5

Estimation, Tracking, and Data

Association

5.1 INTRODUCTION
This chapter covers the concepts of parameter estimation, target tracking, and
data association algorithms used to enable multiple-target tracking in real-world
environments. References [1] through [4] are excellent sources of additional in-
formation on tracking. The topics covered in this chapter include:

• Parameter estimation for radar

• The radar tracking function:
– Waveforms and signal processing
• Types of tracking filters:
– Alpha-beta and alpha-beta-gamma
– Kalman
– Extended-Kalman
– Interacting multiple-model
• Data association algorithms:
– Nearest-neighbor
– Probabilistic data association (PDA)

97
98 Chapter 5

– Joint PDA (JPDA)

– Nearest neighbor-JPDA
– Multi-hypothesis tracking (MHT)
– Other assignment algorithms
• Tracking air targets:
– Aircraft, unmanned aerial vehicles (UAVs)
– Cruise missiles
• Tracking ballistic missile targets:
– TBMs, IRBMs, ICBMs
• Tracking surface targets:
– Ships
– Vehicles

5.2 PARAMETER ESTIMATION FOR RADAR

Parameter estimation is an important function performed by radars. It differs
from radar measurements or observations, inasmuch as that in this chapter esti-
mation refers to extracting parameters not directly measured by radar. The most
common applications of estimation in radar are target tracking and target fea-
ture estimation to enable target classification and discrimination. This chapter
focuses on estimation in the target tracking context, while Chapter 6 treats target
feature estimation for classification, discrimination, and identification.
References [4], [5], and [6] are three sources of theory related to parameter esti-
mation. In general, parameter estimation solves the problem of extracting de-
sired quantities from radar measurements, that is:


x = E{ x z } , (5.1)

where x is the parameter being estimated and z are observations from which es-
timates of x can be extracted. It is assumed that x is observable in the control the-
ory sense, that is, that quantities necessary to estimate x are included in a linear
fashion in measurement z.
 Estimation, Tracking, and Data Association 99

There are three important and desirable characteristics of parameter estimates

and estimators. First, the estimate should be unbiased, that is, the estimation er-
ror should have a mean value of zero. Second, the variance of the estimation er-
ror should be reasonably small or at least understood. Equations (5.2) and (5.3)
are mathematical representations of these concepts:


E{ x− x } = E{ x } = 0 , (5.2)

and

E { ( x − x ) } = E { x } = σ
2 2

x
2
. (5.3)

A third desirable characteristic of an estimator is that the variance of the esti-

mate approaches its theoretical lower bound asymptotically as more measure-
ments are available. Estimators that achieve this property are referred to as
efficient estimators.
There exists a theoretical lower bound on the value of estimation error vari-
ance referred to as the Cramer-Rao bound. When N statistically independent
measurements of x are averaged with a given signal-to-noise ratio (SNR) per ob-
servation, the Cramer-Rao bound on the error variance of the estimate is approx-
imately given by:

σ CRB
2
=
( ∆x ) 2 , (5.4)
N SNR

where ∆x is the measurement resolution of x and N is the number of statistically

independent measurements of x. As will be seen, the lower bound in equation
(5.4) is important for predicting estimation performance, and specifically for pre-
dicting tracking accuracy.

5.3 THE RADAR TRACKING FUNCTION

The radar tracking function is illustrated using the block diagram in Figure 5.1.
As can be seen, the antenna is steered in the direction of the predicted target po-
100 Chapter 5

