Download as pdf or txt
Download as pdf or txt
You are on page 1of 24

I.

INTRODUCTION

A survey of dynamic models used in maneuvering


Survey of Maneuvering target tracking has been reported in [64]. It, however,
does not cover motion models used for tracking
Target Tracking. ballistic and space targets (BT), namely, ballistic
missiles, decoys, debris, satellites, projectiles, etc.
Part II: Motion Models of The primary reason for this omission is that these
models possess many distinctive features that differ
Ballistic and Space Targets vastly from those covered in [64]. To supplement
[64], a survey of these models is presented here. To
X. RONG LI, Fellow, IEEE our knowledge, such a survey is not available in the
literature.
VESSELIN P. JILKOV, Member, IEEE
University of New Orleans Overall, a BT has a less uncertain motion than
many other types of powered vehicles, such as
This paper is the second part in a series that provides a maneuvering aircraft or agile missiles: Most BTs
comprehensive survey of maneuvering target tracking without follow a flight path that is largely predetermined
addressing the so-called measurement-origin uncertainty. It by the performance characteristics specific for a
surveys motion models of ballistic targets used for target tracking.
Models for all three phases (i.e., boost, coast, and reentry) of target type, hence the name ballistic targets, although
motion are covered. some more advanced ballistic missiles can undergo
small maneuvers usually for retargeting. However,
CONTENTS this does not mean that the motion of a foreign BT
I. Introduction can be determined accurately. In fact, tracking a
II. Ballistic (Coast) Flight foreign BT is likely to be plagued with a variety of
A. Gravity
B. Coordinate Accelerations uncertainties, including those concerning trajectory
C. Coast Motion Models loft or depression, thrust profile management, target
III. Reentry weight, propellant specific impulse, sensor bias, and
A. Aerodynamic Forces
B. Ballistic (Nonmaneuvering) RV Motion atmospheric parameters. Many of these uncertainties
C. Maneuvering RV Motion stem from the uncertainty in the target type, the
D. Motion of Endo-Atmospheric Ballistic Targets principal uncertainty in modeling the motion of a
IV. Boost Phase
A. Accelerations foreign BT for the tracking purpose.
B. Kinematic Models The entire trajectory of a BT is commonly
C. Dynamic Models divided into three basic phases: boost (and possibly
D. Example of A Profile-Based Model
V. Integration of BT Motion Models for Entire Trajectory a post-boost phase), ballistic (also known as coast),
VI. Concluding Remarks and reentry. The boost phase is the powered,
References endo-atmospheric flight, which lasts from launch to
Appendix
thrust cutoff or burnout. It is followed by the ballistic
phase, which is an exo-atmospheric, free-flight motion
Part I: Dynamic Models. IEEE Transactions on Aerospace and Electronic for most BTs, continuing until the atmosphere is
Systems, 39, 4 (Oct. 2003), 1333—1364.
Part II: Originally published as Ballistic Target Models. In Proceedings of reached again. The atmospheric reentry begins when
the 2001 SPIE Conference on Signal and Data processing of Small Targets, the atmospheric drag becomes considerable and
vol. 4473, 559—581.
Part III: Measurement Models. In Proceedings of the 2001 SPIE Conference endures until impact.
on Signal and Data Processing of Small Targets, vol. 4473, 423—446. For the tracking purpose, only the most substantial
Part IV: Decision-Based Methods. In Proceedings of the 2002 SPIE
Conference on Signal and Data Processing of Small Targets, vol. 4728, forces that may act on a BT are considered: thrust,
511—534. aerodynamic forces (most notably, atmospheric
Part V: Multiple-Model Methods. IEEE Transactions on Aerospace and
Electronic Systems, 41, 4 (Oct. 2005), 1255—1321.
drag and possibly lift), and gravity. Not all of these
forces are present at a level that significantly affects
Manuscript received August 4, 2007; revised April 8, 2008; released
for publication August 25, 2008.
the motion of a BT in all regimes of the trajectory.
The boost phase is characterized by a large thrust,
IEEE Log No. T-AES/46/1/935930.
which in the case of rocket staging is subject to
Refereeing of this contribution was handled by W. Koch.
abrupt, jump-wise changes. The effects of the drag
This research was supported in part by ARO Grant
and gravity are also essential in this phase. After the
W911NF-08-1-0409, ONR-DEPSCoR Grant N00014-09-1-1169,
Project 863 through Grant 2006AA01Z126, and Louisiana BoR boost, drag is no longer present and thrust vanishes or
Grant LEQSF (2009-12)-RD-A-25. drops to a very low level. During the exo-atmospheric
Authors’ address: Dept. of Electrical Engineering, University of ballistic phase the motion is governed essentially by
New Orleans, New Orleans, LA 70148, E-mail: (xli@uno.edu). the gravity only. Still, small retargeting maneuvers
are possible. The reentry phase features a rapid
c 2010 IEEE
0018-9251/10/$26.00 ° drag-induced deceleration with possible lateral

96 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
accelerations. So, the significant forces in the different same disclaimers as made in the Introduction of
phases are boost (thrust, gravity, and aerodynamic); [64] apply to this survey, including those on the
coast (gravity); reentry (aerodynamic and gravity). restriction of point targets for temporal behaviors and
In general, the total acceleration vector of a BT, in the interrelationship between dynamic models and
an inertial coordinate system (CS) (e.g., the ECI-CS, tracking algorithms. The reader should keep in mind
see Fig. 2 in the Appendix) can be decomposed as a that results of many extensive studies of BT tracking
vector sum: have not been published in the open literature. In
a = aT + aA + aG = aT + aD + aL + aG (1) other words, much of the BT information, particularly
target-type specific information, is classified and not
where aT , aA , and aG denote the accelerations open to the general public. As a result, this survey
induced by thrust, aerodynamic forces, and gravity, covers only those dynamic models for BT tracking
respectively, and aA consists of accelerations due to in the open literature available to us. The emphasis
drag, aD , and lift, aL . In a frame fixed to the Earth, the of the survey is on missile tracking, although some
relative total acceleration also includes those induced models covered are applicable to tracking other
by the Earth’s rotation. targets.
The entire end-to-end motion of a BT can be This paper is a heavily revised and updated version
modeled by a “wide-band” dynamic model (e.g., of [61]. It is organized as follows: Motion models for
nearly constant velocity, acceleration, jerk, or Singer the simplest phase, the ballistic flight, are covered in
model) capable of covering all the phases. Most Section II. This is followed in Section III by a survey
models developed for the boost phase, the most
of the models for reentry vehicles. Section IV then
sophisticated of all three phases, can serve this
describes models for the most sophisticated, the boost
purpose. This is, however, rather crude and not in
phase. Integration of motion models for different
common use. Target dynamics during the different
phases is the topic of Section V. Concluding remarks
phases are substantially different. A succinct outline
are made in Section VI. The Appendix provides
of the main dynamic phases is given in [28]. What is
background information for the most common CSs
more natural and rational, as well as common practice,
used in BT tracking.
is to develop different models specific for each phase
that more fully exploit the inherent characteristics of
the portion. II. BALLISTIC (COAST) FLIGHT
Throughout this paper, let the target position
and velocity vectors be p = [x, y, z]0 and v = p_ = A. Gravity
[x_ , y_ , z_ ]0 , respectively, in whatever coordinate
¡¡! While gravity is present in the entire flight of
system is considered. For example, p = OS P in a BT, it is the dominating, if not sole, force acting
¡!
the (East-North-Up) ENU-CS and p = OP in the on a coast BT. The following three gravity models,
(Earth-centered-inertial) ECI-CS (see Fig. 2). While denoted by a(0) (1) (2)
G , aG , and aG , are commonly used for
models intended for distinct phases differ significantly BT tracking.
in general, they have a common structure. The 1) Flat Earth Model: The simplest possible
kinematic part, with x = [p0 , v0 ]0 , of the state-space model of gravity assumes a flat, nonrotating Earth.
models of a BT has the form In this model, the gravity acting on the target in the
· ¸ ENU-CS is a constant:
v
_x = : (2)
a a(0)
G = [0, 0, ¡g]
0
(3)
Various models differ from one another in 1) the where g is the constant gravitational acceleration of
choice of models of acceleration a by accounting the Earth.
for different acceleration components at varying 2) Spherical Earth Model: Assume that the Earth
accuracy levels, and 2) the additional models needed and the BTs can be represented as point masses at
for describing the evolution of unknowns in the their centers1 and that the gravitational forces of the
acceleration components. moon (and stars) can be neglected. Since the target
We survey dynamic models proposed/used for has a negligible mass relative to the Earth’s mass,
the distinct regimes of a BT. The motion phases are the gravitational acceleration aG is the solution of a
presented in an order according to their sophistication so-called restricted two-body problem, obtained by
levels, rather than their chronology in the trajectory. Newton’s inverse-square gravity law [7] as
For each phase, we describe physics (forces or
¹ ¹
accelerations) first and then motion models in a(1)
G (½) = ¡ u =¡ ½ (4)
different CS. k½k2 ½ k½k3
Compared with [64], this survey is slightly more
tutorial in nature for the benefit of the readers less 1 Thisholds if the Earth and the targets are spherically symmetric
familiar with the ballistics or aerodynamics. The with an even distribution of their masses.

