Survey of Maneuvering Target Tracking. Part I. Dynamic Models

I.
INTRODUCTION
The key to successful target tracking lies in the

Survey of Maneuvering effective extraction of useful information about the
target’s state from observations. A good model of
Target Tracking. the target will certainly facilitate this information
extraction to a great extent. In general, one can say
Part I: Dynamic Models without exaggeration that a good model is worth a
thousand pieces of data. This statement has an even
stronger positive connotation in target tracking where
observation data are rather limited. Most tracking
algorithms are model based because knowledge of
X. RONG LI, Senior Member, IEEE target motion is available and a good model-based
tracking algorithm will greatly outperform any
VESSELIN P. JILKOV, Member, IEEE
University of New Orleans model-free tracking algorithm if the underlying model
turns out to be a good one. As such, it is hard to
overstate the importance of the role of a good model
here.
This is the first part of a comprehensive and up-to-date survey Various mathematical models of target motion
of the techniques for tracking maneuvering targets without have been developed over the past three decades.
addressing the so-called measurement-origin uncertainty. It
They are, however, scattered in the literature. Many
of them have never appeared in any periodical in the
surveys various mathematical models of target motion/dynamics
public domain. As a result, few people have a good
proposed for maneuvering target tracking, including 2D and 3D
knowledge of these models. This is partly due to a
maneuver models as well as coordinate-uncoupled generic models
lack of a comprehensive survey. The importance of
for target motion. This survey emphasizes the underlying ideas such a survey for both practitioners and researchers
and assumptions of the models. Interrelationships among models in the tracking community is evident. The single best
and insight to the pros and cons of models are provided. Some source so far is, in our opinion, the recent book by
material presented here has not appeared elsewhere. Blackman and Popoli [1], which is nonetheless far
from complete. Some more or less standard models
for target motion can be found in established books
CONTENTS on target tracking and/or estimation, such as [2–12].
I. Introduction This paper is the first part of a comprehensive and
II. Mathematical Models for Maneuvering Target Tracking up-to-date survey of the techniques for maneuvering
III. Nonmaneuver Models target tracking. The survey is an ongoing project.
IV. Coordinate-Uncoupled Maneuver Models The conference versions of its first several parts
V. 2D Horizontal Motion Models have appeared in [13–17]. It is well known that the
VI. 3D Motion Models so-called measurement-origin uncertainty and target
VII. Concluding Remarks
motion uncertainty are two major challenges in target
tracking. To limit the scope of the work, this survey
References
deals only with the second uncertainty, leaving the
techniques unique for the data-association problems
untouched.
Target detection, tracking, and recognition are
closely interrelated areas, with significant overlaps.
It is not easy to draw a clear line to separate them. To
Manuscript received September 11, 2002; revised April 22, 2003; be relatively more focused, this part covers mainly
released for publication July 31, 2003. dynamic models of a “point target,” that is, those
IEEE Log No. T-AES/39/4/822061. of the dynamic (temporal) behaviors, rather than
spatial characteristics, of a target. While many of
Refereeing of this contribution was handled by P. K. Willett.
these models are also useful for target detection and
This research was supported in part by ONR Grant recognition, this survey is only concerned with their
N00014-00-1-0677, NSF Grant ECS-9734285, and NASA/LEQSF value for target tracking. This of course does not
Grant (2001-4)-01.
prevent us from developing or applying a model that
Authors’ address: Dept. of Electrical Engineering, University of describes both the temporal evolution and spatial
New Orleans, New Orleans, LA 70148, E-mail: (xli@uno.edu). characteristics of a target.
Needless to say, target dynamic models
0018-9251/03/$17.00
c 2003 IEEE and tracking algorithms have intimate ties. The
IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 39, NO. 4 OCTOBER 2003 1333
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY MADRAS. Downloaded on November 18,2023 at 10:56:45 UTC from IEEE Xplore. Restrictions apply.
applicability of a target dynamic model for a practical of additive noise,
problem can hardly be evaluated without referring to
the corresponding tracking algorithms used. In other xk+1 = fk (xk , uk ) + wk (1)
words, some target models and tracking algorithms zk = hk (xk ) + vk (2)
work well jointly. To be more focused and concise,
however, this interdependence is largely ignored where xk , zk , and uk are the target state, observation,
here. and control input vectors, respectively, at the discrete
This survey emphasizes the underlying ideas time tk ; wk and vk are process and measurement
and assumptions of the models. This should help noise sequences, respectively; and fk and hk are some
the reader understand not only how these models vector-valued (possibly time-varying) functions. Such
work but also their pros and cons. It is hoped that a discrete-time model is often obtained1 by discretizing
a distinctive feature of this survey is that it reveals (sampling) the following continuous-time model [2]2
well the interrelationships among various models. _ = f(x(t), u(t), t) + w(t),
x(t) x(t0 ) = x0 (3)
However, the reader should keep in mind that much
of such discussion is based on our personal views and z(t) = h(x(t), t) + v(t) (4)
preferences, not always accurate or unbiased, although where xk = x(tk ) and it is usually assumed3 that
a great deal of effort has been made toward this zk = z(tk ), vk = v(tk ), hk (xk ) = h(x(tk ), tk ). The control
goal. In addition to such discussions, some material input is often assumed (approximately) piecewise
included in this survey has not appeared elsewhere. constant with uk = u(t), tk t < tk+1 when discretizing
Regrettably, many important issues associated a continuous-time system. In target tracking, the
with the target dynamic models, particularly those control input u is usually not known. Note that
of implementation, cannot be discussed (at least to
a desirable degree) due to space limitation as well wk = w(tk ), fk (xk , uk , wk ) = f(x(tk ), u(tk ), w(tk ), tk ):
as our background and experience. Nevertheless, we
In fact, it is often more appropriate to use the
would appreciate very much receiving comments and
following mixed-time models for most tracking
any missing material that should be covered in this
problems
survey.
This paper is a heavily revised and extended _ = f(x(t), u(t), t) + w(t),
x(t) x(t0 ) = x0 (5)
version of [13]. Its remaining part is organized
zk = hk (xk ) + vk (6)
as follows. Section II gives briefly state-space
representations of target dynamics and observation because while observations are usually available
system. Section III describes nonmaneuver models. only at discrete time instants, the target motion is
The presentation of dynamic models of target more accurately modeled in continuous time. For
maneuvers breaks down as follows: while Section IV example, target motions should not depend on how
deals with those that are uncoupled along different and when samples are taken, which is often the case,
spatial coordinates, 2D and 3D coupled models are however, for a discrete-time model. For a similar
reviewed in Sections V and VI, respectively. The final reason, a discrete-time equivalent model is usually
section provides concluding remarks.
1 One so obtained is referred to as a discretized model regardless
II. MATHEMATICAL MODELS FOR MANEUVERING of its equivalence to the continuous-time system. It is termed
TARGET TRACKING a discrete-time equivalent if the discretization is exact and the
effects of the continuous-time and discrete-time noise processes are
The primary objective of target tracking is to equivalent, which is the case for a linear system without the control
input u with the standard discretization (i.e., sample and hold).
estimate the state trajectories of a target—a moving Note, however, that not all discrete-time models can be obtained
or movable object. Although a target is almost never by discretizing a continuous-time system. A model defined directly
really a point in space and the information about in discrete time following the same principle as one in continuous
its orientation is valuable for tracking, a target is time is referred to as a direct discrete-time counterpart here. We use
the term discrete-time version for either or both of discretized and
usually treated as a point object without a shape in
direct discrete-time models.
tracking, especially in target dynamic models. A target 2 We sacrifice rigor for readability: (3) is not actually well defined
dynamic/motion model describes the evolution of the because x_ = dx=dt may not exist at all since the “continuous-time
target state with respect to time. white noise” has infinite variance and is a specter in the cosmos
Almost all maneuvering target tracking methods that represents the nonexistent derivative of a Wiener process and
the like. We warn, however, that formal manipulation of (3) or (8)
are model based. They assume that the target motion
may easily lead to incorrect results. That is why in the mathematical
and its observations can be represented by some world (3) is replaced by dx(t) = f(x(t), u(t), t)dt + dw(t), x(t0 ) = x0
known mathematical models sufficiently accurately. and certain rules must be followed.
The most commonly used such models are those 3 Although discrete-time measurement z and thus the function h
k k
known as state-space models, in the following form and noise vk are in fact from sensors that do integration over time.
1334 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 39, NO. 4 OCTOBER 2003
more systematic and consistent, and is in many z axes, respectively, and [x,_ y, _ is the velocity
_ z]
cases probably preferable to the corresponding direct vector. When a target is treated as a point object,
discrete-time counterpart. the nonmaneuvering motion is thus described by the
The continuous-, discrete-, and mixed-time vector-valued equation x(t)_ = 0, where x = [x, _ z] .
_ y,
linear counterparts of the above models are the Note that z direction is treated differently because a
corresponding pairs of the following equations nonmaneuvering motion is assumed in the horizontal
xk+1 = Fk xk + Ek uk + Gk wk (7) x–y plane. In practice, this ideal equation is usually
_ = w(t) 0, where w(t) is white
modified as x(t)
_ = A(t)x(t) + E(t)u(t) + B(t)w(t),
x(t) x(t0 ) = x0 noise with a “small” effect on x that accounts for
(8) unpredictable modeling errors due to turbulence, etc.
The corresponding state-space model is given by, with
zk = Hk xk + vk (9) _ z] ,
_ y, y,
state vector x = [x, x,
z(t) = C(t)x(t) + v(t): (10) _ = diag[Acv , 0]x(t) + diag[Bcv , 1]w(t)
x(t) (11)
One of the major challenges for target tracking
where w(t) = [wx (t), wy (t), wz (t)] is a continuous-time
arises from the target motion uncertainty. This
vector-valued white noise process with power spectral
uncertainty refers to the fact that an accurate dynamic
density matrix diag[Sx , Sy , Sz ], Acv = diag[A2 , A2 ], and
model of the target being tracked is not available
to the tracker. Specifically, although the general Bcv = diag[B2 , B2 ] with6
form of the model (1) or (5) is usually adequate,
0 1 0
a tracker lacks knowledge about the actual control A2 = , B2 = : (12)
0 0 1
input u of the target, and possibly the actual form
of f, its parameters, or statistical properties of the The direct discrete-time counterpart of the above
noise w for the particular target being tracked. Target model is [2]
motion modeling is thus one of the first tasks for
maneuvering target tracking. It aims at developing a xk+1 = Fxk + Gwk = diag[Fcv , 1]xk + diag[Gcv , T]wk
tractable model that accounts well for the effect of = diag[F2 , F2 , 1]xk + diag[G2 , G2 , T]wk (13)
target motion.
In this paper, we describe the efforts and where
results in modeling the target motion for tracking a Fcv = diag[F2 , F2 ], Gcv = diag[G2 , G2 ]
maneuvering target without knowing its true dynamic 2
1 T T =2 (14)
behavior. Most of these efforts have been made along F2 = , G2 =
two lines: 1) approximate the actually nonrandom 0 1 T
control input u as a random process of certain wk = [wx , wy , wz ]k is a discrete-time white noise
properties, and 2) describe typical target trajectories sequence and T is the sampling interval. Note that
by some representative motion models with properly wx and wy correspond to noisy “accelerations” along
designed parameters. x and y axes, respectively, while wz corresponds to
Target motions are normally classified into noisy “velocity” along z axis. If w is uncoupled across
two classes: maneuver and nonmaneuver. A its components, then the nonmaneuvering motion
nonmaneuvering motion is the straight and level modeled by the above models is uncoupled across x,
motion at a constant velocity4 in an inertial reference y, and z directions. In this case, the covariance of the
system, sometimes also referred to as the uniform noise term in (13) is given by
motion. Loosely speaking, all other motions belong
to the maneuvering mode. cov(Gwk ) = diag[var(wx )Q2 , var(wy )Q2 , var(wz )]
4 (15)
T =4 T3 =2
III. NONMANEUVER MODELS Q2 = 3 2
:
T =2 T
It is well known that a point moving in our
This model is defined directly in discrete time and is
3D physical world can be described by its 3D
not entirely equivalent to the above continuous-time
position and velocity vectors. For instance,5 x =
_ y, y, _ can be used as a state vector of such
_ z, z] model.
[x, x,
The discrete-time equivalent of the above
a point in the Cartesian coordinate system, where
continuous-time model is [2]
(x, y, z) are the position coordinates along x, y, and
xk+1 = diag[F2 , F2 , 1]xk + wk (16)
4 Note that velocity is a vector and speed is its magnitude.
5 In this survey, the regular symbol x and z are used for simplicity
6 Forconvenience, we use the shorthand notation A =
to denote the state and measurement vectors, respectively, while the
positions along the x, y, and z axes are denoted by sans serif font diag[A1 , A2 , : :: , An ] to denote a block-diagonal matrix A, where Ai
(x, y, z). and A are not necessarily square matrices.
LI & JILKOV: SURVEY OF MANEUVERING TARGET TRACKING. PART I: DYNAMIC MODELS 1335
where constant-velocity, constant-acceleration, and
S Sy S
cov(wk ) = diag x Q̃2 , Q̃2 , z polynomial models.
T T T 2) Markov process models: The control input
4 (17) is modeled as a Markov process, which has a time
3
T =3 T =2
Q̃2 = : autocorrelation. This includes the well-known Singer
T3 =2 T2 model, its various extensions, and some other
Note the difference between the discrete-time models.
equivalent (16) and the direct discrete-time counterpart 3) Semi-Markov jump process models: The
(13). control input is modeled as a semi-Markov jump
In a 2D scenario where the altitude z is not process.
considered, the above models take the more popular Most target maneuvers are coupled across different
form, respectively, coordinates. For simplicity, however, many maneuver
models developed assume that this coordinate
_ = Acv x(t) + Bcv w(t)
x(t)
coupling is weak and can be neglected. This is
xk+1 = Fcv xk + Gcv wk (18) particularly the case for those that model the control
input u as a random process. As a consequence, we
xk+1 = Fcv xk + wk : need to consider only a generic coordinate direction.
_ and ẍ be the target position, velocity, and
Let x, x,
The above models (11), (13), (16), and (18)
acceleration along a generic direction, respectively.
are known as the continuous- and discrete-time
Specifically,
constant-velocity (CV) models, or more precisely,
“nearly-constant-velocity models.” Equation (18) ẍ(t) = a(t): (19)
is in fact a “(small) white acceleration model” The models discussed in this section differ in how the
since accelerations along x and y directions are function a(t) is defined.
modeled as (small) white noise. The term “(nearly) In this section, the state vector is always taken
constant-velocity model” emphasizes that these _ ẍ] along the generic direction, unless
to be x = [x, x,
accelerations are small. Note that the control input stated otherwise explicitly.
u is zero in the nonmaneuver models, although in
reality an actual thrust of the target may be present
A. White-Noise Acceleration Model
to balance other forces so as to maintain the motion.
Also, inclusion of any unnecessary component The simplest model for a target maneuver
(e.g., acceleration) in the state vector would degrade is the so-called white-noise acceleration model
tracking performance. [2]. It assumes that the target acceleration ẍ(t) is
an independent process (strictly white noise). It
IV. COORDINATE-UNCOUPLED MANEUVER
differs from the nonmaneuver model of Section
MODELS
III only in the noise level: the white noise process
The control input u responsible for a target w used to model the effect of the control input u
maneuver is primarily deterministic in nature and has a much higher intensity than the one used in a
most often unknown to the tracker. A natural way nonmaneuver model. A maneuver by its very nature
is to model it as an unknown, deterministic process aims at accomplishing a certain task and thus is
and estimate this process from measurement data rarely independent with respect to time. The main
during tracking. Such deterministic input models are attractive feature of this model is its simplicity. It is
the basis for the so-called input estimation method sometimes used when the maneuver is quite small or
(see, e.g., [16, 18–25]). Due to a lack of knowledge of random. It is also used in some maneuvering target
its dynamics, this unknown process is often assumed tracking techniques, such as the so-called noise-level
to be piecewise constant and treated as an unknown adjustment, discussed in a subsequent part of this
time-invariant parameter over a time window. The survey (see [16] for a preliminary version).
main difficulty then lies in the determination of the
input level and the instants at which the input jumps. B. Wiener-Process Acceleration Model
This method is covered in detail in a subsequent part
of this survey, of which [16] is a preliminary version. The second simplest model is the so-called
An alternative is to model the input u as a random Wiener-process acceleration model [2]. It assumes
process, which is in fact much more popular than the that the acceleration is a Wiener process, or more
above deterministic modeling. Models in this class generally and precisely, the acceleration is a process
proposed in the literature can be largely classified into with independent increments, which is not necessarily
three groups as follows. a Wiener process. It is also referred to simply as the
1) White noise models: The control constant-acceleration (CA) model or more precisely
input is modeled as white noise. This includes “nearly-constant-acceleration model.”
This model has two commonly used versions. The for a sampling period T, it would not be true in
first one, referred to as the white-noise jerk model, general for any other sampling period T unless it is
assumes that the acceleration derivative (i.e., “jerk”) a multiple of T: T = nT. Even if such periods exist,
_ is an independent process (white noise) w(t):
a(t) we would not be so lucky that one of them is used by
_ = w(t), with power spectral density Sw . The
a(t) chance.
_ =
corresponding state-space representation is x(t)
A3 x(t) + B3 w(t), where C. Polynomial Models

0 1 0 0 It is well known that any continuous target
trajectory can be approximated by a polynomial of
A3 = 0 0 1 , B3 = 0 : (20)
a certain degree to an arbitrary accuracy. As such, it
0 0 0 1
is possible to model target motion by an nth-degree
Its discrete-time equivalent is polynomial [2] in the Cartesian coordinates:

1 T T 2 =2 1
x(t) a0 a1 an wx (t)
xk+1 = F3 xk + wk , F3 = 0 1 T (21) t
y(t) = b0 b1 bn
.. + wy (t)
0 0 1 .
z(t) c0 c1 cn wz (t)
where tn =n!

