in X is denoted as XL = {ℓ1, …, ℓn}. Given this, X is a labeled multitarget state-set if |XL| = |X|—i.e., if its elements have distinct labels, in which case targets are uniquely identified. An RFS ⊆ ℑ × L0 is a labeled RFS (LRFS) if |L| = || for all realizations = X of . Consequently, the distribution of an LRFS has the following property: (X) = 0 if |XL| ≠ |X|.
In LRFS theory, labels ℓ are unknown random state variables, which must be Bayes-optimally estimated along with the unknown random kinematic states u1, …, un. By way of contrast, in conventional track-management approaches labels are deterministic, heuristic bookkeeping devices.
The LRFS approach requires appropriate definitions of , , fk(z|u,ℓ) and fk|k−1(u,ℓ|u′,ℓ′) when (u,ℓ), (u′,ℓ′) ∈ ℑ × L0. The primary distinction is that fk|k−1(u,ℓ|u′,ℓ′) = δℓ,ℓ′·fk|k−1(u|u′,ℓ′)—i.e., targets do not change labels. For purposes of multitarget tracking and classification (see Remark 4), these quantities will usually depend on the labels. However, for general tracking it can usually be assumed that , fk(z|u,ℓ) = fk(z|u), and fk|k−1(u|u′,ℓ′) = fk|k−1(u|u′).
3.2. Important Labeled RFSs
These are most simply defined in terms of their PGFLs, where 0 ≤ h(u,ℓ) ≤ 1 are labeled test functions:
Labeled Multi-Bernoulli (LMB) LRFS:
where
J ⊆
L0 is finite, 0 ≤
qℓ ≤ 1 and
sℓ[
h] = ∫
h(
u,ℓ)⋅
s(
x,ℓ)
du and
sℓ[
1] = 1 for all ℓ ∈
J.
Labeled Multi-Bernoulli Mixture (LMBM) LRFS: .
Generalized Labeled Multi-Bernoulli (GLMB) LRFS: where: (a) O is a finite set of indices o; (b) so,ℓ(u) = so(u,ℓ) with ∫so,ℓ(u)du = 1 for each o,ℓ is the spatial distribution corresponding to the target label ℓ and the index o; (c) ωo(L) ≥ 0 for all finite L ⊆ L0; (d) Σo∈ O ΣLωo(L) = 1; and (e) so,ℓ[h] = ∫h(u,ℓ)⋅so(u,ℓ)du.
Any labeled multitarget distribution can be approximated by a GLMB distribution that has the same PHD and cardinality distribution [
26].
Remark 3. A Poisson RFS ⊆ ℑ × L0 is not an LRFS. For, let D(u,ℓ) ≥ 0 be a PHD on ℑ × L0—i.e., ∫ D(u,ℓ)du = 0 for all but a finite number of ℓ ∈ L0 with μ = Σℓ ∫ D(u,ℓ)du. Let (X) be Poisson with PHD D(u,ℓ): (X) = e−μΠ(u,ℓ) ∈ XD(u,ℓ) and let X₀ = {(u1,ℓ₀), …, (un,ℓ₀)} with |X₀| = n. Then, (X₀) ≠ 0 even though the target with label ℓ₀ has n different kinematic states.
3.3. The GLMB Filter
This filter was introduced in 2011 in [
8] and elaborated in [
6,
7]. Suppose that
L0 = {0,1, …,} × {1,…} and if ℓ = (
k,
i) then
tk is the time that trackℓ was initiated and
i ≥ 1 distinguishes it from any other track created at time
tk. At time
tk, a finite number of labels in
Lk = {
k} × {1, …} are assigned to hypothesized newly-appearing tracks. Thus, at time
tk, the set
L[0:k] of all currently assigned track labels is a finite subset of
L0:k = {0,1, …,
k} × {1,…}; and each such label is an unknown discrete random variable ℓ ∈
L[0:k] which must be estimated.
Given this, the family of GLMB distributions solves the labeled multitarget Bayes filter in exact closed form. In particular, at time
tk:
where the summation is taken over all (
α1,…,
αk) ∈
A1 × … ×
Ak; where each
αi:
L[0:i] → {0,1, …, |
Zi|} is a label-to-measurement association—i.e., if
αi(ℓ) =
αi(ℓ′) > 0 then ℓ = ℓ′; where
is a target spatial distribution; and where
Ai denotes the set of all such associations
αi at time
tk.
The GLMB filter is a true Bayesian multitarget tracker because it is guaranteed to propagate target tracks with unique track labels (a “true” tracker), which in turn are realizations of unknown random identity-variables (a “Bayesian” tracker).
