Professional Documents
Culture Documents
SPIEDigitalLibrary.org/conference-proceedings-of-spie
A multiple target
target tracking
tracking(MTT)
(MTT)system
systemdesigned
designedfor
foruse
usebybyaa space
space-based programsuch
-based surveillance program suchasas the
the Strategic
Strategic
Defense
Defense Initiative
Initiative (SDI)
(SDI) will
willbe
berequired
required to
to handle a total number of targets and a target density density far beyond that of any any
previous
previous MTT system. One estimate
system. One estimatedefines
definesthe
thethreat
threattotoinitially
initiallyconsist
consist of
of3000
3000 missile boosters that
that ultimately
ultimately deploy
30,000
30,000warheads
warheads and
and 250,000
250,000decoys.1
decoys.' In addition, the
the tracking
tracking system
system will have
have to
to contend
contend with other objects, such
other objects, such as
as stars,
stars,
resident space objects, andand debris,
debris, as
as well
well as clutter for lookdown conditions. Finally, further complications
conditions. Finally, complications arise
arise during
during
certain segments of the encounter
encounterwhenwhenthe
thesystem
system will
will have to track very closely
closely spaced
spaced clusters of targets that may only only
be partially resolved by thethe viewing
viewingsensors.
sensors. The
The term "cluster" willwill be
be used
used toto refer
refer to a collection of objects with similar
similar
state vectors.
To date, typical MTT system design has been for environments with a relatively small
small number
number of targets
targets and low target
densities.2*3
densities 2 3 Also,
Also, MTT
MTT algorithms have typically
typicallybeen
been developed
developed toto accommodate
accommodate processing data from
processing of data from a single sensor.
With the introduction of multiple sensor configurations, there
there are
are aa variety of processing
processing chains that are possible and
and many
different kinds of MTT algorithms
different algorithms are
arebeing
beingconsidered.4,5,6
considered.4*5'6 Thus,
Thus, SDI
SDI presents an unprecedented challenge
challenge for the
development and evaluation ofof multiple sensor, multiple target
target tracking
tracking algorithms.
algorithms.
244 / /SPIE
SPIEVol.
Vol. 1096
1096Signal
Signaland
andData
DataProcessing
Processingof
ofSmall
SmallTargets
Targets 1989
1989
Data association
association isis the key
key issue
issue for any system. This
any MTT system. is particularly
This is for SDI
true for
particularly true because of
SDI because necessity to
thenecessity
ofthe to
establish pure
establish pure target tracks to achieve target classification
classification (or
(or discrimination)
discrimination) and
and to
to maintain
maintain good
good tracks
tracks thereafter
thereafter for
for
weapon guidance.
weapon guidance. Similarly,
Similarly, data association isis critical
data association trackinglow
to tracking
critical to low observables because misassociations,
observables because well as
as well
misassociations, as
excessive false
excessive and missing
false and tracks, can
missing tracks, be caused
can be clutter or false
caused by clutter signals. Also,
false signals. due to the angle-only
Also, due measurement nature
angle -onlymeasurement nature
EO sensors, there will be
of EO be considerable involvedininthe
considerable difficulty involved thereinitiation oftracks
reinitiationof thatbecome
tracksthat lost due
degraded and lost
become degraded due
to premature deletion.
leads to
multiple targets (including stationary ones or clutter) leads
signals or multiple
The existence of either false signals misassociations.
to misassociations.
misassociation isisthe
A misassociation the incorrect assignment (or
incorrect assignment weighting)of
(or weighting) anobservation
ofan track(state
observationtotoaatrack estimate) for
(stateestimate) filtering. The
for filtering. The
Misassociations do
incorrect observation can be a false signal or from another target. Misassociations do not occur for
not occur single target
for single tracking
target tracking
without false signals.
signals. The numberof
The number misassociations increases
ofmisassociations as the
increases as density of
the density targets and
of targets false signals
andfalse increases.
signals increases.
Related to data
data association
association performance and computational
performance and manner in
requirements, is the manner
computational resource requirements, which the
in which
There are
distributed over multiple platforms. There
processing is distributed associated with partitioning
are obvious computational advantages associated
multiple platforms. However,
across multiple
the tracking problem across maybe
theremay
However, there between the
conflict between
beaaconflict achieve
requirementstotoachieve
therequirements
data association
accurate data
accurate andthe
association and desireto
thedesire distributethe
to distribute anddata
trackingand
thetracking associationprocessing
dataassociation by partitioning
processingby the global
partitioning the
problems. Potential approaches to partitioning are discussed in more detail later in
MTT problem into local tracking problems. in the
the
paper.
Track initiation.
2.3 Track
process in
intensive process
Track initiation is potentially the most resource intensive systemson
MTT systems
in MTT pertrack
onaaper basis. The problem
track basis.
of track initiation with passive sensors in a dense environmentisisdifficult
targetenvironment
dense target two-dimensional
thetwo
becauseofofthe
difficult because -dimensional (2(2-D)
-D)
nature of the measurement
nature of data.One
measurement data. approachisistotoform
Oneapproach angle-only,
formangle -only, sensor -level tracks
sensor-level sensor and then to combine
each sensor
tracks at each
sensor-level tracksto
sensor -leveltracks formthree-dimensional
toform (3 -D)tracks
three -dimensional(3-D) throughthe
tracksthrough processof
theprocess oftriangulation. However, aa major
triangulation. However, problem
major problem
occurrence of
that can result from triangulation is the occurrence of false intersections or ghosts.
ghosts.
One approach
approach to reducingthe
to reducing ghostingproblem
theghosting with triangulation
problemwith to compare
triangulationisis to successive points
compare successive or to
points or use angle rate
to use
the two sensor
information from the -level, angle-only
sensor-level, thatare
tracksthat
angle -onlytracks for combination
considered for
areconsidered track. Alternative
3 -Dtrack.
combination into a 3-D
information from
methods require the simultaneous use of angle information
methods three or more sensors.
from three sensors. Four types of
generictypes
Fourgeneric multiple
of multiple
2.4 Birth
Birth-to-death
-to -death tracking.