Figure 5.1 Block Diagram of the Radar Tracking Function

sition, track waveforms are transmitted to the target, target echoes are received,
these measurements are associated with current targets, the target state vectors
are updated using the new measurement, the state vector is predicted to the time
of the next track update, and this information is used to steer the antenna start-
ing the process again. This functional sequence is referred to as a tracking loop,
and executing it is referred to as closing the track loop.
The two major sub-functions that comprise the radar tracking function are the
data association algorithm and the tracking filter. The remaining functions in
Figure 5.1 are performed by the radar scheduler (schedule the track waveform
and steer the antenna), the radar transmitter (transmit the track waveform), and
the receiver (receive the target echo).
The waveforms and signal processing associated with the tracking function
are as follows. First, narrowband waveforms are used for tracking, that is, radio
frequency bandwidths in the range of 5 to 50 MHz. The signal processing per-
formed is the standard pulse matched filtering, followed by range and ampli-
tude interpolation and peak detection. Monopulse processing is used to estimate
target sine space angles u and v. For cases of low-elevation tracking, multipulse
waveforms and processing are employed (e.g., MTI or pulse-Doppler) to miti-
gate clutter backscatter effects.
The subsequent sections in this chapter are devoted to describing the tracking
filters, data association algorithms, and the specific tracking of air, ballistic mis-
sile, and surface targets. However, first the subject of the coordinate systems and
transformations between them will be discussed briefly in the next section.
 Estimation, Tracking, and Data Association 101

5.3.1 Coordinate Systems

Coordinate systems are an important aspect of the radar tracking function. Since
phased-array radars are the focus of this book, the measurement coordinate sys-
tem is a special form of spherical coordinates: range and the two direction-co-
sines u and v, referenced to the antenna array boresight, referred to as (R, u, v)
coordinates.
However, most tracking systems use an inertial frame of reference to estimate
target position, typically expressed in Cartesian coordinates: (x, y, z). This refer-
ence system is selected since most target classes move most naturally in Carte-
sian coordinates, not in spherical coordinates. Therefore, at a minimum,
transformations between (R, u, v) and (x, y, z) are necessary to relate measure-
ments to the target states used by the tracking filters. These effectively convert
the radar measurements to Cartesian array face coordinates, and then translate
and rotate them to Cartesian inertial coordinates. For fixed-antenna arrays on
stationary platforms, this is all that is required.
The situation is more complicated for moving antenna arrays, whether
mounted on pedestals or mounts, or on a moving platform such as a ship, air-
craft, or missile. In these cases additional rotations and translations are needed
to account for the pedestal and/or platform motion relative to the Cartesian in-
ertial frame in which target tracks are updated and predicted.
Therefore, a sequence of coordinate transformations and translations are a
fundamental component of the radar tracking process. In most cases, Cartesian
inertial coordinates will be used in discussing the tracking algorithms through-
out the remainder of this chapter, with the exception of data association algo-
rithms, which must relate measurements to state vectors. However, it is
important to not lose sight of these coordinate systems when designing and ana-
lyzing radar tracking systems.

5.4 TYPES OF TRACKING FILTERS

5.4.1 Fixed-Gain Filters
There are many types of tracking filters in use for radar tracking applications.
However, there are two fundamental classes of tracking filters: (i) fixed-gain fil-
ters and (ii) computed-gain filters. The first class is the simplest and its imple-
102 Chapter 5

mentations use the least data processing throughput per tracked target. The
most common of these are the α – β and α – β – γ or alpha-beta and alpha-beta-
gamma filters. The mathematical forms are:

T
x  x   α k +1 0  x
 x
 = z −
 k +1      +  , (5.5)
  k +1   x
 k  0 β k +1   x


and

T
x  x   α k +1 0 0  x
 x
    
  =  z k + 1 −  x
  0 β k + 1 0  +  x
 , (5.6)
x  k + 1



 x  k 


0
 0 γ k + 1  
x


where z is the measurement vector at the k + 1 sample time, and α, β, and γ are
the fixed or precomputed weights for sample time k + 1. As can be seen, since
only a small number of adds, subtracts, and multiplies are required for each
track update, the computational requirements for these filters are small.

5.4.2 Computed-Gain Filters

5.4.2.1 Kalman Filters. There are many forms of computed-gain tracking filters.
However, the most common and widely used is the Kalman filter (KF) type.
These filters are in the statistical filtering class, that is, they embody dynamical
models of the specific target motion and use these to propagate the expected-
value of the state estimates and the covariance matrix of the estimation errors.
The gain computations are where the Kalman filters incur the majority of neces-
sary computations. The state vector update equations are very similar to those in
equations (5.5) and (5.6).
The Kalman filter, references [1] through [4], is the optimal filter if the target
dynamics and the measurement-to-state relationships are linear. If one or both of
these relationships are nonlinear, then the Kalman filter is the optimal linear fil-
ter. Theoretically, an optimal nonlinear filter does exist; however, there is no sys-
tematic method to determine its form. Hence, this is the reason that some form
 Estimation, Tracking, and Data Association 103