LI & JILKOV: SURVEY OF MANEUVERING TARGET TRACKING, PART II 97

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
where ½ is the vector from the Earth center to the includes the coordinate accelerations aC = a(1) (2)
C + aC
target, k½k its length, u½ = ½=k½k its unit vector, and induced by the Earth’s rotation [72, 7], which consist
¹ is the Earth’s gravitational constant. of the Coriolis2 a(1) (2)
C and centrifugal aC terms:
3) Ellipsoidal Earth Model: More accurate
expressions for the gravitational acceleration can a(1)
C = 2− £ v = 2M− v, a(2) 2
C = −£(− £ ½) = M− ½
be obtained by replacing the above spherical Earth
(6)
model with an ellipsoidal (or more precisely,
0
spheroidal) Earth model. Such a more precise where − = [−x , −y , −z ] is the Earth’s angular velocity
approximation–accounting for the Earth oblateness vector, v is the target’s linear velocity vector in the
by including the second-order gravitational harmonic noninertial Earth-fixed CS [75, 7] (Fig. 2), and the
term J2 of the Earth’s gravitational field model–is antisymmetric matrix M− is given by
[7, 23] 2 3
( μ ¶ 0 ¡−z −y
(2) ¹ 3 re 2 6 7
aG (½) = ¡ u + J M− = 4 −z 0 ¡−x 5 : (7)
k½k2 ½
2 2 k½k
) ¡−y −x 0
£ [(1 ¡ 5(u0½ uz )2 )u½ + 2(u0½ uz )uz ] For the boost and reentry portions of a BT in a
close vicinity of the Earth and over a short period,
the effect of the Coriolis and centrifugal forces may
(5)
be neglected. During the ballistic flight, however, this
where re is the Earth’s equatorial radius, J2 is a
effect is usually significant and should be accounted
correction constant, and uz is the unit vector along
¡! for, particularly in the case of a long-range BT
OzI (see Fig. 2). The best-known Jeffery constant (relative to the Earth’s radius).
J2 represents the difference between the polar For simplicity of presentation, let
and equatorial moment of inertia. It quantifies the ½ (i)
oblateness of the Earth and is approximately equal (i) aG , ECI
aE = , i = 0, 1, 2 (8)
to one third of the ellipticity of the Earth. Even more gi , noninertial CS
accurate models are available. For example, the
be the total acceleration induced by the Earth, where
specifications of the World Geodetic System 1984
(WGS-84) ellipsoid and the Earth Gravitational Model a(i) (i)
G was given by (3)—(5) and gi = aG ¡ aC is its
(EGM 96) are given in great detail in [77]. expression in an Earth-fixed frame.
The boost and reentry phases are relatively
short in range compared with the Earth radius and C. Coast Motion Models
take place in a close vicinity of the Earth. Thus
a flat, nonrotating Earth model may be adequate, In the exo-atmospheric coast phase, no thrust
particularly in the presence of other more dominating is applied and no drag is experienced. The motion
uncertainties. On the contrary, the coast flight of a may be considered governed by the acceleration aE
long-range BT comprises a much greater range, and induced by the Earth alone, with other factors (e.g.,
thus accounting for the Earth sphericity (and even perturbations) neglected. Thus, the total acceleration is
ellipticity) and rotation is essential. The spherical a = aE . It follows from (2) that the state-space models
model (4) is classical and has been most commonly of a coast target with the state vector x = [p0 , v0 ]0 have
used in a variety of BT tracking applications the form · ¸
v
[90, 51, 55, 85, 29, 99, 56, 5, 9]. It has been used x_ = (i) (9)
for BT tracking over a short range and/or period, for aE
example, as a gravity model as an integral part of
where a(i)
E was given by (8). So, coast modeling of
the total acceleration within a boost (boost-to-coast)
the BT amounts to selection of an appropriate model
motion model [85, 29, 5, 9] or for track initiation
for aE .
purposes [99]. However, it may be inadequate
1) Inverse-Square Model in ECI-CS: In this
for more demanding BT tracking applications, in
particular, precision tracking of coast targets over case, the acceleration part of the state-space model
a long time period or at a low data rate. For a (9) is described by the inverse-square model (4)
long-range coast BT, gravity is the sole or dominating in the ECI-CS
p as a(i) (1) 3
E = aG (p) = ¡¹p=kpk , where
acceleration and thus needs more accurate models kpk = x2 + y2 + z2 .
such as the ellipsoidal Earth or WGS-84 model. A trajectory described by (9) with (4) is confined
to the so-called orbital plane and is a part of a conic
B. Coordinate Accelerations
2 The Coriolis acceleration is induced by the rotation of the Earth
If the target motion is considered in a noninertial that causes the Coriolis effect–the apparent deflection of a body
frame fixed to the Earth (e.g., the ENU-CS, ECF-CS, in motion with respect to the Earth, as seen by an observer on the
and body frames), its relative total acceleration Earth. This effect appears in a noninertial CS [7].

98 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
orbit, governed by the Keplerian motion equation [7]. Similar to (4), the acceleration part a(i) (2)
E = aG (p) of the
A dynamic model is used in target tracking mainly state-space model (9) has the following form [7, 29]:
for propagating the target state (i.e., state prediction),
known as the Kepler problem in astrodynamics, 2Ã Ã μ ¶2 !! 3
ce z
and for covariance prediction. Although the model 6 1 + kpk2 1 ¡ 5 kpk x7
given by (9) with (4) is highly nonlinear, the state 6 7
6 7
propagation with it can be done in an efficient 6Ã Ã μ ¶2 !! 7
¹ 6 6 ce 7
manner, through the algorithm outlined below. (2) z 7
aG (p) = ¡ 3 6 1+ 2
1¡5 y7
Given x(t0 ) = [p00 , v00 ]0 at time t0 the predicted target kpk 6 kpk kpk 7
6 7
state x(t) = [p0 , v0 ]0 at t = t0 + T (with T > 0) is given 6Ã Ã 7
6 μ ¶2 !! 7
by [7] 4 ce z 5
μ ¶ μ ¶ 1+ 2
3¡5 z
¸2 ¸3 kpk kpk
p= 1¡ C p0 + T ¡ p S v0
kp0 k ¹ (10)
p μ ¶
¹¸(³S ¡ 1) ¸2 C where ce = (3=2)J2 re2 .
v= p0 + 1 ¡ v
kp0 k kpk kp0 k 0 This model was chosen in [29], [28] as the coast
model for a 6-state EKF-based ballistic filter in the
where C, S, and ³ are known functions of ¸. The ECI-CS implemented in a multiple-model tracking
variable ¸ is defined through its time derivative by system. However, as pointed out in [29], a model
p
¸_ = ¹=kpk and can be computed accurately and mismatch caused by moderate trajectory perturbations
efficiently via the Newton iteration scheme for solving may lead to a track divergence. This effect can be
a so-called time-of-flight equation, as given below alleviated by introducing small fictitious zero-mean
for the case of an elliptical orbit (i.e., a = 2=kp0 k ¡ process noise. This technique was used in [29] to
kv0 k2 =¹ > 0) [7]: adapt the model for the post-boost phase of the
1. Initialize trajectory, which may involve small maneuvers.
1 ¡ akp0 k p00 v0 In the early work on BT tracking (e.g., satellite
® := p , ¯ := orbit determination [51]), the target dynamics was
¹ ¹
usually considered deterministic (i.e., without process
aT kp k noise). This often led to divergence of the EKF.
¸ := p , ° := p0 :
¹ ¹ Introducing fictitious process noise is an effective
means to account for such factors as model errors,
2. Repeat the following until jT ¡ ¿ j < ²: neglected perturbations, nonlinearities, and computer
p roundoff errors. While the state prediction remains
1 ¡ cos ³
³ := a¸2 , C := unaffected if additive zero-mean process noise input is
³ present, the prediction error covariance is increased
p p
³ ¡ sin ³ by the process noise covariance Q, and thus the
S := p , ¿ := ®¸3 S + ¯¸2 C + °¸ possibility for the EKF to diverge is reduced. The
³ ³
price paid is a possible accuracy degradation of
d¿
:= ®¸2 C + ¯¸(1 ¡ ³S) + ° the filter when the deterministic model is indeed
d¸ adequate. This is closely related to the problem of
· ¸¡1
d¿ noise identification and adaptive filtering. See, e.g.,
¸ := =¸ + (T ¡ ¿ ): [51], [71], [60]. More recent work along these lines