A key concept
concept in SDI
SDI isisthe
the maintenance
maintenance of a continuous
continuous track on thethe initial
initial boost
boost phase
phase missile, the
the bus
bus that
that is
is released
released
from the boost phase missile,
missile, and
and the
the subsequent
subsequent reentry vehicles
vehicles(RVs)
(RVs)and and decoys
decoysthatthatare
are deployed
deployed from
from the
the bus.
bus. This
birth-to-death
concept, called birth -to -deathtracking,
tracking,isisdescribed
describedininFigure
Figure1.1. First,
First, the
thebooster
booster isis tracked. Then the
tracked. Then the track
trackhistory
historyandand
an accurate estimate of bus kinematics are required. This Thisinformation
information isis used
used for
for tracking
tracking the
the RVs
RVs and
anddecoys
decoys that
that are
are
deployed. The
TheRVs
RVsandanddecoys
decoyscouldcouldfirst
firstappear
appearasasan
anextended
extendedobject,
object,then
thenasasaapartially
partiallyresolved
resolvedcluster,
cluster,and
andfinally
finally as
as
individual objects. Once
Once discrimination
discriminationbetween
betweenan an RV
R V and
and the
the associated
associated decoys
decoys is performed,
performed, it is particularly important
important
that a good track on the RV be maintained thereafter.
thereafter.
Another important feature
feature of
ofbirth
birth-to-death
-to -death processing
processing isis that
that itit offers
offers the
the advantage
advantage of substantially
substantially reducing
reducing the
amount of track initiation
initiation processing
processing required
required in a dense environment.
environment. Tracks are initiated early when the threat density
is low.
low. Then, continuity is maintained as additional
additional objects
objects are
are spawned.
spawned. InInthis thisway,
way, the
the amount
amountof
ofcostly
costly track
trackinitiation
initiation
is substantially
substantiallyreduced
reduced during the midcourse phase.
Track spawning
spawning isis not
not to
to be
be confused with
with track
track splitting.
splitting. Track splitting refers to creating additional candidate or
hypothesis
hypothesis tracks
tracks for
for aa single
single target. Track spawning
target. Track spawning refers to creating more than one tracktrack from
from aa previous single track
because additional targets have been deployed or resolved in that location. Track spawning
location. Track spawning is expected to require fewer
resources than so
so-called
-called "cold
"cold start"
start" track initiation.
246 / /SPIE
SPIEVol.
Vol. 1096
1096Signal
Signaland
andData
Data Processing
Processing of
of Small Targets
Targets 1989
Evaluation
Evaluation of algorithms is complex
complex because
because of the diversity
diversity of aspects involved. Typically, the major tradeoff is
involved. Typically,
tracking performance
tracking performance versus
versus the
the amount
amount of processor
processor resources
resources required. The characteristics
required. The characteristics of
of performance
performance versus
versus
for an
required resources for an algorithm
algorithm usually
usually depend onon the
the threat,
threat, the
theoperating
operatingconditions,
conditions,and
andthethealgorithm
algorithmdesign
design
Algorithmsshould
parameters. Algorithms should be
becharacterized
characterizedtotoestablish
establishthese
theserelationships
relationships so
so that
thatcomparisons
comparisonsamong
among algorithms
algorithms
can be made and algorithms selected for a processing chain. For example,
chain. For example, Figure 3 illustrates these characteristics
characteristics for a
family
family of
ofalgorithms
algorithms for
for tracking
tracking an isolated target
target with
with false
false signals.
It is not enough to evaluate performance; an expression for the required resources resources for
for an
an algorithm
algorithm is
is also
also needed,
although this
this isisoften
oftenomitted
omittedfrom
fromreports
reports and
andpapers.
papers. Useful measures of required
required resources
resources are throughput and memory
memory
as a function
function ofoftarget
target density
density and
and the
the rate
rate at which
whichtargets
targets must
mustbe
beprocessed.
processed. However, the ultimate measure
measure ofof processing
resources for on on-board
-board processors
processors isis size, weight,
weight, and power.
The so
so-called
-called assignment
assignment or basis-free
basis -free technique11
technique11isisapplicable
applicable to
to the
the simpler
simpler case where all the targets
targets are
aresimilar,
similar,
but this technique must be extended for SDI where the the distinction
distinction between
betweenreentry
reentry vehicles
vehiclesand
anddecoys
decoysisisimportant.
important. TheThe
assignment approach uniquely pairs the actual target
approach uniquely target states to the estimated states so that the sum of the
the squared
squared estimation
estimation
errors for all the targets is minimized.
minimized. This
This approach
approach may
may tend to underestimate the magnitude
magnitude of the errors. An AnSDI
SDI
tracking panel
tracking panel has
has recently
recently adopted
adopted aavariation
variationofofthis
thisassignment
assignmentapproach
approachfor forevaluating
evaluatingtracking
trackingestimation
estimation
performance 12
performance.12
An alternative to this approach
approach isis to
to pair
pair an
an actual
actual target
target state
statetotoan
anestimated
estimatedstate
statebased
basedonontrack
trackpurity,
purity,namely,
namely,
based on the source of the measurements used in the track. However,However, thisthis approach
approachmay
may be biased
biased toward
toward some
some tracking
algorithms relative to others.
The evaluation
evaluation of tracking
tracking performance
performance isis not not limited
limitedtotothe
theevaluation
evaluationofofstate
stateestimation
estimationand
andprediction
predictionerrors.
errors.AA
variety of
ofmeasures
measures ofofperformance
performance are typically
typicallyneeded
needed to
to characterize
characterize performance for an application,
performance for application, such
such as
as the number
of false
false and missed tracks, probability of misassociation,
misassociation, and accuracy
accuracy of the state covariance
covariance matrix.
matrix. Further
Furtherwork
workisis
needed to establish
establish not only
only meaningful
meaningful evaluation
evaluation criteria
criteria but
but also
also to predict performance.
performance. Furthermore,
Furthermore,algorithm
algorithm
performance should be characterized
characterized so
so that
that itit can
can be
be predicted
predicted asas aa function
function of
of the target
target density,
density, probability of missed
missed
and false signals,
signals, number
number ofof new
newtargets,
targets, and
and other
other error sources.
2.7
2.7 Multiple frame signal processing.
The complexity
complexity of of multiple
multiple frame
frame signal
signalprocessing
processingisisgreater
greaterthan
thanfor
forthe
the traditional
traditional single
single frame
frame approach.
approach. With
multiple frame processing, the target motion must be be considered. In
In a multiple frame matched approachfor
matched filter approach for example,
example,
a bank of filters is
is needed account for
needed to account for the
the full
full range
range of
of target
target velocities.
velocities.