of the Kalman filter is used in many if not most target tracking applications. In
most cases, due to the non-linearities mentioned above, a special form of the
Kalman filter is used in these cases. It is referred to as the Extended-Kalman fil-
ter (EKF). The EKF is a Kalman filter matched to linearized version of the equa-
tions of state dynamics and observations.
The fundamental concept behind the Kalman filter is the minimization of the
estimation error in the mean-squared sense, and so the Kalman filter is often re-
ferred to as a minimum mean-squared estimate (MMSE). Figure 5.2 is a flow-
chart of the sequence of computations required to implement one track update
and prediction of the target state and error covariance matrix.
It should be noted that the Kalman filter is a recursive estimator, that is, it up-
dates its estimates upon receipt of each new measurement. To illustrate the idea
of recursive estimation, consider the problem of recursive estimating the arith-
metic mean of a sequence of values. Figure 5.3 shows the estimation sequence.
In a similar fashion the Kalman filter’s recursive formulation can be derived.
This is shown in Figure 5.4. This class of filter is often referred to as a predictor-
corrector form, as indicated in Figure 5.4d.
Practical tracking issues that the Kalman filter does not handle include:

• Nonlinear motion models (use linearization, i.e., the extended KF [EKF])

• Nonlinear measurement equations (use EKF and a de-biased consistent transfor-
mation [polar-to-Cartesian])
• Unknown inputs to dynamic (system) equation and/or mode changes (differ-
ent motion models, e.g., uniform vs. acceleration or turn)
• Correlated noises (auto and cross-correlated)
• Unknown sensor resolution and multipath propagation
• Unknown number of targets
• Unknown origin of measurements: data association uncertainty.

However, due to its optimality, simplicity, and systematic handling of corre-

lated target dynamics and correlated measurement errors, and since these minor
drawbacks can be mitigated by proven techniques, such as using the Extended-
Kalman filter for nonlinear problems, addition of process noise (the filter’s un-
certainty of the true underlying dynamical models), increasing update rates to
104

Figure 5.2 Kalman Filter Processing Sequence

 Figure 5.3 Recursive Estimate of Arithmetic Mean
105
106

Figure 5.4a Derivation of Recursive Kalman Filter Equations

 Figure 5.4b Derivation of Recursive Kalman Filter Equations (continued)
107
108

Figure 5.4c Derivation of Recursive Kalman Filter Equations (continued)

 Figure 5.4d Derivation of Recursive Kalman Filter Equations (continued)
109
110 Chapter 5

ameliorate residual nonlinear effects, and so on, it is the ubiquitous “workhorse”

of tracking filters.

5.4.2.2 Interacting Multiple-Model Filters. Interacting multiple-model (IMM) filters

consist of a bank of parallel Kalman filters that implement different target dy-
namical models. These are blended or fused based on a Bayesian approach as
shown in Figure 5.5. The IMM filters provide the capability for a tracking filter to
use different target models during a target’s trajectory to minimize tracking er-
rors. Unlike tracking approaches like MHT, which propagate all possible target
hypotheses at each track update, the IMM only propagates a single step at each
update. This yields a computational load of approximately M times that of a sin-

Figure 5.5 IMM Filter Block Diagram

 Estimation, Tracking, and Data Association 111

gle Kalman filter, where M is the number of target models employed. The de-
tailed filtering logic is described in the following paragraphs.
The steps of the IMM approach are given below.

1. Model Conditioned Reinitialization

Each EKF is reinitialized with a composite state and covariance whose con-
stituent components are weighted by the conditional probability that the tar-
get is transitioning to the particular model.

The conditional probability that the target is transitioning to (or remains

as) model i is:

( k − 1)
π ji µ (jk −1)
µ = , where
j |i
µi( k , k −1)

M
µ ( k , k − 1)
i
= ∑ π ji µ (jk −1) is the total probability of being in model i,
j =1

πij is the a priori Markov transition probability that the target is transitioning
from model i to model j, and M is the number of models.

The states and covariance matrices are mixed using the conditional proba-
bilities for each respective model as weights.