for tracking an orbital target can be found in [44],
A useful program-like pseudocode (with a few where Q was tuned adaptively using the most recent
small typographical errors) of the above algorithm state estimate and covariance in a gravity-gradient
for all conic (i.e., elliptical, parabolic, and hyperbolic) model with the inverse-square law. Although a
trajectories can be found in [99] (see also [89], [19]), number of ways to choose Q have been proposed
along with the computation of the Jacobian F(t, t0 ) = in the literature, in practice Q still remains a design
@x(t)=@x(t0 ) necessary for an extended Kalman filter parameter, which is adjusted/tuned based mostly on
(EKF) [51, 4]. Explicit evaluation of the Jacobian by engineering experience and intuition [23, 12]. More
direct differentiation of (4) can be found in [55]. For a comprehensive discussions of this issue can be found
comprehensive treatment of the theoretical background in [63] and [64].
and solutions to the Kepler problem, the reader is 3) Model with J2 Correction in ENU-CS: It is
referred to [7]. straightforward to obtain the state-space form of the
2) Model with J2 Correction in ECI-CS: A acceleration model in the ENU-CS associated with
refined model for the gravitational acceleration is this gravity model as
based on the ellipsoidal Earth model (5). Note that
up = [x, y, z]0 =kpk and uz = [0, 0, 1]0 in the ECI-CS. aE = g2 = a(2) 2
G (½) ¡ 2M− v ¡ M− ½ (11)

LI & JILKOV: SURVEY OF MANEUVERING TARGET TRACKING, PART II 99

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
where ½ = p + r = [x, y, z + krk]0 , with r being the for application. It is used for state prediction
vector from the Earth center to the origin of the to get x̂(tk+1 j tk ). For the propagation of the
ENU-CS, and krk = re + zS is the sum of the Earth associated error covariance needed in, e.g.,
radius and the altitude of the origin. Kalman filtering, it is proposed in [56] to add, in
Conversion to the radar face CS and the effect, component-uncoupled small noise w(tk ) =
corresponding RUV-CS (see the Appendix) and of its [wx , wy , wz ]0 (tk ) to [ẍˆ , ÿˆ , z̈ˆ ]0 (tk j tk ) in the dynamics
application in the EKF can be found in [23]. A similar equation, where cov(w) = diag(qx , qy , qz ) is a
model is given in [70] with inverse-square gravity, design parameter. As a result, what is proposed
an ellipsoidal Earth, and Coriolis and centrifugal is a coordinate-“uncoupled,” state-dependent,
acceleration terms in the radar face CS, along with piecewise-constant, non-zero-mean white-noise
the respective EKF with RUV measurements. acceleration model, using the terminology of [4],
4) Inverse-Square Model in ENU-CS: With the [64]. There seems room for improvement here by a
spherical Earth’s gravity, the model (11) is simplified better way of adding noise or by adding temporally or
to spatially correlated noise, but at a price of increased
aE = g1 = a(1)
G (½) ¡ 2!©v ¡ ! © ½
2 2
(12) complexity.
Note that the motions along the x, y, z directions
where ! is the Earth rotation rate, and
described by this model are not actually uncoupled
2 3
0 ¡ sin Á cos Á over multiple time steps. The coupling arises from the
6 7 dependence of the acceleration on the common target
© = 4 sin Á 0 0 5 (13)
state in (14). This model has been reported in [56] to
¡ cos Á 0 0 provide performance-competitive with that of the EKF
where Á is the latitude of the origin of the ENU-CS. based on the fully coupled nonlinear inverse-square
This follows from − = ![0, cos Á, sin Á]0 and M− = !© model-for several BT tracking scenarios with a high
in this case (i.e., −x = 0). sampling rate (of 4 Hz). It seems worthwhile to try
Like the models considered before, the standard this model in conjunction with some more precise
EKF technique is directly applicable to this model acceleration models, rather than (12), as originally
in the ENU-CS. The linearization needed for error proposed in [56].
covariance propagation may not be sufficiently The underlying idea of this “acceleration
accurate for relatively long propagation time periods. compensation” [35] approach is not restricted to
For cases with high sampling rates, however, a simple coast models or the specific form of the acceleration
yet accurate piecewise-constant acceleration model model used: The piecewise-constant approximation
[56] could be used, as discussed next. can be directly applied in any situation where the
5) Piecewise-Constant Acceleration Model: The total acceleration is available as a function of the
idea of this approach, proposed in [56], is simple position and velocity. Indeed, the same approach
and natural. At each propagation time step tk ! tk+1 ; was used in [52] to account for gravity in a simple
first sample (i.e., compute) the continuous-time kinematic model of gravity turn motion during
total acceleration process at the estimated point boost (see Section IVB2); similarly, in [35] a
thrust acceleration term derived from a (nearly)
x̂(tk j tk ) = [x̂, xˆ_ , ŷ, yˆ_ , ẑ, zˆ_ ]0 , that is, â(tk j tk ) = a(x̂(tk j tk )). constant-acceleration (CA) filter estimates was
For example, if (12) is used to model the total included in the total acceleration to predict the
acceleration, then boost motion of a BT; an additional drag-induced
â(tk j tk ) = g1 (x̂(tk j tk )): (14) acceleration term was included in [31], along with
the gravity and thrust terms. In general, this approach
As such, under the assumption that the target motion is simple for implementation, but its accuracy
is uncoupled along x, y, and z directions, and the may be inadequate in the case of a low sampling
acceleration a(t) = [ẍ(t), ÿ(t), z̈(t)]0 = â(tk j tk ) is rate.
constant over the time period [tk , tk+1 ), the following 6) Models in Other Coordinates: Models in
discrete-time dynamics model along the x direction is other CS (e.g., radar face CS and spherical CS) can
clearly obtained be obtained by conversion from the ENU-CS. For
· ¸ · ¸· ¸ example, an explicit expression of the inverse-square
x(tk+1 ) 1 tk+1 ¡ tk x(tk )
= model (neglecting the Coriolis and centrifugal forces)
x_ (tk+1 ) 0 1 x_ (tk ) can be found in [15].
· ¸
(tk+1 ¡ tk )2 =2 ˆ
+ ẍ(tk j tk ):
tk+1 ¡ tk III. REENTRY

Likewise for y and z directions. During reentry, two significant forces are
This state-dependent piecewise-constant always present: the gravity and atmospheric drag. If
acceleration model is simple and straightforward maneuvers are possible, aerodynamic lift must also