Several
Several different
different types
types of multiple
multiple frame signal
signal processing
processing approaches are under
under investigation13
investigation13 and further com-
com-
parisons
parisons and tradeoffs are needed. While
Whilemultiple
multiple frame
frame signal
signal processing methods require additional resources, theythey
might
might substantially
substantially improve
improve performance
performance by
by reducing
reducing the
the probabilities
probabilities of missed
missed and
and false
false signals.
signals. This could be vital
vital to
to
SPIE
SPIE Vol.
Vol. 1096
1096 Signal
Signal and
andData
DataProcessing
ProcessingofofSmall
SmallTargets
Targets1989
1989/ / 247
247
Concurrent
Concurrent oror parallel
parallel processing
processing isis an
an increasingly
increasinglyimportant
importantaspect
aspectofofMTT.
MIT. As As the number and density of the the
potential targets that must be
be processed
processed increases,
increases, the
the processing
processing load
load increases
increases substantially
substantially and
and not
not in
in aasimple
simple linear
linear
fashion. High
fashion. High performance
performance tracking
tracking of
oflow
low observables
observables is also very demanding in
very demanding in terms
terms of the processing resources
resources
required.
"robust" can
The term "robust" can be
be applied
appliedtototracking
trackingalgorithms
algorithms that
that perform
performwellwell or
or with
with graceful degradation
degradation ifif subjected
subjected
to threat or operating
operating conditions
conditions that
that vary
vary substantially
substantially from the design
design conditions.
conditions. This is another issue that cannot yet
be addressed in an established quantitative
quantitative fashion.
fashion. AAmajor
major concern
concernisis to
to what extent should an algorithm design depend
on a priori information?
information? The degree
degree of of dependence
dependence on a priori information such as typical
typical booster
booster thrust patterns and
and
booster launch sites must be considered
consideredin inalgorithm
algorithm design
design and
andthe
therisks
risks carefully
carefully weighed.
DATA ASSOCIATION
3. DATA
3.1 Choice
3.1 Choice of data association algorithm.
algorithm.
It has long
longbeen
been recognized
recognized that
that misassociation can lead
lead to
to tracking
tracking errors
errors that
that far
far exceed
exceed those
those caused by measurement
measurement
noise and target dynamics.14
noise dynamics.14 To To illustrate
illustrate this
this problem,
problem, Figure
Figure 44 shows
showsthethe ratio
ratio of
of the
the root mean square of position
estimation errors with and without the presence of
estimation of false
false signals
signals in
in the
the track gate of a single established track.6 The top
curve in Figure
Figure 44 shows
showsthe
the error
error ratio for the case where the nearest-
nearest-neighbor
neighbor data
data association method is used. Clearly,
Clearly,
potential problems
there are potential problemsin in aa dense
dense environment
environment that
thatmay
may arise with this simple,
simple, standard
standard approach
approach to
to data
dataassociation.
association.
The nearest-
nearest-neighbor
neighborapproach
approach to to data association determines aa unique
association determines unique pairing
pairing so that atat most
most one
one observation
observation is
is
paired with a given
paired given track. Then, observation
track. Then, observation-to-track
-to -track pairings
pairings are
are irrevocably
irrevocablymade.
made. The method
method is based
based upon likelihood
theory, with
theory, with the
the goal
goal being
being the
the minimization
minimization of ofaa distance
distance function
functionfromfromthe
theobservation
observationtoto the
the predicted
predicted position of the
target. An
An alternative
alternative approach,
approach,multiple
multiplehypothesis
hypothesistracking
tracking(MHT),
(MHT),postpones
postponesdifficult
difficultdata
dataassociation
associationdecisions
decisionsuntil
until
more information, in the form of subsequent observations, is is received.
received. An extreme example of this approach is the gated
optimal method whose performance
performance isis shown
shown by the lower curve in Figure Figure 4.
4. Using the gated
gated optimal
optimalmethod
methodall
allfeasible
feasible
satisfaction of
(as determined by the satisfaction of aa gating
gating relationship)
relationship) observation
observation-to-track
-to -trackpairings
pairings (or
(or branches)
branches) are propagated.
propagated.
Then, an estimate of the true target
target position
position is is formed from a probabilistically weighted
weighted sum of the the branches.
Under the condition that
that all
all feasible branches are formed
formed and
and maintained, the results
results shown
shown in Figure 4 for the gated
optimal method
method represent
represent a lower bound for the tracking error
error in the
the presence of potential misassociation.
misassociation. In practice, a
feasible
feasible implementation
implementation ofof the
the MHT
MHT approach willwill continually
continually form
form multiple
multiple data
data association
association hypotheses
hypotheses but
but maintain
maintain
only
only aa subset
subset of
ofthe
the more
more likely
likelyhypotheses.
hypotheses. The discussions
discussions of MHT
MHT given
givenin
inReferences
References 22 and
and 15
15 address
address these issues
issues in
in
more detail.
248 //SPIE
SPIEVol.
Vol. 1096
1096Signal
Signaland
andData
Data Processing
Processing of Small Targets 1989
Targets 1989
Several studies
studies have
have recently
recently been
been performed
performed to quantify data data association
association performance
performance as a function
function of the
the environment
environment
when multiple targets are considered.
considered. These
Thesestudies
studiesexamined
examinedthe theprobability
probabilityofofmisassociation
misassociationfor
forexisting
existingtarget
targettracks.
tracks.
A Monte Carlo simulation
simulation was conducted for for the
the global
global optimal
optimal assignment
assignment of observations to existing tracks. All
existing tracks. All the
observations in the field of view
view of
of the
the sensor
sensor for for a single
single frame
frame ofof data
data were
were available
available to
to the
the assignment
assignment process.
process. In
In the
the
first study
study the results were obtained as as aa function
function of of target
target density,
density, as
as measured
measured byby the
the average
average number
numberof ofobservations
observations
(hits) in the track gates,
(hits) gates, and
and the
thepercentage
percentageofofobservations
observations obtained
obtained from
from targets
targets without
without existing
existing tracks. The
Thelatter
latter
criterion, number ofof observations
observations without tracks,tracks, is of interest because
because itit quantifies
quantifies the
the necessity
necessity to
to maintain
maintain tracks
tracks on
on all
all
example, known
targets (even, for example, known decoys)
decoys) in in the
the sensor
sensorfield
field of
ofview.
view.