M
x (i )
= ∑ µ (jk|i −1) x (j
k − 1, k − 1 k − 1, k − 1)
, is the state used to reinitialize the ith model.
j =1

M
Pi( )
= ∑ µ (jk|i −1)  Pi(
k − 1, k − 1)
+ ∆P(ji
k − 1, k − 1 k − 1,k − 1) 
, is the covariance matrix used
j =1
 
to reinitialize the ith model, where:

∆Pij = ( x i − x j )( x i − x j )T is added to account for the differences between

models i and j.
112 Chapter 5

2. Model-Conditioned Filtering and Prediction

Update the state and covariance of all EKFs using the new measurement z k
for the (k + 1)th innovation.

3. Model Probability Update

The composite state and covariance matrix are blended based on the poste-
rior probability for each model, given below:

Lki µi( k , k −1)

µ (k )
i
= M
where,
∑L µ k
j
( k , k − 1)
j
j =1

M
µi( k , k −1) = ∑ π ji µ (jk −1) is the total probability of model i calculated in step 1.
j =1

1
Li = exp( −ξi2 / 2) is the likelihood of the target being in model
( 2π )
m
Si
i, and

ξi2 = yiT Si−1 y i is the chi-square statistical distance of the measured state
from the predicted state.

y =  z k − H k ( x k , k −1 )  ,

Si =  H k Pk , k −1 H kT + R  ,

where m is the dimension of the state.

4. Estimate Fusion

This step produces the output of the tracking system used to determine the
associated measurement.
 Estimation, Tracking, and Data Association 113

M
x ( ) = ∑ µi( k ) x (i ) , is the state used to reinitialize the ith model.
k ,k k ,k

i =1

M
(k ,k )
= ∑ µi( k )  Pi( ) + ∆Pi( )  is the covariance used to reinitialize the ith
k ,k k ,k
P
i =1
 
model, where the term:

∆Pi = ( x − x i )( x − x i )T corrects the covariance matrix for the difference be-

tween the fused estimate and model i.

The output of the IMM tracking system is the composite state and covari-
(k ,k ) (k ,k ) 
ance  x ,P , which are used in the association algorithm to de-
 Composite
termine the measurement for the next innovation.

5.5 DATA ASSOCIATION ALGORITHMS

There are two fundamental classes of data association algorithms: (i) Non-Baye-
sian, and (ii) Bayesian approaches. These will be discussed in the following sec-
tions.

5.5.1 Nearest Neighbor

The “Achilles heel” of tracking is the correct association of new radar measure-
ments to their originating target. Good data association, therefore, is the neces-
sary condition for achieving good target tracking performance. The data
association problem is especially challenging in environments with residual
clutter echoes, densely spaced targets or target complexes, missed detections,
and high false alarm rates. Therefore, the type of data association algorithm
needed for a given tracking problem depends on these factors.
In benign situations with a few widely spaced targets, very high probability
of detection, very low probability of false alarm, in a clear environment (i.e., no
114 Chapter 5

clutter, jamming, or other interference), almost any data association algorithm

can be made to operate adequately. If these conditions are “guaranteed” to be
the case, then a simple data association algorithm can and should be used. One
such non-Bayesian algorithm is the nearest-neighbor (NN) technique. This al-
gorithm updates each track with the new target detection closest to it (in the
statistical “distance” sense). The NN algorithm can be expressed mathemati-
cally as:

( Rtrack − Rmeasured ) 2 +
(θtrack − θ measured ) 2 +
(φtrack − φmeasured ) 2 ≤ D , (5.7)
σ R2 σ θ2 σ φ2

where the sum of normalized squared-errors in range, azimuth angle, and eleva-
tion angle, for example, is compared to a threshold and the track yielding the
smallest of those sums for a given target measurement vector compared to the
track state vector quantities would be updated using that measurement. This
process is repeated for all measurements.
If the sum in equation (5.7) exceeds the threshold, D, then no measurement-to-
track assignment is made, since exceeding D corresponds to a very small proba-
bility of correct target-to-track association. If the errors (differences in the nu-
merators) in equation can be modeled as zero-mean Gaussian random variables,
then the sum results in a chi-square distribution with three degrees-of-freedom.
This allows D to be calculated to edit associations with any arbitrary probability
of false association. When using a Kalman-type filter, the estimation error vari-
ances in the denominators can be obtained from the filter’s computed error co-
variance matrix.

