practical fusion of quantized measurements via particle filtering
TRANSCRIPT
Practical Fusion of Quantized Measurements via
Particle Filtering 1
Yanhua Ruan, Peter Willett
ECE Department
University of Connecticut
Storrs, CT 06269
Alan Marrs
Qinetiq Ltd.
Malvern Technology Centre
Great Malvern, UK WR14 3PS
Francesco Palmieri
Dipartimento di Ingegneria
dell’Informazione, Seconda Universita degli Studi di Napoli
Casa Reale dell’Annunziata, via Roma 29
I-81031 Aversa (CE), Italy
Stefano Marano
DIIIE
Universita degli Studi di Salerno
via Ponte don Melillo
I-84084, Fisciano (SA), Italy
Submitted: January 2003
Revision: September 2005
Second Revision: June 2006
1This work was supported by the UK Ministry of Defence Corporate Research Program. The authors
also wish to acknowledge Murat Efe, Orhan Arikan and Yaakov Bar-Shalom for sharing their useful ideas.
Abstract
Most treatments of decentralized estimation rely on some form of track fusion, in which local track
estimates and their associated covariances are communicated. This implies a great deal of commu-
nication; and it was recently proposed that by an intelligent quantization directly of measurements,
the communication needs could be considerably cut.
However, several issues were not discussed. The first of these is that estimation with quan-
tized measurements requires an update with a non-Gaussian distribution, reflecting the uncertainty
within the quantization “bin”. In general this would be a difficult task for dynamic estimation,
but Markov-Chain Monte-Carlo (MCMC, and specifically here particle filtering) techniques appear
quite appropriate since the resulting system is, in essence, a nonlinear filter. The second issue is that
in a realistic sensor network it is to be expected that measurements should arrive out-of-sequence.
Again, a particle filter is appropriate, since the recent literature has reported particle filter mod-
ifications that accommodate nonlinear-filter updates based on new past measurements, with the
need to re-filter obviated. We show results that indicate a compander/particle-filter combination
is a natural fit, and specifically that quite good performance is achievable with only 2-3 bits per
dimension per observation.
The third issue is that intelligent quantization requires that both sensor and fuser share an
understanding of the quantization rule. In dynamic estimation this is a problem since both quantizer
and fuser are working with only partial information; if measurements arrive out of sequence the
problem is compounded. We therefore suggest architectures, and comment on their suitabilities for
the task — a scheme based on delta-modulation appears to be promising.
1 Introduction
1.1 Measurement Fusion
It would appear that, for reasons of price, of survivability, and of performance, future generations
of military (and civilian) surveillance systems will comprise many heterogeneous sensors. At its
simplest level one might envision a pair or small suite of sensors each checking one another’s reports
and thereby avoiding the masking of detections that is inevitable with a fluctuating target; and
indeed, even with co-located and homogeneous sensors, there is some benefit to splitting one’s
remote sensing resources in such a way [22]. But a more adventurous interpretation of multi-sensor
surveillance involves many very cheap and not co-located sensors, perhaps of the DADS (deployable
autonomous distributed system) sort discussed in [11, 15]: there are obvious advantages in terms
of robustness, and creation of a decentralized “field” of sensors is a nice way to overcome the r−4
power-return law.
Along with the vision of a smart “web” of sensors come a number of issues, however: there is that
of deployment and layout [15], control [19], and, of greatest interest here, application to estimation.
Although distributed detection has been studied reasonably extensively (e.g., [6, 18, 20]), in most
applications the focus is, and should be, on the acquiring and tracking of threats using a distributed
array of sensors. Issues of concern here include the method for transport of information between
sensors, and the eventual fusion of information from disparate sources.
Ideally, there would be no problem at all in either respect: communication would be complete
and perfect, and there would likewise be no need for specialized estimation algorithms, since all
information from all sensors would arrive promptly. A fused estimate would therefore be no different
from that of a normal “centralized” tracking system, just with a richer set of observations on which
it would be based. Realistically, however, communication is over bandlimited channels.
What should be communicated? Blair in [4] gives a helpful taxonomy of several different fusion
approaches:
• Reporting Responsibility. Each target is assigned to a particular sensor, which makes allthe measurements and maintains the track. Presumably this sensor is that having the best
measurements (i.e., probably, is nearest), and otherwise is not so heavily loaded that such a
task is beyond its means; responsibility may be handed off to another sensor when appropriate.
Reporting responsibility is simple. But it is not really data fusion at all, saving that the
responsible sensors’ estimates can be shared; and it has very few of the above advantages of
fusion.
• Centralized Composite Tracking. Here each sensor’s observations – more likely, associated
measurement reports (AMRs) – are shared with a centralized “fusion center” tracker. The fu-
sion center digests these and broadcasts its current state estimates (composite tracks), prepara-
tory to association by the sensors’ next scan. This is ideal data fusion; its disadvantage, and
perhaps unreality, is its reliance on an excellent communications backbone.
• Distributed Track Fusion. Sensors each maintain all tracks, and information about tracks (i.e.,state estimates and covariances) is shared. This is, presumably, the lightest-loaded system as
1
regards communication. Its main disadvantage is that the state estimates thereby maintained
are necessarily correlated with each other (although it is possible to reduce this via the tracklet
approach 1[8]), and it is computationally complicated to remove this even in the idealized case
of a distributed Kalman filter.
• Distributed Composite Tracking. In this case each sensor maintains its own track of all targets,but, unlike Distributed Track Fusion, AMRs are shared.
Measurements 1 Measurements 2
Tracker 1 Tracker 2
False alarm
Platform 1 Platform 2
True detection
Measurements 1 Measurements 2
Tracker 1
Tracker 2
False alarm
Platform 1 Platform 2
True detection
Figure 1: Illustration of two fusion philosophies. Left: Distributed Track Fusion (3rd bullet); Right:
Distributed Composite Tracking (4th bullet).
The idea of the final two fusion schemes are illustrated in Figure 1. Most research seems to have
been on the penultimate scheme, despite its disadvantages (e.g. [12]). One suspects that this may
be a “cultural” bias, in that within a community whose specialty is tracking, it is not unexpected
that the item to be communicated is most comfortably a track. And, indeed, it would at first appear
that the sharing of track estimates is logical since they comprise (ideally, in a linear/Gaussian world)
a sufficient statistic of measurement history – the complexities of their combination are not too
onerous a price for miserly communication.