100 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
be considered, although lift may be significant also force of the atmosphere acting on the target has in
for some endo-atmospheric nonmaneuvering targets general two components: the lift, which is normal to
such as artillery projectiles. A reentry vehicle (RV) the velocity, and the drag, which is opposite to the
is referred to as a maneuvering RV if it involves velocity. The total drag is the sum of zero-lift drag
maneuver but a ballistic RV if it does not. aD and lift-induced drag a(r)D (also known as residual
Formally the total acceleration of an RV is drag), where the zero-lift drag is present even in the
given by (1) with a zero thrust (aT = 0). Most RV absence of lift (e.g., for zero angle of attack). Like
motion models can be viewed as a coast model plus the zero-lift drag, both lift aL and its induced drag a(r)
D
dominating terms to account for drag and possibly have a magnitude proportional to (1=2)´(h)kvk2 . This
lift. In fact, during reentry aerodynamic forces are factor quantifies the free-stream dynamic pressure [72]
usually much greater than gravity: drag in excess of and is indispensable in the RV dynamic models.
20 g can be present over most of the RV trajectory The complete aerodynamic acceleration is the
below 30 km, and possibly as high as 100 g, and lift vector sum of drag and lift:
and other transverse loads can easily exceed 10 g. On aA = aD + a(r)
D + aL : (16)
the other hand, the effect of the Coriolis, centrifugal
forces, and the Earth oblateness on the target motion Since the aerodynamic accelerations are w.r.t. the
is much smaller than during the coast flight. relative velocity, it is more convenient to use some
We first describe the models for drag and lift, and Earth-fixed noninertial CS, instead of the ECI-CS.
then present motion models for an RV. Note that [81] As such, unless otherwise stated, all RV models and
is a useful reference but is largely overlooked by the our descriptions are w.r.t. the ENU-CS (in particular,
tracking community so far. v = vr ).
3) Air Density: The air density ´ is usually
A. Aerodynamic Forces approximated as an exponential function [76, 81, 11]
or more accurately as a locally exponential function4
1) Drag: Drag depends on target speed, altitude, of the target altitude [70, 88, 102]:
shape, size, angle of attack, etc., as well as flow
parameters of a particular flow regime. It is a well ´(h) = ´0 e¡·h (17)
studied, yet complex, topic [27, 93, 37]. For most BT where ´0 and · are known constants. See, e.g., [102]
tracking problems, however, there is no need to be for the parameters of the local exponential functions.
involved with so many detailed factors. More precise but nonanalytic models of the air density
The drag is opposite to the target velocity v are available [81].
relative to the atmosphere in a noninertial frame
(e.g., the ENU-CS or ECF-CS),3 with a magnitude B. Ballistic (Nonmaneuvering) RV Motion
proportional to the air density ´ and the square of
the target speed kvk. Specifically, the drag-induced In simple terms, a nonmaneuvering RV travels
acceleration is given by (see, e.g., [72]) along a “purely” ballistic trajectory in the atmosphere,
hence the name ballistic RV (BRV). Such a motion
1 1
aD = ¡ ®´(h)kvk2 uv , uv = v (15) does not involve lift or lift-induced drag.
2 kvk BRV motion models can be classified into two
where h is the target altitude and the drag parameter groups: with and without knowledge of the drag
® = ScD =m depends on the target mass m, reference parameter ®.
area S (i.e., the target body cross-sectional area 1) Kinematic Models (Known Drag Parameter): If
perpendicular to the velocity), and the so-called drag ® or ¯ is (assumed) known, modeling of a BRV is
coefficient cD . The coefficient cD generally depends straightforward: use a coast model from Section II
on S and v, but is sometimes assumed constant [9]. plus an additional drag model (15). The state
The inverse of the drag parameter, ¯ = 1=®, is known vectors in such a kinematic model involve only the
as the ballistic coefficient (BC). Consequently, for the kinematic variables of the RV, hence the name. More
tracking purpose, all uncertainties associated with specifically, the state-space model is given by (2) with
the drag (other than the air density and velocity) are the acceleration part
generally aggregated in (the dynamic models of) the a = gi ¡ 21 ®´0 e¡·h kvkv (18)
drag parameter ® or the ballistic coefficient ¯ within q
the simple drag formulation (15). 2 2 2
where kvk = x_ + y_ + z_ is the speed, and gi for
2) Lift: When a BT is moving in the air (e.g.,
i = 0, 1, 2 are defined in the coast models for flat,
with a non-zero angle of attack, defined as the angle
spherical, and ellipsoidal Earth models, respectively.
between the thrust and velocity vectors), the reactive
4 “Local” in the sense that the function is piecewise exponential,
3 Thisvelocity is denoted by vr in the ECI-CS and whenever it may with parameters depending on the target height; that is, each piece
be confused with an absolute velocity. corresponds to some layer of the atmosphere.

LI & JILKOV: SURVEY OF MANEUVERING TARGET TRACKING, PART II 101

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
For the flat and spherical models in (18) the target to augment the kinematic state vector of an RV by
height h = z; for the ellipsoidal model, h can be some time-varying parameters used to model ® [3].
approximated by the distance between the target The dynamic equation of these parameters, together
position and the intersection point of the vector with a kinematic model as given above, thus forms a
½ with the reference ellipsoid (Fig. 2) [12]. For dynamic model of the RV.
low-altitude targets, most often a constant gravity Other than using drag tables [12], probably the
gi = a(0)
G is assumed [17, 88, 11, 38, 33, 32, 104], simplest model is this state augmentation in, e.g.,
which greatly simplifies (18). More details of the the Cartesian coordinates as xa = [x, y, z, x_ , y_ , z_ , ®]0
ellipsoidal Earth-based model (18) can be found in with an assumed unknown but (nearly) constant ®(t);
[23], along with the calculation of the Jacobian and that is, model ®(t) as a Wiener (or more generally
the successive transformations to the radar face and independent increment) process: ®_ = w® (t), where
RUV-CS. For an analysis of this model in the radar w® (t) is zero-mean white noise whose variance is
face CS, the reader is referred to [96, 70]. Usually a design parameter. Then a complete RV dynamic
model noise is also introduced in (18). model is obtained by concatenating this equation with
A motion model in the Cartesian CS is most some RV kinematic model, e.g., (18). This simple
natural. In the RV tracking, however, actual drag model is probably the most common for BRV
measurements of a dish or phased array radar are tracking. It was used in combination with various
typically expressed in some other CS, such as the RV kinematic models in [96, 70, 17, 2, 12, 74]. It is
spherical or RUV-CS. These measurements have a possible to use ¯, instead of ®, for state augmentation
highly nonlinear relationship with the target state in [30, 84, 42, 83]; however, it was argued in [2], [12]
the Cartesian CS. Often (e.g., EKF-based) nonlinear that ®, which leads to a linear problem, is preferable
filtering techniques [70, 23] are employed to handle to ¯.
both dynamics and measurement nonlinearities. (Of In reality, ¯ is not constant, depending on several
course, other nonlinear filtering techniques such as factors. Assuming an exponential air-density model
those surveyed in [65] can also be used.) This can (17) with a fixed set of ´0 and ·, the ballistic
be done in a mixed coordinate system–the state coefficient ¯ is approximated in [88] as a linear
vector in the Cartesian CS and measurements in the function of the height z: ¯ = ¯0 + ¯1 z, where ¯0
sensor CS. Alternatively, either the measurements and ¯1 are unknown parameters, incorporated into
can be converted to the Cartesian CS or the state the state vector to be estimated. This linear model,
can be converted to the sensor CS. As remarked however, turns out to be inadequate over some altitude
in [34], measurement conversion, along with
ranges. To overcome this problem, a new variable ° =
proper initialization and prediction, yields tracking
´=¯ = ®´ with a dynamic equation °_ = ¡·°(t)z_ (t) was
performance comparable to state conversion. The
introduced in [88] based on the locally exponential
study reported in [70] concluded that the measurement
model of ´ (17). Along with a = gi ¡ 21 °kvkv, they
nonlinearities play a significant role. This is true
form a dynamic BRV model. In this model, unknown
especially for not-so-accurate sensors. Along these
constant ·, specified over several altitude ranges, is
lines, it was argued in [88] that modeling the RV
a key parameter. Its uncertainty was handled in [88]
motion in the sensor CS rather than the Cartesian
by means of an interval Kalman filtering technique
CS could be advantageous. A BRV motion model
developed in [18], assuming it is in an interval of a
assuming an ellipsoidal Earth was used in the
length 2¢· center at a nominal value ·0 .
spherical CS of a dish radar in [12]. While the
Another idea by which to model the drag
motion model is too lengthy to be included here, the
parameter is to consider its deviation from a known
measurement model is the simplest possible. It allows
one to circumvent such critical issues as linearization nominal value ®¯ in terms of the ratio ®=®¯ [23].
Assuming its logarithm ± = ln(®=®) ¯ is constant (i.e.,
of the measurement model and CS conversion. Based
_®=® = 0 with ® = ®e ¯ ± ), an RV kinematic model
on this, a practical system for tracking incoming
can be augmented with the model ±(t) _ = 0 in an
ballistic targets was implemented in [12]. For a
complete coverage of this topic, the reader is referred obvious manner. In [23] this is done for (18) with
to Part III [62]. the ellipsoidal Earth model. In our opinion, it seems
2) Dynamic Models (Unknown Drag Parameter): more effective to assume ±(t) _ = w (t) with white
±
The assumption of a known ® or ¯ greatly simplifies noise w± (t) or model ±(t) as a first-order Markov
the modeling of the RV motion. However, it process. Note that the use of a fixed nominal value
is not necessarily valid in practice–® may be ®¯ makes the model practicable only for a known
unknown, often time varying or too complex to class of targets with ® corresponding to the chosen
model adequately, and needs to be estimated, ¯ The uncertainty associated with the drag is a rather
®.
especially in the case of an unknown target type. The complex issue, analyzed in great detail in [81]. This
prevailing approach to handling this uncertainty is uncertainty was coped with in [103] recently by using