Results given in Figure 5 show the manner in which the probability of misassociation dramatically increases with the
number of hits in the gate and with the percentage of observations without without tracks.
tracks. The desirability
desirability of having
having aa complete
complete
global approach, with all
approach, with all observations
observationsand
andtracks,
tracks,for
fordata
dataassociation
associationwaswasalso
alsoshown
shownby byaasecond
secondstudy
studyininwhich
which the
the sensor
sensor
field of view
viewwas
wasrestricted.6
restricted.6Then,
Then,observations
observations were
were only
onlyreceived
received for
for aasubset
subsetof
ofthe
thetracks
tracksthat
that were
were previously
previously established
established
on a group of closely
closely spaced
spaced targets. Thus, in
targets. Thus, in this case, there were tracks without observations. The Theresult
resultwas
wasan
anincrease
increase
of about 50 percent in thethe fraction
fractionof
ofobservations
observations that
thatwere
weremisassociated,
misassociated, whenwhen compared
comparedwithwiththe
thecondition
conditionininwhich
which
the sensor field of view
view was
wasincreased
increased so
so that
that observations werewere received
received from
from all
all targets.
targets.
The
The term "preferential tracking"
tracking" has
has been coined to refer to the
been coined the selective
selective use of tracks
tracks or
or observations in data
observations in
association.8 InInpreferential
association.8 preferential tracking
tracking only
onlyaa subset
subset of
of the potentially available
the potentially availabletracks
tracks or
or observations
observationsisis used
used in
in the
the
association processing of the data
data from
from the
the field
field of
of view of
of aasensor.
sensor. These studies provide some indication
indication of
of the reduced
reduced
performance that can result from preferential tracking.
3.3 threat density.
3.3 SDI threat density.
SPIE Vol.
SPIE Vol. 1096
1096 Signal
Signal and
andData
DataProcessing
ProcessingofofSmall
SmallTargets
Targets1989
1989/ / 249
249
There are
are many
many approaches
approaches totocombining
combining data from multiple
data from sensors and distributing
multiple sensors the processing
distributing the multiple
over multiple
processing over
platforms or vehicles. The selection
vehicles. The of a specific architecture can have a substantial impact on tracking performance and
selection of
Majorconsiderations
the processors. Major
total weight of the datamisassociation,
includedata
considerationsinclude load, and
communicationload,
misassociation, communication survivability.
andsurvivability.
This section examines some of the basic types of processing architecture applicable applicable to aerospace-based
toaerospace surveillance.
-based surveillance.
One approach
approach isis centralized processingin
centralized processing which, in effect, a single global
in which, track file
global track formed using
file isisformed observations from
the observations
usingthe from
sensors. In this approach there
all sensors. only be a single
should only
there should foreach
trackfor
single track An alternative approach forms local track
target. An
eachtarget. track
files based
files individual sensor data or
based upon individual upon only
or upon those observations
only those determined to
observations determined bewithin
tobe surveillance
specifiedsurveillance
withinaaspecified
volume. Using this approach,
volume. thereisisthe
approach, there potentialtotoform
thepotential morethan
formmore trackon
onetrack
thanone anygiven
onany target.Thus,
giventarget. track-to-track
Thus,track -to -track
association and track
association and track fusion
fusion may
maybe be aa very importantpart
veryimportant the system
of the
part of forthe
logicfor
systemlogic latterapproach.
thelatter approach. The processing
The processing
encompasses track
architecture encompasses initiation and
track initiation termination, as
and termination, well as track maintenance.
as well
have been
Four generic types of processing chains for track maintenance have aerospace -based surveillance
been identified for aerospace-based surveillance
fourtypes,
Thesefour
applications such as SDI.6 These whichhave
types,which sincebeen
havesince adoptedby
beenadopted anSDI
byan trackingpane1,12
SDItracking are as
panel,12 are follows.
as follows.
Independent Sensor
Type I: Independent sensor independently of
Sensor Processing - Tracks are processed for each sensor the data
of the from the
data from the
Frame-to-frame
other sensors. Frame -to -frame association and are performed
and filtering are any sensor
without any
performedwithout -to- sensor processing.
sensor-to-sensor processing.
1p-II:
Type II: Hierarchical Sensorlevel
Processing-- Sensor
Hierarchical Processing followed by track fusion. Frame
processing isis followed
level processing -to -frame association
Frame-to-frame
are followed
and filtering are -to- sensorprocessing.
sensors-to-sensor
followed by sensors processing. Each involved in two association
observation isis involved
Each observation processes.
association processes.
position estimation
target position
Type III: Observation Fusion - Multiple sensor, observation association, and target followed
estimationisis followed
by frame -to -frameassociation
frame-to-frame and filtering.
association and filtering. Sensor processing precedes frame
-to- sensor processing
Sensor-to-sensor -to -frame processing.
frame-to-frame processing. EachEach
subjected to two association processes.
observation isis subjected
Type IV: Centralized Observation-to-track
Processing--Observation
Centralized Processing association is followed by mutiple
-to -track association Associa-
mutiple sensor filtering. Associa-
of data
tion is performed on each frame of each sensor as it becomes available.
data from each per target
available. Only one track file record per is
target is
required in the entire
required system. Only
entire system. association process is performed on each observation.
one association
Only one
chain and the physical
A clear distinction can be made between the functional (logical) processing chain of the
distribution of
physical distribution the
processing. With multiple platforms andand on -board processing, there are
on-board number of
are aa number ways in which the processing can be
ofways
physicallydistributed
physically type of
distributed over the platforms for a given type chain. Thus
processingchain.
ofprocessing each basic
for each
Thus for there are alternative
type, there
basic type,
implementation methods some of discussed below.
of which are discussed
4.1
4.1 Centralized data association.
(Type IV)
approach (Type
A major advantage of the centralized approach accuracy is maximized. A
tracking accuracy
that tracking
IV) isis that targettrack
A target that
trackthat
consists of
consists fromall
observationsfrom
ofobservations shouldbe
sensorsshould
allsensors accuratethan
moreaccurate
bemore thetracks
thanthe couldbe
thatcould
tracksthat established on
be established the partial
on the partial
the individual
data received by the sensors. Then, as a further important factor in performance, increased
individual sensors. accuracy will
trackaccuracy
increasedtrack
decrease misassociation. However, the
misassociation. However, the centralized requiresbackup
approachrequires
centralized approach processingand
backupprocessing trackfiles
andtrack survivability
forsurvivability
filesfor
the platform
case the
in case the central
containing the
platform (or platforms) containing file fails. Figure 7 shows
track file
central track method for implementing the
one method
shows one
centralized processing approach. Here, Here,all targettracks,
alltarget wellas
tracks,asaswell as aastar file, are
starfile, maintainedwithin
are maintained processor.
centralprocessor.
withinaacentral
All sensor observations are sent
observations are to the
sent to central processor for
the central association and
for association newtrack
and new initiation. Upon
trackinitiation. association, track
Upon association, track
filter update is performed.