T5 =20 T 4 =8 T3 =6 (25)

Q = cov(wk ) = Sw Q3, Q3 = T 4 =8 3
T =3 2
T =2 : with a certain choice of the coefficients ai , bi , ci , where
3
T =6 T 2 =2 T (x, y, z) are the position coordinates and (wx , wy , wz )
are the corresponding noise terms. Such an nth-degree
(22) polynomial model amounts to assuming the nth time
Note that Sw is the power spectral density, not the derivative of the position is (nearly) constant (i.e., the
variance, of the continuous-time white noise w(t). position deviation from such a constant nth derivative
The second version can be called Wiener-sequence motion is equal to the noise w). The CV and CA
acceleration model. It assumes that the acceleration models described above are special cases (for n = 1, 2,
increment is an independent (white noise) process. respectively) of this general nth-degree model with
An acceleration increment over a time period is the white noise w(t). Note that this model is coordinate
integral of the jerk over the period. This model is uncoupled if wx , wy , wz are uncorrelated. Also, an
most conveniently expressed in discrete time directly, nth-degree polynomial has (n + 1) parameters per
given by coordinate. That is why a model of an nth-degree
2 polynomial is often called an (n + 1)th-order model.
T =2 This model in its general setting does not appear

xk+1 = F3 xk + G3 wk , G3 = T : (23) very attractive for tracking for several reasons. Such
1 models are usually good for fitting to a set of data,
that is, for smoothing problem; however, the primary
Note that its noise term has a covariance different purpose of tracking is prediction and filtering, rather
from that of the white-noise jerk model: than fitting or smoothing. It is difficult to develop an
4 uncomplicated and efficient method to determine the
T =4 T3 =2 T2 =2
coefficients ai , bi , ci systematically in a general setting.
Q = cov(G3 wk ) = var(wk ) T3 =2 T2 =2 T : Nevertheless, many special polynomial models have
T2 =2 T 1 been developed for target tracking. In fact, most of
the models discussed in this section can be viewed as
(24)
special cases of this general polynomial model with
The above models are simple but crude. Actual different models for the noise w(t).
maneuvers seldom have (nearly) constant accelerations
that are uncoupled across coordinate directions. D. Singer Acceleration Model—Zero-Mean First-Order
As explained before, a continuous-time model is Markov Model
more accurate than its discrete-time versions for most
practical situations since a target moves continuously In stochastic modeling, a random variable is used
over time. The assumption of the direct discrete-time to represent an unknown time-invariant quantity,
CA model (i.e., the second version above) that the while an unknown time-varying quantity is modeled
acceleration increment ¢ak = ak+1 ak = a(tk+1 ) a(tk ) by a random process. As far as temporal properties
is independent across different sampling intervals are concerned, white noise constitutes the simplest
is hardly justifiable, except for its simplicity and class of random processes. The second simplest
mathematical tractability. Were this assumption true class is either the processes with independent
increments, represented by the Wiener processes, or
the so-called Markov processes, which include the
Wiener processes and white noise as special cases.
White noise is “isolated” in time since its value
at one time is uncoupled of any other time, while a
Markov process is “local” in time because its value
at one time depends on values at other times only Fig. 1. Ternary-uniform mixture pdf.
through its nearest neighbors. Consequently, it is
natural to consider a Markov process model whenever equal probability Pmax ; or accelerate or decelerate at a
white noise models are not good enough. rate uniformly distributed over (amax , amax ). It turns
The Singer model [26] assumes that the target out that
acceleration a(t) is a zero-mean first-order stationary a2
¾ 2 = max (1 + 4Pmax P0 )
Markov process with autocorrelation Ra (¿ ) = 3
E[a(t + ¿ )a(t)] = ¾ 2 e®M¿M , or equivalently, power where Pmax , P0 , and amax are design parameters. We
spectrum Sa (!) = 2®¾ 2 =(!2 + ®2 ). Such a process a(t) emphasize that Singer model is a maneuver model
is the state process of a linear time-invariant system7 and thus P0 should be probability of having a zero
_ = ®a(t) + w(t),
a(t) ®>0 (26) acceleration during a maneuver, rather than probability
of a nonmaneuvering motion. Note also that this
where w(t) is zero-mean white noise with constant ternary-uniform mixture distribution of acceleration
power spectral density Sw = 2®¾2 . Its discrete-time can obviously be used for other maneuver models and
equivalent is it is used here only to determine ¾ 2 .
ak+1 = ¯ak + wka (27) It is clear from (28)–(29) that in the limit:
where wka is a zero-mean white noise sequence with 1) As the maneuver time constant ¿ increases (i.e.,
variance ¾ 2 (1 ¯ 2 ) and ¯ = e®T . The state-space ®T decreases), the Singer model reduces to the CA
representation of the continuous-time Singer model model [more precisely, to the white-noise jerk model
is since cov(wk ) reduces to Q of (22) instead of Q of

0 1 0 0 (24)]. If the following direct discrete-time counterpart
x(t)

_ = 0 0 1 x(t) + 0
w(t): (28) of (28) were set up

0 0 ® 1 1 T T2 =2

Its discrete-time equivalent is xk+1 = 0 1 T xk + G 3 wk (30)
0 0 ¯
1 T (®T 1 + e®T )=®2
then its limit as ¿ increases would be the
xk+1 = F® xk + wk = 0 1 (1 e®T )=® xk + wk :
Wiener-sequence acceleration model. This relationship
0 0 e®T between the Singer and CA models makes sense since
(29) the deterministic part of the acceleration in the Singer
The exact covariance of wk is a function of ® and T model becomes constant in the limit as ¿ increases.
and can be found in, e.g., [26, 1, 2]. 2) On the other hand, as the maneuver time
The success of the Singer model relies on an constant ¿ decreases (i.e., ®T increases), the Singer
accurate determination of the parameters ® and ¾ 2 model reduces to the CV model. In this case, the
[27]. The parameter ® = 1=¿ is the reciprocal of acceleration becomes white noise.
the maneuver time constant ¿ and thus depends on Consequently, for a choice of 0 < ®T < , the
how long the maneuver lasts. For example for an Singer model corresponds to a motion in between
aircraft, ¿ 60 s for a lazy turn and ¿ 10–20 s of (nearly) constant velocity and (nearly) constant
for an evasive maneuver, as suggested in [26]. The acceleration. It should thus be clear that the Singer
parameter ¾ 2 = E[a(t)2 ] is the “instantaneous variance” model has wider coverage than the CV and CA
of the acceleration. It was proposed in [26] to model models.
the distribution of the acceleration by the following Many other models have been proposed (see,
ternary-uniform mixture (see Fig. 1): the target may e.g., [28–33]) which are equivalent to or are simple
move without acceleration with probability P0 ; variants of the Singer model. It has been applied
accelerate or decelerate at a maximum rate amax with in [33] for angular acceleration as well as linear
acceleration. It is interesting that the same model,
7 A stationary Markov process with a rational power spectrum (as is
except that ® is data-dependent and time-varying,
the case here) is equivalent to the state of an asymptotically stable
was obtained approximately in [34] for the angular
linear time-invariant system excited by strictly white noise: Every velocity components (pitch and yaw) of the relative
such process can be represented as the state of such a system and motion between the target and seeker based on
the state of such a system is such a Markov process. well-known relationships in classical mechanics.
The Singer acceleration model is a popular model excluding the time instants at which samples are
for target maneuvers (see, e.g., [35–40], [1, 2]). It taken since ā(t) is assumed piecewise constant. The
was the first model that characterizes the unknown discrete-time equivalent is
target acceleration as a time-correlated (i.e., colored) 2 ®T 2

T =2 (®T 1 + e )=®
stochastic process, and has served as a basis for % &
further development of many effective maneuver xk+1 = F® xk + # T (1 e®T )=® $ āk + wk
models. 1 e®T
The Singer model is in essence an a priori model
since it does not use online information about the (33)
target maneuver, although it can be made adaptive where F® was given in (29). These two equations
through an adaptation of its parameters ®, Pmax , P0 , differ from the Singer model only in the additional
T, and/or amax . We cannot reasonably expect any terms associated with ā(t) and āk , respectively. For
a priori model to have a remarkable effectiveness example, the noise w(t) and wk are identical to
for the diverse acceleration situations of actual those in the Singer model. Note also that this model
target maneuvers. As a consequence of its a priori corresponds to a Singer model with noise w of a
nature, the Singer model is also symmetric in that non-zero adaptive mean.
the assumed ternary-uniform mixture distribution A key underlying assumption of the “current”
of the acceleration is symmetric. One of the main model in [41] (but not so stated explicitly) is that
shortcomings of the Singer model stems from this āk+1 = âk , or more specifically,
symmetry; that is, the target acceleration has zero ¢ ¢
mean at any moment. Indeed, this is almost the best āk+1 = E[ak+1 M z k ] = E[ak M z k ] = âk (34)
one can do a priori without online information about
where z k stands for all measurements through time
the target maneuver. Otherwise towards where should
tk . This is questionable and can actually be avoided.
the mean be? However, nothing really prevents us
Since the last equation of (33) reads
from using online information if we can withstand a
little more sophistication. Several more sophisticated ak+1 = e®T ak + (1 e®T )āk + wka (35)
acceleration models have been proposed to remedy
this shortcoming. They are described next. we propose to improve the “current” model by
replacing āk+1 = âk with the following recursion
E. Mean-Adaptive Acceleration Model āk+1 = E[ak+1 M z k ] = e®T E[ak M z k ] + (1 e®T )āk

= e®T âk + (1 e®T )āk : (36)
An acceleration model, called the “current” model
by its authors, proposed in [41], is in essence a Singer Such a relationship makes better sense because āk+1
model with an adaptive mean; that is, a Singer model depends on the current estimate âk as well as past
modified to have a non-zero mean of the acceleration: information āk . However, what is resulting is no
a(t) = ã(t) + ā(t), where ã(t) is the zero-mean Singer longer a purely current model.
acceleration process, defined by (26), and ā(t) is the In the “current” model, the a priori (unconditional)
mean of the acceleration, artificially assumed constant probability density f(ak+1 ) of the acceleration
over each sampling interval. Such a non-zero-mean ak+1 at time k + 1 in the Singer model is replaced
acceleration satisfies by a conditional density f(ak+1 M âk ). Clearly, this
conditional density carries more accurate information
_ = ®ã(t) + w(t)
a(t) or _ = ®a(t) + ®ā(t) + w(t)
a(t) and is better to be used than the a priori density, as
explained before. The following conditional Rayleigh
(31) density (see Fig. 2) was proposed in [41] for ak+1 :
_ over any sampling interval. The
_ = ã(t)
since a(t) f(a M âk ) =
estimate âk of ak from all available online information
(amax a)2
2
(i.e., the sequence z k of observations through time k) !
ck (amax a) exp 1(amax a) âk > 0
2c2k
is taken to be the “current” value of the mean āk+1 ,
hence, the name. It is potentially more effective than !
c2 (a a (a a max )2
k
) exp
max 1(a a max ) âk < 0
the Singer model. 2ck2
The state-space representation of this model is where 1() is the unit step function; a max is the
negative acceleration limit, not necessarily equal to
0 1 0 0 0 amax ; and ck is an âk -dependent parameter. Note
_ =
x(t) 0 0

1 x(t) + 0 ā(t) + 0 w(t) that the ternary-uniform distribution assumption in
the Singer model was made only out of the need
0 0 ® ® 1
to calculate the error variance of the acceleration.
(32) Similarly, this conditional Rayleigh assumption was
Fig. 2. Conditional Rayleigh model.
Fig. 3. Asymmetric pdfs of normal acceleration.
made for the sole purpose of obtaining the variance of Gauss-Markov process b(t):
the acceleration prediction, which turns out to be
4¼ an (t) = ® + ¯e°b(t) (37)
(amax âk )2 âk > 0
¢ ¼
¾k2 = E[(ak+1 āk+1 )2 M âk ] = : where ®, ¯, ° are design parameters, depending on
4¼
(a max + âk )2 âk < 0 the particular target type; b(t) satisfies the equation
¼
of the Single model b(t)_ = (1=¿ )b(t) + w(t), where
This result is valid only under the simplifying
assumption that E[ak+1 M z k ] = E[ak+1 M âk ], which is ¿ is the correlation time constant of b(t), which in
the case when âk is a sufficient statistic of z k for ak+1 . general differs from that of an (t). As shown in Fig. 3,
Applications of the “current” model can be found three typical choices of ®, ¯, ° and the corresponding
in the literature, e.g., to a benchmark tracking problem (highly asymmetrical) probability density functions
[42]. (pdfs) were given in [43], including one [(®, ¯, °) =
(8, 4, 0:5)] that is considered typical of modern
piloted aircraft in evasive maneuvers.
F. Asymmetrically Distributed Normal Acceleration
Both pros and cons of this model stem from the
Model
fact that the parameters ®, ¯, ° are target-type specific.
Target acceleration can be decomposed along It is more accurate than the Singer model at the
two directions: lift (normal to the target velocity and cost of designing these parameters, which requires
wing directions for an aircraft) and thrust or drag knowledge of the target type, obtained either a priori
(along the velocity direction). Each component can or a posteriori. The shortcoming of the Singer model
be modeled by a time-correlated random process. with a symmetric pdf is overcome by the use of
The nonnormal component can be modeled well by additional information of target type in this model.
the Singer model. However, this is often not the case Note that the temporal correlation of the
for lift, which is usually the dominant one, especially acceleration in this model does not have a rational
during a maneuver. Its direction is determined by the power spectrum and is much more complicated
target aspect angle and its magnitude can be modeled than in the Singer model. It would be interesting
as a colored random process with an asymmetrical to compare this model with a Singer model of the
distribution. It was proposed in [43] that the normal normal acceleration an having the initial asymmetric
acceleration an (t) be modeled as an asymmetric and pdf an (t0 ) = ® + ¯e°b(t0 ) with the same parameters
deterministic function of a zero-mean first-order ®, ¯, ° as above.
G. Markov Models for Oscillatory Targets Note that the autocorrelation Ra (¿ ) is not periodic
anymore if ³ 1. More generally, if the target
It is not uncommon in reality that acceleration acceleration has autocorrelation
along one coordinate direction is oscillatory due to, S
¾2
e.g., wind sway or platform roll. The Singer model is Ra (¿ ) = a e³!n M¿ M cos(!n 1 ³ 2 M¿ M µ) (42)
not very suitable for such practical maneuvers. The cos µ
autocorrelation of such acceleration may be described then the state-space model [44, 12, 11] is given by
by (41) with Bw(t) replaced by [0, 0, b, c] w(t), where
Ra (¿ ) = ¾a2 e®M¿ M cos(!c ¿ )
S 2¾a2
b= ! sin(' µ)
= ¾a2 e³!n M¿ M cos(!n 1 ³ 2 ¿ ) (38) cos µ n
!n2 = ®2 + !c2 , ³ = ®=!n 2¾a2 3
c= ! sin(' + µ) (43)
cos µ n
where ¾a2 ,
®, !c , ³, and !n are average power, ³
damping coefficient, actual (damped) frequency, ' = tan1 S
damping ratio, and undamped natural frequency of 1 ³2
the target acceleration, respectively. and the corresponding transfer function of the
Such an acceleration process is the response of a prewhitening system is
second-order prewhitening system to zero-mean white
bs + c
noise input w(t) with power spectral density 2®¾a2 . It H(s) = : (44)
has transfer function s2 + 2³!n s + !n2
s + !n This model was utilized in a multiple-model tracker in
H(s) = 2 (39)
s + 2³!n s + !n2 [45].
The discrete-time equivalent of model (41) is
and is described by (see, e.g., [44, 12, 11])
found to be
_
a(t) 0 1 a(t) 1 1 T F13 F14
= + w(t)
_
d(t) !n2 2³!n d(t) (1 2³)!n 0

xk+1 = xk + wk (45)
(40) 0 F(!c , ®)
where d(t) = a(t)_ w(t), which can be called
0
acceleration drift. The state-space model for x =
_ ẍ, d] is
[x, x, where
2®!c e®T (2®!c cos !c T + (®2 !c2 ) sin !c T) !c e®T (!c cos !c T + ® sin !c T)
1
!c (®2 + !c2 ) !c (®2 + !c2 )

e®T (!c cos !c T + ® sin !c T) e®T
sin !c T
F(!c , ®) = 0
!c !c

(®2 + !c2 )e®T sin !c T ®T

e (!c cos !c T® sin !c T)
0 !c (46)
!c
!c (3®2 + !c2 + 2®3 T + 2®!c2 T) e®T [!c (3®2 + !c2 ) cos !c T ®(®2 3!c2 ) sin !c T]
F13 =
!c (®2 + !c2 )2
!c (2® + ®2 T + !c2 T) + e®T [2®!c cos !c T + (®2 !c2 ) sin !c T]