Moreover, because it is an exact closed-form solution of the labeled multitarget Bayes filter, the GLMB filter has provably Bayes-optimal track-management. At time tk−1, an (approximate) Bayes-optimal multitarget state estimate Xk−1|k−1 is extracted from fk−1|k−1(X|Z1:k−1). At time tk, a similar estimate Xk|k is extracted from fk|k(X|Z1:k). If (u,ℓ) ∈ Xk−1|k−1 and (u′,ℓ) ∈ Xk|k then (u,ℓ) and (u′,ℓ) both belong to the track with label ℓ. If (u′,ℓ) ∉ Xk|k for any u′ then track ℓ has been dropped. If (u,ℓ) ∈ Xk|k but (u′,ℓ) ∉ Xk−1|k−1 for any u′ then a track with label ℓ has been initiated or reacquired.
Due to the number of association-vectors (
α1, …,
αk) increases without bound, the summation in Equation (25) must be pruned at every time-step. The information loss due to pruning can be characterized exactly—i.e., the
L1 norm between the pruned and unpruned distributions is the sum of the weights of the pruned terms [
6] (Proposition 5).
Using Gibbs stochastic sampling techniques, the GLMB filter can be implemented with computational order
O(
n2m) where
m is the current number of measurements and
n the current number of tracks [
5]. This is particularly advantageous when clutter is dense. The most recent such implementations can simultaneously track over a million 2D targets in significant clutter using off-the-shelf computing equipment [
27].
Remark 4. Every target has a unique identity state variable [25] (pp. 135, 196–197). A track label is a provisional identity assigned to a target in lieu of its actual identity. The GLMB filter can be generalized from joint multitarget detection and tracking to joint multitarget detection, tracking, and identification. This is accomplished by incorporating identity information into target labels [9].
3.4. The LMBM Filter
In [
10] it was shown that the family of LMBM distributions solves the labeled multitarget Bayes filter in exact closed form. The corresponding LMBM filter is, therefore, a true Bayesian multitarget tracker with provably Bayes-optimal track management. It is somewhat less computationally expensive than the GLMB filter, but also less accurate since LMBM distributions are less accurate approximations of labeled multitarget distributions than GLMB distributions.
4. The PMBM Filter
There are at least three successive versions of the PMBM filter. The section is organized as follows: the “unlabeled” PMBM (U-PMBM) filter (
Section 4.1); the “undetected targets” interpretation of this filter (
Section 4.2); the “label-augmented” PMBM (LA-PMBM) filter (
Section 4.3); “hybrid labeled-unlabeled” RFSs (
Section 4.4); the “hybrid labeled-unlabeled” PMBM (H-PMBM) filter (
Section 4.5); and theoretical issues with the H-PMBM filter (
Section 4.6).
4.1. Unlabeled PMBM (U-PBMB) Filter
In this original 2012 version [
11], all RFSs are unlabeled—LRFSs are never mentioned. All target-birth RFSs are assumed to be Poisson—in our notation,
for all
k ≥ 1—as is the initial RFS:
. Given this, the PMBM filter propagates PMBM distributions in exact closed form. Specifically,
and
where
and
are Poisson and where
and
are MBM. The demonstration of this fact in [
11,
12] is somewhat sketchy. The following PGFL-based verification of it will be useful in the sequel.
U-PMBM Filter Time-Update. According to Equation (20) and substituting
, the PGFL prediction formula is:
where, by assumption,
Gk−1|k−1[
h] is PMBM:
Thus, predicted PGFL is easily seen to be PMBM:
where:
U-PMBM Filter Measurement-Update. Let
Z = {
z1, …, z
m} with |Z| =
m be collected at time
tk. According to Equation (21) the measurement-updated PGFL is:
where, by Equation (22),
F[
g,
h] =
eκ[g−1]⋅
Gk|k−1[
h(1 +
pDLg−1)] and, by assumption,
Gk|k−1[
h] is PMBM:
Thus,
, where:
and the measurement-updated PGFL is:
where:
Since each Fl[g,h] has the same Poisson factor , it is sufficient to show that the measurement-update of a PMB PGFL is a PMBM PGFL Accordingly, in what follows we neglect the index l in Fl[g,h].