However, it must be recalled that in a dynamic estimation situation with measurements of
uncertain provenance, covariances as well as state estimates must be shared. Thus, in a system
estimating an N -dimensional state, there are N +N(N + 1)/2 numbers to be communicated each
time: for example, with position, velocity and acceleration being tracked in three dimensions, this
means 54 separate numbers require transmission whenever data is fused2. Further, although it may
not be necessary to transmit each as a 32-bit number, due to the processing it is not clear how
relevant each number (for example, the (2,4) element of the covariance matrix) is to the eventual
fused estimation accuracy, and hence the best strategy to use for quantization is murky.
This appears to have been noticed in [5] by Blair, who compares the communication load of
Distributed Track Fusion (3rd bullet) to that of Distributed Composite Tracking (4th bullet) –
1There are limitations for the tracklet approach in track initialization and tracking maneuvering targets.2In some track fusion schemes, communication load can be lessened in that local track information does not need
to be shared for all scans; an information filter can be used if the state dimension is much larger than the measurement
dimension; information other than track estimate can be shared in non-standard distributed fusion.
2
the scheme in which associated measurements are directly shared. He finds that they are, rather
surprisingly, comparable when each represents its numbers with the same high (32 bit) precision.
Aside from a number of studies of decentralized detection, comparatively little research has
focused on the architecture in which measurements are shared directly. However, some recent work
[16] has found that the data communication requirements of the two schemes could be similar
and that the latter could achieve further bandwidth improvement through the use of an intelligent
quantizer. Therefore, the measurement-sharing schemes may be preferable, since data fusion and
tracking is in principle straightforward. Further, it is intuitively appealing, since measurements
which might appear locally at an individual sensor to be spurious, can, if several sensors report a
measurement near the same “surprising” location, be accorded credibility.
1.2 Issues and Plan of Paper
In the preceding section we have tried to make the case that direct sharing of measurements may
be preferable to the sharing of local estimates: in the latter case fusion is at first blush “easy”
since it amounts to estimation of a Gaussian process with Gaussian measurement errors, but in
fact the correlation due to process noise roils the matter when out-of-sequence measurements are
incorporated – this putative advantage may become moot in the case of a nonlinear model.
There is a tendency to think of measurement fusion as a “bandwidth hog”, since a great many
observations need to be shared. We disagree, and counter that although there are many measure-
ments, the direct coupling of each measurement’s location to its information content leads naturally
to schemes by which it can be communicated cheaply via an appropriate quantization. Specifically,
we intend to show that even with transmission of only a few bits per dimension per measurement,
quite good performance is achievable.
Now, even beyond standard multi-sensor concerns such as gridlock and registration, there are
most certainly issues:
• sensors with limited energy and communication capability – paradigms for appropriate quan-
tization and (lossy) data sharing must be explored
• data fusion – the admission that quantized data is being fused calls into question the use of
fusion schemes based on a Gaussian assumption
• architecture – in a network of sensors one is left more free to decide on the structure of
information communication (star, ring, tree, etc.)
• measurements arriving out-of-sequence – it is inevitable, given the communication delays
within even a moderate-scale fusion network, that occasionally a measurement will arrive to
be fused after another measurement with a later time stamp.
There has been little treatment in the literature of practical low-bandwidth schemes for estimation.
Much is known about moderate-bandwidth track fusion, and about low-bandwidth decentralized
detection. There are a number of clever results on low-bitrate decentralized estimation – well
represented in [10], for instance – that promise to be helpful. However, most results on distributed
3
estimation are at least one of: asymptotic, static or Gaussian. In tracking the situation is non-
Gaussian due to measurement-origin uncertainty, it is dynamic (since interesting targets tend to
move), and the number of sensors/observations is very finite.
We shall try to visit each of the four bulletized concerns above, as follows:
• Section 2. The first concern in a nonuniform quantization scheme is the quantizer itself,
and here we propose to use the robust compander idea from communications. We explore
its performance analytically: when the measurement noise is small (compared to the process
noise and to the uncertainty from quantization) its benefits can be substantial.
• Section 3. The analytical efforts from the previous section are not based on a particular fusionscheme. However, quantized measurements yield measurement noise that is non-Gaussian, and
a practical means of fusion is required. Particle filters offer a natural means to estimate in
non-Gaussian noise, and we explore this. We further note that in a practical sensor network
measurements arrive out-of-sequence; fortunately, recent research has shown that particle
filters lend themselves to this problem as well.
• Section 4. A nonuniform quantizer implies a subtlety: both quantizer and fusion center mustagree on where the quantizer is centered. Note that this is not (necessarily) the registration
problem, and ideally does not arise if the quantization is uniform. We suggest several archi-
tectures, of which our favorite is that based on the communications concept of ∆-modulation
[17].
• Section 5. The compander/particle-filter/∆-modulator idea is investigated by simulation.As was found analytically for the compander alone in Section 2, adequate performance is
achievable with only a few bits per dimension per sensor.
We do not pretend to give comprehensive results on a low-bandwidth measurement-fusion tracker
in this paper. But by our analytic treatment of quantization under measurement uncertainty and
our suggestion for an integrated scheme we hope to stimulate interest in this practical viewpoint.
2 Benefits of Non-Uniform Quantization
2.1 Analysis of the Companding Quantizer
We shall adopt the notation that the random variable to be estimated is X, that the observations
of X are {Yi}, and that the quantized versions of these are {Zi}, where the subscript i denotes theindex of the measurement. Specifically, we have that
fX(x) =1√2πe−x
2/2 (1)
for the pdf of the target variable — note that we shall assume this to be unit normal without loss
of generality, since any other Gaussian situation can be modeled via translation and scaling. We
further have that
Yi = X + wi (2)
4
in which {wi} are iid Gaussian with mean zero and variances σ2w. We further have, assuming thatthere is no measurement-origin uncertainty,
Zi = Q [Yi] = Yi + vi (3)
where Q[·] denotes a quantization operation of the observation and vi denotes the quantization error(or noise). We anticipate by assuming the quantization is sufficiently fine that an additive noise
models its effect adequately.
The approach taken here is as follows.
1. All work will be in one dimension. It is not expected that extension to several dimensions
(the more realistic case, and that explored in the previous section) will present qualitative
differences, but we leave that for further research.