F14 =
!c (®2 + !c2 )2
5T
0 1 0 0 and cov(wk ) = Sw 0 eA¿ BB (eA¿ ) d¿ can be found (e.g.,
0 0 1 0 by Mathematica) but is too tedious to be given here.
_ = Ax(t) + Bw(t) =
x(t) x(t) A similar but slightly more sophisticated Markov
0 0 0 1
0 0 !n2 2³!n
model was used in [11, 46] for a target trajectory with
oscillations due to wind-induced bending. Assume that
0 the bending is modeled by a second-order Markov
0 system with undamped natural frequency !n and
+

w(t):
(41)
1 damping ratio ³, along with a standard model of
(1 2³)!n winds (i.e., a Singer model with correlation time
constant ¿ = 1=¹), such that the prewhitening system nonstationary because its second-order statistics
has the transfer function R(t + ¿ , t) depend on t as well as ¿ . The major
1=2 1=2 difference between this model and the Singer model
Sw s2 Sw s2 is that here ¾ 2 (t, ¿ ) is a function of t and ¿ . Simply
H(s) = =
(s + ¹)(s2 + 2³!n s + !n2 ) s3 + ®s2 + ¯s + ° put, this model is a modified Singer model for a
(47) typical motion, referred to as constant turn (with
constant speed and turn rate). Its nonstationarity is
!n2
where ® = ¹ + 2³!n , ¯ = + 2¹³!n , and ° = ¹!n2 .
a consequence of the constant-turn constraint. It is
This system is described by [11, 46]
possible to be more accurate than the Singer model

a_ 0 1 0 a 0 for a constant turn. The price paid is that the precise
_ state-space form of this model is complicated. To
ä (t) = 0 0 1 a + 0 w(t)
... have a time-invariant model of ax (t), it is necessary
a ° ¯ ® ä 1 to consider only those distributions of Á for which
(48) R(t + ¿ , t) = R(¿ ). The uniform distribution of Á
2 2 over (¼, ¼] is one of such distributions. To obtain
where w(t) has power spectrum Sw = c3 ¾ and ¾ =
a state-space model of ax (t), its power spectrum is
R(0) is the mean-square value of the target trajectory
derived, which is, however, not of a rational form
with the following autocorrelation
S and is dependent on the distribution of !. In [47], it
R(¿ ) = c1 e¹M¿M + c2 e³!n M¿M cos(!n 1 ³ 2 M¿ M µ) was assumed that the constant turn rate is uniformly
distributed over a given interval [!max , !max ] and the
(49) heading angle is uniformly distributed over (¼, ¼],
which is a combination of a first-order term and a and they are independent.8 It was proposed in [47]
second-order term. Here c1 , c2 , c3 , and µ are constants, to approximate the irrational power spectrum of
depending on ®, ¯, °, and ¾ [46]. Its discrete-time ax (t) by an nth-order rational spectrum to obtain a
equivalent can be found in [46]. simplified state-space model. This yields an nth-order
Markov model. Such a rational approximation may
be obtained numerically, as in [47]. For instance, if
H. Markov Acceleration Model for Constant Turns
the power spectrum of ax (t) has the following general
A typical target maneuver, such as a turn, often second-order rational approximation
has an approximately constant speed and turn rate. ¯1 s + ¯ 2
The original Singer model is not very good for such S(s) = H(s)H(s) with H(s) =
s2 + ®1 s + ®2
motion, referred to as a constant turn in this survey.
Let the acceleration a along the x and y directions of (51)
the Cartesian coordinates be ax and ay , respectively: then the causal and stable prewhitening system is
a = [ax , ay ] . Denote by v the constant speed, Á(t) given by H(s). The corresponding state-space model
the (velocity) heading angle, and ! = Á_ the constant _ ẍ, dx ] ,
is,9 for x = [x, x,
turn rate. Under the assumption that both Á and ! as
random variables have symmetric distributions and 0 1 0 0 0
are mutually independent, it can be easily shown 0 0 1 0 0
_ =
x(t)

x(t) +

wx (t)
that the zero-mean processes ax (t) and ay (t) are 0 0 0 1 ¯1
uncorrelated [47]. Note, however, that while the
symmetry assumption is fairly reasonable in practice 0 0 ®2 ®1 ¯2 ®1 ¯1
for a priori models, the independence assumption is (52)
questionable since the turn rate ! and the heading Á where wx (t) is white noise with unity power spectral
are actually related by ! = Á_ and thus dependent. density. Note that a horizontal constant turn is covered
Only the model for ax (t) is considered below
since the model for ay (t) can be obtained similarly.
8 The assumptions made in [47] for this model are not consistent
It can be easily shown from the constant-turn equation
ax = ẍ = !y_ = !v sin Á and Á(t + ¿ ) = Á(t) + !¿ that with each other. For example, Á(t + ¿ ) = Á(t) + !¿ indicates that
if Á(t) 8 [¼, ¼], ! 8[!max , !max ] and they are independent,
the autocorrelation of ax (t) is given by then Á(t + ¿ ) has a trapezoidal distribution rather than a uniform
distribution. Note that all simplifying assumptions are actually
R(t + ¿, t)
incorrect. However, such incorrect assumptions may still be used in
= 12 v 2 E[! 2 Icos !¿[1 cos 2Á(t)] + sin !¿ sin 2Á(t)J]e®M¿ M complex situations if it leads to great simplicity, such as that of the
PDAF for approximating a Gaussian mixture by a single Gaussian.
= ¾2 (t, ¿)e®M¿ M (50) What is important is to exercise care, be reasonable, and to judge by
the final outcomes.
assuming that the magnitude of the total acceleration 9 The model given here appears simpler and more reasonable than
a(t) obeys the Singer model. Clearly, ax (t) is what was given in [47].
by the use of two uncoupled 4-dimensional models among these levels is a semi-Markov process (a
here, but actually it has a more accurate single sojourn-time dependent Markov chain) with known
5-dimensional model (see Section V). transition probability PIu(tk ) = āj M u(tk1 ) = āi J,
The second-order model given by (40) is a special i, j = 1, : : : , n, and sojourn-time probability distribution
case of this general second-order model (52) with function Pij (¿ ) = PI¿ij ¿ J, where ¿ij = tk tk1 is
¯1 = 1, ¯2 = !n , ®1 = 2³!n , ®2 = !n2 . It follows from the sojourn time at level āi before it jumps to level
(44) that if ¯1 = b, ¯2 = c, ®1 = 2³!n , ®2 = !n2 the āj . Although the concept of such a semi-Markov
target acceleration of this second-order Markov jump process formalism was introduced in [50, 51],
model has an oscillatory yet stationary autocorrelation only exponentially distributed sojourn time was
given by (42), but the “exact” autocorrelation is considered therein. Note that if the sojourn time has
given by (50), which is nonstationary. Note also an exponential distribution, the input process u(t)
that the above uniform distribution assumption of is in fact Markov—the semi-Markov formulation
the turn rate is better replaced by those of [48], is not needed.11 Moreover, the following model
the Singer’s ternary-uniform mixture or its variants of the acceleration a(t) as a combination of the
(e.g., a binary-uniform mixture or a single-point and above jump-mean model and the Singer model was
uniform mixture) in many practical situations with proposed:
additional information of the turn rate. a(t) = ¯v(t) + u(t) + ã(t) (53)
where ã(t) is the Singer acceleration of (26), v is

I. Semi-Markov Jump Process Models the velocity, u(t) is the unknown acceleration mean,
and ¯ is a drag coefficient. The corresponding
The Singer model approximates the target continuous-time state-space representation is, for
acceleration as a continuous-time zero-mean Markov _ ẍ] ,
x = [x, x,
process. In practice, many target maneuvers involve
an acceleration of a non-zero mean that may be 0 1 0 0 0
reasonably assumed piecewise constant. However, the _ =
x(t) 0 ¯

1 x(t) +

1 u(t) +

0 w(t):
difficulty here is that neither the time intervals over
0 0 ® 0 1
which the acceleration mean is piecewise constant
nor the corresponding constant levels of the non-zero (54)
mean are known to a tracker. In other words, the target acceleration is modeled
One of the simplest piecewise-constant random as the Singer acceleration with non-zero mean
processes, known as jump processes, is the so-called that is intended to be a semi-Markov (but actually
semi-Markov jump process (see [49] for an Markov) jump process. This model can be referred
engineering-oriented description). It differs from a to as a Markovian jump-mean acceleration model.
Markov jump process only in that it has the Markov The unknown acceleration mean u(t) is estimated
property10 if we consider only time instants of a jump, by a weighted sum of the quantization levels:
3
but not necessarily at other times, while a Markov û(t) = ni=1 āi PIu(t) = āi M z(s), s tJ, where as
jump process has the Markov property at all times. in a multiple-model formulation, the weight is the
In other words, the past and the future states of a posterior probability of each level being the correct
semi-Markov process may be coupled through the one, using all online measurements z(s), s t, as well
time interval it stays in a state (called sojourn time) as the initial model probabilities, model transition
as well as the present state—its value at one time may probabilities, and the sojourn time distribution.
depend on values at other time instants through not Fig. 4 depicts some representative random processes
only its nearest neighbor state but also the sojourn as models of the target acceleration. Note that a
time in the state. semi-Markov jump process model for the acceleration
Several Markovian jump-mean acceleration is more effective than a white noise model with a
models have been proposed. The first was the one mean of random jumps.
given in [50–53]. In this method, the unknown Three important issues associated with this model
input u(t) (assumed equal to the non-zero mean are the design of the input quantization levels, the
of the acceleration a) is modeled as a finite-state transition probabilities, and the sojourn time. This is
semi-Markov jump process. Specifically, it was
assumed that the possible mean values of the 11 A semi-Markov process with exponentially distributed sojourn
acceleration are quantized into n known levels time in each state is actually a Markov process. This is a
ā1 , : : : , ān , and the sequence Ku(tk )L of the input consequence of the unique memoryless property of the exponential
distribution—knowledge of time already spent in a state does
not alter distribution of a future state. Similarly, a discrete-time
10 Thatis, the past and future states are independent given the semi-Markov process is Markov only if the sojourn time has a
present state. geometric distribution.
Fig. 4. Representative random acceleration process models. (a) Zero-mean white noise. (b) Zero-mean colored noise. (c) Jump-mean
white noise. (d) Jump-mean colored noise.
Fig. 5. Gamma densities with unity mean value.
similar to model-set design for the multiple-model [57] that the sojourn time be assumed an independent
method [54]. This should not be a surprise because process having a gamma distribution (see Fig. 5),
this jump-mean model actually amounts to multiple which includes exponential as a special case with
models of the (quantized) input in a degenerated form. ® = 1. Such a semi-Markov process and its embedded
The model proposed in [50, 51] for the unknown point process are both known as gamma-renewal
input u is a discrete-time, finite-state (semi-)Markov processes [58] and are generalization of the above
process. A continuous-time counterpart was proposed Markov process with exponential sojourn time and
in [55]. A sojourn-time dependent Markov chain its embedded Poisson process. The word “renewal”
model for target motion was proposed in [56] in the was coined in stochastic processes with regard to
context of the multiple-model approach (see [17] the failure times of certain equipment (i.e., sojourn
also). times in our case), and thus the value of the counting
For simplicity, the sojourn time in a state is oft process models the number of renewals that must be
modeled as having an exponential distribution in made to keep the equipment in working order.
the above “semi-Markov” formalism. The point (or Although it appears more representative of
counting) process associated with such an independent the actual maneuvers, the renewal model is rather
sojourn time process is a Poisson process. Such complicated due to its non-Markovian nature.
a sojourn time ¿ has a monotonically decreasing However, similar to modeling of the non-Markov
probability density. This implies that very small ¿ state process of a system driven by colored noise, the
would occur most frequently of all ¿ within a time prewhitening technique can be used here. The transfer
interval of the same length. This is not consistent with function of this (usually first- or second-order)
the distribution of the durations of practical target prewhitening system or its approximation can be
motions. To correct this deficiency, it was proposed in obtained from the equilibrium power spectrum of
the semi-Markov maneuver process. This leads to Non-zero-Mean Jerk Model: Note first that
an (approximate) state-space representation of the although the Singer model is usually regarded as
renewal model. Simulation results of an application a model for the target acceleration, it can also be
of the renewal model to an agile target that executes interpreted as a zero-mean jerk model in which the
constant turns indicates that the performance _ is defined by ã(t)
target jerk ã(t) _ = ®ã(t) + w(t)
improvement is not commensurate with the model and with ã(0) = 0, which is a zero-mean process. A
algorithm complexity [57], but is significant for an non-zero-mean jerk model was proposed in [62] by
image-enhanced tracking system based on the fusion introducing an additional term in the Singer-model
of a microwave radar and an infrared imaging sensor equation
[59]. _ = ®a(t) + w(t) + a¯_
a(t) (55)
The renewal model was proposed in [57] for
the turn rate. It can clearly be applied for target where a¯_ is a non-zero expected jerk. Note that this
is not equivalent to assuming a(t) _ + a¯_ directly,
_ = ã(t)
acceleration, its mean value, etc. In fact, following the
idea of [50, 51] for an acceleration process, the turn _
where ã(t) is the zero-mean Singer jerk process. It
rate can be modeled as the sum of its mean that obeys was proposed in [62] to determine the mean jerk
the above renewal model and a zero-mean first-order a¯_ adaptively from the most recent estimates of the
Markov process that obeys the Singer model. It would target velocity and acceleration under an assumed
be interesting to compare its applicability with that of coordinated-turn maneuver, as discussed in detail in
the renewal model above. Section VIC. The noise level of w was also proposed
to be determined adaptively from the most recent
J. Jerk Models estimate of target orientation and expected maneuver
level. This model and the “current” model are the
Acceleration Models versus Jerk Models: same in spirit—they aim at improving the Singer
Jerk is the derivative of acceleration. In most model by adding a non-zero-mean term to equations
coordinate-uncoupled models, it is the target that describe the acceleration. As is clear from a
acceleration that is chosen to be the descriptor of a comparison of (32) and (55), they would be equivalent
target maneuver and modeled as a random process. if a¯_ = ®ā(t), where ā(t) is the assumed piecewise
This is most natural from mechanics, kinematics, and constant mean of the acceleration. Consequently, their
vehicle dynamics since acceleration is directly related difference lies mainly in how mean jerk a¯_ and mean
to the force acting on the target. That is why target acceleration ā are obtained.
acceleration is usually taken to be the control input u. The above model was refined in [63] for the
However, for some targets, particularly agile targets, it coordinated-turn maneuvers by removing the
may be more convenient to use a random jerk process acceleration mean to yield
to model the target maneuvers. A jerk model differs ¯_
from an acceleration model usually in simplicity, _ = ®[a(t) ā(t)] + w(t) + a:
a(t) (56)
depending on whether the target motion is better On the other hand, since mean jerk is equal to the
described by a random process model of the jerk or derivative of the expected acceleration, the noiseless
the acceleration. On the other hand, however, jerk in a part of this equation also follows from taking
jerk model must be estimated, the accuracy of which derivative on both sides of a(t) = ã(t) + ā(t), where
is usually by far poorer than acceleration estimates ã(t) is the Singer acceleration. Thus, this model and
since only position (and Doppler) measurements are the “current” model differ only in that the expected
usually available. acceleration ā(t) here is not assumed constant over
First-Order Markov Jerk Model: The first-order each sampling interval.
Markov model as proposed by Singer is for target
acceleration. However, the same modeling method can
V. 2D HORIZONTAL MOTION MODELS
be applied to other target functions, e.g., target jerk.
This method was indeed applied in [60, 61] to the jerk Most 2D and 3D target maneuver models are
process as the maneuver forcing function; that is, the naturally turn motion models. These models are
jerk is modeled as a zero-mean first-order Markov usually established relying on target kinematics, in
process, exactly in the same manner as the Singer contrast to those of the previous section that are based
model for acceleration. The derivation is in a manner on random processes. This is understandable since
completely analogous to that of the Singer model. random processes are more natural for modeling time
This model has a higher dimension than the Singer correlation than describing spatial trajectories where
acceleration model and the performance improvement kinematics is a more appropriate tool.
shown in [60, 61] is not convincing because the test 2D horizontal motion models are described in this
scenario does not appear realistic. section generally in an order from the simplest to the
The following variant of (60) was used in [65, 66]:
Á_ = an =v, where an = an sin(Ã) is the projection of
the actual normal acceleration an on the horizontal
plane and Ã is the roll (i.e., bank) angle. This variant
makes explicit the relationship between the normal
acceleration in the horizontal plane and the actual
Fig. 6. Geometry of 2D target motion. normal acceleration in a bank-to-turn horizontal
motion.
most sophisticated. 3D motion models, including those A. CT Models with Known Turn Rate
for (nonhorizontal) planar motion in the 3D space, are
described in the next section. These models presume that the target moves with
Coordinate-coupled target models are highly (nearly) constant speed v and (nearly) constant angular
dependent on the choice of the state components. (turn) rate !. Assuming ! is known leads to (only
The choice of the state components (and implicitly four-dimensional) state vector, e.g., x = [x, x, _ , in
_ y, y]
the respective kinematic model) is not a trivial
the Cartesian coordinates. It follows immediately from
problem [64], where target dynamics, accuracy of
(57)–(60) that such a circular motion is described by
approximations, sensor coordinate system, among _ ! y,
_ y, _ ; that is,
_ !x]
(5) with f(x, u, t) = [x,
others, must be taken into account.
Various (noiseless part of) kinematic models _
x(t)
proposed for tracking of a target moving in the _
! y(t)
horizontal plane can be comprised from the following _ =
x(t) + Bw(t) = A(!)x(t) + Bw(t)
standard curvilinear-motion model from kinematics _
y(t)
(see Fig. 6):
_
! x(t)
_ = v(t) cos Á(t)
x(t) (57) (61)
0 1 0 0 0 0
_ = v(t) sin Á(t)
y(t) (58)
0 0 0 ! 1 0
_ = at (t)
v(t) (59) A(!) =