Two cases must be considered:
N = 0 and
N > 0. If the former, then
and so from the chain rule for functional derivatives [
14] (Equation (11.280)), we find that
Gk|k[
h] is PMB:
Now assume
N ≥ 1. Applying the general product rule for functional derivatives [
14] (Equation (11.274)) to Equation (33):
where
θ(
z) =
k(
z) +
D[
hpDLz]. The
i-th fraction in the rightmost product is nonzero only if
Wi is empty or a singleton; and if
Wi =
then it is 1. Thus, each list
W0,
W1, …,
WN is mathematically equivalent to an association
α:{1, …,
N}→{0,1, …,
m}—
i (i.e.,
α(
i) =
α(
i′) > 0 implies
I =
i′. Setting
g = 0, Equation (38) becomes, after some algebraic manipulations:
where the summation is taken over all associations. Therefore, the measurement-updated PGFL for a PMB predicted PGFL is:
This can be rewritten as the PMBM PGFL:
where:
4.2. Undetected-Target Interpretation of the U-PMBM Filter
The PMBM filter therefore, solves the unlabeled multitarget Bayes filter in exact closed form. However, the PMBM approach goes beyond this to adopt a specific physical interpretation of PMBM RFSs. Let:
be the PMBM PGFL at time
tk. It is clear from Equations (29) and (41) that the time- and measurement-updates for the Poisson factors are, respectively,
and
. The formulas for
Dk|k−1(
x) and
Dk|k(
x) thus involve
,
, and
, but not
,
,
κi,
Z1:i−1,
Z1:i.
This fact has led to the interpretation of
as a model of the “undetected targets” at time
tk [
4,
11,
12]. According to [
11] (p. 1103), these are “…targets that have never been detected”—i.e., not detected at times
t1, …,
tk. It was subsequently stipulated that “…detected targets cannot become undetected targets” [
4] (p. 246).
The primary justification for the PMBM approach is the following: “One significant benefit of the inclusion of a Poisson component is in initialization of the tracker…The Poisson distribution provides a convenient mechanism for specifying a prior distribution on the number and position of targets when little information is available” [
12] (p. 1670).
However, this potential advantage is negated by a major theoretical obstacle: Poisson RFSs require non-unique labels and so are not LRFSs (see Remark 3). Due to this, they cannot be used in any theoretically rigorous, true multitarget tracker.
A more subtle obstacle is this:
it is theoretically impossible to prune PMBM distributions in a practically useful manner. When a GLMB distribution (Equation (25)) is pruned, the pruned distribution is a GLMB distribution. When a PMBM distribution is pruned, however,
it is usually not even a multitarget density function. First consider an LMB distribution
({
x1, …,
xn}) [
14] (Equation (11.133)). Any term in it has the form
where
x1, …,
xn are distinct, 1 ≤
i1 ≠ … ≠
in ≤
ν, and
f1(
x), …,
fν(
x) are distinct density functions. Since
is not symmetric in
x1, …,
xn—and therefore, not a multitarget density—neither is any other pruning of
. Now, let
be a PMB distribution:
where
α: {1, …,
ν} → {0, 1, …,
n} is an association. Its terms have the same form as before, except that the
fican be equal to
D but those
fithat are not
D are distinct. Since
is symmetric only when
fj =
D for all
j = 1, …,
n, no pruning of
other than this case is a multitarget density. What
is theoretically permissible is to prune an MBM (resp. PMBM) PGFL by eliminating one or more of its MB (resp. PMB) PGFL terms. However, pruning the individual terms of the corresponding MB (resp. PMB) distributions is not permissible—which is exactly what is required to eliminate specific small-weight hypotheses.
4.3. “Label-Augmented” PMBM (LA-PMBM) Filter
As was noted at the beginning of
Section 3.1, unlabeled RFS-based filters, such as the GM-CPHD filter, can heuristically propagate tracks even though they are not true multitarget trackers. The U-PMBM filter can propagate tracks using similar heuristics, but it—like the GM-CPHD filter—is not a true multitarget tracker since it is unlabeled. Accordingly, in 2015 it was modified as follows: “… [the Vo-Vo paper [
7]] shows that the labelled case can be handled within the unlabeled framework by incorporating a label element in to the underlying state space” [
12] (p. 1675). That is, it was claimed that the PMBM filter can be extended to the labeled case by replacing the unlabeled single-target state space ℑ with the labeled state space ℑ ×
L0.
This modified PMBM filter will be referred to as the “label-augmented” PMBM (LA-PMBM) filter. It must propagate PMBM RFSs of ℑ ×
L0 with PGFLs:
Now, however, the PHD Dk|k and spatial distributions must have the respective forms Dk|k(u,ℓ) and , where ∫Dk|k (u,ℓ)du = 0 and ∫ for all but a finite number of ℓ and where for all l,i.
There is a serious theoretical difficulty, however: the are not track distributions. For if otherwise, would imply that , a contradiction. Therefore, appears to be physically meaningless.
Beyond this, the above claim—that “the labelled case can be handled within the unlabeled framework”—is untrue. As was noted in Remark 3, a Poisson RFS on ℑ ×
L0 is not an LRFS since it
requires nondistinct target labels. Consequently, it is not possible for the LA-PMBM filter to be a true multitarget tracker. Instead, it “…is able to maintain track continuity implicitly based on the information provided by metadata” [
4] (p. 245)—that is, only heuristically.