2. The quantizers will function as companders [17]. That is, with reference to (3), we have
Zi = g−1 (Qu [g(Yi)]) (4)
in which Qu denotes a uniform quantizer with resolution δ.
3. As indicated previously, the quantization is assumed to be of reasonable fidelity that the quan-
tization operation can be modeled as adding independent noise, and that standard companding
analysis techniques can be employed.
4. There must be a “gate”: measurements within this gate are quantized, coded, and transmitted,
and those outside are ignored. We take this gate as −A ≤ Y ≤ A, with a typical3 value A = 6.
5. The criterion to be optimized (minimized) is the Cramer-Rao lower bound (CRLB) [21],
reflecting the eventual accuracy of the estimation of X.
Our goal is to see whether there is some benefit to a non-uniform quantizer (versus a uniform A/D)
in estimation of a random quantity laboring under missed detections and false alarms. Ideally we
would have an explicit performance metric for an optimal quantizer and associated fusion scheme.
It may be possible to specify the optimal quantization, but experience suggests that its per-
formance gains versus a merely “good” scheme are evanescent. Borrowing from communication
practice, we adopt as a good, easy and robust approach: the logarithmic compander. The compan-
der has the right shape, in that its levels are finest where the measurement is most probable and
informative, but similar to the communication application there is no implication of optimality. In
the following we simplify by assuming that whatever the optimal fusion scheme may be, its perfor-
mance is close to the CRLB. We find that the performance gains from an intelligent non-uniform
quantization scheme can be impressive.
The idea of a companding quantizer is illustrated in Figure 2. With reference to that figure, we
note that
g(y +∆y) − g(y) = δ (5)
3The necessity of this gate is in our opinion a weakness of our work, in that one would prefer an algorithm that
gated automatically.
5
in which y is the variable to be quantized, g denotes a nonlinear function, δ and∆y are the resolution
of the uniform quantizer and the quantization resolution in y domain, respectively. Using standard
companding analysis, we get
∆y ≈ δ
g(y)(6)
where g denotes differentiation. The error arising from quantization has a uniform distribution with
variance ∆2y/12. Thus, we have
Var(Zi|X) = σ2w +∆2Yi12
(7)
We approximate this as
Var(Zi | X) ≈ s ≡ σ2w +∆2X12
= σ2w +δ2
12g(x)2(8)
We shall note that the error {Zi −X} is generally non-Gaussian. However, to derive the analyticexpression, we assume in the next section that {Zi −X} is Gaussian and independent.
Figure 2: Above: illustration of the companding idea. An observation is nonlinearly transformed and
then uniformly quantized; following this the nonlinearity is undone. Below: a typical “compressive”
nonlinearity, with quantization resolution in the observation domain (i.e. ∆) defined.
The measure we shall adopt to represent the quality of a quantizer is the CRLB. Based on (1,2,3)
and (8), plus our free assumptions of Gaussianity, we get
CRLB = EX½s2
2s2+1
s+ 1
¾−1(9)
where
s = σ2w +δ2
12g(x)2
s = −δ2g(x)
6g(x)3. (10)
Detail of the derivation is given in [16].
6
00.5
11.5
22.5
3
2
3
4
5
60
5
10
15
20
25
σw
for measurement# bits per measurement
CR
LB
no
rma
lize
d t
o u
nq
ua
ntize
d (
dB
)
2
3
4
5
6
0
2
4
6
8
100
5
10
15
# bits per measurement
A = 6, Pr(valid)= 0.1, σw
= 0.05
log2(µ) for µ−law compander
CR
LB
ga
in r
ela
tive
to
un
iform
A/D
(d
B)
Figure 3: Left: Loss in terms of CRLB for use of uniform quantization; Right: Gain from µ-law
companding, σw = 0.05 and πv = 10%.
2
3
4
5
6
0
2
4
6
8
100
1
2
3
4
5
6
7
# bits per measurement
A = 6, Pr(valid)= 0.1, σw
= 0.25
log2(µ) for µ−law compander
CR
LB
ga
in r
ela
tive
to
un
iform
A/D
(d
B)
2
3
4
5
6
0
2
4
6
8
100
0.1
0.2
0.3
0.4
0.5
0.6
0.7
# bits per measurement
A = 6, Pr(valid)= 0.1, σw
= 1
log2(µ) for µ−law compander
CR
LB
ga
in r
ela
tive
to
un
ifo
rm A
/D (
dB
)
Figure 4: Gain from µ-law companding. Left: σw = 0.25 and πv = 10%; Right: σw = 1 and
πv = 10%.
7
2.2 Results
Before examining any specific companders, let us ask where there may be improvement to be had
at all. In Figure 3 (Left) we show the quantization loss in terms of the CRLB (9) as a function of
the number of bits assigned per measurement and in terms of the relative measurement standard
deviation (recall that the prior standard deviation is unity). The quantization loss is defined as
the ratio of the CRLB with a uniform quantizer to the CRLB without quantization; if there is any
benefit to be derived from a non-uniform quantizer it must be less than this, since an unquantized
system provides an upper-bound on performance. It is readily seen that the “interesting” cases are
those in which the number of bits is small (else the quantization noise is negligible) and/or when
the measurement noise standard deviation σw is small — for a large σw any additional error accruing
from quantization is a minor concern4.
We first consider the popular µ-law companding structure. In this case we have for z ≥ 0
g(z) =A ln (1 + µz/A)
ln (1 + µ)(11)
with g(−z) = −g(z). Results are plotted in Figures 3 (Right) and 4 for various values of σw anda fixed πv, where πv denotes the a-priori probability that a given measurement is target-generated.
Note that if the measurement does not arise from a target, then it is considered a false alarm
and afforded a uniform distribution over the “gate”. As expected from Figure 3 (Left), gains are
interesting only for few bits and small σw; but these gains can be quite substantial. Note that a
reasonably large value of µ is most appropriate; in many companding systems the value µ = 255 is
used.
3 Tracking via the Particle Filter
The non-Gaussian error introduced by the measurements’ quantization presents a difficult task for
dynamic estimation. However, the particle-filtering techniques appear quite appropriate for dealing
with this kind of problem. In the sequel we present the basic idea of particle filtering and address
some important issues that come up in implementation.