Due to the large number of
large number ofpotential asindicated
tracks, as
potentialtracks, centralizedimplementation
earlier,centralized
indicatedearlier, Figure 77 isis
shownininFigure
implementationshown
surveillancesystems
large -scale surveillance
probably impractical for large-scale suchasasSDI.
systemssuch Thus,aapartially
SDI.Thus, distributed processing method is
partiallydistributed
amount of processing on
order to reduce the peak amount
desirable in order anyone
on any both the
processor. Either or both
oneprocessor. processing and the
the processing
track files Manydifferent
can be distributed. Many
files can typesof
differenttypes processing/track distributions
ofprocessing/track can be
distributions can identified; one
be identified; these,
ofthese,
one of
distributed central
distributed -level processing, is discussed in Section 4.3.
central-level
Distributedsensor
4.2 Distributed -level tracking.
sensor-level
Vol.1096
SPIEVol
250 //SPIE andData
Signaland
1096Signal ofSmall
Processing of
DataProcessing 1989
Targets 1989
Small Targets
the other
On the otherhand,
hand,points
pointstypically
typically cited
cited ininfavor
favorofofthe
thesensor -level approach
sensor-level approach areare reduced
reduceddata -bus loading
data-bus loading
(presumably
(presumably because
because only
only tracks
tracks and
and not
not all
all observation
observation data
data need
need be transmitted), distribution of the computational
computational
and high survivability.
loading and survivability. Also, if one sensor becomes degraded,
degraded, its
its observations
observationswill
will not
not affect
affect the
the sensor
sensor-level
-level tracks
tracks
of other sensors.
Figure 8 shows aa sensor-level
Figure sensor -leveltracking
trackingapproach
approachthatthatcould
couldbe beapplied
appliedtotothe
theSDI
SDIproblem.
problem.For
For this
thisapproach,
approach, separate
separate
track files are formed independently
independently using
using the observations from each sensor. Note Note that
thattrack
trackfusion
fusionisis required
requiredwhen
whenthe
the
sensor-level
sensor -leveltracks
tracks are
are formed central-level
formed into a central -level track
track file
file (by
(by the
the function
function called
called Multi Source Data
Data Fusion).
Fusion).
The feedback from the
feedback from the central
central-level
-level track file, formed by
by track
track fusion,
fusion, isisdenoted
denoted as
as optional
optional in Figure 8.6»8
8ó,8 However,
some form of feedback willwill probably
probably bebe required track initiation
required to reduce track initiation requirements
requirementsas astargets
targetspass
passfrom
fromthe
thefield
fieldof
of
view
view of
of one
one sensor
sensor to
to that
that of another. This
This problem
problemisis addressed
addressedin in Reference
Reference55by by the
the inclusion
inclusion of
of special
special local fusion
fusion
nodes which,
which, in
in addition
addition to performing track fusion,
fusion, also
also distribute
distribute tracks
tracks among platforms.
platforms.
A modified version of the sensor-level
modified version sensor -level tracking
tracking approach
approach shown in Figure
Figure 88would
wouldpartition
partition the
the global
global tracking
tracking problem
problem
into smaller local tracking problems. Under
tracking problems. Underthisthisapproach,
approach,a agiven
giventracking
trackingplatform
platformwould
wouldonly
onlybe
beresponsible
responsibleforforaasubset
subset
of the
the targets
targets or a small segment of the
the entire surveillance volume.
volume. Thus, this approach is considerably
considerably more complex than
pure sensor
sensor-level
-leveltracking
trackingbecause
because ofofthe
the potential
potential requirement
requirement to to transfer observations. This requirement
observations. This requirement would
would occur
occur
if the
the positions
positions of the tracks being maintained or the subset subset of the surveillance
surveillance volume
volume for
for which
which aa sensor
sensor platform
platform has
has
responsibility does not exactly
exactly coincide
coincide with
withitsitssearch
searchvolume.
volume.However,
However,this
thisapproach
approachcould
couldlead
leadtotopreferential
preferential tracking,
tracking,
which was discussed earlier.
Figure 9 shows
shows an approach thatthat maintains
maintains the
the accuracy
accuracyof ofcentral
central-level
-level tracking,
tracking, but also effectively
effectively distributes
distributes the
the
processing and the track
track files
files in a manner
manner similar
similar to that
that of
of sensor
sensor-level
-leveltracking.
tracking. In this approach, the global
approach, the global track
track file
file is
partitioned among M tracking platforms. Thus,Thus, the
the track
trackfiles
files are
are partitioned
partitionedso so that
thateach
eachtracking
trackingplatform
platformonly
only keeps
keeps aa
proportion 1/M
1/M of the total track
track file.
file.
approach shown
The approach shown in
in Figure
Figure 99 also partitions the processing. Observation
Observation-to-track
-to -track data
data association
association isis done
done on the
sensor platforms where the observations are received while
while track
track filter
filter update
update and gate formation are done atat the tracking
platforms that contain the tracks.
should be noted
It should noted that
thatthe
theapproach
approachshown
shownininFigure
Figure99isisvery
veryflexible
flexible because
because any
any given platform can function as as
either a sensor platform or aa tracking
tracking platform,
platform, or as as both. However,
However, as as previously
previously mentioned, this approach requires
other platforms to function
function asas backups
backups so
so that
that aa significant
significant proportion
proportionofofthe
thetrack
trackfile
file information
informationwill
will not
not be
be lost
lost should
should
a tracking platform fail.
5. CONCLUSIONS
5. CONCLUSIONS
SPIE
SPIE Vol.