5.5.2 Probabilistic Data Association

The Probabilistic Data Association (PDA) algorithm uses a computed likelihood
or estimated association probabilities to aid in solving the return-to-track prob-
lem rather than a simple proximity rule as for the NN algorithm. It is a Bayesian
approach to the data association problem. Although it represents a higher com-
putational load, it provides superior association performance in the presence of
densely spaced targets (or clutter, etc.) to the NN method.
 Estimation, Tracking, and Data Association 115

At each step in the association process, the PDA computes probabilities of cor-
rect measurement-to-track association for each measurement-track pair. The pair
with the highest probability determines the assignment of that measurement.
This process is repeated for all measurements. The detailed equations and pro-
cessing logic are provided in detail in references [1] through [3] and will not be
repeated here.

5.5.3 Joint Probabilistic Data Association

The Joint Probabilistic Data Association (JPDA) algorithm also uses a computed
likelihood or estimated association probabilities to aid in solving the return-to-
track problem rather than a simple proximity rule as for the NN algorithm. It is
again a Bayesian approach to the data association problem. It represents a higher
computational load compared to the simpler PDA but provides superior associ-
ation performance in the presence of densely-spaced targets (or clutter, etc.) to
the PDA method.
At each step in the association process, the JPDA computes probabilities of
correct measurement-to-track association for each measurement-track pair con-
sidering all possible pairings in a joint-probabilistic sense. Similar to the PDA
approach, the highest probability determines the assignment of measurements.
The detailed equations and processing logic for JPDA are also provided in detail
in references [1] through [3] and are not provided here.

5.5.4 Nearest-Neighbor JPDA

This technique developed by R. Fitzgerald at Raytheon is a combination of the
NN and JPDA as its name suggests. It involves preprocessing the measure-
ments using the NN to edit out unlikely measurement-track pairs. The JPDA
algorithm is then applied to those pairs surviving this screening process. This
can yield excellent performance at much lower computational cost than JPDA
alone. Many phased-array radars use this data association technique. It pro-
vides association performance that can approach that of the more theoretically
optimal multiple-hypothesis track (MHT) association technique in certain situ-
ations, while being much less computationally-intensive than MHT methods.
MHT is briefly discussed next.
116 Chapter 5

5.5.5 Multiple-Hypothesis Track

Multiple-hypothesis track is also a Bayesian approach to data association. Un-

like the previously described methods that compare measurement-track pairs
using statistical distance or likelihood measures, MHT creates and maintains a
history of all possible data association hypotheses, creating a new and larger set
of hypotheses at each track update time. As can be imagined, if no “hypothesis
pruning” is performed, the number of hypotheses maintained grows exponen-
tially as a function of time. Hence, use of a poorly designed MHT approach can
have severe computational load consequences.
The source of MHT’s optimality is also the source of its computationally in-
tense nature. Since, in the unconstrained MHT approach, all possible measure-
ment-to-track hypotheses are carried along, the correct hypothesis is available to
the algorithm (the issue is selecting the correct one) and it can therefore be con-
sidered to be theoretically near-optimal. However, unless some pruning of un-
likely hypotheses is performed, MHT is not a practical solution to the data
association problem.
That being said, MHT, with suitable pruning logic, has been successfully used
for many tracking applications. [3] is an excellent source to learn more about
MHT and its practical applications.

5.5.6 Other Assignment Algorithms

Although the algorithms described in Sections 5.1 through 5.5 are commonly im-
plemented techniques for data association, there exist many other approaches,
and more continue to be developed. Table 5.1 is a partial listing of current tech-
niques, including other data association algorithms other than those previously
discussed compiled by F. Daum [3, 8, 9, 10].
As can be seen, the algorithms are rated based on a number of performance
criteria, including:

• Ability to handle unresolved measurement data

• Performance in densely spaced target (e.g., clutter) environments

• Computational complexity (i.e., computer throughput usage).