3.1 Particle Filter
The particle filtering techniques, also known as the sequential or Markov chain Monte Carlo (MCMC)
methods, have drawn an increasing level of interest in recent years. The essence of the method is
that the pdfs of interest are represented by a set of random numbers, namely the “particles” –
essentially any inference about the state, including of course its mean value and the correspond-
ing covariance, are fully derivable given the pdfs. The particles can be propagated according to
the system’s dynamic equation and updated to form a new set of posterior particles using newly
arrived observations, and this process can continue. Particle filtering methods are well suited to
non-Gaussian/non-linear estimation problems in which results from the traditional approaches such
as the extended Kalman filter (EKF) are unsatisfactory. In the sequel, we present our view of the
method.4In the former case (3) is of concern.
8
Consider a general nonlinear filtering problem. It is desirable to find the pdf f (xn|Zn1 ), whereZn1 ≡ {z1, z2, . . . , zn}. In the linear/Gaussian case this is “easy”, since the sufficient statistic for¡xn−1|Zn−11
¢is the mean xn−1|n−1 and covariance Pn−1|n−1 from the Kalman filter. But in general,
this is not so. We have
f(xn|Zn1 ) =f (zn|xn)f¡zn|Zn−11
¢ Z f(xn|xn−1)f(xn−1|Zn−11 )dxn−1 (12)
where f (zn|xn) is the observation pdf ; f¡zn|Zn−11
¢does not depend on xn; f(xn|xn−1) is the state
transition pdf and f(xn−1|Zn−11 ) is available from previous iteration.
The idea of the particle filter begins with a set of point-mass estimates of the posterior state
pdf :
f(xn−1|Zn−11 ) =1
N
NXi=1
δ(xn−1 − xn−1[i]) (13)
where N denotes the number of particles. Our goal is to generate a set of particles {xn[i]}Ni=1 thatrepresents the pdf of f(xn|Zn1 ) from the current particles {xn−1[i]}Ni=1. The Metropolis-Hastingsalgorithm provides us a tool to generate random samples which follow a specific pdf from a Markov
chain. We apply a Metropolis-Hastings special case known as the “independence sampler” to the
generation of xn: Let the transition pdf be
q(xn) = f(xn|Zn−11 ) (14)
which can be approximated by choosing j ∼ Uniform(1,N) and generating xn from f(xn|xn−1 =xn−1[j]). The true pdf is
π(xn) = f(xn|Zn1 ) ∝ f(xn|Zn−11 )f(zn|xn) (15)
and fortunately we do not need to deal with this equation directly to obtain the samples of π(xn).
Instead, we apply the Metropolis-Hasting algorithm and the rejection test is defined by:
α(x,xn) = min
½1,π(xn)q(x)
π(x)q(xn)
¾= min
½1,q(x)
π(x)κf(zn|xn)
¾(16)
Since xn does not depend on x, we can choose x however we want, so let
Pr(Reject) = 1− κZxn
f(zn|xn)dxn (17)
and choose κ to keep exactly N particles after this new scan n. In fact, we use a slightly modified
implementation:
1. For j = 1, 2, . . . , N , propagate xn−1[j] to xn[j] using f(xn[j] | xn−1[j]).
2. For j = 1, 2, . . . , N , calculate
wn[j] = κf(zn|xn[j]) (18)
where κ ensures that wn’s add to unity.
9
3. Draw from xn[j]’s N times (with replacement) according to probabilities wn[j].
This is, of course, the importance sampling/resampling or “bootstrap” filter of Gordon, Salmond
and Smith [9].
The insertion of the quantized measurement to the particle filter is straightforward. All remains
intact, except that we simply update the current particles by the quantized measurements and take
into account the extra error introduced by the quantization given in (7). This latter error is of
course non-Gaussian but is easily inserted to (18).
3.2 Out of Sequence Measurements
Out of sequence measurements (OOSMs) must be accounted for in a multisensor tracking system
where measurements from multiple sensors are sent to a processing center. Due to (random) delays
in ad-hoc network transmission, or indeed due to polling delays in more deterministic communication
networks, an earlier measurement relating to a specific target may arrive at the central processor
after a later one. In other words, the central processor receives a measurement from a particular
target with time tag tk, and it updates its state estimate using this. Later, another observation
with time tag tj (j < k) from the same target arrives at the processing center. The question is how
to incorporate the measurement taken at an earlier time in a track that has already been updated
with a later observation.
There are several potential approaches that traditional tracking algorithms use to address this
problem [3]. Among these, probably the simplest is simply to discard the delayed measurement
and hope that the lost information is not substantial. Other approaches involve using the stan-
dard Kalman filter and make a backward prediction to the time of delayed measurement. These
approaches are suboptimal as they neglect the process noise in retrodiction. In [1, 2] an optimal
OOSM solution, which took process noise into account, was developed and compared with two of the
suboptimal algorithms on some realistic cases. More recently, interest has been drawn to a general
approach to cope with OOSMs on nonlinear and non-Gaussian models – including, of course, a
model with data association. One idea is to absorb the OOSMs in the particle filter framework [13],
and we now give a brief overview of that.
Suppose at time tn we have an ordered set of in-sequence measurements Zn1 . The next measure-
ment zκ arrives and is found to be out-of-sequence, such that τ < tn (τ is the time with which the
OOSM zκ was tagged). It is further assumed that
tb < τ < ta (19)
where b and a are the contiguous time indices (tags) respectively before and after the OOSM in the
ordered sequence Zn1 . Then the updated posterior is
p∗(Xn0 |Zn1 ) = p(Xb
0|Zb1)p(zκ|xκ)p(xκ|xb)p(xa|xκ)
p(xa|xb)p(zκ) p(Xna |Zna) (20)
which can be simplified to
p∗(Xn0 |Zn1 ) = p(Xn
0 |Zn1 )p(zκ|xκ)p(xκ|xb)p(xa|xκ)
p(xa|xb)p(zκ) (21)
10
Then the particle weight (18) can be updated by
w∗n[i] = wn[i]p(zκ|xκ[i])p(xκ[i]|xa[i],xb[i])
π(xκ[i]|Xn0 [i],Z
n1 )
(22)
In [14], tracking performance was investigated in a simulated wireless network for a wide range of
communication delays. Results show that position estimation accuracy close to the lower bound
should be possible for communication intervals up to four times the average inter-scan time. How-
ever, the authors did not consider the communication bandwidth problem: they assumed that
measurements were shared without loss of quality due to communication.