Vol. 1096
1096 Signal
Signal and
and Data
DataProcessing
Processingof
ofSmall
Small Targets
Targets1989
1989// 251
Although group
Although group tracking
tracking may
may be certain segments
be appropriate for certain segments of the tracking
tracking history,
history, the requirements
requirements for
for
performing and maintaining discrimination, situation
situation assessment and weapons allocation are are ultimately
ultimately based
based upon
upon having
having
accurate
accurate individual
individualtarget
target tracks. Here, itit appears
tracks. Here, appearsthat
thatatatleast
leasta alimited
limitedapplication
applicationofofMHT -type methods
MHT-type methods maymay be
necessary.
necessary. References
References 44 and 5 discuss
discuss potential
potential approaches
approaches for the SDISDIproblem
problem in
inmore
moredetail.
detail. However,
However, asaswas
wasillustrated
illustrated
by the results
results shown
shownininFigure
Figure3,3,there
therewill
willbebea atradeoff
tradeoffbetween
betweenperformance
performanceand andcomputational
computationalresources
resourcesrequired.
required. Due
to the nature
nature of the
the SDI
SDI problem
problem in which
which all
allcomputing
computingmachinery
machineryshould
shouldbebeplaced
placedon
onthe
thesensor
sensorplatforms
platforms and maintained
maintained
there, the available
available computational resources will will place fundamental practical limitations
limitations on the algorithms
algorithms that
that can
can be
be
implemented.
reportedin
As reported inReference
Reference1,1,the
theFletcher
Fletchercommission
commission"concluded
"concludedthat
thatnew
newtrack
track-formation algorithmswould
-formation algorithms wouldhave
havetoto
be developed
developed to make a defense
defense feasible."
feasible." Also,
Also, the
theapparent
apparenttracking
trackingproblems
problems associated
associated with
with the
the Delta
Delta181
181 test16
test16
indicate that tracking may be a major issue eveneven for
for relatively
relativelybenign
benignfuture
futurespace
spaceexperiments.
experiments. Thus,
Thus, there clearly is
is the
the
necessity for
for aa long-term
long -termtechnology
technologyprogram
programtoto develop
develop andand evaluate
evaluatetracking
tracking and
and data
data association methods
methods for
for the
the SDI
SDI
near-term
near -term and
and far-term
far -term threats and for
threats and for other
otheraerospace
aerospace-based surveillance systems.
-based surveillance systems.
Another important
important issue
issue that
thatwill
will have a major impact on MTT performance
performance is is the
the distribution
distribution ofofprocessing
processing among
among
the tracking platforms. From
From thethecomputational
computationalpoint
pointofofview,
view, it would
would bebe desirable
desirable to to partition
partition the global MTT problem
into a number of smaller local tracking problems associated with with each
each platform. approachmay
platform. This approach may either
eithertake
takethe
theform
form
of sensor-level
sensor -leveltracking,
tracking,asasshown
shownininFigure
Figure8,8,or
ormay
mayassign
assignthe
theresponsibility
responsibilityfor
foraaparticular
particularset
set of
oftargets
targets or
or clusters
clusters to
to
given platform.
a given appealing from
platform. However appealing from aa purely
purely computational
computational point of of view this
this approach
approach may be, there are severe severe
potential data association problems thatthat may arise from this
this method
method duedue to
to crossing
crossingtarget
target clusters
clusters and
and the
the expansion
expansion andand
ultimate overlapping
overlapping of target
target clusters. As
Asshown
shown byby results presented
presented previously,
previously, the lack of a global data association,
association,
containing all tracks and observations
observations in the
the field
field of
of view for
for aa sensor,
sensor, can
can lead
lead to a dramatic increase
increase inin misassociation.
misassociation.
A related intuitively appealing general concept
concept is is that
that of
of tracking only certain targets of of particular
particularinterests,
interests,namely,
namely,
preferential tracking. For
Forexample,
example, onceonce discrimination
discrimination is is performed, it would be preferable to maintain tracks only on on
RVs, and not on all the decoys. Also, because
because ofof their predicted
predicted destination,
destination, some
some targets
targetsmay
maybe of more
be of more tactical
tactical interest
interest
However,asasdiscussed
than other targets. However, discussedabove,
above, due
duetotothe
theoverlapping
overlappingof oftarget
target clusters
clusters inin the
the22-D
-D measurement space
necessity to maintain an
and the necessity an accurate
accurate global
global track
track file,
file, this
this approach
approach would
would likely
likely lead toto aasignificant
significant increase
increase in
in
misassociation. The result
misassociation. result of
of this
this increase
increase in
in misassociation
misassociation wouldwould be discrimination
discrimination errors
errors and substantially degraded
tracking accuracy
accuracy onon those
those preferred require high
preferred targets that require high performance tracking.
Figure 9 shows
shows an
an architecture
architecture in which
which both
both the
the track
track files
files and
and the
theprocessing
processing are
are distributed
distributed while
while the
the advantages
advantages of
of
centralized processing
centralized processing are
are maintained. However, this
maintained. However, this approach
approach requires
requires considerable
considerable communication
communication and
and the
the use of
backup
backup processing
processing and
and track
track file
file maintenance. Clearly, the question of how
maintenance. Clearly, how to distribute the processing for the MTT
problem
problem requires careful study with an emphasis
emphasis on determining the impact of various processing distributions
distributions upon
tracking performance,
performance, particularly data association,
association, and the
the processor resources required.
252 //SPIE
SPIEVol.
Vol. 1096
1096 Signal
Signal and
and Data
Data Processing of Small Targets
Targets 1989
1989
6. ACKNOWLEDGMENT
6. ACKNOWLEDGMENT
Many colleagues at
at Hughes
Hughes Aircraft
Aircraft Company,
Company, tootoo numerous
numerous to
tolist,
list,have
havecontributed
contributedto
tothis
thispaper.
paper. The Monte Carlo
simulation results were funded
funded by IR&D
IR &D and
and other
other internal
internal Hughes funds. results were
funds. These results were obtained
obtainedwith
with the assistance
of P.
P. Birnbaum,
Birnbaum, T.
T. George-Shores,
George- Shores,J.J.Hyder,
Hyder, K.
K.Hell,
Hell,F.F.Paulsen,
Paulsen,T.
T.Pham
Phamand and M.
M. Trainoff.
Trainoff.
7. REFERENCES
7. REFERENCES
1. Adams,
1. Adams, J.A.
J.A. et
et al,
al, "SDI:
"SDI: The
The Grand
GrandExperiment,"
Experiment,"IEEE
IEEESpectrum,
Spectrum.September
September1985,
1985,pp.
pp.34 -46.
34-46.