4.4. “Hybrid Labeled-Unlabeled” RFSs
Like the U-PMBM filter, the LA-PMBM filter is not a true multitarget tracker—a fact that was pointed out in 2017 in [
23] (Section XI-E). Apparently to address this issue, it was modified in 2018 as follows [
4] (p. 246): A single common label—ℓ*, say—is assigned to all “undetected targets” at all times, whereas “detected targets” are uniquely labeled as in LRFS theory. Additionally, the “undetected-target” RFS at any time is assumed to be a Poisson RFS on ℑ × {ℓ*} (a slightly later paper, [
3], also appears to employ the H-PMBM approach, except that ℓ* is implicit rather than explicit.)
No careful theoretical foundation for the hybrid approach was provided in [
4]. It is, therefore, necessary to construct one here. The label space is
L0 =
L\*∪{ℓ*} where
L\* =
L0 − {ℓ*}. Given a finite subset
X = {(
u1,ℓ
1), …, (
un,ℓ
n)} ⊆ ℑ ×
L0, as usual let
XL = {ℓ
1, …, ℓ
n} denote the set of labels in
X. Additionally, let
X* = {(
u,ℓ) ∈
X|ℓ = ℓ*} be the subset of
X of targets that are “undetected”; and let
X\* =
X −
X* be the targets in
X that are “detected.” Then it is assumed that the only legitimate state-sets
X are those such that |
X\*| = |
XL − {ℓ*}|—i.e., those for which the detected targets have distinct labels other than ℓ*. Let us refer to these as “hybrid state-sets.” Let
be a “hybrid RFS”—i.e., an RFS of ℑ ×
L0 whose instantiations are hybrid. Then it must be the case that
(
X) = 0 if
X is not hybrid. Thus, every distribution defined for hybrid
X must include the factor
.
The goal of Sections XI-XIII of [
4] is to apply the PMBM filter, Equations (28) and (41), to the hybrid state space ℑ ×
L0. This will be addressed in the next section. First, however, we must reformulate
,
,
fk(
z|
u,ℓ),
, and
fk|k-1(
u,ℓ|
u′,ℓ′) when (
u,ℓ), (
u′,ℓ′) ∈ ℑ ×
L0 are hybrid. For “detected targets” (i.e., ℓ ≠ ℓ* and ℓ′ ≠ ℓ*), the usual LRFS formulation applies. For “undetected targets,” it is reasonable to define
,
and
fk(
z|
u,ℓ*) =
fk(
z|
u). It also makes sense to define
for some
since, by definition, targets that have just appeared cannot have been detected yet (to wit: “…these states of…newborn targets will be described by…a Poisson RFS…[whose] elements have label [ℓ*]”[
4] (p. 247).
Finally, consider
fk|k−1(
u,ℓ|
u′,ℓ′) =
pk|k−1(ℓ|
u′,ℓ′)⋅
fk|k−1(
u|
u′,ℓ,ℓ′) when ℓ = ℓ* or ℓ′ = ℓ*. “Undetected targets” must retain the label ℓ*, since they cannot become “detected targets” in the absence of a detection process that has not yet occurred, and if ℓ′ ≠ ℓ* then
pk|k−1(ℓ*|
u′,ℓ′) = 0 since “…detected targets cannot become undetected targets” [
4] (p. 246). Thus,
pk|k−1(ℓ|
u′,ℓ*) =
δℓ,ℓ*⋅
pk|k−1(ℓ*|
u′,ℓ*), in which case it is reasonable to assume that
pk|k−1(ℓ*|
u′,ℓ*) =
pk|k−1(ℓ*|ℓ*) = 1 and
fk|k−1(
u|
u′,ℓ*,ℓ*) =
fk|k−1(
u|
u′). (Note that if target identity is to be taken into account as per Remark 4, these simplifications are no longer appropriate.)
4.5. “Hybrid Labeled-Unlabeled” PMBM (H-PMBM) Filter
Now let:
be the PGFL of a PMBM RFS
on ℑ ×
L0 with ℓ* ∈
L0, where the MBM factor in Equation (45) has been replaced by an LMBM factor since “detected targets” are now assumed to be uniquely labeled; and where
are finite subsets of
L0 and
for all
. Since the Poisson factor
applies only to “undetected targets” with common label ℓ*, it must be the case that
for every
l = 1,..,
Nk|k and that
for some
. We will refer to Equation (48) as an “H-PMBM PGFL”.
Given this,
and
“….can be propagated in parallel, in both cases by carrying out a prediction step and an update step…” [
4] (p. 248); where the undetected-target distribution
with
Xu ⊆ ℑ × {ℓ*} is Poisson and the detected-target distribution
with
Xd ⊆ ℑ × (
L0 − {ℓ*}) is LMBM.