,
B=

0 0 0 1 0 0
_ = a (t)=v(t)
Á(t) (60)
n 0 ! 0 0 0 1
where (x, y), v, Á are the target position in Cartesian where white noise w = [wx , wy ] has the power
coordinates, ground speed (airspeed plus wind speed), spectral density diag[Sw , Sw ]. This CT model is linear
and (velocity) heading angle, respectively, and at and since ! is known. Its discrete-time equivalent is
an are the target tangential (along-track) and normal found to be
(cross-track) accelerations in the horizontal plane,
respectively. This model is fairly general—it accounts xk+1 = Fct (!)xk + wk
for along- and cross-track accelerations, and reduces
sin !T 1 cos !T
to the following special cases: 1 0
! !
1) an = 0, at = 0—rectilinear, CV motion,
0 cos !T 0 sin !T
2) an = 0, at = 0—rectilinear, accelerated motion
= xk + wk
(CA motion if at = constant), 1 cos !T sin !T
0 1
3) an = 0, at = 0—circular, constant-speed motion ! !
(CT motion if an = constant). 0 sin !T 0 cos !T
The last case above with a constant an , which
(62)
has a constant speed and a constant turn rate, is
where
referred to as a constant turn (CT) in this survey
but is often referred to as a coordinated turn (CT) Q = cov(wk ) =
in target tracking, due to an abuse of terminology in 2(!T sin !T) 1 cos !T !T sin !T
0
our opinion. The motion referred to as a coordinated !3 !2 !2
turn in aviation is in fact not so simple and limited,
1 cos !T !T sin !T
which refers to certain constraints in terms of flight T 0
!2 !2
Sw :
dynamics, rather than this purely kinematic model. !T sin !T 2(!T sin !T) 1 cos !T
(This is addressed in more detail in Section VIC, 0
!2 !3 !2
including a more reasonable definition in target !T sin !T 1 cos !T
0 T
tracking.) Such motion is preferably specified in terms !2 !2
of the turn rate ! = Á._ (63)
Its direct discrete-time counterpart (see [67, 1, 2]) is B. CT Models with Unknown Turn Rate
better known, given by (62), where wk is replaced by
diag[G2 , G2 ]wk with G2 defined in (14) and a directly These models differ from the above CT models
defined cov(wk ). The (zero-mean, Gaussian, white) only in that the turn rate is included as a state
noise w in the above is used to model the perturbation component, to be estimated. As such, they are
of the trajectory from the ideal CT motion. described by (61) in continuous-time or (62) in
An approximation12 of the noiseless part of (62) is discrete-time plus an additional equation for !. The
[68, 69]: two most popular models for ! are the Wiener process
model
1 T 0 !T 2 =2
0 1 (!T)2 =2 0 !T _ = w! (t),
!(t) in continuous time (65)

Fct (!)xk xk
0 2
!T =2 1 T !k+1 = !k + w!,k , in discrete time (66)
2
0 !T 0 1 (!T) =2
(64) and the first-order Markov process model

x + [x_ (1=2)y!T]T
_
x_ y!T
_ _ 2 1
(1=2)x(!T) _ =
!(t) !(t) + w! (t), in continuous time
= ¿!
y + [y_ + (1=2)x!T]T
_ (67)
y_ + x!T
_ _
(1=2)y(!T) 2
!k+1 = e T=¿!
!k + w!,k , in discrete time (68)
which is a second-order polynomial in !. It provides
where ¿! is the correlation time constant for the turn
a simple but less accurate alternative to the exact CT
rate, and w is zero-mean white noise of a suitable
model. It is of certain value when a nonlinear tracker
level, which can be determined exactly the same
(e.g., extended Kalman filter (EKF)) is designed with
way as for the corresponding models for acceleration
a state vector that includes the (unknown) turn rate.
described in the previous section. Some other models
However, it is valid only for !T 0, which may be
described in the previous section can also be used for
violated in many cases with large sampling interval T,
the turn rate. For example, the renewal model of [57]
such as in an air traffic control (ATC) system.
is one of them.
In the rare cases where the constant turn rate is
Discretization of a continuous-time model here
(approximately) known a priori, the above CT model
has a unique issue: which discrete-time ! should be
gives good tracking performance. The necessity of
used in Fct (!) of (62)? The most popular way is to use
an exact knowledge about the value of the turn rate
!k here. Alternatively, it was proposed in [71] to use
makes the direct use of this model unrealistic for most
!k+1 in Fct (!) instead. This implicit method is more
practical applications. A natural idea is to replace
stable. The difference of the two schemes is somewhat
the above ! by its estimate, based on, e.g., the latest
similar in spirit to that of approximating a derivative
velocity estimates, as used in [70, 67, 4]. However,
by the forward difference and backward difference
this may inject unacceptably large errors into the
in the finite difference method. With this rough
system. Additional efforts are obviously requisite to
analogy in mind, it appears that replacing ! in (62)
model the motion with an unknown turn rate within
by !¯ = 12 (!k + !k+1 ) would further improve accuracy
this framework.
since the center point is usually a better approximation
Multiple Known Turn-Rate Models: Another
than either of the end points, more or less like the
natural solution is based on the use of multiple
central difference is more accurate than the forward
models with different, fixed turn rates. This approach
difference and backward difference at the cost of more
alleviates the effect of the uncertainty in the turn rate
computation (and possibly numerical instability). For
and takes advantage of the simple and linear form of
the current problem, however, no extra computation
the dynamic model (73) given the turn rate. In this
is required by this new scheme if the procedure of
approach, the sequence of turn rates K!k L is modeled
solving the linearized equation as proposed in [71] is
as a Markov (or semi-Markov) chain taking values
used, although possible numerical problems should be
in the set I!1 , !2 , : : : , !n J, governed by the transition
checked. These three schemes lead to the following
probabilities PI!k = !i M !k1 = !j J, i, j = 1, : : : , n, as
three different linearized models by the first-order
well as initial probabilities. This approach has been
Taylor series expansions at [x, !] = [x̂kMk , !ˆ kMk ] :
well established (see, e.g., [67, 4], and [68, 69]),
mostly for ATC tracking applications. Therefore,
a main application of this CT model with known Fct (!k )xk = Fct (!ˆ kMk )xk + F! (!ˆ kMk )x̂kMk (!k !ˆ kMk )
turn rate is serving as one or more (with different ! Fct (!k+1 )xk = Fct (!ˆ kMk )xk + F! (!ˆ kMk )
values) models in a multiple-model architecture. (69)
x̂kMk (!k+1 !ˆ kMk )
12 By expanding sin µ and cos µ up to the second-order terms:
¯ k = Fct (!ˆ kMk )xk + F! (!ˆ kMk )x̂kMk (!¯ !ˆ kMk )
Fct (!)x
sin µ µ and cos µ 1 µ2 =2 for µ 0.
where F! (!ˆ kMk ) = (@=@!)Fct (!) M!=!ˆ kMk . If (66) is As stated before, the choice of the state vector is
used, however, these models have an identical state not trivial for the turn models. Essentially two classes
prediction because have been proposed. They differ in the representation
of the velocity vector: in the Cartesian and polar
E[xk+1 M z k ] = E[Fct (!)xk M z k ] = Fct (!ˆ kMk )x̂kMk (70) coordinates, respectively.
where z k stands for all measurements through time 1) CT Models with Cartesian Velocity: In
tk . This follows from E[!¯ M z k ] = E[!k+1 M z k ] = this model, the state vector is chosen to be x =
_ !] ; that is, the velocity vector (x,
_ y, y,
[x, x, _
_ y)
E[!k M z k ] = !ˆ kMk , a consequence of (66). Nevertheless,
the corresponding covariances differ because different is represented in the Cartesian coordinates.
models are used. That is why these models result Consequently, a direct discrete-time version of the
in very similar tracking performance, as reported model is given by (see [67, 2]):
in [72] without explanation, for two IMM-EKF
Fct (! ) 0
(interacting multiple model) algorithms using the first xk+1 = xk + diag[G2 , G2 , 1]wk (73)
0 ¯
two schemes above with real ATC data. The state
prediction would also be different if model (68) is where ¯ = e®T , ! = !k , !k+1 , !,
¯ or something
used. similar, G2 was given in (14), wk = [wx , wy , w! ]k is
We point out that it is theoretically more appealing zero-mean white noise with suitable statistics—they
to obtain a discrete-time model for state vector are noise terms for acceleration in the x and y
x = [x, x, _ !] by discretizing the following
_ y, y, directions and for turn rate.
continuous-time equations jointly: This model is known to be used successfully
[x, _ ÿ] (t) = (x,
_ ẍ, y, _ ! y,
_ y, _ (t) + w(t)
_ !x) (71) as one of models in numerous multiple-model
configurations (see, e.g., [74, 67, 2, 75, 4]).
_ = ®!(t) + w! (t)
!(t) (72) 2) CT Models with Polar Velocity: Obviously, the
where ® = 0 or 1=¿! . It is reasonable to expect an velocity vector can also be represented in the T polar
2 2
improvement in performance with this alternative at coordinates as [v, Á] in terms of speed v = x_ + y_
the price of an increased complexity. and (velocity) heading angle Á = tan1 (y= _ x).
_ The state
A continuous-time model is more accurate vector is thus x = [x, y, v, Á, !] . The corresponding
than its discrete-time versions, but the latter are differential equation is given by (5) with f(x, u, t) =
needed for most applications. In the case where [v cos Á, v sin Á, 0, !, 0] ; that is
the continuous-time model is nonlinear, such as
the CT models with unknown turn rate, there are _ = [v cos Á, v sin Á, 0, !, 0] + w(t)
x(t) (74)
in general at least two approaches to acquiring its
discrete-time linear approximate models. Such models which follows immediately from (57)–(60) by setting
are needed in the application of a linear (e.g., Kalman) ! and v constant. By linearization first and then
filter based nonlinear filtering technique. The first discretization, its discrete-time linearized model is
is to linearize the nonlinear equation for the state given in [76] (see also [1]):
first and then discretize the linearized differential
x + (2=!)v sin(!T=2) cos(Á + !T=2)
equation (by finding the solution via integration). y + (2=!)v, sin(!T=2) sin(Á + !T=2)
This approach is more commonly used because it

is easy. An alternative is to discretize the nonlinear xk+1 =
v +w
k
state-space equation first and then linearize the
Á + !T
discrete-time model. These two approaches are
referred to as discretized linearization and linearized ! k
discretization, respectively, in [73, 64]. The second (75)

approach appears more accurate in general than the where wk is white noise with covariance
first since linearization usually lose more accuracy 3 2
than discretization and thus should be done later. 0 0 2 2
T ¾!_ =3 T2 ¾!2_ =2
Q = diag , T ¾v_ , 2 2 :
However, the second approach may not be tractable 0 0 T ¾!_ =2 T 2 ¾!2_
in general because discretization requires solving
the nonlinear differential equation, which is often a (76)
great challenge. Fortunately, for the CT motion with This model was successfully used as a building block
an unknown turn rate, the corresponding differential of a multiple-model algorithm for aircraft tracking
equations for the state are simple and the solution application in an air defense system [77, 78].
can be readily obtained [73, 64]. The equations of the The above model uses (65) for the turn rate
nearly CT model, obtained by the first approach, and per se. Instead, it may be better to use (67). The
some performance comparison results are also given corresponding discrete-time model need be modified
in [73, 64]. accordingly.
The discretized linearization (i.e., discretization in terms of the sensor coordinate system, which is
after linearization) alternative of this model can be inherently a nonlinear problem. The following simple
found in [73, 64], along with a comparison of its geometrically oriented procedure of estimating the
performance with that of the linearized discretization center was proposed in [80]. Assume that each target
based on a theoretical error analysis13 and Monte position measurements are points on the circle; replace
Carlo simulations. It was concluded therein that the chord between any two consecutive measurement
whenever possible linearized discretization is points with the straight line segment connecting them;
preferable to discretized linearization. For the two the center can then be determined from the (average)
CT models with polar velocity, the former slightly intersection of the perpendicular bisectors of two or
outperforms the latter. It was also claimed therein that more such straight line segments. An essentially the
the CT model with polar velocity outperforms the CT same procedure was used in [81] for estimating the
model with Cartesian velocity. center. Note that using the center estimates injects
The use of the following variant of the above additional nonlinearities into the system, which are
CT model with polar velocity was reported in [79]. not accounted for in the above linear model.
Let the state vector be x = [x, y, v, Á, an ] ; that is,
replace the turn rate ! in the above model with the D. Curvilinear Motion Model
normal (transversal) acceleration an . Then, it follows
immediately from (57)–(60) that the continuous-time This model, proposed in [82], is more general than
state-space model for the CT motion is given by those considered so far in this section. It accounts
(5) with f(x, u, t) = [v cos Á, v sin Á, 0, an =v, 0] . A for possibly non-zero normal (cross-track) and
discrete-time linear approximation of this model was tangential (along-track) target maneuver accelerations
obtained based on a refined Euler-Cauchy scheme as simultaneously. For the Cartesian state vector x =
[79] _ , it follows from the standard equations
_ y, y]
[x, x,
xk+1 = xk + Tf(xk + 12 Tf(xk )) + wk (77) of curvilinear motion (57)–(60)14 that this model in
continuous time is given by
followed by the standard EKF linearization, where
wk is white noise, obtained by a backward difference _ = Acv x(t) + B(x(t))a(t) + w(t)
x(t) (79)
of (the derivative of) the continuous-time white
noise. Note that xk+1=2 xk + 12 Tf(xk ). A set of where a = [at , an ] is the acceleration for the maneuver,
such unknown acceleration models was included Acv was given by (11), and
in a tracker for ATC in [79] to account for possible
0 0
horizontal motions.
_ _
x(t) y(t)
S S
C. Circular Motion Models x(t) _ _
2 + y(t) 2 x(t) + y(t)
_ 2 _ 2

B(x(t)) = :
0 0
For a circular motion of a target, if its center were

known, the simplest model would be to represent the _
y(t) _
x(t)
S S
circle in the polar coordinates and place the origin at x(t) _ 2
_ 2 + y(t) _ 2 + y(t)
x(t) _ 2
the circle center. In this coordinate system, the target
_ (80)
dynamic model is linear for x = [½, µ, µ]
Pretending the solution formula for the linear system
xk+1 = diag[1, F2 ]xk + diag[1, G2 =T]wk (78) works for this actually nonlinear case, its discretized
version (by the standard discretization) in the form
where F2 and G2 were given by (14) and wk is white
noise. The corresponding measurement equation xk+1 = Fcv xk + Gk (x)ak + wk (81)
is pseudo-linear because the noise covariance is
actually state dependent, described in detail in a is highly nonlinear because of its dependence on the
subsequent part (see [15] for a preliminary version). target state via matrix B:
As a result, the Kalman filter is not optimal but = T
can be nonetheless implemented straightforwardly. Gk (x) = eA(T¿) B(x(kT + ¿ ))d¿ (82)
This maneuver-centered CT model was introduced 0
in [80]. While the idea underlying this model is and the integration involved in Gk (x) is hard to
intuitively appealing, the inherent nonlinearity of evaluate exactly. It was shown in [82] that Gk (x) can
the problem is not avoided. It obviously relies on ¢
be found approximately to be, with 'k+1 = Ák + !k T,
an accurate determination of the center of the turn
13 Based on estimation of the Frobenius norm of the neglected terms 14 Theheading angle Á defined in [82] differs from the traditional
in the approximations. one, which is adopted here.
Fig. 7. Curvilinear motion trajectories.