4 Architecture of the Sensor Network
The measurements of all the sensors among the wireless sensor network are shared directly with
the fusion center, and communication issues must be considered. A practical protocol known as
the Bluetooth architecture [7] has been developed. In that case a frequency hopping Time Division
Multiple Access (TDMA) approach is taken and the sensor nodes form piconets where one of the
nodes becomes the “master”. The master node, which plays the role of fusion center, broadcasts and
other nodes (sensors) receive during odd time slots and the master receives while the other nodes
broadcast during even time slots. Intra-piconet collisions are avoided by only permitting one node
to communicate with the master during each time slot. This leads to the concept of the master
polling a node, which then replies in the following time slice, which in turn implies that a node
must in general wait for several time slots before being able to communicate with its master again.
Piconets can be networked together by forming their master nodes into a net with one node being
nominated as the overall master.
Intelligent quantization requires that both sensor and fuser share an understanding of the quan-
tization rule, and in dynamic estimation this is a problem, particularly so if measurements arrive out
of sequence. There must be a “measurement gate” to implement quantization schemes: measure-
ments within this gate are quantized, coded, and transmitted, and those outside are ignored. We
take this gate as xn−A ≤ Y ≤ xn+A, and use the typical value A = 6σm and x the estimate of thetarget state. Aside from the gate, the quantization rule consists of quantization scheme selection,
which in the case of a compander means the nonlinearity. Now, in a dynamic system all these can
be pre-defined and known to all sensors in the network; all, that is, except for the center of the
gate – the location of the “middle” of the non-uniform quantizer.
This may seem at first trivial. However, since all information is processed at the fusion center
and only there is the repository of the global estimate of target state. Sensors need to be constantly
informed of this information – presumably by the fusion center – in order to implement the
quantization schemes. The fusion center could broadcast this information over its allocated time
slots; however, such transmission would consume bandwidth, and would hence appear incompatible
with the original intent to save on bandwidth. There are probably many approaches to deal with
this. Here we propose four: the first three do not require “feedback” broadcast from the fusion
center to the sensors, while the last one requires a low-bandwidth broadcast.
• In the first scheme (A), a fusion center broadcast is avoided by arranging each sensor itself totrack based upon the (local) measurements available to it, with the center of its gate obtained
11
from this local estimate. The fuser uses the global estimate as an approximate center of the
gate to decode the quantized measurements. This scheme would appear to work well if the
difference between the local and the global estimates are much less than the error introduced
by the process noise, measurement noise or quantization noise; that is, when the local track
is decent.
• The second scheme (B) is an overlay to the first. Since the fusion center has access to allmeasurements and is, presumably, powerful in computational capability, it can not only calcu-
late the global estimate of the targets, but also infer what each sensor knows about the target
state, i.e., the local estimate of each sensor. Thus, when decoding the quantized measurements
the fuser can use its estimate of each local estimate as the center of each local gate, instead
of its own global estimate.
• A third approach (C) is suggested by the ∆-modulation scheme [17] from communication
theory: each sensor will use its local track estimate to update the center of its gate, and
will provide this information to the fusion center along with its measurement data. This
communication is most efficient when it is specified in terms of movement in integer numbers
of the local sensor’s own quantization grid, and the special case of single-bit transmission is
perhaps the most interesting.
• In the fourth idea (D) the ∆-modulation scheme of the previous is altered such that the fusioncenter commands the movement (gate re-alignment) of each sensor by a broadcast message.
We note that in schemes C & D there is no need to constrain grid-movement communication
to be only of a single bit. There is some comparison of schemes in Table 1. The last row of that
table, the expected performance, is subjective: we would expect that schemes A & B will suffer
when the fusion center’s understanding of the quantization grid deviates significantly from those of
the local sensors, and that scheme D is preferable to C when tracking is challenging due to the grid
movement’s being controlled by more-accurate fused-track information. Scheme C is perhaps the
“cleanest” of the group, since grid-movement information arrives at the fusion center along with all
data; however, an OOSM will perturb estimates, since already-arrived data must be re-examined in
light of a previously undeclared grid movement. In the following sections we pay particular attention
to scheme D, but we note that the other schemes are interesting and deserve further study.
Scheme: A B C D
FC → LS no no no yes
LS → FC no no yes no
local tracker yes yes yes no
performance poor ? fair good
Table 1: Comparison of four schemes to manage fusion center grid movement, as given in the text.
The first and second rows relate to the need for fusion center (FC) to local sensor (LS) and reverse-
path communication of grid-movement information. The third row shows whether there is a need
for a local tracker. The fourth row gives expected level of performance in challenging tracking
situations.
12
5 Simulation Results
The target motion is modeled as constant velocity using the standard kinematic equations
xn+1 = Fxn + vn
yn = h(xn, tn) +wn (23)
for n = 1, 2, . . . , N . Here, as usual, xn represents the state of the target at time tn, and yn its
corresponding observation; the matrix F and function h(·) are known, and are assumed to representan observable and controllable system. The random sequences {vn,wn} are assumed white, zero-mean, Gaussian, and mutually independent, with E{vnvTn} = Qn, E{wnwTn } = Rn.
Suppose the time interval between two consecutive sampling point is T , then for one dimension
F =
"1 T
0 1
#, Q = q
"T 3
3T 2
2T 2
2 T
#(24)
and for two dimensions
F =
1 T 0 0
0 1 0 0
0 0 1 T
0 0 0 1
, Q = q
T 3
3T 2
2 0 0T 2
2 T 0 0
0 0 T 3
3T 2
2
0 0 T 2
2 T
(25)
where q is the (continuous-time) process noise power spectral density. The false alarms, if present,
are assumed uniformly distributed in the surveillance region V and their number follows a Poisson
distribution with mean λV , where λ denotes the spatial density.
We test the algorithms in two different types of scenarios:
1. one dimensional motion, linear measurement model and two asynchronous (random) sensors,
with and without false alarms.
2. two dimensional motion, bearing only measurement model and in the wireless sensor network
of [14], without false alarms.
In the following we present the simulation results obtained from these scenarios and take the mean
squared error (MSE) of the estimate as the measure of the tracking performance.