Moreover, the following claim is made about these two filters: “In the following development of the prediction and update steps, we use the fact that the posterior pdf of the overall multitarget state RFS…factorizes as…” (in current notation):
However, Equation (49) is untrue. For by Bayes’ rule:
It then follows from Equation (49) that
. Likewise,
Thus, RFSs
Xu and
Xd are statistically independent of each other. This means that the filter for
and the filter for
are
statistically decoupled. However, this is not the case, since “…the update step for
Xd involves the prediction results for both
Xd and
Xu…” [
4] (p. 248). (This is because the LMBM component of
Gk|k[
h|
Z1:k] in Equation (41) depends on the Poisson component
of
Gk|k−1[
h|
Z1:k−1] via
θk(
z) =
κk(
z)+
Dk|k−1[
hpDLz]. That is,
.)
4.6. Theoretical Issues With the H-PMBM Filter
From
Section 4.1 we know that the PMBM filter on ℑ ×
L0 is guaranteed to propagate PMBM distributions on ℑ ×
L0 in exact closed form. However, does it propagate
hybrid PMBM distributions in exact closed form? This does not appear to be the case.
Consider, for example, Equation (36) with
k = 1 and with the single-target state space being ℑ ×
L0 (with ℓ*∈
L0) rather than ℑ:
Here,
since
and
are independent of labels, and:
Given that , D1|1(u,ℓ) has the correct form for an H-PMBM PGFL However, the product is not an LMB PGFL This is because the individual Bernoulli factors are indexed by the elements of Z1, not by labels in L0.
Thus, as a heuristic workaround, a distinct label ℓ
z≠ ℓ* is assigned to each
z ∈
Z1: “For each measurement…a new Bernoulli component is created, to which [a] unique label…is assigned” [
4] (following Equation (107)); and: “…[e]ach measurement at each time step gives rise to a new potentially detected target. That is, there is the possibility that a new measurement is the first detection of a target, but it can also correspond to another previously detected target or clutter, in which case there is no new target. As this target may exist or not, its resulting distribution is Bernoulli and we refer to it as [sic] ‘potentially detected target’” [
3] (p. 1885).
It follows that the labeled track distribution of the Bernoulli representation of the “potentially detected target” corresponding to z∈ Z1 must be .
This workaround results in at least three theoretical difficulties:
There is an inherent theoretical conflict between labeling using ℓ∈ L0 and labeling using z∈ Z1. Since and since if ℓ ≠ ℓ* and since ℓz ≠ ℓ*, it follows that s1|1(u,ℓz) = 0 identically for all z ∈ Z1—a contradiction. One could sidestep this difficulty by redefining , but this would be another heuristic workaround.
In [
4] (pp. 245–246) the following was stated: “In cases of limited prior birth information, one typically uses a heuristic to generate new Bernoulli components based on measurements from the previous time step (Reuter et al. [
28]). Such heuristics can be avoided with the MB-Poisson model…” This is untrue on both counts. First, and as was noted following Equation (17), approaches that dynamically estimate the target-birth process “based on measurements from the previous time step”—i.e., based on
Zk-1—are theoretically permissible. Examples include [
29] and [
30]—and [
28]. Second, note that the “MB-Poisson model” employs a “heuristic to generate new Bernoulli components based on measurements from”: the current time-step! Thus, how is it conceptually different from the approach in [
28]?
More seriously, the dynamical transition of undetected targets to detected ones occurs during the measurement-update, as mediated by fk(Zk|X), rather than—as theoretically should be the case—during the time-update, as mediated by fk|k−1(X|X′,Z1:k−1). Thus, fk(Zk|X) has been implicitly assumed to have the form fk(Zk|X,X′)—which is not the case (see Equation (54)).
The H-PMBM filter therefore, does not appear to have a theoretically rigorous, closed-form mechanism for assigning labels to newly-detected “undetected targets.” And this fact is a direct consequence of the Poisson factor in Equation (48).
However, there is a far more fundamental theoretical and phenomenological difficulty:
the hybrid approach has no basis in physical reality. Targets are physically real entities regardless of whether or not they are detected. They have distinct (but unknown) real-world identities and therefore, inherently have distinct (unknown) labels. As was noted in Remark 4, target labels in
L0 are provisional identities assigned in lieu of more precise identifying information. LRFS labels are, therefore, not “artificial variables that are added to the target states” [
3] (p. 1884). Rather, they are
standbys for the realizations of a physically real random state-variable: target identity. The H-PMBM approach, by way of contrast, requires targets with label ℓ* to have multiple kinematic states simultaneously—a
physical impossibility.