1 1 1
sin 'k+1 sin Á k T cos Á k
!k2 !k2 !k

1 1
cos 'k+1 cos Ák
!k !k

Gk (x) Ga (Ák , !k ) = Gat (Ák , !k ), (83)

1 cos ' + 1 cos Á 1 T sin Á
k+1 k k
!k2 !k2 !k

1 1
sin 'k+1 sin Ák
!k !k

1 1 1
2 cos 'k+1 + 2 cos Ák T sin Ák
!k !k !k

1 1
sin 'k+1 sin Ák
!k !k

Gat (Ák , !k ) = (84)

1 sin ' + 1 sin Á + 1 T cos Á
2 k+1 2 k k
!k !k !k

1 1
cos 'k+1 + cos Ák
!k !k
under the following simplifying assumptions: 1) in (85) the effect of the tangential acceleration at
the acceleration a is piecewise constant over each [i.e., the Gat (Ák , !k )atk term] is “added” to the CT
sampling interval [kT, kT + T) and 2) the speed (constant-speed constant turn-rate) motion. Equation
change over a sampling interval is much smaller than (85) makes it clear the capability of this model to
the speed itself: atk T ¡ vk . An analytically equivalent account for “relatively small” tangential accelerations
form15 of (81) is, with at being the only explicit as well as the normal accelerations, while the
forcing term, popular CT model accounts only for the latter (see
Fig. 7). Both forms are applicable to tracking targets
xk+1 = Fct (!k )xk + Gat (Ák , !k )atk + wk : (85) performing maneuvers with concurrent non-zero
Note that while in (81) the maneuver acceleration along- and cross-track accelerations, but attention
term Ga (Ák , !k )ak is “added” to the CV motion, should be paid to the approximating assumptions of
the models.
This model, combined with a suitable model
15 Itcan be obtained by using x_ k = (an =!k ) cos Ák and y_ k = for turn rate, is one of the most sophisticated target
(an =!k ) cos Ák : maneuver models for 2D horizontal motions.
E. Ship Models ¯ = 0 is given by [88, 89, 66]
The maneuver models surveyed in this paper, xk+1 = xk + Tvk sin Ák (94)
especially those for 2D horizontal motions and yk+1 = yk + Tvk cos Ák (95)
all 3D models of the next section, are particularly 1 Tvk ¿
suitable for aircraft. These models were developed for Ák+1 = Ák + Tvk [k + 2 (k
0 )Tvk ¿ e ] (96)
target tracking purposes. For other (e.g., navigation) k+1 = k eTvk ¿ + 0 (1 eTvk ¿ ) (97)
purposes, numerous, more precise dynamic models
can be found in the literature on vehicle dynamics vk = Kv0 = v0 (1 + 1:9k2 L2 )1 (98)
(see, e.g., [83–86]), which is beyond the scope of S
this survey. These models, however, require better where ¿ = (p=2 + p2 =4 q)=L and 0 = =v0 .
knowledge about the vehicle than what is available The time constant ¿ was set to zero in [88, 90],
to a tracker. A few of these models have been adapted resulting in a constant turn rate (i.e., k+1 = k ),
to tracking applications, resulting in vehicle dynamics to eliminate the dependence of the model on the
based models. An example is those for aircraft based ship-specific hydrodynamic constants. However,
on flight dynamics described in Section VID. Note, the ship length L was actually treated as known
however, that they are not really point-target models. therein. The unknown k was assumed to take on
Likewise, a number of precise ship motion models one of the three possible values I0, c , c J with a
based on ship dynamics are available in the literature, preset constant c , representing rectilinear, left-turn,
which depend on the particular ship’s form and and right-turn motions, respectively, and the tracker
size. There exist also less precise but more generally presented therein was based on a multiple-model
applicable models. The following model from [87] is algorithm using these three models for k .
one in the latter class The above nonlinear model has been proposed for
ship tracking [88, 89, 66]. Other ship dynamic models
x_ = v sin(Á ¯) (86) are available (see, e.g., [91, 33, 92]), some of which
appear simpler and more popular.
y_ = v cos(Á ¯) (87)
Á_ = ! (88) VI. 3D MOTION MODELS
! = K (89) Many of the 2D horizontal models reviewed

above have been considered for application to 3D
v2 qL tracking of civilian aircraft in ATC systems. Such
_ = 0 2 + s31 ± (90)
2pL v0 targets maneuver mostly in a horizontal plane with
v nearly constant speed and turn rate and have little
¯_ = 0 [q¯ + s21 ±] (91) or limited vertical maneuver, usually performed not
2pL
at the time of a horizontal turn. Thus, the altitude
v = Kv0 (92) changes are most often modeled independently by a
1 (nearly) CV model or a random walk model along
1:9 2 L2
K = 1+ (93) z direction, leading to an acceptable accuracy in
v02 practice. However, when the task is to track agile
military aircraft, capable of performing “high-g”
where model noise is not included for simplicity.
turns in the 3D space (e.g., for tracking in air defense
Here (x, y), Á, !, , ¯, ± are ship position, heading,
systems) rather than just horizontally, decoupled
(heading) turn rate, velocity vector turn rate,
models may be inadequate. Many efforts have been
drift angle, and control ruder angle deviation,
devoted to solving this problem, and more accurate
respectively;16 v = v() and v0 = v(0) are ship speeds
models have been developed, which are surveyed next.
at turn rate and = 0 (i.e., at the onset of the turn),
respectively; the hydrodynamic constants p, q, s21 , s31
depend on ship geometry and size, in particular, ship A. Basic Kinematic Relations
length L. The main feature of this model of a ship,
Let p = OP, v = p,_ a = v_ = p̈ be target position,
which is a sizable object, is revealed by (89) that velocity, and acceleration, respectively, in the inertial
relates two turn rates and by (92) that relates two (Cartesian) frame , = Oxyz , where P is the target
speeds. The discretized version of this model with center. Denote % = P»´³ as the target body frame.
The angular velocity vector of the target is defined
16 Heading is the angle of the longitudinal axis and velocity heading in the body frame as % = (p, q, r) and in an arbitrary
is the angle of the velocity vector. We use either term if they frame (e.g., the inertial frame ,) as [93, 86]
coincide or one does not exist (strictly speaking, a point target = p» + q´ + r³
without a shape has no heading). Turn rate is usually defined as the (99)
heading change rate. with p = ´_ ³, q = ³_ », r = »_ ´
where », ´, ³ are expressed in that arbitrary frame.
The fundamental relation of kinematics (FRK) [86,
94] states that for any time-varying vector u(t) we
have
du, du%
= + %, u (100)
dt dt
where u, and u% are u expressed in the inertial frame
, and the body frame %, respectively, and %, is the
angular velocity vector of % with respect to ,. This
relation holds true in all frames. For example, Fig. 8. Geometry of a 3D orthogonal-velocity motion.
, , % ,
du du ,
= + %, u, (101) this is the case even if _ = 0 (i.e., turn rate is not
dt dt
, % % % constant) provided has a constant direction.
du du % Equation (105) plays a central role in most 3D
= + %, u% (102)
dt dt kinematic models. Virtually all models in the literature
(e.g., [95, 10, 96, 97, 70, 6, 98, 1, 99]) presume
where superscripts , and % denote quantities explicitly or implicitly that (105) holds. As shown
expressed in , and % frames, respectively. In the above, a necessary and sufficient condition for its
sequel, for simplicity we drop superscript , and
validity is the orthogonality between the linear and
write u_ and ü for du, =dt and d2 u, =dt2 , respectively.
angular velocity vectors. In this sense, models that
However, we maintain the less compact notations
obey (105) are orthogonal-velocity models, although
du% =dt and d 2 u% =dt2 , rather than the misleading u_ %
and ü% because the latter are likely to be incorrectly they are widely referred to as “coordinated turn”
interpreted as (du=dt)% and (d2 u=dt2 )% , respectively. models, which is actually a misnomer. The geometry
Note that since the target body frame % rotates as the of these models is illustrated in Fig. 8.
target does, we have = %, ,
, where is the angular The FRK (100), (103), and (104) (or (105)) are
velocity vector of the target in the inertial frame ,. the key to understand various 3D maneuver models
In what follows, we consider only the important proposed in the literature.
case where the »-axis of the body frame is aligned with Since the target tracking community is more
the velocity vector, that is, » = v=v, where v = NvN is familiar with the matrix language than the vector
the target speed. Then by the Poisson’s formula [93] analysis and we are more concerned with state-space
»_ = », we have models, in the sequel we use the following
relationship repeatedly. Let = [x , y , z ] and
_ vv_
vv v
= _ y,
v = [x, _ . Then
_ z]
v2 v
that is, x_ 0 z y
v v v_
v_ = v_ + v= v+ v (103) v = M y_ , M = z 0 x :
v v2
z_ y x 0
meaning that the total acceleration is the vector sum
of a linear acceleration (v=v)v_ and an acceleration (106)
v responsible solely for turning. This relation also
follows directly from FRK (100) with u = v and » = B. Constant-Turn Models
v=v. Further, it follows from (103) by straightforward
1) 3D Constant-Turn Models: It is clear from
calculus17 that
(103) that a constant speed (i.e., v_ = 0) motion
v v a corresponds to a v = 0 (i.e., a " v) and is described
= 2 v+ 2 : (104)
v v equivalently by
It turns out that (103) and (104) are equivalent. a = v: (107)
It is clear from (104) that the angular velocity is Since is unknown in general, one of the simplest
given by
v a possible way is to augment the state vector by its
= 2 (105) components x , y , z in the inertial frame, to be
v
_ y,
estimated. That is, let x = [x, y, z, x, _ x , y , z ] .
_ z,
if and only if v = 0, that is, " v. It follows
Then (107) becomes
from (105) that " a and that v and a are in a plane _
(known as the maneuver plane) orthogonal to ẍ x x_ x_ y z z y_
and thus the motion is planar, but not necessarily _
ÿ = y y = M y_ = z x_ x z_ :
horizontal, if has a constant direction. Note that
z̈ z z_ z_ x y_ y x_
17 v v_ = v v) = (v v) ( v)v = v 2 ( v)v.
a=v ( (108)
If, in addition, is modeled as nearly constant Therefore, the CT motion can be modeled by a
_ is white noise), then the continuous-time
(i.e., second-order Markov process
3D (nearly) CT model—which has (nearly) constant
a_ = ! 2 v + w (112)
angular velocity and speed in the 3D space—is given
in the state space by where w is white noise. As explained in footnote
21 right after (123), this motion is (nearly) planar
0 I3 3 0 0
for small w. This fact, along with _ = 0 and (105),
_ =
x(t)

0 M 0 x(t) + 0 w(t): (109)
implies that this is a planar constant speed and
0 0 0 I3 3 constant turn rate (or angular velocity) motion. Hence,
This model is nonlinear since M of (106) depends on (112) is a (nearly) planar CT model.
the state components x , y , z . This popular model can be traced back to [95, 100].
Note the difference between this model and A discussion of the vector models (107) and (110),
those based on (105): while " v under (105), along with an analysis and comparison with similar
of (109), albeit nearly constant, is not necessarily models, can be found in [99]. These models, along
orthogonal to v, although v and a are orthogonal to with several implementations, were reviewed in [1].
each other under (107). " a is true for both models. The state-space form of this model in the Cartesian
To our knowledge, this simple 3D-CT model has not coordinates is clearly given by, with state x =
_ y,
[x, y, z, x, _ ẍ, ÿ, z̈] ,
_ z,
been considered in the tracking literature, despite
its generality: it can describe constant-speed motion
0 I3 3 0 0
along non-zero torsion (nonplanar) curves.
_ =
x(t) 0 0

I3 3 x(t) + 0 w(t)
This model has a (nearly) constant speed and
angular velocity. An even more general model can 0 ! 2 I3 3 0 I3 3
be developed by replacing _ = 0 with !_ = 0, where
(113)
! = NN, resulting in a general constant-turn (constant _ ẍ, y, y, _ z̈] ,
_ ÿ, z, z,
or with state x = [x, x,
speed and constant turn rate) model.
2) Planar Constant-Turn Models: Under an _ = diag[A(!), A(!), A(!)]x(t) + diag[B, B, B]w(t)
x(t)
additional assumption " v [or equivalently, (105)]
_ = 0, it follows from differentiation of (114)
as well as
(107) that 0 1 0 0

A(!) = 0 0 1, B = 0 (115)
2
a_ = a= ( v) = ( v) ( )v = ! v 2
0 ! 0 1
(110)
where the turn rate ! is given by where the white noise w(t) has power spectral density
matrix diag[Sx , Sy , Sz ]. Its discrete-time equivalent
¢ Nv aN NvNNaN a model [70] with state x = [x, x, _ ẍ, y, y, _ z̈]
_ ÿ, z, z,
! = NN = = = : (111) 18
v2 v2 v is
xk+1 = diag[F(!), F(!), F(!)]xk + wk (116)

sin !T 1 cos !T
1
! !2
sin !T
F(!) = , cov(wk ) = diag[Sx Q(!), Sy Q(!), Sz Q(!)] (117)
0 cos !T
!
0 ! sin !T cos !T