 Table 5.1 Partial List of Data Association Algorithms, after [3, 8, 9, 10]

Relative Performance in Dense Computational

Multiple Target Environments Complexity

Time horizon
Considered Number of Data Unsolved Data
(no. of Association Modeled in Unresolved Resolved Exact Approximate
Algorithm Samples) Hypotheses Algorithm Data Data Solution Solution

Nearest neighbor 1 1 No Poor Poor Low Low

Probabilistic data 1 1 No Poor Fair Low Low

association (PDA)

Joint probabilistic data 1 1 No Fair Fair Poly Medium

association (JPDA)

Nearest-neighbor JPDA 1 1 No Fair Good Poly Medium

Assignment 1 1 No Fair Good Poly Medium

Multidimensional N 1 No Good Excellent Poly High

assignment

Multiple hypothesis N Many No Good Excellent Exp High

tracker (MHT)

Koch MHT N Many Yes! Excellent! Excellent Exp High

117
118 Chapter 5

5.6 TRACKING AIR TARGETS

Tracking air targets poses several challenges. This is in part due to the variety of
target types contained within the air target class, including:

• Manned aircraft

• Unmanned aerial vehicles (UAVs)

• Helicopters

• Cruise missiles.

Tracking manned aircraft is problematic compared with tracking ballistic mis-

siles for the reason that the pilot in the aircraft can choose to maneuver at any
time, and can select a number of distinct types of maneuvers. The ballistic mis-
sile after rocket burnout is effectively traveling in a fixed-parameter parabolic
trajectory with only gravity acting upon it until it reenters the earth’s atmo-
sphere and is subject to drag deceleration.
UAVs may fly trajectories similar to a manned aircraft or not depending on
the UAV capabilities and the control strategy. Similarly, there are several
types of cruise missiles, some that fly at very low altitudes at subsonic speeds
and can be terrain-following, and some that can be launched from high alti-
tudes, travel at supersonic speeds, and can dive at steep angles when attack-
ing.
For these reasons, a single target model cannot be used for tracking the broad
class of air targets. The IMM filter described in Section 4.2.2 is one method to ac-
commodate a broad range of possible target dynamics. Consider the following
distinct behaviors for air targets:

• Constant-speed, straight and level flight

• Constant-acceleration, straight and level flight

• 3-G turn in plane

• 3-G climb or dive.

Possible state transitions (Figure 5.6) might be:

 Estimation, Tracking, and Data Association 119

Figure 5.6 Possible Air Target State Transitions

π 11 π 12 0 0 
 
 0 π 22 π 23 0 
Π= , where
0 0 π 33 π 34 
 
 0 0 π 43 π 44 

π11 = probability of staying at constant-speed,

π12 = probability of transitioning constant-speed to constant-acceleration,
π22 = probability of remaining at constant-acceleration,
π23 = probability of transitioning constant-acceleration to a 3-G turn,
π33 = probability of staying in the 3-G turn,
π34 = probability of transitioning 3-G turn to 3-G climb or dive,
π43 = probability of transitioning from 3-G climb to dive to 3-G turn,
π44 = probability of remaining in a 3-G climb or dive.

5.7 TRACKING BALLISTIC MISSILE TARGETS

The basic ballistic missile tracking problem is illustrated in Figure 5.7.
120 Chapter 5

Figure 5.7 Basic Ballistic Missile Initial Trajectory

As can be seen, there are five phases during early flight:

• Initial rocket burn

• Pitch over maneuver

• Object separations

• Gravity turn

• Boost cessation.

After the missile reaches apogee and enters the descent portion of its flight, ad-
ditional phases of flight are:

• Ballistic flight under force of gravity

• Slow-down due to drag when within the earth’s atmosphere.

 Estimation, Tracking, and Data Association 121

Use of an IMM tracking filter for ballistic missile targets is illustrated in Figure
5.8. The transition probabilities used in the IMM filter allow a well-organized
method of adding a priori knowledge of the target trajectory characteristics to be
incorporated into the tracking system. As missiles transition from boost to ballis-
tic, ballistic to reentry, and eventually transition from reentry to a possible ma-

Figure 5.8 Example of Parallel Filters in IMM Approach

122 Chapter 5

neuver and back to reentry a number of distinct state transitions can be defined.
The associated dynamical models are described in Table 5.2. Possible target state
transitions are given in the following transition matrix.