5. Alien, T.G.
5. Allen, T.G. et
et al,
al, "Multiple
"Multiple Information
Information Set
Set Tracking
Tracking Correlator
Correlator (MISTC)
(MISTC) Concept
Concept Development,"
Development," Proceedings
Proceedings
of 1987
1987 Tri-Service
Tri -Service Data Fusion Symposium,
Data Fusion Symposium, Laurel,
Laurel,Maryland,
Maryland,June
June9 9-11,1987, pp.342
-11, 1987, pp. 342-353.
-353.
6. Drummond,
Drummond,O.E.,
O.E.,"Lecture
"LectureNotes
Noteson
onMultiple
MultipleTarget
TargetTracking,."
Tracking,."October
October1985;
1985;revised
revised10
10April
April1989,
1989, Technol-
ogy
ogy Training
Training Corp.,
Corp., Torrance,
Torrance, Ca.
7. Drummond, O.E.,
7. Drummond, O.E., and
andS.S.S.
S. Blackman,
Blackman, "Multiple
"Multiple Sensor,
Sensor, Multiple Target Tracking Challenges
Challenges of the Strategic
Strategic
Defense Initiative," Proceedings
Proceedings ofof the
the 1st
1st National
National Symposium
Symposium on Sensor Fusion.
Fusion, Orlando, FL, April
April 1988.
1988.
8. Drummond,
Drummond, O.E.,
O.E.,"The
"TheAlgorithm
AlgorithmDevelopment
Development Challenge
Challenge of Tracking the SDI Dense Threat," IST Workshop
1ST Workshop
on SDI: BM/C3, IDA, Alexandria,
Alexandria, VA,
VA,November
November24,
24,1987.
1987.
9. Frenkel,
Frenkel,G.,
G.,T.S.
T.S. Paterson
Patersonand
andM.E.
M.E.Smith,
Smith, "SDI
"SDI Battle
Battle Management/C3
Management/C3 Algorithms Technology Program Plan,"
PaperPP-2068,
IDA Paper -2068, Institute Defense Analyses,
Institute for Defense Analyses, Alexandria,
Alexandria, VA,
VA, April
April 1988.
1988.
10. Drummond, O.E.,
10. Drummond, O.E., S.S.
S.S. Blackman, and K.C.
K.C. Hell,
Hell, "Multiple Sensor
Sensor Tracking
Tracking of
of Clusters
Clusters and
and Extended
Extended Objects,"
Objects,"
Proceedings1988
Technical Proceedings 1988 Tri
Tri-Service Data Fusion
-Service Data Fusion Symposium,
Symposium, Laurel,
Laurel, Maryland,
Maryland,1717-19
-19 May 1988.
1988.
11. Drummond,O.E.,
11. Drummond, O.E.,Multiple-
Multiple-Object Estimation, UCLA
Object Estimation, UCLAPh.D.
Ph.D. dissertation,
dissertation,1975.
1975. Xerox
XeroxUniversity
UniversityMicrofilms
Microfilms No.
75-26,954.
75 -26, 954.
12. Frenkel,G.
12. Frenkel, G.(Editor),
(Editor),Proceedings
Proceedingsofofthe
theSDI
SDITracking
TrackingPanels,
Panels,January
January1989
1989Sessions,
Sessions,Institute
InstituteofofDefense
DefenseAnalysis,
Analysis,
Alexandria, VA, January
January1989.
1989.
13. Schaum, A.
13. A. (Editor),
(Editor), Proceedings
Research Laboratory,
Laboratory, Monterey,
Monterey,CA,
- - , I of
s the
h NRL Seminar: Algorithms for Autonomous n
ate n_ Passive Surveillance,
r i Naval
CA,22 22-23 1989.
-23 February, 1989.
14.
14. Singer, R.A. and J.J. Stein, "An Optimal Tracking Filter for Processing
Processing Sensor
Sensor Data
Data of Imprecisely
Imprecisely Determined
Determined
Origin
Origin hi
in Surveillance
Surveillance Systems,"
Systems,"Proceedings
Proceedingsof
of 1971
1971IEEE
IEEE Conference
Conference on
on Decision
Decision and Control, Miami Beach,
Beach, FL,
FL,
December 1971, pp
December 1971, pp171
171-175.
-175.
16. Covault, C,
16. Covault, C., "SDI Delta 181
181 Achieves
Achieves Battle Satellite
Satellite Data Goals,"
Goals," Aviation
Aviation Week
Week and
and Space
SpaceTechnology,
Technology,
February15,
February 15,1988, pp14
1988, pp 14-17.
-17.
SPIE
SPIE Vol.
Vol. 1096
1096 Signal
Signal and
andData
DataProcessing
Processingof
ofSmall
SmallTargets
Targets1989
1989// 253
p®
DEPLOYMENT
POINTS
/__ i ,g
BUS
W.
'I)UNRESOLVED
\ t
O`
SHROUDS 4 O\N \ PARTIALLY
RESOLVED
(22
SECO \ RE-ENTER1Nß
RE-ENTERING
O O00\
ON
/ 1
RESOLVED TRAIN
®
1® BOOST TRACK.