102 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
Conversion to the ENU-CS yields6
ENU VTC
aA (®) = 21 ´kvk2 TVTC ® : (20)
The VTC-to-ENU coordinate transformation matrix
ENU
TVTC can be given in two alternative forms in terms of
v and (μ, Á), respectively:
2_ 3
x=kvk ¡y_ =kvg k ¡x_ z_ =(kvk kvg k)
ENU 6 7
TVTC = 4 y_ =kvk x_ =kvg k ¡y_ z_ =(kvk kvg k) 5
z_ =kvk 0 kvg k2 =(kvk kvg k)
(21)
Fig. 1. Maneuvering RV model. 2 3
cos μ cos Á ¡ cos Á sin μ cos Á
6 7
= 4 cos μ sin Á sin Á sin μ sin Á 5 (22)
a multiple-model approach to quantize an assumed ¡ sin μ 0 cos μ
range of the ballistic coefficient.
Additional references addressing modeling/ where Á = tan¡1 (y_ =x_ ) and μ = tan¡1 (¡z_ =kvg k) are
estimation issues include [20], [40], [1], [46]. the velocity “heading” and “elevation” angles,
respectively.
This model allows a simple physical interpretation:
C. Maneuvering RV Motion
®t > 0 (or ®t < 0) corresponds to a left (or right)
While the above BRV models take into account ground turn and ®c > 0 (or ®c < 0) to climbing (or
only the zero-lift drag, a maneuvering RV (MaRV) diving). Such a clear interpretation could facilitate
model also accounts for possible lift and residual drag the adjustment of the model parameters and RV
tracking by multiple models, among other things.
in an appropriate coordinate system.
An MaRV model so specified involves in general
1) Models of Aerodynamic Forces:
four parameters: ®, ¸0 , ®t , and ®c . As pointed out
a) Models in velocity-turn-climb frame: The
in [17], not all of them are observable since (19)
lift may be decomposed into two components for depends on only three scalar parameters: ®d , ®t , and
convenience (see Fig. 1) [17, 16]: turn (along unit ®c . Loosely speaking, we can consider ®VTC as a
vector ut ) and climb (along unit vector uc ). The turn vector parametrizing MaRV maneuvers, although ®d
direction ut is horizontal5 and perpendicular to the is actually dominated by the nonmaneuvering zero-lift
velocity vector (i.e., ut kOxy and ut ? uv ), and the drag.
climb direction uc is perpendicular to the turn and A similar model was presented in [81] based
velocity vectors (i.e., uc ? ut and uc ? uv ). Formally, on a zero-lift nominal trajectory, which has an
ut = uz £ uv and uc = uv £ ut , where uz = [0, 0, 1]0 is 8-dimensional state vector involving the lift parameter
the unit vertical vector. This triplet of unit vectors ®L = ¡[®t , ®c ]0 as well as the kinematic variables. It is
uv , ut , uc defines an orthogonal body frame, centered in a similar coordinate system, which differs from the
at the target position, which can be referred to as VTC frame only in the axial orientation.
the “velocity-turn-climb” (VTC) frame. Analogical b) Models in LOS-CS: Alternatively, the
to (15), it follows from (16) that the complete residual drag a(r)
D and lift aL in (16) can be expressed
aerodynamic acceleration aA is expressed as [23] in a line-of-sight (LOS) coordinate system,7
defined in the ENU-CS by the unit vectors u» = p=kpk
aA (®) = 21 ´kvk2 (¡®d uv + ®t ut + ®c uc ) (along the position vector), u´ = h £ p=kh £ pk (along
the angular momentum vector), and u³ = u» £ u´
= 12 ´kvk2 [uv , ut , uc ]®VTC (19) (right-hand rule), where h = ½ £ v (see Fig. 2). Denote
the vector sum of the lift and the residual drag in
where ®VTC = [¡®d , ®t , ®c ]0 , ®d = (1 + ¸0 )®, and ®t
this LOS-CS as ®» u» + ®´ u´ + ®³ u³ . It is thus clear
and ®c quantify the lift along the turn and climb
directions, respectively. The ratio ¸0 of the lift to the that ®LOS = [®» , ®´ , ®³ ]0 quantifies the maneuvers of
an MaRV in the LOS-CS. After conversion to the
critical lift (i.e., the lift at the maximum lift-to-drag
ratio) characterizes the residual drag parameter 6 That is, evaluate T = [u , u , u ] in the ENU-CS: u = [x_ , y_ , z_ ]0 =kvk,
in terms of the zero-lift drag parameter ® as ¸0 ®. v t c v
ut = [0, 0, 1]0 £ uv = [¡y_ , x_ , 0]0 =kvg k, uc = uv £ ut = [¡x_ z_ , ¡y_ z_ , x_ 2 +
p
x_ 2 + y_ 2 are the target speed
y_ 2 ]0 =(kvk kvg k), where kvk and kvg k =
5 Although a flat Earth model is assumed in [17], ut can be thought and ground speed, respectively.
of as tangential to the reference ellipsoid in an ellipsoidal Earth 7 Not to be confused with the LOS aligned with the sensor

model, i.e., ut kOxy in the ENU-CS. measurement axes [9].