6!T 8 sin !T + sin 2!T 2 sin4 (!T=2) 2!T + 4 sin !T sin 2!T
4! 5 !4 4! 3

4
2 sin (!T=2) 2!T sin 2!T 2
sin !T
Q(!) =

:
(118)
!4 4! 3 2! 2

2!T + 4 sin !T sin 2!T sin2 !T 2!T + sin 2!T
4! 3 2! 2 4!
18 Thecovariance cov(wk ) given here appears more reasonable than

what was suggested in [70, 1].
The motions in the x, y, z directions in this model condition " v, the angular velocity vector is
are coupled only through the common turn rate !. given by (105). Substituting it in (120) results in
This model specifies a constant-turn motion in the modeling the target acceleration as a second-order
so-called maneuver plane,19 defined by the velocity Markov process with state dependent coefficients20
and acceleration vectors. in the inertial frame
This model relies on knowledge of !. If ! is
unknown, a straightforward way is to augment the d 2 v% _
a_ = 2®a (2®2 + !2 )v + w, w= + v
state vector by ! and estimate ! (see Section VB). dt2
This, however, increases the state dimension and (122)
makes the model highly nonlinear. An alternative is where
to compute ! from the best speed and acceleration
estimates by (111) so as to take advantage of the ¢ Nv aN va
! = NN = , ®= : (123)
linear structure of (116) at a given !. The use of v2 v2
speed and acceleration estimates will, however, This model, along with one in a slightly different
inject additional errors into the model and result form in rotating line-of-sight (LOS) coordinates, was
in accuracy degradation, especially when the proposed in [96]. Both forms were utilized therein for
orthogonality property a v = 0 is severely violated two highly accurate trackers.
for a constant-speed motion. A possible rescue The noise term w in (122) reflects the effect of
is to impose a v = 0 as a kinematic constraint,
the forces and moments applied to the target. It was
which, however, turns the dynamic model into a
modeled in [96] as zero-mean (Gaussian) white noise
highly nonlinear one if it incorporates the constraint
with power spectral density matrix diag[Sx , Sy , Sz ]
directly. Mainly to avoid this nonlinearity, an
that is to be designed. The damping coefficient ® is
alternative approach was adopted in [101, 70, 102,
a normalized target drag (i.e., the ratio of negative
103] where the constraint was incorporated into a
pseudomeasurement model, described in detail in the tangential acceleration to target speed). Both ® and
subsequent part of this survey on measurement models turn rate ! can be interpreted as unknown model
(its conference version is [15]). parameters. As stated in [96], it can be shown that
(122) defines zero-torsion (planar) trajectories (see
Fig. 9) if and only if w = 0, that is, non-zero noise
C. Variable-Turn Models components w account for out-of-plane motion.21 As
such, (122) is a (nearly) planar variable-turn model if
The above models based on the constant-speed
its noise term w is small.
condition (107) and the constant turn-rate assumption
A possible enhancement of this model would be to
are restrictive in describing the variety of possible
model w in the body frame, which more accurately
maneuvers. More sophisticated and potentially more
reflects the nature of the acting forces, and then
accurate models are surveyed next. They do not
transform it to the inertial frame.
assume that a target moves with a constant turn
and are thus capable of describing more complex This planar variable-turn model is more general
maneuvers. and perhaps more accurate than the planar CT model
1) Planar Variable-Turn Model: By applying the of (112). It reduces to the model (112) [with ! given
FRK to the target velocity vector v, we have by (111)] for a CT motion under the constant-speed
condition (i.e., a v = 0). When the speed is not
dv% constant, the damping term due to non-zero ® allows
v_ = + v: (119)
dt the model to automatically adapt itself to target
Differentiation and applying the FRK again yield maneuvers. Both ® and ! vanish for an unaccelerated
motion.
d 2 v% _ It is straightforward from (122) to obtain the
v̈ = + v + 2 v_ ( v):
dt2 following continuous-time state-space form of
(120) this model in the inertial frame with state x =
This equation describes a general motion of a rigid
20 The formulation given here follows from [96, (5a) and (5b)] in
body in space [96]. Under the orthogonal-velocity
view of the identity
2 2 2
a2 va va Nv aN
19 Inthe special case where the maneuver plane is horizontal, this = (cos2 µ + sin2 µ) = + = ®2 + ! 2 :
v2 v2 v2 v2
model simplifies greatly and it is more commonly used than (the
more general) model (57)–(60) for the derivation of some horizontal (121)
CT models [1]. In this case (i.e., at = 0 and ! = an =v), however, 21 Thisimplies that as a special case with ® = 0, (112) with w = 0
both models are essentially equivalent. describes a planar motion.
Fig. 9. 3D planar variable-turn trajectories.
_ ẍ, y, y,
[x, x, _ z̈]
_ ÿ, z, z, Like the planar CT model (116), motions in
different coordinates in this model are coupled only
_ = diag[A(!c , ®), A(!c , ®), A(!c , ®)]x(t)
x(t) through the common ® and !c . This is another
difference from that of (40) with acceleration replaced
+ diag[B3 , B3 , B3 ]w(t) (124)
by velocity, where motions in different coordinates are
not coupled (i.e., different values of ® and !c may be
0 1 0 0 1 0
involved).
A(!c , ®) = 0 0 1 = 0 0 1 We found the discrete-time equivalent model as
0 (®2 + !c2 ) 2® 0 !n2 2³!n
xk+1 = diag[F(!c , ®), F(!c , ®), F(!c , ®)]xk + wk (127)
(125)
0 Q = cov(wk ) = diag[Sx Q(!c , ®), Sy Q(!c , ®), Sz Q(!c , ®)]

B3 = 0 (128)
1 where F(!c , ®) was defined by (46) and Q(!c , ®) is
given by
where !n2 =® 2
+ !c2 , !c2 2 2
= ® + ! , and ³ = ®=!n , and
the white noise w(t) has power spectral density matrix A+B +C
diag[Sx , Sy , Sz ]. q11 =
4®! 2 (®2 + ! 2 )3
It can be seen by comparing (124) with (40) that
if the white noise term of this model were replaced A = e2®T [(®2 3! 2 )c + (! 2 3®2 )s (®2 + ! 2 )2 ]
by that of (40), the velocity component along each B = 8e®T ®![2®!c0 + (®2 ! 2 )s0 ]
Cartesian coordinate would be implicitly modeled as
a random process having an oscillatory exponentially C = ®2 ! 2 (4®T 11) + ! 4 (1 + 4®T)
decaying autocorrelation
e2®T (!c0 + ®s0 e®T !)2
q12 =
Rx_ (¿ ) = Ry_ (¿ ) = Rz_ (¿ ) = ¾ 2 e®M¿M cos(!c ¿ ) 2! 2 (®2 + !2 )2
S e2®T (c s ®2 + ! 2 ) + 4e®T ®!s0 ! 2
= ¾ 2 e³!n M¿ M cos(!n 1 ³ 2 ¿ ) (126) q13 =
4®! 2 (®2 + ! 2 )
Note that both the actual oscillation frequency !c and e2®T (c s ®2 ! 2 ) + !2
q22 =
the undamped natural frequency !n of the velocity 4®! 2 (®2 + ! 2 )
autocorrelation functions increase with the damping
coefficient ® for a given physical turn rate ! of the e2®T s02
q23 =
target; the frequencies and the turn rate are equal 2! 2
(i.e., !c = !n = M!M) only if there is no damping; the e2®T (c + s ®2 ! 2 ) + !2
autocorrelations Rx_ (¿ ) = Ry_ (¿ ) = Rz_ (¿ ) are not periodic q33 =
4®! 2
when damping is large (i.e., ³ 1). This makes good
c = ®2 cos 2!T, s = ®! sin 2!T
sense in view of the physical meanings of ®, !, !c ,
and !n . ! = !c , c0 = cos !T, s0 = sin !T:
2) Coordinated-Turn Models: By the FRK (100) This coordinated turn model involves unknown
¢
with u = ag = a g, where g stands for the constant parameters T, L, q, r, which can be obtained as
gravitational acceleration in the inertial frame, we follows [62]. By the FRK (100) with u = v, we have
have in the body frame v_ = dv% =dt + v, which in the body frame is
¥ %
% v_ 0 r q v
%
da%g dv
(a_ g ) = + ( ag )% : (129) = 0 + r 0 p 0 (133)
dt dt
0 q p 0 0
Let T be the thrust (minus drag) and L the lift, along
» and ´ directions, respectively. Then, we have a%g = where v is the target speed. On the other hand,
(dv=dt)% is the total acceleration a% in the body frame,
[T, L, 0] , % = [p, q, r] , and
that is,
T 0 r q % T g»
dv %
( ag )% = M% L , M% = r 0 p : = a = L + g´ (134)
dt
0 q p 0 0 g³
(130) where g% = [g» , g´ , g³ ] is the gravitational acceleration
%
Since (a_ g ) = TIFBF a_ g , where TIFBF is the transform from in the body frame. It thus follows that
the inertial frame to the body frame, (129) becomes r = (L + g´ )=v, q = g³ =v
_
T 0 r q T (135)
_ T = v_ g» , L = a g T:
BF _
TIF ag = L + r 0 p L : (131)
The target body frame is determined by the
0 q p 0 0
(estimated) velocity—to which the thrust T is
An equivalent version of this general model in the parallel—and lift vector L.
body frame for the so-called coordinated turn motion Modeling the target accelerations in the body
was used in [62] to derive an equation for mean jerk. frame is simpler than in the inertial frame. However,
A class of motion in the 3D space, known as the fact that it is in a frame different than the sensor
coordinated turn, is defined (for aircraft) by the inertial frame induces difficulties. In essence, all
following conditions, introduced originally in [62] ways around these difficulties involve explicitly or
and refined in [98]: 1) (mean) thrust T (minus implicitly conversion from one frame to the other,
drag)—acceleration along the velocity direction—and which can be done accurately in general only if
(mean) lift L (acceleration normal to the aircraft accurate knowledge of the target position and model
wing plane) are constant; 2) (mean) roll angle Ã parameters is available. In addition, while this model
(angle around the longitudinal axis) is constant; and is intuitive appealing, its implementation is rather
3) (mean) angle of attack and sideslip22 are zero. complicated due to, e.g., its dependence on thrust, lift,
These conditions confine the (average) target motion and other parameters.
to a plane, known as the maneuver plane. They are Furthermore, for the purpose of predicting target
consistent with the bank-to-turn characteristics of position in an antiaircraft fire control application
fixed-wing aircraft. Clearly such an (average) motion (not as part of any filter), the following model in
does not necessarily have a constant turn rate or the maneuver plane was employed in [62]. The
constant speed. aircraft motion is described by a standard curvilinear
Under these coordinated turn conditions, we have motion model of (57)–(60) in the maneuver plane with
T_ = 0, L_ = 0, p = ´_ ³ = ( ´) ³ = 0 (because , acceleration (at , an ) determined by the thrust, lift, and
´, ³ are in the same plane orthogonal to »), and thus gravity as
(131) becomes
at = T + gt , an = L cos Ã + gn (136)
rL
IF where L cos Ã is the projection of the lift along the
a_ = a_ g = TBF rT , IF
TBF = (TIFBF )1 = (TIFBF ) :
normal direction and Ã is the roll angle. Here, gt and
qT gn are the gravity components along target velocity
(132) and normal directions, respectively, given by
The expectation of this model is the mean
jerk equation obtained in [62]. Its state-space gt = g» cos Á + g´ sin Á, gn = g» sin Á + g´ cos Á
representation can be obtained straightforwardly.
(137)
22 The angle of attack is the angle between the heading and the
where Á is the target velocity heading in the fixed
velocity vector projected onto the longitudinal symmetric plane Cartesian frame in the maneuver plane. Likewise, the
and the sideslip is the angle between the velocity vector and the target position (x, y) (and speed v) in (57)–(60) are all
longitudinal symmetric plane. in the fixed Cartesian frame in the maneuver plane.
Clearly, the approximate curvilinear model [p , v , a , u ] is [104]
of Section VD in the state-space form can be
v 0
used here in the maneuver plane. However, the
a 0
small speed-change assumption as well as the _x = v a x +