5. A Statistical Theory of Undetected Targets
The meaning of the “undetected target” concept is extremely unclear. Thus, the purpose of this section is to devise a statistically rigorous—and yet intuitive—theory of undetected targets. As stated in the Introduction, our ultimate goal is to construct a concrete formula for the posterior “undetected targets” PGFL . The argument presented is as follows.
Section 5.1: We surmise that the “undetected target” concept is meaningful only at the instant that an observation process—in the form of the standard multitarget likelihood function
fk{
Zk|
X)—is applied.
Section 5.2: A more useful formula for
fk{
Zk|
X), Equation (57).
Section 5.3: Thedetected-target likelihood function
, Equation (69).
Section 5.4: The detected-target density
, Equation (75).
Section 5.5: The undetected-target likelihood function
, Equation (83).
Section 5.6: The undetected-target density
, Equation (90).
Section 5.7: The measurement-updated PGFL and its associated detected-target and undetected-target PGFLs, Equations (97), (100), and (101).
Section 5.8: Analysis of the “undetected target” interpretation.
Section 5.9: The detected-target and undetected-target PGFLs when the prior PGFL is Bernoulli, Equations (110) and (112).
5.1. The “Undetected Target” Concept
At its most elemental level, the concept of “detected” vs. “undetected” target at time tk is independent of previous measurement history. The multitarget predicted distribution fk:k-1(X|Z1:k-1) determines how probable any given multitarget state-set X will be at time tk. However, only the current multitarget likelihood function fk(Z|X) determines which elements of X are detected vs. undetected at time tk.
The question then becomes: Given a finite subset X ⊆ ℑ, which elements of X generated measurements in Zk and which did not? The most that we can say is that, for each Y ⊆ X, there is some probability that all elements of Y ⊆ X generated measurements in Zk. The detected-target set is, therefore, a discrete RFS ⊆ X. Likewise, there is some probability that no elements of V ⊆ X generated measurements in Zk. The undetected-target set is therefore, a discrete RFS ⊆ Xwith . Given this, the following questions will be addressed:
How do we extend the discrete distributions and to posterior probability densities and ?
What are the PGFLs and of and ?
If Gk|k−1[h|Z1:k−1] has a given algebraic form, then what forms do and have?
5.2. The “Standard” Multitarget Likelihood Function
This is, for
Zk = {
z1, …,
zm} with |
Z| =
m and
X = {
x1, …,
xn} with |
X| =
n, [
14] (Equation (12.139)), [
13] (Equation (7.21)):
where
α:{1, …,
n}→{0,1, …,
m} is a measurement-to-track association (MTA) and
Ak is the set of all such associations at time
tk. That is,
α is such that
α(
i) =
α(
i′) > 0 implies
i =
i′. As usual,
κk(
z) denotes the intensity function of the Poisson clutter process,
λk = ∫κ
k(
z)
dz,
Lz(
x) =
fk(
z|
x) is the single-target likelihood function, and
is the state-dependent probability of detection. We will abbreviate
κk(
z) =
κ(
z),
λk =
λ,
, and
fk(
z|
x) =
f(
z|
x).
Remark 5. In general a multitarget state-set X will be labeled. However, the labeled version of Equation (54) is almost identical in form to Equation (54): fk(Z|X) = e−λκZ if X is not labeled (not just if X =). Additionally, the “undetected target” concept was originally raised in the context of unlabeled RFSs. Thus, it is sufficient to use Equation (54).
Choose a particular α ∈ A. Then Xd:α = {xi∈ X|α(i) > 0} is the set of xi∈ X that—according to the hypothesis α—generated measurements in Z. Likewise, Xu:α = {xi ∈ X|α(i) = 0} is the set of those that did not. Now note the following:
MTAs are in one-to-one correspondence with pairs (Y,τ) where Y ⊆ X with |Y| ≤ |Zk| and where τ:Y ⇒ Zk is a one-to-one function (i.e., τ(y) = τ(y′) implies y = y′).
For on the one hand, let us be given α. Then define the pair (Yα,τα) where Yα = {xi∈ X|α(i) > 0} and τα(xi) = zα(i) for xi∈ Y. On the other, let us be given a pair (Y,τ). Then for each i ∈ {1, …, n} define α(Y,τ)(i) = j if xi∈ Y and τ(xi) = zj; but α(Y,τ)(i) = 0 if otherwise—i.e., if there is no j ∈ {1, …, m} such that τ(xi) = zj. It is easily verified that the transformations α ↦ (Yα,τα) and (Y,τ) ↦ α(Y,τ) are inverses of each other.
Now define:
where the unitless ratio:
is a measure of how “target-like” vs. “clutter-like” the measurement
τ(
y) is (additionally, note that the “∗” in “
” and “
” no longer refers to the label “ℓ*” in
Section 4.4.)
Given this, note that the multitarget likelihood function can be rewritten as:
It therefore, follows that for all X. Thus, is the same thing as , but under perfect-detection conditions.