LI & JILKOV: SURVEY OF MANEUVERING TARGET TRACKING, PART II 103

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
ENU-CS, the total aerodynamic acceleration is [23] governing parameters ®, such as ®VTC = [¡®d , ®t , ®c ]0
μ ¶ if (20) is used or ®LOS = [®» , ®´ , ®³ ]0 if (23) is used.
1 2¯ 1 ENU LOS
aA (®) = ´kvk ® ¡ v + TLOS ® (23) These parameters are unknown and thus need to
2 kvk
be estimated. Like the dynamic BRV models, the
where ®¯ is a nominal value of the zero-lift state augmentation is commonly used: augment
drag parameter, discussed in Section IIIB, and the kinematic state vector x = [x, y, z, x_ , y_ , z_ ]0 by the
ENU
TLOS = [u» , u´ , u³ ] is the LOS-to-ENU coordinate unknown ®. Thus a dynamic model for ® (®VTC ,
transformation matrix, determined by evaluating ®LOS , or ®RVC ) is needed. Various models are possible
u» , u´ , u³ in the ENU-CS. here, although only a few have been proposed.
c) Discussion: Although the above two a) Wiener process models: The simplest
models are conceptually similar, there are two main possible model for ® is to assume it is nearly
differences. First, the model in the VTC frame lump constant–model ® as a Wiener process [17]: ®_ =
the drag and lift-induced drag, while the LOS-CS w® (t), where w® (t) is zero-mean white noise. As such,
model uses a nominal value ®¯ for the zero-lift drag the state-space MaRV model is given by (2) with
parameter. They are theoretically equivalent for the
pure lift part. For the drag part, the former amounts to a = aA (®) + aE(i)
estimating the total drag and the latter to estimating ®_ = w® (t)
its variation relative to the nominal value. The
latter should be better if this nominal ®¯ is known where models for aA (®) are given in Section IIIC1.
(approximately), as in the case of a known RV type. Such a model with ® = ®VTC = [¡®d , 0, 0]0 and a
Sometimes ® is assumed constant if it is deemed constant gravity (i.e., a(i) (0)
E = aG ) was proposed and
not to vary much over a wide range of altitude investigated in [17]. For better accuracy, a spherical
[81]. The other main difference lies in the choice or ellipsoidal Earth model a(i) (i)
E = g1 of (12) or aE = g2
of CS to parameterize the lift and residual drag. In of (11) could be used. These models expressed in the
general, the use of the observer-dependent LOS-CS ENU-CS are somewhat more complicated, due to the
for parametrization of target maneuver is more Coriolis and centrifugal terms, than in the ECI-CS,
convenient from the observer (sensor) point of view, which, however, requires an ENU-to-ECI coordinate
ECI
but it gives rise to a dependence of the maneuver transformation TENU for the aerodynamic accelerations.
parameters on the specific observer-target geometry, Such a model with ® = ®VTC = [¡®d , ®t , ®c ]0 was
with some undesirable consequences, including an developed in [29], [28], where a(i) (1)
E = aG and aA (®) =
increased complexity of adjusting design parameters. 1 2 ECI ENU VTC
¡ 2 ´kvr k TENU TVTC ® [75]. Obviously, the more
In contrast, the VTC representation depends only (2)
accurate model aG of (10) could be used for the
on the target-Earth geometry and is invariant of the gravity part instead.
sensor location. Furthermore, in the VTC frame the For further information of these models, such as
lift (responsible for maneuvers) always lies in its y-z adjustment or adaptation of the design parameters
plane, while the nonmaneuvering force (i.e., the drag) and performance results, the reader is referred to [17],
is always along the x-axis. Such a decomposition [16], [29], [28], [53].
makes the tracking task relatively easier. As far as b) Markov models: Alternatively, ® can be
error coupling and ill conditioning are concerned, modeled as a slightly more sophisticated, first-order
while the VTC frame is good in terms of process error Markov process (known as the Singer model for
decoupling, LOS-CS is better regarding measurement aircraft maneuver): ®_ = ¡K®(t) + w® (t), where w® (t)
error decoupling [26]. In this regard, it appears that is zero-mean white noise and K is a design parameter.
the range-velocity Cartesian (RVC) coordinates [26] If vector ® is associated with a single physical
should be considered. Consequently, if a reasonably acceleration, K should be a scalar, otherwise it can
good nominal value ®¯ is available, we think the be a (diagonal) matrix. As such, the state-space MaRV
following model deserves to be considered model is given by (2) with
μ ¶
1 2¯ 1 ENU RVC a = aA (®) + a(i)
aA (®) = ´kvk ® ¡ v + TRVC ® : E
2 kvk
®_ = ¡K®(t) + w® (t):
2) Motion Models: Given models of the
aerodynamic accelerations aA in the ENU-CS, Such a model was proposed in the ENU-CS [23]
development of a motion model for an MaRV is without any driving noises for ® = ®LOS , K =
straightforward: choose a coast model depending on diag(·» , ·´ , ·³ ) (with ·´ = ·³ ), a(i)
E = g2 , and aA (®) of
the underlying assumptions concerning the gravity, (23). This is the most comprehensive dynamic model
and then incorporate into it a model for aA , such for MaRV tracking in the literature known to us. It
as (20) or (23) in the ENU-CS. This can also be utilizes the truly “dynamic” model in conjunction
done in the ECI-CS, where aA must be transformed with the accurate ellipsoidal Earth gravity model.
to the ECI-CS. Such a model involves maneuver Its (essentially) equivalent variants in other CS

104 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
can be developed in an obvious manner. Similarly, As far as the maneuver-related accelerations
the Singer model was adopted in [81] for the lift are concerned, this unknown-input model can be
parameter ®L = ¡[®t , ®c ]0 in an aerodynamic frame loosely thought of as “nonparametric” (model-free)
similar to the above VTC frame. Note that, albeit while the other models above are “parametric” (i.e.,
seemingly the same, modeling ®VTC with (20) [17] model-based) in some sense. Their relative pros
and modeling ®LOS with (23) [23] as a white or and cons are parallel to those of parametric and
Markov process are practically different (see the nonparametric models in general–the parametric ones
discussion in Section IIIC1c). Consequently, as our are more preferable if reasonably good knowledge
general description here alluded to, it is meaningful to of the parameters involved is available, otherwise the
develop and investigate a dynamic model using ®VTC nonparametric models should be considered.
or ®RVC , parallel to the above model with ®LOS . Some other dynamic models similar to those for
The above dynamic models of an MaRV rely maneuvering aircraft tracking [64] have also been
on the following idea: Quantify the unknown developed/used for ballistic targets, such as those for
aerodynamic accelerations through a parameter scaled weaving (spiraling) motions in [13], [101], [102].
by a kinematic state-dependent factor 12 ´kvk2 , model
the parameter as a random process, and augment the D. Motion of Endo-Atmospheric Ballistic Targets
kinematic state vector by this parameter.
c) Unknown input models: Instead of modeling In contrast to the ballistic missiles that have a
the uncertain lift and residual drag explicitly as above, purely exo-atmospheric coast flight (Section II),
their effects on the RV motion can be lumped and some small artillery projectiles, such as mortars and
represented by an unknown input u = [ux , uy , uz ]0 in howitzers, have a ballistic motion that is entirely
the kinematic model of the MaRV: a = aD + gi + u. within the atmosphere. Their motion is governed
Specifically, with a flat Earth model, this model takes primarily by gravity, drag, and possibly lift (similar
the following form in the ENU-CS [57] to an MaRV motion), even though they do not
maneuver. However, due to the intrinsic properties
a = ¡ 12 ®´kvk2 [cos μ cos Á, cos μ sin Á, ¡ sin μ]0 + g0 + u
(e.g., shape, stabilization mode, or aerodynamics) of
(24) such a projectile, there are also other forces producing
which is a simple variation of (18) (with the dynamics such as drift and Magnus moment [69] that
addition of u). Here the zero-lift drag parameter may be significant and should be accounted for. The
® is assumed known, the drag acceleration is in a mode of stabilization is critical in this regard [79].
slightly different but clearly equivalent form in view Fin-stabilized projectiles (e.g., mortars) achieve
of the first columns of (21)—(22), and Á and μ are the stability through aerodynamic forces caused by fins
velocity “heading” and (negative) “elevation” angles, located at the aft end of the projectile. The dominating
respectively, as defined before. forces are gravity and drag. All other forces can
This model uses the additional input u to cover the be neglected for some fin-stabilized projectiles
(lift and its induced drag) accelerations responsible for [79]. In such a case, the kinematic model is (2)
maneuver. It is idealistic because the zero-lift drag with the acceleration part given by (18), the same
parameter is not known in practice. This unknown as for a BRV. This model with a constant gravity
parameter is better replaced in practice by a (known) was used in [41], [21] and [80] for initial trajectory
nominal value ®, ¯ although with this replacement any estimation based on radar measurements, followed
deviation from ®¯ will give rise to an additional term by impact point prediction (IPP). To model the
in the input and thus may trigger a state estimate target dynamics for mortars with subsonic speeds,
correction in the input-estimation method as applied the drag parameter ® is assumed nearly constant in
in [57]. In a similar manner an unknown input u was [80] with a hypothetical initial value in a predefined
also used in [39] to account for the thrust acceleration set. By running multiple autonomous filters for
during boost. different initial ® values, the most probable one (target
The unknown input u as well as the kinematic class) is identified and then used for IPP. The drag
state variables is estimated in [57] by the well-known parameter is highly dependent on the Mach number.
input-estimation method for adaptive filtering A “nearly constant” dynamic model of a variant of
[14, 6, 63]. Albeit simple, the effectiveness, if not the ballistic coefficient was used in [21] for trajectory
the applicability, of the input-estimation method here estimation, and the IPP was done using templates
is more or less questionable due to its fundamental of ® as a function of the Mach number. It was also
assumption that the unknown input is constant over established in [79] that for some types of mortars
the period of interest. In fact, u here is inherently time the pitch-over motion can be delayed significantly,
varying and state dependent, as can be seen from, leading to a significant lift. This lift can be modeled
e.g., (23). Nevertheless, this is a simple model to as in Section IIIC, but it might be an overkill since
which many adaptive filtering techniques effective for this lift is not as significant as in the case of an
uncertain (time-varying) input can be applied. MaRV. The lift was assumed in [79] to affect only the