wb (141)
piecewise-constant acceleration assumption on which
(a g) + T(v, a)u 0
2
this approximate model is based should be kept in v
°u I3 3
mind. These assumptions are not valid for a large
thrust and sampling interval. where p, v, a are the position, velocity, and
3) Generalized Coordinated-Turn Model: acceleration vectors, respectively, in an inertial frame
Applying the FRK (100) to the target acceleration and T(v, a) = [t, b, n] is the coordinate transformation
(with the gravity subtracted) ag = a g gives matrix from the target body frame to the inertial
frame, where t = v=v, b = (a g) v=N(a g) vN,
da%g n = b t are the unit tangential, normal, and binormal
a_ = + (a g): (138) vectors, respectively.
dt
In this model, the (unperturbed) motions in the
Under the orthogonal-velocity condition " v three spatial directions are coupled and confined to
[i.e., (105)] and the constant acceleration condition the maneuver plane, but not necessarily at a constant
da%g =dt = 0, (138) becomes speed. The time-correlated perturbation u accounts for
the (mostly out-of-plane) unmodeled motions and is
v a
a_ = (a g) (139) modeled in the body frame to be more precise. This,
v2 however, necessitates a coordinate transformation.
which is the kinematic equation of [98]. It can be This in turn introduces a dependency of the model on
verified that the curvilinear trajectories generated by velocity and acceleration estimates, which may lead
(139) have zero torsion, that is, they are confined to to a model error and thus accuracy degradation. This
the maneuver plane. generalized coordinated turn model does not estimate
If the gravity were not extracted, then the the turn rate explicitly and is highly nonlinear.
orthogonal-velocity condition (105) and the constant While the constant-turn models of Section VIB
acceleration condition da% =dt = 0 (not equivalent to and the coordinated-turn models of Section VIC are
da%g =dt = 0) would lead to the following second-order similar, they are based on two quite different ideas.
model in terms of the damping coefficient ® and turn The former are of kinematic type, which aim at fitting
typical target trajectories during a maneuver, while
rate !
the development of the latter models relies on flight
v a va a2 dynamics explicitly.
a_ = a= a v = ®a (®2 + ! 2 )v
v2 v2 v2 The generalized coordinated turn model (139) is
(140) more convenient than the coordinated turn model
which is similar to (122). (e.g., without the need to determine the thrust and
The conditions " v and da%g =dt = 0 are closely lift). However, the planar VT model (122) is even
related with coordinated turns. The constant thrust more convenient for most applications. The underlying
and lift condition validates assumption da%g =dt = 0; assumptions for the constant-turn models of Section
the zero angle of attack and sideslip condition aligns VIB are that (105) holds and that the turn rate or
the »-axis of the real body frame (as defined in flight the angular velocity vector is constant. Note that it
dynamics) with the assumed velocity-based body follows from (107) and _ = 0 that a_ = a. It differs
frame; and the constant roll angle condition validates from the generalized coordinated turn model only in
assumption v = 0 since the roll angular rate is the lack of the g term, which is constant if is
the projection of the angular velocity to the heading constant. In particular, the two models coincide when
direction: Ã_ = (v=v). As such, a coordinated turn the maneuver plane is horizontal.
motion can be readily described by (139) and thus Some other 3D target maneuver models are hard
(139) is a generalized model for coordinated turn and to be grouped into the above classes. For example, the
quasi coordinated turn motions. maneuver model b̈ = 0, ë = 0 was proposed in [97]
This model was originally proposed in [98], _ e, e_ ] , where (x, y, z) and v
with a state x = [x, y, z, v, b, b,
implemented in [104], and later analyzed and are target position and speed, respectively, and b and
validated in [99] by using real data of aircraft e are the bearing and elevation of the velocity vector,
trajectories. respectively.
It was proposed in [104] to use a (vector-valued)
Singer process u to model the perturbations in D. Flight Dynamics Based Models
acceleration in the body frame. As a result, the
state-space representation of this generalized An analysis of real trajectory data presented in
coordinated turn model for the state vector x = [99] reveals that a target motion is not necessarily
confined to a plane in reality, but all the above models the relationship between the Euler angles and the
are exclusively or primarily for planar motions angular velocity. The aircraft body rotational motion
and may be inadequate for more sophisticated ((144)–(145)) strongly affects the translational motion
target motions. Models that incorporate sufficient ((142)–(143)) through the dependence of the model
out-of-plane motions may be more desirable in such coefficients c(©), c® (©), c¯ (©), C± (©), cT (©) on the
situations. A class of more sophisticated models, attitude angles © [105].
based on flight dynamics, appears to be a better Given all the model coefficients, the uncertainty
alternative in this regard. The price is, however, a of a maneuver boils down to that of the control input:
dramatic increase in nonlinearity, state dimension, thrust T and control angles ±. A state-space model
and number of design parameters largely uncertain can be developed in a trivial manner. For example,
in a noncooperative tracking environment. Also, they augment the state vector by T and ± and model their
are not really point-target models. We review these variations as zero-mean white noise processes [105].
models next. The resulting state vector x = [p , v , , © , ± , T] is
The basic idea of these models is to enhance 16-dimensional and the state-space model is given
the modeling accuracy by not only exploring more by (142)–(145), with white noise added to (143) and
detailed flight-dynamics relationships but also (144), plus
explicitly accounting for the aircraft’s controls. Were T_ = wT , ±_ = w± : (147)
the controls known to trackers, these models would
describe an aircraft motion more accurately than other The model (142)–(147) is quite general. What
models. So a fundamental issue here is the uncertainty is crucial here is the determination of the model
in the pilot’s control actions in a noncooperative coefficients c(©), c® (©), c¯ (©), C± (©), cT (©) and d® ,
environment. Modeling these controls as random d¯ , D , D± , dT . This was done in [105] heuristically
processes is again a “standard” means to account for based on an analysis of a particular real trajectory
this uncertainty. We consider the model of [105] to of a given aircraft. Caution has to be exercised in the
illustrate features of these models. application of such results to other situations.
The basis for this model is the rigid-body flight A more rigorous model, albeit not general,
dynamics model of aircraft motion [83], given in the was proposed in [106]. Under some simplifying
form assumptions essentially the same as those for the
coordinated turn, an analytical expression of the lift as
p_ = v (142) a function of the attitude angles © was derived, which
2 explicitly describes the dependence of the acceleration
v_ = (c + c® ® + c¯ ¯+C± ±)v + cT T + g (143)
v_ on © in (143). The effects of the zero-mean thrust
_ = f() + (d ® + d ¯ + D + D ±)v 2 + d T
(minus drag) term T D and deflection angles were
® ¯ ± T
(144) collectively modeled as a first-order Markov (Singer)
process °, governed by °_ = (1=¿ )° + w° , in place of
_ = G(©)
© (145) (147) as above. The state vector x = [p , v , , © , ° ]
is 15-dimensional.
with
In general, prediction accuracy of these models
(Ixx Iyy )qr=Ixx for a target state can be dramatically improved
by accounting for the impact of the target attitude
f() = (Ixx Izz )rp=Ixx
angles on the target acceleration. In fact, these
(Iyy Ixx )pq=Ixx models are clearly unobservable under position-only
(146)
1 sin Á tan µ cos Á tan µ measurements. Estimating the attitude angles is
possible if, in addition to the standard position (and
G(©) = 0 cos Á sin Á
possibly velocity) measurements, direct attitude
0 sin Á cos1 µ cos Á cos1 µ measurements are available. The use of attitude
where p = [x, y, z] is the position in the inertial frame, measurements was first suggested in [43] and further
= [p, q, r] the angular velocity in the body frame, developed in [107, 106, 108, 105], and others.
® the angle of attack, ¯ the sideslip, ± = [±A , ±R , ±E ]
the angles of aileron, rudder, and elevator deflections, VII. CONCLUDING REMARKS
respectively, © = [Ã, µ, Á] an attitude vector of the
Euler angles: roll (bank angle) Ã, pitch µ, yaw The target maneuver models developed for target
(heading) Á, T the thrust, g the gravity, I the moment tracking can be classified into 1D, 2D, and 3D
of inertia, and c, c® , c¯ , C± , cT , d® , d¯ , D± , dT are categories, according to the coupling between motions
coefficients. Equations (142)–(143) describe the along different coordinates.
aircraft body translational motion in the inertial frame, The 1D models have no (or weak) coupling among
(144) is a moment equation describing the angular coordinates. Most of them are based on modeling
acceleration in the body frame, and (145) provides the driving force (usually acceleration) of the target
maneuver as a random process without recourse to the ACKNOWLEDGMENTS
actual target dynamics or kinematics directly. These
models form three families, corresponding to three The authors would like to thank the following
classes of random processes: white noise, Markov people for their comments on the conference version
processes, and semi-Markov jump processes. The [13] and/or draft of this paper: Yaakov Bar-Shalom,
simplest white-noise family models the target position Edward Beadle, Jeff Bell, Bob Bishop, Henk Blom,
derivative (or difference) of a certain order as white Jean Dezert, Bob Fitzgerald, Keith Kastella, Genter
noise. It includes the classical CV and CA models. van Keuk, Ron Mahler, and Dave Sworder.
In the simple and widely used family of Markov
processes, the target acceleration or jerk is modeled REFERENCES
as a Markov process of various degrees of complexity.
The most well-known representative of this family [1] Blackman, S. S., and Popoli, R. F. (1999)
is the Singer model. The semi-Markov jump process Design and Analysis of Modern Tracking Systems.
Boston, MA: Artech House, 1999.
models are the most sophisticated family of the three.
[2] Bar-Shalom, Y., Li, X. R., and Kirubarajan, T. (2001)
It has quite good potential, but the intricacy involved Estimation with Applications to Tracking and Navigation:
in these models makes it hard for the practitioners to Theory, Algorithms, and Software.
apply them. New York: Wiley, 2001.
The 2D and 3D models differ from the 1D models [3] Bar-Shalom, Y., and Li, X. R. (1993)
by not only their correlation across coordinates, but Estimation and Tracking: Principles, Techniques, and
Software.
also their explicit dependence on the target kinematics Boston, MA: Artech House, 1993. (Reprinted by YBS
and possibly dynamics. These models try to capture Publishing, 1998).
the important characteristics of the target behavior [4] Bar-Shalom, Y., and Li, X. R. (1995)
during maneuvers, which typically involve various Multitarget-Multisensor Tracking: Principles and
turning motions. Many versions of (nearly) CT models Techniques.
Storrs, CT: YBS Publishing, 1995.
have been developed. Few models are available that
[5] Bar-Shalom, Y. (Ed.) (1990)
are valid for other maneuver motions. Apart from the Multitarget-Multisensor Tracking: Advanced Applications.
popular CT models, it seems that the 2D approximate Norwood, MA: Artech House, 1990.
curvilinear motion model and the planar variable-turn [6] Bar-Shalom, Y. (Ed.) (1992)
model of (122) deserve more attention than they have Multitarget-Multisensor Tracking: Applications and
received due to their attractive simplicity, applicability, Advances, Vol. II, Norwwod, MA: Artech House, 1992.
and flexibility. [7] Bar-Shalom, Y., and Fortmann, T. E. (1988)
Tracking and Data Association.
More accurate models are intuitively appealing, but New York: Academic Press, 1988.
they also pose problems and challenges. For example, [8] Blackman, S. S. (1986)
they are usually highly nonlinear and thus require Multiple Target Tracking with Radar Applications.
effective nonlinear filtering algorithms; they may lead Norwood, MA: Artech House, 1986.
to poor tracking performance due to their dependence [9] Farina, A., and Studer, F. A. (1985)
on target state in the case of inaccurate state estimates, Radar Data Processing, Vol. I: Introduction and Tracking,
Vol. II: Advanced Topics and Applications.
particularly for agile targets at a low data rate. That Letchworth, Hertfordshire, England: Research Studies
is partly why simple but less accurate models have Press, 1985.
their values and reasons to exist. While the choice of [10] Maybeck, P. S. (1982)
a particular model depends certainly on the particular Stochastic Models, Estimation and Control, Vols. II, III.
application, a good understanding of pros and cons of New York: Academic Press, 1982.
each model is nonetheless extremely helpful. [11] Maybeck, P. S. (1979)
Stochastic Models, Estimation and Control, Vol. I.
Most of the models developed in the literature, New York: Academic Press, 1979.
especially 2D and 3D models, are particularly [12] Gelb, A. (1974)
suitable for aircraft, although they are more or less Applied Optimal Estimation.
applicable to many other targets. Many of these Cambridge, MA: MIT Press, 1974.
aircraft models are fairly accurate, but they have a [13] Li, X. R., and Jilkov, V. P. (2000)
common omission—wind effect is not considered A survey of maneuvering target tracking: Dynamic
models.
explicitly. A reasonably large number of dynamic In Proceedings of the 2000 SPIE Conference on Signal and
models have also been developed for ballistic targets Data Processing of Small Targets, Vol. 4048, Orlando, FL,
(e.g., missiles), which will be covered separately in Apr. 2000, 212–236.
a subsequent part of this survey, of which [14] is [14] Li, X. R., and Jilkov, V. P. (2001)
a preliminary version. To our knowledge, however, A survey of maneuvering target tracking—Part II:
Ballistic target models.
few models have been developed that are particularly In Proceedings of the 2001 SPIE Conference on Signal and
suitable for other targets, such as submarines, ships, Data Processing of Small Targets, Vol. 4473, San Diego,
and ground targets. This deserves more effort. CA, July–Aug. 2001, 559–581.
[15] Li, X. R., and Jilkov, V. P. (2001) [29] Pearson, J. B., and Stear, E. B. (1974)
A survey of maneuvering target tracking—Part III: Kalman filter applications in airborne radar tracking.
Measurement models. Transactions on Aerospace and Electronic Systems, AES-10
In Proceedings of the 2001 SPIE Conference on Signal and (1974), 319–329.
Data Processing of Small Targets, Vol. 4473, San Diego, [30] Fitts, J. M. (1973)
CA, July–Aug. 2001, 423–446. Aided tracking as applied to high accuracy pointing
[16] Li, X. R., and Jilkov, V. P. (2002) systems.
A Survey of maneuvering target tracking—Part IV: Transactions on Aerospace and Electronic Systems, AES-9
Decision-based methods. (May 1973).
In Proceedings of the 2002 SPIE Conference on Signal and [31] Landau, M. (1976)
Data Processing of Small Targets, Vol. 4728, Orlando, FL, Radar tracking of airborne targets.
Apr. 2002. Presented at the National Aerospace and Electronics
[17] Li, X. R., and Jilkov, V. P. (2003) Conference (NAECON), Dayton, OH, 1976.
A survey of maneuvering target tracking—Part V: [32] Blair, W. D., Watson, G. A., and Rice, T. R. (1991)
Multiple-model methods. Tracking maneuvering targets with an interacting
In Proceedings of the 2003 SPIE Conference on Signal and multiple model filter containing exponentially correlated
Data Processing of Small Targets, Vol. 5204, San Diego, acceleration models.
CA, Aug. 2003. Southeastern Symposium on Systems Theory, Columbia,
[18] Chan, Y. T., Hu, A. G. C., and Plant, J. B. (1979) SC, Mar. 1991.
A Kalman filter based tracking scheme with input
[33] Mandzuka, S. (2000)
estimation.
Ship tracking control: Optimal estimation of navigation
Transactions on Aerospace and Electronic Systems,
parameters.
AES-15, 2 (Mar. 1979), 237–244.
In Proceedings of the 42th International Symposium
[19] Korn, J., Gully, S. W., and Willsky, A. S. (1982) ELMAR, Zadra, 2000.
Application of the generalized likelihood ratio algorithm
[34] Ekstrand, B. (2001)
to maneuver detection and estimation.
Tracking filters and models for seeker applications.
In Proceedings of the 1982 American Control Conference,
Transactions on Aerospace and Electronic Systems,
Arlington, VA, June 1982.
AES-37, 3 (2001), 965–976.
[20] Bogler, P. L. (1987)
Tracking a maneuvering target using input estimation. [35] McAulay, R. J., and Denlinger, E. J. (1973)
Transactions on Aerospace and Electronic Systems, A decision-directed adaptive tracker.
AES-23, 3 (May 1987), 298–310. Transactions on Aerospace and Electronic Systems, AES-9,
2 (Mar. 1973), 229–236.
[21] Bogler, P. L. (1990)
Radar Principles with Applications to Tracking Systems, [36] Cloutier, J. R., Evers, J. H., and Feeley, J. J. (1989)
New York: Wiley, 1990. Assessment of air-to-air missile guidance and control
[22] Farooq, M., and Bruder, S. (1990) technology.
Information type filters for tracking a maneuvering target. IEEE Control Systems Magazine, CSM-9 (Oct. 1989),
Transactions on Aerospace and Electronic Systems, 26, 3 27–34.
(1990), 441–454. [37] Easthope, P. F., and Heys, N. W. (1994)
[23] Whang, I. H., Lee, J., and Sung, T. (1994) Multiple-model target-oriented tracking system.
Modified input estimation technique using In Proceedings of the 1994 SPIE Conference on Signal and
pseudoresiduals. Data Processing of Small Targets, Vol. 2235, Apr. 1994.
Transactions on Aerospace and Electronic Systems, 30, [38] Costa, P. J. (1994)
1 (1994), 220–228; also published in 30, 2 (1994), Adaptive model architecture and extended Kalman-Bucy
591–598. filters.
[24] Sung, T. K., and Lee, J. G. (1997) Transactions on Aerospace and Electronic Systems, 30, 2
A decoupled adaptive tracking filter for real applications. (Apr. 1994), 525–533.
Transactions on Aerospace and Electronic Systems, 33, 3 [39] D’Souza, C. N., McClure, M. A., and Cloutier, J. R. (1994)
(1997), 1025–1030. Spherical target state estimators.
[25] Lee, H., and Tahk, M-J. (1999) In Proceedings of 1994 American Control Conference,
Generalized input-estimation technique for tracking Baltimore, MD, June 1994, 1675–1679.
maneuvering targets. [40] Pulford, G., and La Scala, B. (1996)
IEEE Transactions on Aerospace and Electronic Systems, A survey of manoeuvring target tracking methods and
35, 4 (1999), 1388–1402. their applicability to over-the-horizon radar.
[26] Singer, R. A. (1970) Technical Report CSSIP 14, Cooperative Research Center
Estimating optimal tracking filter performance for for Sensor Signal and Information Processing, Australia,
manned maneuvering targets. July 1996.
Transactions on Aerospace and Electronic Systems, AES-6 [41] Zhou, H., and Kumar, K. S. P. (1984)
(July 1970), 473–483. A “current” statistical model and adaptive algorithm for
[27] Singer, R. A., and Benhke, K. W. (1971) estimating maneuvering targets.
Real-time tracking filter evaluation and selection for AIAA Journal of Guidance, 7, 5 (Sept.–Oct. 1984),
tactical applications. 596–602.
Transactions on Aerospace and Electronic Systems, AES-7 [42] Jilkov, V. P., Mihaylova, L. S., and Li, X. R. (1998)
(Jan. 1971), 100–110. An alternative IMM solution to benchmark radar tracking
[28] Pearson, J. B. (1970) problem.
Basic Studies in Airborne Radar Tracking Systems. In Proceedings of International Conference on
Ph.D. dissertation, University of California at Los Multisource-Multisensor Information Fusion, July 1998,
Angeles, 1970. 924–929.
[43] Kendrick, J. D., Maybeck, P. S., and Reid, J. G. (1981) [57] Sworder, D. D., Kent, M., Vojak, R., and Hutchins, R. G.
Estimation of aircraft target motion using orientation (1995)
measurements. Renewal models for maneuvering targets.
Transactions on Aerospace and Electronic Systems, AES-17 IEEE Transactions on Aerospace and Electronic Systems,
(Mar. 1981), 254–260. 31, 1 (Jan. 1995), 138–149.
[44] Fitzgerald, R. J. (1967) [58] Cox, D. R. (1962)
Filtering horizon-sensor measurements for orbital Renewal Theory.
navigation. London: Methuen, 1962.
Journal of Spacecraft and Rockets, 4, 4 (Apr. 1967), [59] Sworder, D. D., Singer, P. F., and Hutchins, R. G. (1993)
428–435. Image-enhanced estimation methods.
[45] Wheaton, B. J., and Maybeck, P. S. (1995) Proceedings of the IEEE, 81, 6 (June 1993), 797–812.
Second-order acceleration model for an MMAE target [60] Mehrotra, K., and Mahapatra, P. R. (1997)
tracker. A jerk model to tracking highly maneuvering targets.
Transactions on Aerospace and Electronic Systems, 31, 1 Transactions on Aerospace and Electronic Systems, 33, 4
(1995), 151–166. (1997), 1094–1105.
[46] Ozkaya, B., and Arcasoy, C. (1998) [61] Mahapatra, P. R., and Mehrotra, K. (2000)
Analytical solution of discrete colored noise ECA Mixed coordinate tracking of generalized maneuvering
tracking filter. targets using acceleration and jerk models.
Transactions on Aerospace and Electronic Systems, 34, 1 Transactions on Aerospace and Electronic Systems, 36, 3
(1998), 93–102. (2000), 992–1001.
[47] Helferty, J. P. (1996)
[62] Berg, R. F. (1983)
Improved tracking of maneuvering targets: The use of
Estimation and prediction for maneuvering target
turn-rate distriburions for acceleration modeling.
trajectories.
Transactions on Aerospace and Electronic Systems, 32, 4
IEEE Transactions on Automatic Control, AC-28 (Mar.
(Oct. 1996), 1355–1361.
1983), 294–304.
[48] Li, X. R., Zhao, Z-L., Zhang, P., and He, C. (2002)
[63] Song, T. L., Ahn, J., and Park, C. (1988)
Model-set design for multiple-model estimation—Part II:
Suboptimal filter design with pseudomeasurements for
Examples.
target tracking.
In Proceedings of 2002 International Conference
Transactions on Aerospace and Electronic Systems, 24 (Jan.
on Information Fusion, Annapolis, MD, July 2002,
1988), 28–39.
1347–1354.
[64] Gustafsson, F., and Isaksson, A. J.
[49] Howard, R. A. (1964)
Best choice of state variables for tracking coordinated
System analysis of semi-Markov processes.
turns.
IEEE Transactions Military Electronics, MIL-8 (Apr.
In Proceedings of the 35th IEEE Conference on Decision
1964), 114–124.
and Control, Kobe, Japan, Dec. 1996, 3145–3150.
[50] Moose, R. L., and Wang, P. L. (1973)
An adaptive estimator with learning for a plant containing [65] Semerdjiev, E., Mihaylova, L., and Li, X. R. (1999)
semi-Markov switching parameters. An adaptive IMM estimator for aircraft tracking.
IEEE Transactions on Systems, Man, Cybernetics, SMC-3 In Proceedings of 1999 International Conference on
Information Fusion, Sunnyvale, CA, July 1999, 770–776.
(May 1973), 277–281.
[51] Moose, R. L. (1975) [66] Semerdjiev, E., Mihaylova, L., and Li, X. R. (2003)
An adaptive state estimator solution to the maneuvering An adaptive IMM estimation technique based on
target problem. perturbation models.
IEEE Transactions on Automatic Control, AC-20, 3 (June Submitted to Transactions on Aerospace and Electronic
1975), 359–362. Systems.
[52] Gholson, N. H., and Moose, R. L. (1977) [67] Li, X. R., and Bar-Shalom, Y. (1993)
Maneuvering target tracking using adaptive state Design of an interacting multiple model algorithm for air
estimation. traffic control tracking.
Transactions on Aerospace and Electronic Systems, AES-13 IEEE Transactions on Control Systems Technology, 1, 3
(May 1977), 310–317. (Sept. 1993), 186–194; special issue on air traffic control.
[53] Moose, R. L., VanLandingham, H. F., and McCabe, D. H. [68] Dufour, F., and Mariton, M. (1991)
(1979) Tracking a 3D maneuvering target with passive sensors.
Modeling and estimation of tracking maneuvering targets. Transactions on Aerospace and Electronic Systems, 27, 4
Transactions on Aerospace and Electronic Systems, (July 1991), 725–739.
AES-15, 3 (May 1979), 448–456. [69] Dufour, F., and Mariton, M. (1992)
[54] Li, X. R. (2002) Passive sensor data fusion and maneuvering target
Model-set design for multiple-model estimation—Part I. tracking.
In Proceedings of 2002 International Conference on In Y. Bar-Shalom (Ed.), Multitarget-Multisensor Tracking:
Information Fusion, Annapolis, MD, July 2002, 26–33. Applications and Advances, Vol. II, Norwood, MA: Artech
[55] Lim, S. S., and Farooq, M. (1991) House, 1992, ch. 3.
Maneuvering target tracking using jump processes. [70] Watson, G. A., and Blair, W. D. (1992)
In Proceedings of the 30th IEEE Conference on Decision IMM algorithm for tracking targets that maneuver
and Control, Brighton, UK, 1991, 2049–54. through coordinated turns.
[56] Campo, L., Mookerjee, P., and Bar-Shalom, Y. (1991) In Proceedings of the 1992 SPIE Conference on Signal and
State estimation for systems with sojourn-time-dependent Data Processing for Small Targets, Vol. 1698, 236–247.
Markov model switching. [71] Kastella, K., and Biscuso, M. (1996)
IEEE Transactions on Automatic Control, 36, 2 (Feb. Tracking algorithms for air traffic control applications.
1991), 238–243. Air Traffic Control Quarterly, 3, 1 (Jan. 1996), 19–43.
[72] Wang, H., Kirubarajan, T., and Bar-Shalom, Y. (1999) [87] Pershitz, R. (1973)
Precision large scale air traffic surveillance using Ship’s maneuverability and control.
IMM/assignment estimators. Leningrad, USSR: Sudostroenie, 1973.
Transactions on Aerospace and Electronic Systems, 35, 1 [88] Semerdjiev, E., Mihaylova, L., and Semerdjiev, T. (1998)
(Jan. 1999), 255–266. Manoeuvring ship model identification and interacting
[73] Gustafsson, F. (2001) multiple model tracking algorithm design.
Adaptive Filtering and Change Detection. In Proceedings of 1998 International Conference on
New York: Wiley, 2001. Information Fusion, Las Vegas, NV, July 1998, 968–973.
[74] Vacher, P., Barret, I., and Gauvrit, M. (1992) [89] Semerdjiev, E., and Mihaylova, L. (2000)
Design of a tracking algorithm for an advanced ATC Variable- and fixed-structure augmented interacting
system. multiple model algorithm for manoeuvring ship tracking
In Y. Bar-Shalom (Ed.), Multitarget-Multisensor Tracking: based on new ship models.
Applications and Advances, Vol. II, Norwood, MA: Artech International Journal of Applied Mathematics and
House, 1992, ch. 1. Computers, 10 (2000), 591–604.
[75] Lerro, D., and Bar-Shalom, Y. (1993) [90] Semerdjiev, E., and Mihaylova, L. (1998)
Interacting multiple model algorithm with target Adaptive interacting multiple model algorithm for
amplitude feature. manoeuvring ship tracking.
Transactions on Aerospace and Electronic Systems, 29, 2 In Proceedings of 1998 International Conference on
(Apr. 1993), 494–509. Information Fusion, Las Vegas, NV, July 1998, 974–979.
[76] Gertz, J. L. (1989) [91] Tzeng, C-Y., and Chen, J-F. (1999)
Multisensor surveillance for improved aircraft tracking. Fundamental properties of linear ship steering dynamic
Lincoln Laboratory Journal, 2, 3 (1989), 381–396. models.
[77] Busch, M., and Blackman, S. (1995) Journal of Marine Science and Technology, 7, 2 (1999),
Evaluation of IMM filtering for an air defence system 79–88.
application. [92] Skjetne, R., and Fossen, T. I. (2001)
In Proceedings of the 1995 SPIE Conference on Signal and Nonlinear maneuvering and control of ships.
Data Processing of Small Targets, Vol. 2561, 435–447. In Proceedings of Oceans 2001 MTS/IEEE Conference,
[78] Blackman, S. S., Busch, M. T., and Popoli, R. F. (1999) Vol. 3, Nov. 2001, 1808–1815.
IMM/MHT solution to radar benchmark tracking [93] Miele, A. (1962)
problem. Flight Mechanics: Theory of Flight Paths, Vol. 1.
Transactions on Aerospace and Electronic Systems, 35, 2 Reading, MA: Addison-Wesley, 1962.
(Apr. 1999), 730–737. Also appeared in Proceedings of [94] Hauser, W. (1965)
1995 American Control Conference, Seattle, WA, June Introduction to the Principles of Mechanics.
1995, 2606–2610. New York: Wiley, 1965.
[79] Blom, H. A. P., Hogendoorn, R. A., and van Doorn, B. A. [95] Bryan, R. S. (1980)
(1992) Cooperative estimation of targets by multple aircraft.
Design of a multisensor tracking system for advanced air M.S. thesis, Air Force Institute of Technology,
traffic control. Wright-Patterson AFB, Ohio, June, 1980.
In Y. Bar-Shalom (Ed.), Multitarget-Multisensor Tracking: [96] Asseo, S. J., and Ardila, R. J. (1982)
Applications and Advances, Vol. II, Norwood, MA: Artech Sensor independant target state estimator design and
House, 1992, ch. 2. evaluation.
[80] Roecker, J. A., and McGillem, C. D. (1989) In Procceedings of the National Aerospace and Electronics
Target tracking in maneuver centered coordinates. Conference (NAECON), 1982, 916–924.
Transactions on Aerospace and Electronic Systems, 25 [97] Bullock, T. E., and Sangsuk-Iam, S. (1984)
(Nov. 1989), 836–843. Maneuver detection and tracking with a nonlinear target
[81] Kawase, T., Tsurunosono, H., Ehara, N., and Sasase, I. model.
(1998) In Proceedings of the 23rd IEEE Conference on Decision
Two-stage Kalman estimator using an advanced circular and Control, Las Vegas, NV, Dec. 1984.
prediction for tracking highly maneuvering targets. [98] Bishop, R. H., and Antoulas, A. C. (1994)
In Proceedings of ICASSP98, 1998. Nonlinear approach to the aircraft tracking problem.
[82] Best, R. A. and Norton, J. P. (1997) AIAA Journal of Guidance, Control and Dynamics, 17,
A new model and efficient tracker for a target with 5 (1994), 1124–1130. Also in Proceedings of the 1991
curvilinear motion. AIAA Guidance, Navigation and Control Conference, New
Transactions on Aerospace and Electronic Systems, 33, 3 Orleans, LA, Aug. 1991, 692–703.
(July 1997), 1030–1037. [99] Nabaa, N., and Bishop, R. H. (2000)
[83] Roskam, J. (1979) Validation and comparison of coordinated turn aircraft
Airplane Flight Dynamics and Automatic Control, Part I. maneuver models.
Roskam Aviation and Engineering Corp., Lawrence, KS, Transactions on Aerospace and Electronic Systems, 36, 1
1979. (Jan. 2000), 250–259.
[84] Regan, F. J., and Anandakrishnan, S. M. (1993) [100] Maybeck, P. S., Worsley, W. H., and Flynn, P. M. (1982)
Dynamics of Atmospheric Re-Entry. Investigation of constant turn-rate dynamics for airborne
New York: AIAA, New York, 1993. vehicle tracking.
[85] Etkin, B., and Reid, L. D. (1996) In Proceedings of the National Aerospace and Electronics
Dynamcis of Flight: Stability and Control. Conference (NAECON), 1982, 896–903.
New York: Wiley, 1996. [101] Tahk, M., and Speyer, J. L. (1990)
[86] Boiffier, J-L. (1998) Target tracking subject to kinematic constrants.
The Dynamics of Flight: The Equations. IEEE Transactions on Automatic Control, 35 (Mar. 1990),
New York: Wiley, 1998. 324–326.
[102] Alouani, A. T., and Blair, W. D. (1991) [106] Andrisani, D., Kuhl, F. P., and Gleason, D. (1986)
Use of a kinematic constraint in tracking constant speed, A nonlinear tracker using attitude measurements.
maneuvering targets. Transactions on Aerospace and Electronic Systems, 22
In Proceedings of 30th IEEE Conference on Decision and (Sept. 1986), 533–539.
Control, Brighton, England, Dec. 1991, 1055–1058. [107] Lefas, C. C. (1984)
[103] Alouani, A. T., and Blair, W. D. (1993) Using roll-angle measurement to track aircraft maneuvers.
Use of a kinematic constraint in tracking constant speed, Transactions on Aerospace and Electronic Systems, AES-20
maneuvering targets. (Nov. 1984), 672–681.
IEEE Transactions on Automatic Control, 38, 7 (July [108] Andrisani, D., Kim, E. T., Schierman, J., and Kuhl, F. P.
1993), 1107–1111. (1991)
[104] Nabaa, N., and Bishop, R. (1999) A nonlinear helicopter tracker using attitude
Solution to a multisensor tracking problem with sensor measurements.
registration errors. Transactions on Aerospace and Electronic Systems, 27 (Jan.
Transactions on Aerospace and Electronic Systems, 35, 1 1991), 40–47.
(1999), 354–362.
[105] Mook, D. J., and Shyu, I-M. (1992)
Nonlinear aircraft tracking filter utilizing control variable
estimation.
AIAA Journal of Guidance, 15, 1 (Jan.–Feb. 1992),
228–237.
X. Rong Li (S’90—M’92—SM’95) received the B.S. and M.S. degrees from
Zhejiang University, Hangzhou, Zhejiang, PRC, in 1982 and 1984, respectively,
and the M.S. and Ph.D. degrees from the University of Connecticut, Storrs, in
1990 and 1992, respectively.
He joined the Department of Electrical Engineering, University of New
Orleans in 1994, where he is now university research professor, department chair,
and director of information and systems laboratory. During 1986–1987 he did
research on electric power at the University of Calgary, AB, Canada. He was an
assistant professor at the University of Hartford, West Hartford, CT, from 1992
to 1994. His current research interests include signal and data processing, target
tracking and information fusion, stochastic systems, statistical inference, and
electric power.
Dr. Li has authored or coauthored four books: Estimation and
Tracking (with Yaakov Bar-Shalom, Norwood, MA: Artech House, 1993),
Multitarget-Multisensor Tracking (with Yaakov Bar-Shalom, YBS Publishing,
1995), Probability, Random Signals, and Statistics (CRC Press, 1999), and
Estimation with Applications to Tracking and Navigation (with Yaakov Bar-Shalom
and T. Kirubarajan, Wiley, 2001); six book chapters; and more than 160 journal
and conference proceedings papers.
Dr. Li has served the International Society of Information Fusion as the
president (2003), vice president (1998–2002) and a member of board of directors
(since 1998); served as general chair for 2002 International Conference on
Information Fusion, and steering chair or general vice-chair for 1998, 1999, and
2000 International Conferences on Information Fusion; served IEEE Transactions
on Aerospace and Electronic Systems as an associate editor from 1995 to 1996
and as editor since 1996; served Communications in Information and Systems
as an editor since 2001; received a CAREER award and an RIA award from
the U.S. National Science Foundation. He received 1996 Early Career Award
for Excellence in Research from the University of New Orleans and has given
numerous seminars and short courses in U.S., Europe and Asia. He won several
outstanding paper awards, is listed in Marquis’ Who’s Who in America and Who’s
Who in Science and Engineering, and consulted for several companies.
Vesselin P. Jilkov (M’01) received his B.S. and M.S. degree in mathematics
from the University of Sofia, Bulgaria in 1982, the Ph. D. degree in the technical
sciences in 1988, and the academic rank senior research fellow of the Bulgarian
Academy of Sciences in 1997.
He was a research scientist with the R&D Institute of Special Electronics,
Sofia, (1982–1988) where he was engaged in research and development of radar
tracking systems. From 1989 to 1999 he was a research scientist with the Central
Laboratory for Parallel Processing—Bulgarian Academy of Sciences, Sofia,
where he worked as a key researcher in numerous academic and industry projects
(Bulgarian and international) in the areas of Kalman filtering, target tracking,
multisensor data fusion, and parallel processing. Since 1999 Dr. Jilkov has been
with the Department of Electrical Engineering, University of New Orleans, where
he is currently an assistant professor, and is engaged in teaching and conducting
research in the areas of hybrid estimation and target tracking. His current research
interests include stochastic systems, nonlinear filtering, applied estimation, target
tracking, information fusion.
Dr. Jilkov is author/coauthor of over 50 journal articles and conference papers.
He is a member of ISIF (International Society of Information Fusion).