5.1 Linear Motion and Linear Measurement
We first test the algorithm in the simplest case — one target moves in a one dimensional Cartesian
coordinate, and its position is observed directly. We can write the measurement equation as
yn = Hxn +wn (26)
and
H = [ 1 0 ] , R = σ2m (27)
Two sensors are at different locations and are arranged to take scans alternately. Sensor 1 is
designated as a local sensor (relative to the fusion center) and sensor 2 is a remote sensor. We
13
assume that every measurement from sensor 2 is sent to the fusion center and the transmission is
subject to a random delay t, where t follows an exponential distribution with mean τ .
We attempt to achieve bandwidth improvement via intelligent quantization schemes and dealing
with the OOSMs. The parameters are:
• process noise power spectral density q = 0.01 m2/s3;
• measurement noise standard deviation σm = 30 m;
• sampling time interval T = 30 s;
• initial target state x0 = [0 5]T (m m/s);
• probability of detection pD = 1.0.
• clutter (false alarms) density λ = 0 or λ = 10−3 1/m.
In Figure 5 we compare the particle filter with the Kalman filter when the quantized measure-
ments are used. It can be seen that the particle filter outperforms the Kalman filter significantly
when the quantization bits are few. This is readily explained since the small number of quanti-
zation bits introduces large quantizational non-Gaussian errors into the measurement uncertainty
and the Kalman filter is designed only to handle Gaussian noise. When the non-Gaussian measure-
ments noise components are not ignorable, the performance deterioration of the Kalman filter is
inevitable. On the contrary, such a problem does not exist in the particle filter due to its capability
to incorporate the non-Gaussianity and nonlinearity of the model.
1 2 3 4 5 6 7 8200
250
300
350
400
450
500
550
600
MS
E
number of the quantization bits
PF w/ µ-law quantized measurementsKF w/ µ-law quantized measurementsKF w/ unquantized measurements
Figure 5: Comparison of the MSEs of the particle filter and the Kalman filter using µ-law quantizerin the false alarms free scenarios. (Pd = 100%, q = 0.01m
2/s3, σm = 30 m , T = 30 s.)
Simulation results are shown in Figures 6 (Left) and 7 (Left) for clutter-free scenarios and in
Figures 6 (Right) and 7 (Right) for scenarios with clutter. Figures 6 display the results of the µ-law
compander versus the uniform quantizer. We can see a similar pattern exhibited in both figures:
the µ-law compander outperforms the uniform quantizer in ranges from 2 bits to 6 bits, and the
improvement is most significant (around 10% difference) at 3 bits; the two schemes yield almost the
same MSEs at 1 bit and the ranges over 7 bits. To understand the figures, note the two quantizers
14
are identical for 1-bit quantization, and when bit-rates are high enough, the quantizations are very
fine and the errors introduced by them are negligible compared to the measurement uncertainty.
We also display the results for the unquantized (ideal) case, which means perfect transmission and
infinite bandwidth, with estimation via the Kalman filter (in Figure 6 (Left)) and the PDAF (in
Figure 6 (Right)), as the performance references. In Figure 6 (Left) we can see when the number
of bits increases, the performance of both quantizers approaches that of the Kalman filter.
1 2 3 4 5 6 7 8190
200
210
220
230
240
250
260
MS
E
number of the quantization bits
PF w/ uniform quantizerPF w/ µ-Law quantizerKalman filter w/o quantization
1 2 3 4 5 6 7 8220
240
260
280
300
320
340
MS
E
number of the quantization bits
PF w/ uniform quantizerPF w/ µ-Law quantizerPDAF w/ µ-Law quantizationPDAF w/o quantization
Figure 6: Comparison of the MSEs of the particle filters using µ-law and uniform quantizers,
(Pd = 100%, q = 0.01m2/s3, σm = 30 m , T = 30 s). Left: In the false alarms free scenarios. Right:
In the presence of false alarms with spatial density λ = 10−3 m−1.
We are also interested in how much improvement is offered through processing the OOSMs as
opposed simply to ignoring them; the result is shown in Figures 7. As upper and lower baselines,
we show in Figure 7 (Left) respectively the results of the Kalman filter that processes only local
measurements, and of the clairvoyant KF that has access to all measurements taken in sequence. In
Figure 7 (Right), the process is repeated except that we use the PDAF to replace the Kalman filter,
since false alarms are present; there is a similar pattern. In both cases the improvement of MSE
ranges from 5% when the average delay is 0.5T0 to around 15% when the average delay increases
beyond T0 – T0 denotes the time interval between two closest sensor scans. This indicates that we
can improve our estimation accuracy significantly by processing OOSMs. When τ gets larger, the
difference between with and without processing OOSMs diminishes, since both curves will converge
to the upper bound, that being the local-sensor-only case. Note that in the figures the MSEs of the
particle filter that processes OOSMs increase (in a relatively slow manner) with the increase of τ .
This is due to the following two factors:
1. The first and the dominant one, when τ increases, more measurements become out-of-sequence
and therefore fewer measurements are available to the current estimate. To amplify, suppose
the measurement yn reflects the target state at time tn but arrives at the processor at a later
time tn+k. Consequently yn fails to provide information to the estimate xn, although it can
be used by the estimate xn+k via OOSM processing. In other words, after all, OOSMs lead
to information loss, and any OOSM algorithm can only alleviate this.
2. Secondly, we set the maximum lag (8 seconds here) when dealing with OOSMs – the mea-
15
surements arriving with a time lag greater than the maximum lag will not be processed. Thus
when τ becomes larger than this, more measurements are likely to arrive “too late”, and will
be discarded and we lose more information.
With false alarms present, an interesting phenomenon is seen in Figures 6 (Right) and 7 (Right):
the particle filter outperforms the PDAF in the ranges of large number of bits or small τ . Presumably
this is due to the PDAF being a suboptimal algorithm, while the particle filter is (asymptotically)
optimal. In contrast, from Figures 6 (Left) and 7 (Left), we see that the particle filter always
converges in performance to the (optimal) Kalman filter (the small gap between them shown in
Figure 7 (Left) is due to the 6-bit quantizer used in the particle filter.)
0 0.5 1 1.5 2
210
220
230
240
250
260
270
280
290
6 bits quantizer
MS
E
average delay τ0 (nomalized by T0)
KF1 with local meas. inputPF w/ OOSM rejectedPF w/ OOSM processedPF w/ OOSM prcsd and ∆-modulationKF2 with all meas. input
0 0.5 1 1.5 2240
260
280
300
320
340
360
3808 bits quantizer
MS
E
average delay τ0 (nomalized by T0)
PDAF1 with local meas. inputPF w/ OOSM rejectedPF w/ OOSM processedPDAF2 with all meas. inputPF w/ OOSM prcsd and ∆-Modul.