LI & JILKOV: SURVEY OF MANEUVERING TARGET TRACKING, PART II 105

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
target altitude, leading to a simple form aL = [0, 0, l]0 aN = aT + aD , rather than its components separately.
with a nearly constant l. More specifically, the total Here a key issue is how to model the net acceleration
acceleration is a = g0 ¡ 12 ®´0 e¡·h kvkv + [0, 0, l]0 with aN .
_ modeled as white noise.
l(t)
Spin-stabilized projectiles (e.g., howitzers)
B. Kinematic Models
achieve stability through the gyroscopic effect due
to the spinning motion of the projectile imparted 1) Stochastic Models:
at launch. The gyroscopic effect, however, causes a) Constant net-acceleration: The
a lateral motion which is significant for IPP [69]. nongravitational net acceleration is assumed constant
Consequently, a model for spin-stabilized projectiles (a_ N = 0) [85] or nearly constant (a_ N = wN ) [29] in this
should include an additional term of drift acceleration model. The corresponding state-space representation,
a¢ . In [41], [21], a¢ was modeled as nearly constant with the 9-state vector x = [p0 , v0 , a0N ]0 , is given by
[a_ ¢ = w¢ (t)] and included in the state vector to p_ = v and
be estimated. In [79] a¢ was derived based on v_ = aN + aG
the gyroscopic effect as an explicit function of (26)
a_ N = 0 or a_ N = wN
the so-called spin coupling parameter s, resulting
in the model for the total acceleration: a = g0 ¡ where wN is small zero-mean white noise with a
1 ¡·h
2 ®´0 e kvkv + [0, 0, l]0 + a¢ (s). Assuming s as nearly covariance for design. In the ECI-CS, aN =
constant [i.e., s_ = ws (t)] then leads to a dynamic model [aN , aN , aN ]0 if aN is expressed in the Cartesian
x y z
with only one extra state variable s to be estimated. coordinates or aN = kaN k[sin μN cos ÁN , sin μN sin ÁN ,
sin μN ]0 if aN is expressed in the spherical coordinates
IV. BOOST PHASE (kaN k, ÁN , μN ) [85]. Clearly, other forms of a state
vector can be considered, such as in spherical
During boost, the target is subject to a large thrust coordinates with position (altitude, latitude, longitude),
as well as drag (and lift) and gravity. The thrust velocity (radial, northward, eastward), and thrust
acceleration profiles (i.e., its magnitude and direction vector (magnitude, in-plane, out-of-plane directions
as functions of time) vary largely from one type of thrust direction) [73], or position in spherical
of target to another: While some targets maintain coordinates and speed, heading, and flight path angle
an essentially constant thrust acceleration, others, with respect to the local horizontal [54, 95, 51].
multistage targets in particular, have an abruptly The “nearly constant” version (a_ N = wN ) is suitable
(jump-wise) changing thrust. Moreover, some more to cover small maneuvers for motions with a low
advanced targets can also have a considerably variable level of thrust, and indeed was designed [29, 28]
thrust during later stages [86, 9, 28]. for a post-boost phase of motion, which is primarily
exo-atmospheric with nongravitational net acceleration
of the order of 1 m=s2 .
A. Accelerations
Although conceptually the same as the CA model
The thrust acceleration is determined by the for aircraft motion [64], which assumes a 3D straight
thrust magnitude T, its direction unit vector uT , and line motion with a constant acceleration, this model
target mass m as aT = (T=m)uT [7]. The associated is fairly nonlinear due to the gravitational term aG
uncertainty includes its magnitude ®T = kaT k = T=m unless the constant gravity model a(0) G is used. Further,
and direction uT . this model allows in general a target motion with a
The total target acceleration during boost is non-zero angle of attack, in contrast to the widely
used gravity turn (GT) models, discussed later. As
T D
a = aT + aD + aG = u ¡ u + a(i) (25) such, this model may be more appropriate than the
m T m v G
GT models for tracking maneuvering (e.g., staging)
where uv = vr =kvr k is the unit vector of the target BTs for which the GT assumption is invalid.
velocity vr relative to the atmosphere. Note that the As explained in [64], a CA model is in fact a
thrust and drag are usually much larger than the second-order polynomial (i.e., quadratic) model,
gravity during the boost phase. where the motion is usually assumed uncoupled
It would be desirable to augment the kinematic along the three dimensions [92]. As reported in [86],
state vector (position and velocity) with the thrust however, such a quadratic model falls far short of the
vector (T and uT ), drag magnitude D = 21 ®´kvr k2 required accuracy in modeling even the first stage of
(or the drag parameter ® provided the air density ´ a long-range target. Similar and simple models–a
is known), and mass m so that they are estimated CA model with a constant gravity in a multiple-model
separately. However, such a state vector would be framework and second/third order polynomials–have
unobservable [9]. A more feasible approach is to been applied in [49], [50], [48] for trajectory tracking
handle directly the net (thrust minus drag) acceleration and backfitting for launch point estimation. A CA

106 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
model was also used in [87] for the boost phase in a form:
comparison of different filters for BT tracking. 2 3 2 32 3
x(t) a0 a1 ¸a1 1
b) Correlated acceleration: The above model 6 7 6 76 7
assumes that the net acceleration is constant or has 4 y(t) 5 = 4 b0 b1 ¸b1 5 4 t ¡ t0 5 (29)
an independent increment. In reality, however, the net 2
z(t) c0 c1 ¸c1 (t ¡ t0 ) =2
acceleration is dominated by the thrust, which has a
where the model parameters to be estimated (i.e., ai ,
strong temporal correlation. In such a case, although
bi , ci , i = 0, 1, and ¸) are assumed time invariant. Note
the above nearly constant model can still be used
that a0 , a1 , and ¸a1 are actually position, velocity, and
with larger noise at a cost of significantly reduced
acceleration, respectively, along the x-axis at time t0
accuracy, it appears more appropriate to model the
(likewise for the y and z directions). Many parameter
nongravitational net acceleration as a first-order
estimation techniques, such as the least-squares
Markov process [29]:
(LS) method [86], can be used to estimate these
a_ N = ¡(1=¿ )aN + wN (27) parameters. For unknown ¸, however, this is a
nonlinear estimation problem, a simplified special case
where wN is zero-mean white noise. The covariance
which nevertheless has an exact LS solution as given
of wN and the correlation-time constant ¿ are design
in [86].
parameters.
Although generally more appealing than the
Alternatively, this model can be used directly
quadratic model of Section IVB1a for the GT
for the total rather than net acceleration a: a_ =
motions, this model appears to lack the accuracy
¡(1=¿ )a + w. This amounts to a Singer model for
required for more demanding tracking applications.
the gravity as well as the net acceleration, and the
As pointed out in [86], while performing well for
resulting model with the 9-state vector x = [p0 , v0 , a0 ]0
constant thrust accelerations (typically valid only over
is linear. It is less accurate than the above Singer
a short time period), it is not generally adequate in the
model for the net acceleration due to the accuracy
case of rocket staging. For such cases, a generalization
difference in the gravity models involved. However,
was suggested in [86]: augment the state vector with
the gravitational acceleration aG is less significant
two additional angles (yaw and pitch) representing,
than the net acceleration aN in the boost phase. Both
respectively, the horizontal and vertical orientations of
of these correlated acceleration models have been
the acceleration vector relative to the velocity vector.
reported [28] to demonstrate quite good performance
This generalized model is in essence the quadratic
for tracking boosting targets with real data.
model of Section IVB1a with aN expressed in the
The models discussed here are fairly general–no
spherical coordinates (kaN k, ÁN , μN ) [85] subject to
information specific to the boost phase is utilized.
the GT constraint aN = ¸v.
Therefore, they also apply to the other two phases, but
The above GT model (29) is given in a time-series
they cannot be expected to perform as well as good
polynomial form. Its state-space form is [52]
models specific to a flight phase.
· ¸
2) Gravity Turn Models: During the boost 0 1
(post-boost) phase, a BT typically maintains a small x_ = diag[A(¸), A(¸), A(¸)]x with A(¸) =
0 ¸
angle of attack and thus follows approximately a
trajectory known as a gravity turn (GT), which (30)
is characterized by thrust being parallel to the where x = [x, x_ , y, y_ , z, z_ ]0 . Its usual discrete-time
Earth-relative velocity vector vr [72]. Making use of equivalent is
this motion pattern can improve modeling (and hence
xk+1 = diag[F(¸k ), F(¸k ), F(¸k )]xk with
tracking) accuracy considerably if the pattern is indeed 2 3
followed. 1 ¸T
1 (e k ¡ 1) (31)
A kinematic booster model that utilizes the GT F(¸k ) = 4 ¸k 5
assumption was proposed in [86]. It is a quadratic 0 e ¸k T
model (see [64]) with a modification such that
the acceleration and velocity vectors are parallel. where ¸(t) is assumed piecewise constant and equal to
Specifically, in the ECF-CS the state vector includes ¸k over any sampling interval [kT, (k + 1)T).
the kinematic components [x, x_ , y, y_ , z, z_ ]0 and possibly This model has only one parameter, the
the ratio of the net acceleration magnitude to the acceleration-to-speed ratio (ASR) ¸k , and is linear
target relative speed, ¸ = kak=kvk. As far as only if ¸k is (assumed) known. In such a case Kalman
nongravitational net accelerations are concerned, the filtering can be applied directly and therefore the
GT constraint reads as a = ¸v or state-space form (31) is preferable to the polynomial
form (29) for tracking purposes. In reality, however,
ẍ = ¸x_ , ÿ = ¸y_ , z̈ = ¸z_ : (28)
¸k is not known and has to be determined (e.g.,
The (nearly) constant net-acceleration model subject modeled a priori, estimated from data, or both).
to this GT constraint takes the following time-series Moreover, the assumption that ¸ is constant over a