For future reference note that if
Y = {
y1, …,
yn} with |
Y| =
n, then:
where the third summation is taken over all {
z1, …,
zn} ⊆
Z of cardinality
n = |
Y| ≤ |
Z|.
5.3. The General Detected-Target Likelihood Function
Given these preliminaries, let
X be a fixed finite subset of ℑ and define:
where, note, (
1X)
Y = 1 if
Y ⊆
X and (
1X)
Y= 0 otherwise. This is a continuous density in
Z and a discrete distribution in
Y:
where the final equation follows from Equation (11). Equation (61) is the probability that all of the elements of the subset
Y of
X generated measurements; and is largest when
pD(
x) ≈ 1 for
x ∈
Y and
pD(
x) ≈ 0 for
x ∈
X −
Y, where “≈” denotes approximate equality. It is the distribution of the
detected-target RFS in
X:
Equation (57) has the following interpretation: LZ(X) is the unweighted average of hypotheses regarding the likelihood that subset Y of X generated measurements in Z. The factor quantifies the “raw detectability” of Y, whereas measures the degree to which detectability is degraded by clutter under perfect-detectability conditions.
We need to transform
so that it becomes a continuous density
with respect to
Y. This is accomplished as follows. For
X = {
x1, …,
xn} with |
X| =
n and
Y = {
y1, …,
yν} with |
Y|=
ν, define:
where the summation is taken over all one-to-one functions
τ:{1, …,
ν} ⇒ {1, …,
n}. This is a multitarget density function with respect to Y.
Note that Equation (64) can be rewritten in the same form as Equation (59):
where the first summation is taken over all one-to-one functions
τ:
Y ⇒
X and where we define
. Given this, in
Section 6.1,
Section 6.2 and
Section 6.3 it is respectively shown that:
where Equations (66)–(68) are true for all finite
X,Y ⊆ ℑ and all multitarget densities
f(
X).
We are now in a position to define the
general detected-target likelihood function:
It is the likelihood that, given a target-set
X, the following are simultaneously true:
Zk is the measurement-set collected at time
tk; and
Y ⊆
X is a subset of targets in
X that generated measurements in
Zk. In
Section 6.4 the following is verified:
5.4. The General Detected-Target Density
Let us be given the prior distribution
fk|k−1(
X|
Z1:k−1). Since
does not depend on
Z1:k−1, then
and so from Bayes’ rule and the total probability theorem we obtain:
This is the probability (density) that, at time tk, the measurement-set Zk will be collected, and that the elements of Y ⊆ ℑ generated measurements in Zk.
Additionally, from Bayes’ rule, we get the general detected-target posterior density—i.e., the probability (density) that all of the elements of
Y ⊆ ℑ generated measurements in
Zk:
It is the distribution of the
general detected-target RFS at time
tk.
We thereby end up with the following specific formulas:
For, substituting Equation (69) for
and applying Equation (67):
where the final equation follows from Equation (11.251) of [
14].
5.5. The General Undetected-Target Likelihood Function
The detected-target RFS
was defined in Equation (63). By definition, the undetected-target RFS in
X is
. For all
V ⊆
X, note that:
(Note that (
1X)
V should be used rather than (
1X)
X-V—the latter is incorrect because it does not force
V to be a subset of
X. For example, let
V = {
u} where
x ∉
X. Then it should be the case that
. However, if we instead use (
1X)
X-V, then since (
1X)
X-{x} = (
1X)
X = 1 we would get the incorrect result
.)
Additionally, note that
is largest when
pD(
x) ≈ 0 if
x ∈
V and
pD(
x) ≈ 1 if
x∈
X −
V—in which case all of the elements of
X −
V should generate measurements in the manner described by
. The undetected-target analog of Equation (60) is, therefore:
From here on, the analysis for
proceeds in the same manner as that for
. That is, replace
with:
This is the
general undetected-target likelihood function—i.e., the likelihood that, given a target-set
X, the following are true:
Zk is the set of generated measurements at time
tk; and
V ⊆
X is a subset of targets in
X that generated no measurements. Thus, by Equations (58) and (68):
5.6. The General Undetected-Target Density
Let
fk|k−1(
X|
Z1:k−1) be the predicted multitarget distribution at time
tk. Since
, from Bayes’ rule and the total probability theorem we obtain:
This is the probability (density) that, at time
tk, the measurement-set
Zk will be collected; and that none of the elements of
V ⊆ ℑ generated measurements in
Zk. Thus, the general undetected-target density—i.e., the probability (density) that none of the elements of
V ⊆ ℑ generated measurements in
Zk—is:
This leads to the following specific formulas:
For, using Equation (67):
and:
where the final equation results from Equation (11.251) of [
14].