Figure 7: Comparison of the MSEs of the particle filter with OOSM processed versus the particle
filter with OOSM discarded, (Pd = 100%, q = 0.01m2/s3, σm = 30 m , T = 30 s). Left: In the false
alarms free scenarios. Right: In the presence of false alarms, (λ = 10−3 m−1 and ∆-modulation isalso included).
In the above simulations we assume that the quantization rule is shared perfectly by the two
sensors. This is of course not the case in practice. We adopt ∆-modulation as the way of sharing
the estimate (scheme D), and the results are shown in Figures 7; they are also compared to the case
of perfect sharing of estimate (i.e., both sensors have the same understanding of the quantization
center). It can be seen the performance degrades by the surprisingly small factor of 4% through the
use of 1-bit-per-coordinate communication. This factor is scenario dependent and can presumably
be reduced even further by increasing the bandwidth usage; but it is encouraging for scheme D.
5.2 A Wireless Sensor Network
The plant equation in (23) is defined in Cartesian coordinates by xn ≡ [ px(tn) vx(tn) py(tn) vy(tn)],with observations yin ≡ [ θ(tn) ] in bearing for the ith sensor given in polar coordinates. The mea-surement model is given by
yin = arctan
Ãpy(tn)− siypx(tn)− six
!+wn (28)
where [siy, siy] denote the position of the i
th sensor in Cartesian coordinates.
16
We consider a small sensor network given in [14]. It comprises 11 sensors and a tree architecture
which essentially is a scatternet formed by three piconets. The master node for the scatternet
represents the “command” node where data fusion and tracking will be performed. The sensors are
modeled as simple bearings-only sensor types, representative of an acoustic array. The fusion center
obtains measurements from other sensors according to the Bluetooth polling protocol. A schematic
overview of the architecture is shown in Figure 8.
Fusion Center
Sensor
Figure 8: The architecture of the sensor network.
The observation rate for each of the sensor nodes in the scenario was fixed at one per second, and
the scenario was re-run for a variety of communication rates to investigate their effects on tracking
performance. The scenarios were also run for a variety of quantization rates. The master shares its
knowledge of target state estimate by 1 bit ∆-modulation transmission (i.e., scheme D).
The parameters of the scenario are:
• process noise power spectral density q = 0.1 m2/s3;
• measurement noise standard deviation σm = 1◦;
• sampling time interval T = 1 s;
• initial target state x0 = [88.3 − 4.0 766.6 − 4.0]T (m m/s m m/s);
• probability of detection pD = 1.0.
To minimize power consumption, we attempt to extend the time intervals between transfer of
data packets between nodes and to choose an appropriate intelligent quantization scheme. At the
same time, in order to maintain tracking performance, we need to process OOSMs. Whichever
quantizer we choose, the master node needs to share the updated target state estimate with all
sensors since this knowledge is vital to the quantization task. In order to inform all the sensors
of this information, we opt for the ∆-modulation idea. The fuser broadcasts 1 bit per coordinate
message to all sensors during odd time slots. Upon receiving this message, the sensors update their
current estimate of the target state by adding/subtracting a steplength and take it as the “window”
center for the quantization.
Simulation results are shown in Figures 9. In the left figure, we compare the uniform quantizer
with the µ-law compander for different quantization rates. The time intervals between transfer of
17
data packets is set to 1 second and the OOSMs are processed by the two schemes. Both quantizers
adopt the ∆-modulation scheme which needs 1 bit per coordinate bandwidth for communication:
the uniform quantizer requires grid-movement commands too, since to be efficient it must spread its
levels within a gate. We can see that the µ-law compander offers 20% and 6% MSE improvements
versus the uniform quantizer for 2 and 3 bits respectively; the difference becomes negligible when the
number of bits exceeds 3. In the right figure, we also show the MSEs of the particle filters with and
without processing of OOSMs in which µ-law compander is used and the number of quantization
bits is set to 6. It can be seen that the processing of the OOSMs improves the performance quite
significantly, and this is especially so in large communication-delay situations.
1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 625
30
35
40
45
50
55
MS
E
number of the quantization bits
PF w/ uniform quantizerPF w/ µ-Law quantizer
1 1.5 2 2.5 3 3.5 420
40
60
80
100
120
140
1606 bits µ-law quantizer
MS
E
Communication interval T
PF w/ OOSM rejected and ∆-Modul.PF w/ OOSM processed and ∆-Modul.
Figure 9: Left: Comparison of the MSEs of particle filters using µ-law and uniform quantizer.
(Pd = 100%, q = 0.1m2/s3, σm = 1
◦, T = 1 s). Right: Comparison of the MSEs of particle filterswith OOSM processed versus particle filter with OOSM discarded. (Pd = 100%, q = 0.1m2/s3,
σm = 1◦).
6 Summary
Here we consider the multi-sensor tracking problem with energy and/or bandwidth constraints –
both amount to a preference for reduced communication. We are also particularly interested in
large sensor networks of irregular (or ad-hoc) structure, since these offer attractive robustness and
economy. A concern with such networks, however, is the inevitability of out-of-sequence measure-
ments (OOSMs), due either to the randomness of information propagation or to the necessity for a
sensor-polling architecture.
For estimation fusion most researchers appear to favor track fusion architectures in which local
state estimates and their associated covariances are shared directly. As compared to measurement
fusion, track fusion seems appealing in its simplicity and parsimonious communication. However,
in track fusion, one has to address the cross-correlation due to process noise and prior information,
which is much harder than measurement-to-track association and fusion. Furthermore, OOSMs
are difficult to absorb, and it has recently been shown that the claim that track fusion has lower
communication needs may be open to debate.
18
In fact, we have indicated in the first part of this paper that via an intelligent quantization
scheme (such as a compander) the communication needs of measurement fusion can be cut still
further. Similarly, a recent paper by one of the authors found a neat procedure for the processing of
OOSMs based on particle filters – this scheme is applicable to the non-Gaussian measurement noise
accruing from the quantization operation, but also subsumes nonlinear models (such as bearings-
only measurements) and data association. Fusion of quantized measurement via particle filtering is
thus a natural marriage of the two ideas.