Survey of Maneuvering Target Tracking. Part I. Dynamic Models

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Survey of Maneuvering Target Tracking. Part I. Dynamic Models

Uploaded by

Copyright:

Available Formats

I.

The key to successful target tracking lies in the

E. Mean-Adaptive Acceleration Model āk+1 = E[ak+1 M z k ] = e®T E[ak M z k ] + (1 e®T )āk

Fig. 3. Asymmetric pdfs of normal acceleration.

!c (2® + ®2 T + !c2 T) + e®T [2®!c cos !c T + (®2 !c2 ) sin !c T]

where ã(t) is the Singer acceleration of (26), v is

Fig. 5. Gamma densities with unity mean value.

discretization, respectively, in [73, 64]. The second (75)

! = K (89) Many of the 2D horizontal models reviewed

xk+1 = diag[F(!), F(!), F(!)]xk + wk (116)

18 Thecovariance cov(wk ) given here appears more reasonable than

You might also like

Survey of Maneuvering Target Tracking. Part I. Dynamic Models

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Survey of Maneuvering Target Tracking. Part I. Dynamic Models

Uploaded by

Copyright:

Available Formats

I.

The key to successful target tracking lies in the

E. Mean-Adaptive Acceleration Model āk+1 = E[ak+1 M z k ] = e®T E[ak M z k ] + (1  e®T )āk

Fig. 3. Asymmetric pdfs of normal acceleration.

!c (2® + ®2 T + !c2 T) + e®T [2®!c cos !c T + (®2  !c2 ) sin !c T]

where ã(t) is the Singer acceleration of (26), v is

Fig. 5. Gamma densities with unity mean value.

discretization, respectively, in [73, 64]. The second (75)

! = K­ (89) Many of the 2D horizontal models reviewed

xk+1 = diag[F(!), F(!), F(!)]xk + wk (116)

18 Thecovariance cov(wk ) given here appears more reasonable than

You might also like

E. Mean-Adaptive Acceleration Model āk+1 = E[ak+1 M z k ] = e®T E[ak M z k ] + (1 e®T )āk

!c (2® + ®2 T + !c2 T) + e®T [2®!c cos !c T + (®2 !c2 ) sin !c T]

! = K (89) Many of the 2D horizontal models reviewed