BOOST TRACK, BOOST
BOOST TRACKING
TRACKING SYSTEM
SYSTEM INITIATES
INITIATES TRACKS
® SHROUDS,
@ INITIATE
SHROUDS. TRACKS
INITIATE WHEN
TRACKS WHENVIEWED
VIEWEDBY
BYMIDCOURSE
MIDCOURSE TRACKING
TRACKING SYSTEM
SYSTEM
® SECO,
(D BOOST
SECO. TRACKER
BOOST TRACKERDATA
DATACEASES
CEASES
®
(4)DEPLOYMENT, TRACK
DEPLOYMENT. BUS
TRACK BUSAND
ANDINITIATE
INITIATETRACKS
TRACKSON
ONNEW
NEW CLUMPS
CLUMPS
®
(§)FRAGMENTS, INITIATE
FRAGMENTS. TRACKS
INITIATE TRACKSON
ONPIECES
PIECES
® RESOLVED
(6) CLUSTERS,
RESOLVED CLUSTERS.TRACK
TRACKINDIVIDUAL
INDIVIDUALOBJECTS
OBJECTS FOR
FOR DISCRIMINATION, KILL
KILL
ASSESSMENT, AND BATTLE
ASSESSMENT. BATTLE MANAGER
0
@ RE-RE-ENTERING
ENTERING TRAINS, HAND
TRAINS. OVER
HAND OVERTO
TOTERMINAL
TERMINALTRACKING
TRACKING SYSTEM
SYSTEM
Figure 1. Birth-to-death
1. Birth -to -death tracking.
40
25
20
-D LOCATION
TARGET 22-D LOCATION
ESTIMATION ERROR
ERROR is
15
RESOLUTION - 250 M
RATIO =
2 FALSE SIGNALS GATE
MONTE CARLO
MONTE CARLO . 10
10 RUNS
RUNS
RSS ERROR
RSS ERROR WITH
WITH
LEGEND
FALSE SIGNALS
FALSE SIGNALS 10
o == U(0.0,2.01
U[0.0,2.0] M/S RSS ERROR
RSS ERROR WITHOUT
1 5
2400
00 55 10 15 20
o 800 1200 1600
TIME SINCE
TIME SINCE DEPLOYMENT (SEC) RESOURCES:
RESOURCES: THROUGHPUT
THROUGHPUT AND MEMORY FACTOR -NUMBER
MEMORY FACTOR NUMBEROF
OFTRACKS
TRACKS
0.
I
®= 2.0
H=
I
LEGEND
LEGEND
2.0 OTHER
OTHER HITS
uIN GATE
HITS IN
3(=
X= 1.5
1.5 OTHER
OTHER HITS IN
IN GATE
GATE
0.2
20 ®== 1.0
1.0 OTHER
OTHER HITS IN GATE
HITS IN
O ®= 0.5
9= 0.5 OTHER
OTHER HITS IN GATE
HITS IN
TARGET 2 -D LOCATION
ESTIMATION ERROR
RATIO: 0.2
15
i
RSS ERROR WITH < 0.1
FALSE SIGNALS
10
á
RSS ERROR WITHOUT
FALSE SIGNALS
O
o. II
0.25
0.25 0.50 0.75 1.0
1.0 1.25 1.50 1.75 20
AVERAGE NUMBER OF FALSE SIGNALS
SIGNALS IN
IN A
A 99%
99% GATE
GATE
t 5 10 15 20 25 30 35 40 45 50 55 6(
showing
4. RSS position estimation performance showing
Figure 4. PERCENT OBSERVATIONS WITHOUT TRACKS
OBSERVATIONS WITHOUT TRACKS
seven scans
after seven
results after signals.
false signals.
of false
scans of Figure 5.
5. Observations established tracks
Observations without established
increase misassociations.
254 //SP/E
SPIEVol.
Vol. 1096
1096Signal
Signaland
andData
Data Processing of Small
Processing of Small Targets 1989
Targets 1989
\\sz-~
I MITS TRACK
* MAINTENANCE
UNASSOCIATED
HITS
ASSOCIATION
ASSOCIATION
/ \ TR ACK FILE
s TAR AND
STAR /CENTRAL V" CORDS
~~l
T RACK FILE
TRACK FILE
ECORDS
RECORDS V F|LE /
1 TRACK P~"
S TAR AND
STAR AND
RACK FILE
TTRACK
/cTyytN SAR FILE f
CORDS 1
'
fRECORDS
ECORDS
NITS
rs U-
yfr
r
HI
250 500 750 1000 1250 1500 1750 2000 • CENTRAL PROCESSOR
TIME SINCE DEPLOYMENT,
DEPLOYMENT. SEC
SEC
Figure 6. Number of
6. Number oftargets
targets in
in aa 99%
99% gate for Figure 7. Centralized data
7. Centralized dataprocessing
processing (Type
(Type IV).
IV).
a target in a cluster.
i
tfP a TRACK
TRACK AND TRACK
TRACK
TRACK r£>*^*
STAR FILE j£r5^
INITIATION RECORDS
5OG1P INITIATION RECORD "1 VILE' )
SENSOR 1 RECORDS
RECORDS N^XV?*
JNPS Nits
SIGNAL
(PROCESSOR
PROCESSO TRACK MIT /TRACK
HIT/TRACK TRACK
TRACK DATA
DA TA BASE
BASE
--M FILTER 8 MANAGER
kJA
NAGER
1__ HITS
^
^
MAINTE NANCF
MAINTENANCF
ASSOCIATION
ASSOCIATION
PAIRS
PAIRS --
* UPDATE RECORDS
U ( TRACK J
fI t
LOCAL PROCESSOR
PRO CESSOR 1
A
TRACK K>
0 ¡ OPTIONAL
RECORDS ^,/jJ OPTIONAL
* K/IULTI SOURCE
SOURC E
MULTI
CDATA
)ATA FUSION
O TRACK
TRACK
TRACK 1 ' ^~^
TRACK AND RECORC>S
INITIATION RECORDS V FILE )
SENSOR N
FILE
STAR FILE
RECORDS JNPS
O*
Ds
U
SIGNAL
PROCESSOR
1___ ^
-»
f
_^ TRACK
TRACK
MAINTENANCE
ASSOCIATION
ASSOCIATION
HIT TRACK
HIT TRACK
PAIRS *E
PAIRS
FILTER
UPDATE
TRACK
RECORDS
RECORDS
DA TA BASE
TRACK DATA
4-- - MANAGER
BASE
NAGER
(TRACK ]
HITS
L ^^s
LOCAL PROCESSOR
PRO CESSOR
Distributedsensor
Figure 8. Distributed sensor-level
-level processing (Type II).
UNASSOCIATEO
UNASSOCIATED OBSERVATIONS
II
LOCAL PROCESSOR
LOCAL PROCESSOR 1,1, SENSOR
SENSOR11 TRACK CENTRAL PROCESSOR A
GATES
PROCESSING,
SIGNAL PROCESSING, FILTER UPDATE TRACK
TRACK MAINTENANCE INITIATION
ASSOCIATION
ASSOCIATED
OBSERVATIONS e
LOCAL PROCESSOR
PROCESSOR 7, SENSOR
SENSOR 7
V
\7
CENTRAL PROCESSOR X
PROCESSING,
SIGNAL PROCESSING, FILTER UPDATE
TRACK MAINTENANCE 1 TRACK
ASSOCIATION GATES
UNASSOCIATED OBSERVATIONS
Distributedversion
Figure 9. Distributed version of
of centralized
centralized processing
processing (Type IV).
SPIE Vol.
SPIE Vol. 1096
1096 Signal
Signal and
andData
DataProcessing
ProcessingofofSmall
SmallTargets
Targets1989
1989/ / 255
255