In the second part of this paper, we have discussed some practical problems of target tracking in
sensor networks and suggested solutions by making the particle filtering techniques to work with the
OOSMs and quantizers. We have implemented the particle filtering/OOSM/quantization algorithms
in two types of scenarios and the simulation results have shown that exploiting OOSMs in wireless
sensor network can effectively improve tracking performance. The improvement margin yielded by
OOSMs enables us to extend the communication intervals between sensors for the purpose of energy
saving while not compromising the overall performance. We have further shown that the intelligent
quantization of measurements is more effective than the uniform quantization in terms of bandwidth
efficiency.
The compander we used here is perhaps only “partially intelligent”. That is, we assumed a fixed
number of quantization bits as well as a fixed size (length, area or volume) of quantization gate.
The questions naturally arisen are:
• Should we make the quantizer data-adaptive?
• How much improvement in performance will that yield?
• When, presumably in terms of target maneuvering index, is the improvement most interesting?
• Is there a structure that is preferable to the one that we have adopted?
As regards the last, let us note that to use a non-uniform quantizer it is necessary that both quantizer
and fuser share a common understanding of the scheme: where should the quantization levels be
most dense and where most sparse? In this work we have adopted the ∆-modulation paradigm, in
which to save bandwidth the fusion center broadcasts changes in its state estimate using only one
bit per coordinate. This appears to work reasonably well, but other approaches are possible.
Here is a brief list of our findings (and musings):
• There is little loss from quite coarse quantization of measurements: only 2-3 bits per dimension.
• Particle filtering is an attractive way to allow for the non-Gaussian measurement noise accruingfrom quantization.
• Use of both non-uniform and uniform quantization requires a common understanding of a
quantization grid. This grid must be modified dynamically, and in a manner that is synchro-
nized between sensors and fusion center.
• There are a number of ways to achieve this synchronized movement. The one that we haveinvestigated, and which seems to hold most appeal, is that in which movements are according
to a ∆-modulation command from the fusion center.
19
• OOSMs ought not, in general, to be rejected. Again, the particle filtering methodology offersa simple means to incorporate them.
References
[1] Y. Bar-Shalom, “Update with Out-of-Sequence Measurements in Tracking: Exact Solution”,Proceedings of the 2000 SPIE Conference on Signal and Data Processing of Small Targets, April2000.
[2] Y. Bar-Shalom, M. Mallick, H. Chen and R. Washburn, “One-Step Solution for the GeneralOut-of-Sequence Measurement Problem in Tracking”, Proceedings of 2002 IEEE AerospaceConference, Big Sky MT, March 2002.
[3] S. Blackman and R. Popoli, Design and Analysis of Modern Tracking Systems, Artech House(Boston), 1999.
[4] W.D. Blair, “Practical Aspects of Multisensor Tracking”, in Multitarget-Multisensor Tracking:Applications and Advances (Volume III), Bar-Shalom & Blair eds., Artech-House, 2000.
[5] W.D. Blair, “Multiplatform/Multisensor Tracking System Architectures and Technical Issues”,presented at the Third ONR/GTRI Workshop on Target Tracking and Data Fusion, May 17,2000.
[6] R. Blum, “Necessary Conditions for Optimum Distributed Detectors Under the Neyman-Pearson Criterion”, IEEE Transactions on Information Theory, May 1996.
[7] J. Bray, C.F. Sturman, Bluetooth 1.1: Connect Without Cables, Prentice Hall.
[8] O. Drummond, “Track and Tracklet Fusion Using Data from Distributed Sensors”, Proceedingsof the Workshop on Estimation, Tracking and Data Fusion: A Tribute to Yaakov Bar-Shalom,Monterey CA, May 2001.
[9] N. Gordon, D. Salmond and A. Smith, “Novel Approach to Nonlinear/Non-Gaussian BayesianState Estimation”, IEE Proceedings F, Vol. 140, No. 2, pp. 107-113, 1993.
[10] T. Han and S. Amari, “Statistical Inference Under Multiterminal Data Compression”, IEEETransactions on Information Theory, pp. 2300-2324, October 1998.
[11] E. Jahn, M. Kaina and M. Hatch, “Fusion of Multisensor Information from an AutonomousUndersea Field of Sensors”, Proceedings of the International Conference on Information Fusion,Sunnyvale CA, July 1999.
[12] X. Li and K. Zhang, “Optimality and Efficiency of Optimal Distributed Fusion”, Proceedingsof the Workshop on Estimation, Tracking and Data Fusion: A Tribute to Yaakov Bar-Shalom,Monterey CA, May 2001.
[13] M. Orton and A. Marrs, “A Bayesian Approach to Multi-Target Tracking and Data Fusion withOut-of-Sequence Measurements”, IEE International Seminar on Target Tracking: Algorithmsand Applications, October 2001.
[14] M. L. Hernandez, A. D. Marrs, S. Maskell, M. R. Orton, “Tracking and Fusion for Wire-less Sensor Network”, Proceedings of the 5th International Conference on Information Fusion,Annapolis, MD, July 2002.
[15] M. Owen, D. Klamer and B. Dean, “Evolutionary Control of an Autonomous Field”, Proceed-ings of the International Conference on Information Fusion, Paris, July 2000.
[16] F. Palmieri, S. Marano, P. Willett, “Measurement Fusion for Target Tracking Under BandwidthConstraint”, Proceedings of the 2001 IEEE Aerospace Conference, Big Sky MT, March 2001.
[17] P. Peebles, Digital Communication Systems, Prentice-Hall NJ, 1987.
20
[18] A. Reibman and L.W. Nolte, “Optimal Detection and Performance of Distributed Sensor Sys-tems”, IEEE Transactions on Aerospace and Electronic Systems, January 1987.
[19] P. Shea, M. Owen, “Fuzzy Control in the Distributed Autonomous Deployable System”, Pro-ceedings of the SPIE Signal Processing, Data Fusion and Target Recognition, Orlando FL, April1999.
[20] J. Tsitsiklis, “Decentralized Detection”, in Advances in Statistical Signal Processing, vol. 2 —Signal Detection, Poor and Thomas, editors, JAI Press, 1990.
[21] H. Van Trees, Detection, Estimation and Modulation Theory, Part I, Wiley, 1968.
[22] P. Willett, M. Alford and V. Vannicola, “The Case for Like-Sensor Pre-Detection Fusion”,IEEE Transactions on Aerospace and Electronic Systems, October 1994.
21