5.7. PGFLs of the Detected/Undetected Target Densities
The PGFL corresponding to
(Equation (75)) is:
The PGFL corresponding to
(Equation (87)) is simpler than
:
where the final equation results from Equation (11.251) of [
14].
Finally, in
Section 6.5 it is shown that:
Thus, posterior PGFL at time
tk is an amalgam of the undetected-target and detected-target PGFLs. Note that if
pD = 1 (all targets are perfectly detectable) then
(there are no undetected targets) and
(all targets are detected).
5.8. Analysis of the "Undetected Target" Interpretation
Let us now apply the preceding analysis to the “undetected-target” interpretation, in which:
“undetected targets” are those that “…have never been detected…” [
11] (p. 1103); and
“…detected targets cannot become undetected targets” [
4] (p. 246).
In what follows it will be demonstrated that the second claim leads to a contradiction, whereas the first one appears to be consistent with the formal theory of undetected targets.
Claim (2) Leads to a Contradiction. According to Equation (100), is the PGFL of targets that are undetected only at time tk. According to Claim (2), if a target is undetected at tk then it was also undetected at times t1, …, tk−1. Given this, it must be the case that —and, in particular, that is always Poisson.
However, this is not true. For, examine the first steps of the U-PMBM filter. Begin with
G0|0[
h] = 1—i.e., no targets are initially present in the scene. Then
where
is the PHD of the Poisson target-appearance RFS; and so
by Equation (20); and so
G1|1[
h|
Z1] is a PMB PGFL as in Equation (36):
These equations are consistent with the “undetected targets” interpretation, Equation (45), since:
From Equation (20), the next predicted PGFL is PMB:
It can be shown that
is PMB, not Poisson. The claim that “detected targets” cannot become “undetected targets”, therefore, leads to a contradiction. The proof of this fact for general
will not be proved here, since it suffices to address the following informative special case. Let
,
, and |
Z1| = 1, so that
G2|1[
h|
Z1] is Bernoulli. Then in
Section 5.9 we will determine
and
and show that the former is Bernoulli—i.e., not Poisson.
Claim (1) is Consistent with the Formal Theory of Undetected Targets. Alter the preceding argument as follows. Instead of predicting G1|1[h|Z1], predict its Poisson factor to obtain . Next, determine the undetected-target posterior PGFL of —i.e., the PGFL of those targets undetected at times t1,t2. According to Equation (105), since is Poisson it is equal to . Then predict to obtain the Poisson PGFL . Determine the undetected-target posterior PGFL of —i.e., the PGFL of those targets undetected at times t1,t2, t3. According to Equation (105), it is equal to . Repeat in this manner. At time tk, determine the unpredicted-target posterior PGFL of the Poisson PGFL . It is the PGFL of those targets undetected at times t1, …, tk and is equal to .
5.9. Undetected/Detected-Target PGFLs for a Bernoulli Prior
Suppose that
Gk|k−1[
h|
Z1:k−1] = 1 −
q +
q·
s[
h] is Bernoulli. Then the measurement-updated PGFL is Bernoulli [
14] (p. 520):
In
Section 6.9 it is shown that the detected-target posterior PGFL is Bernoulli:
In
Section 6.8 it is shown that the undetected-target posterior PGFL is not Poisson:
Here, qu is the probability that the posterior undetected-target RFS is nonempty—i.e., it is the target’s composite probability of undetectability. Note that qu + qd = q+.
Additionally, note that qu parses the distinction between nonexistent vs. undetectable targets. If q> 0 (the target exists) and pD = 0 (it is undetectable), then no information can be collected about it and so its composite undetectability is qu = q. For example, if q = 1 then qu = 1—i.e., if a definitely-existing target is undetectable then it is compositely undetectable.
At the other extreme, if q = 0 (it does not exist) then qu = 0 (it is compositely detectable: 1 − qu = 1). This seems counter-intuitive since a nonexistent target would seem to be inherently undetectable. However, a nonexistent target is neither detectable nor undetectable. Whereas an existent target can generate either an actual measurement z or the null measurement , a nonexistent target cannot generate any measurement. Thus, a nonexistent target has been “detected” if, as must be the case, it has not generated any measurement. In this sense, all nonexistent targets are compositely detectable.
Now suppose that
q = 1—i.e., that the target definitely exists. Then:
That is, the target’s composite undetectability qu is a composite of its “raw undetectability” s[1 − pD] and the degree to which clutter density impairs its detectability. It varies between qu = 0 when κ = 0 and qu = 1 when κ = ∞. That is, the composite undetectability of a definitely-existing target is 0 if there is no clutter; and its composite detectability 1 − qu is 0 if the clutter density is infinite (and, thus, the signal-to-noise ratio is extremely small).