π 11 π 12 0 0 
 
 0 π 22 π 23 0 
Π= , where
0 0 π 33 π 34 
 
 0 0 π 43 π 44 

π11 = probability of remaining in the boost phase,

π12 = probability of transitioning from boost to ballistic,
π22 = probability of remaining in the ballistic phase,
π23 = probability of transitioning from ballistic to reentry,
π33 = probability of remaining in the reentry phase,
π34 = probability of transitioning from reentry to maneuver,
π43 = probability of transitioning from maneuver to reentry,
π44 = probability of remaining in the maneuver phase.

Figure 5.9 illustrates the track transitions with a state transition diagram. The envi-
ronment in which ballistic missile target tracking occurs is much more challenging
than for air target tracking due to the large number of possible closely spaced ob-
jects associated with the missile complex under track. Figure 5.10 illustrates a sim-
plified view of the source multiple objects to be tracked in this environment.

Table 5.2 Ballistic Missile Dynamics Models versus Trajectory Phase

Object Type Dynamics Models Key Required Parameters

Boosting object within Earth’s Acceleration, drag, and gravity Missile acceleration during
atmosphere are modeled boost, drag parameters, and
missile mass

Ballistic object within Earth’s Drag and gravity are modeled Missile drag parameters, mas,
atmosphere and velocity at booster burnout
(Vbo)

Ballistic object outside Earth’s Gravity is modeled Missile mass, and velocity at
atmosphere exit of Earth’s atmosphere
 Estimation, Tracking, and Data Association 123

Figure 5.9 Possible State Transition Diagram for Ballistic Missile Targets

Figure 5.10 Simplified Ballistic Missile Tracking Environment

124 Chapter 5

5.8 TRACKING SURFACE TARGETS

Surface targets are the least challenging from a target-dynamics perspective.
However, since they are typically immersed in sea clutter, consistent detection of
ship targets becomes problematic. Also, the occurrence of detections due to re-
sidual clutter after Doppler processing via either MTI or pulse-Doppler process-
ing may be “spiky” or noise false alarm-like in nature.
Due to the possibility of slowly moving real target and clutter returns in close
proximity (and perhaps unresolved in range), the data association problem is
more challenging for ship-tracking, especially in high sea-state conditions. Use
of MHT techniques is highly recommended to mitigate loss of slow-moving
ships in heavy sea clutter, as well as to possibly implement track-before-detect
methods when signal-to-clutter ratios are marginal after Doppler processing.
Due to the slow target dynamics involved for ship tracking, relatively low-rate
track-while-scan techniques can usually be employed with fixed-gain or simpli-
fied Kalman filters, as opposed to the much higher data rates and more complex
tracking filters required for air and ballistic missile tracking.

5.9 REFERENCES
[1] Y. Bar-Shalom, Multitarget/Multisensor Tracking: Applications and Advances, Artech
House, 2000
[2] Y. Bar-Shalom & X. Li, Multitarget-Multisensor Tracking, YBS, 1995
[3] S. Blackman & R. Popoli, Design and Analysis of Modern Tracking Systems, Artech
House, 1999
[4] E. Brookner, Tracking and Kalman Filtering Made Easy, Wiley-Interscience, 1998
[5] A. Gelb, Applied Optimal Estimation, MIT Press, 1974
[6] S. Haykin & A. Steinhardt, Adaptive Radar Detection and Estimation, Wiley-Inter-
science, 1992
[7] H. Van Trees, Detection, Estimation and Modulation Theory, Part 1, Wiley-Interscience,
2001
[8] F. Daum, “A System Approach to Multiple Target Tracking,” Chapter 6 in Multi-Tar-
get Multi-Sensor Tracking, Volume II, edited by Yaakov Bar-Shalom, Artech House
1992
[9] W. Koch & G. van Keuk, “Multiple Hypothesis Track Maintenance with Possibly
Unresolved Measurements,” IEEE Transactions on Aerospace & Electronic Systems, Vol.
33, pages 883–892, 1997
[10] F. Daum, “Book Review: Multiple Target Multisensor Tracking,” IEEE AES Systems,
September 1996