LI & JILKOV: SURVEY OF MANEUVERING TARGET TRACKING, PART II 107

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
sampling period of length T is not valid if T is not GT motion is
small enough. a = ®vr =kvr k + aG : (32)
The GT kinematic model (31) or (30) has much
Since the drag is in the opposite direction of the
in common conceptually with the constant-turn (CT)
velocity,
model [64] widely used for tracking maneuvering T¡D
aircraft. Both motions are defined by kinematic ® = kaN k = (33)
m
constraints: the acceleration and velocity vectors are
where T, D, and m are the thrust, drag, and target
parallel (akv) in a GT but orthogonal (a ? v) in a CT.
mass, respectively.
They are two possible extreme cases. The motions
It is often assumed that during boost the thrust
along the three dimensions are coupled only through
the ASR ¸ = kak=kvk in a GT and only through the magnitude T is constant (T_ = 0) and the target mass
m is linearly decreasing at a constant rate ± > 0 (i.e.,
turn rate ! = kak=kvk in a CT (see (111) of [64]).
Many techniques have been applied or devised for m_ + ± = 0) [72]. Alternative to the assumption T_ = 0,
handling the uncertainty regarding the turn rate ! assume now that the magnitude of the nongravitational
in aircraft tracking and can be transplanted to the net force F(t) = T(t) ¡ D(t) = F(t0 ) is constant,8 where
problem of unknown ¸ in the GT motions. As studied t0 is an arbitrary initial time. Then for t ¸ t0 ,
in [52], they include Wiener or first-order Markov F(t) F(t0 )
®(t) = =
process modeling, multiple models, ASR adaptation, m(t) m(t0 ) ¡ (t ¡ t0 )±
and a kinematic constrained pseudomeasurement
®(t0 )
method. It should be kept in mind, however, that the = (34)
dynamics of ! and ¸ are quite different–while ! 1 ¡ ¯(t0 )(t ¡ t0 )
can be nearly constant over a relatively long period of ± ±
time (e.g., for civilian aircraft turning), ¸k is normally ¯(t) = =
m(t) m(t0 ) ¡ (t ¡ t0 )±
very rapidly (exponentially) decreasing right after
¯(t0 )
a rocket launch and changes in a jumpwise manner = (35)
during rocket staging. 1 ¡ ¯(t0 )(t ¡ t0 )
The GT models discussed here account only _
where ¯ = ¡m=m. From differentiation of (34)—(35) it
for the nongravitational net acceleration. In the follows that ®(t) and ¯(t) satisfy
absence of gravitational effects and with a constant
¸, they clearly describe a 3D rectilinear motion ®_ = ®¯ (36)
with a constant acceleration. Use of these models ¯_ = ¯ 2 : (37)
implicitly presumes that the gravity can be neglected
or handled separately. For the model (29), this was Then, an 8-state GT model for the state vector x =
done in [86] by a correction of the target position [p0 , v0 , ®, ¯]0 is given by
offset (in sensor measurements) due to the gravity x_ = [v0 , a0 , ®¯, ¯ 2 ]0 (38)
calculated at a nominal target altitude. For the model
(31), a simple yet effective technique is to use a with
gravity compensation similar to the piecewise-constant a = ®vr =kvr k + aG : (39)
acceleration model described in Section IIC5 [52]. Such a model was given in [9] with the inverse-square
Such an approach was also used in [35]. gravity a(1) _ _
G (5) in the ECI-CS, where vr = [x + ! y, y ¡
_ 0
! x, z] [see (54)].
D. Dynamic Models The effect of the net (thrust minus drag) force is
accounted for by (36)—(37) in this model. A similar
The booster models discussed so far are essentially model was given in [5, p. 138] as
kinematic–they do not account for the flight
dynamics specific for rocket-powered vehicles. ®_ = ®2 (t)=c (40)
However, the rocket-specific flight dynamics can with an assumed known initial state ®(t0 ) and known
be utilized to develop more accurate models. Such constant c, where the 7-dimensional state vector is
models are surveyed next, beginning with several GT x = [p0 , v0 , ®]0 . The model (40) was also used in [54]
models. within a state in spherical coordinates, along with two
1) Gravity Turn Models: In a GT motion, the options to handle unknown c: includec (with c_ = 0) in
thrust, drag, and hence their net acceleration, aN = the state vector as well as ®, or use multiple quantized
aT + aD , are all parallel to the Earth-relative velocity values of c. It can be shown that for c = F(t0 )=± =
vr = v ¡ − £ p, where p and v are target position
and velocity in the ECI-CS [see also (54)]. That is, 8 Thisdiffers from [9], which in effect assumes that the drag
aN = ®vr =kvr k, where ® = kaN k is the magnitude of acceleration kaD k = (1=2m)ScD ´(h)kvk2 can be written in the form
the net acceleration. Thus, the total acceleration in a of (34).

108 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
®(t0 )=¯(t0 ), the only solution of (40) is given by (34) Rather than in terms of ¯(t) of (35), the model
exactly; that is, both models describe the same net (36)—(37) can be expressed alternatively in terms of
acceleration and are thus theoretically equivalent with
1 m(t) m(t0 ) ¡ ±(t ¡ t0 )
this value of c. Equations (36)—(37) are preferable to ¿ (t) = = = = ¿ (t0 ) ¡ (t ¡ t0 )
(40) since the latter requires exact knowledge of the ¯(t) ± ±
constant c while the former allow indirect estimation which was referred to as the normalized time in [31].
of it (at the expense of an extra state variable ¯). Then the equivalent model is ®_ = ®(t)=¿ (t), ¿_ = 1
Actually, in view of the rocket thrust equation, the in the continuous time and ®k+1 = ®k =(1 ¡ T=¿k ),
thrust is T = ¡mU, _ where m _ < 0 is the mass flow rate ¿k+1 = ¿k ¡ T in the discrete time.
and U the rocket exhaust velocity (U = Isp g0 with Isp While the key GT parameter ® is modeled
being the specific impulse [101]), and the magnitude explicitly based on rocket dynamics in the dynamic
of the thrust acceleration is [47]: ®T = kaT k = T=m = GT models here, such a parameter ¸ is either assumed
_
¡mU=m. Differentiation of it yields known or modeled as a random process without
T_ 1 T_ recourse to rocket dynamics in the kinematic GT
®_ T = ®T ¯ + = ®T2 + models of Section IVB2.
m U m
2) General Boost Models: The GT dynamic
with ¯ = ¡m=m,_ which for constant thrust (i.e., T_ = 0) models rely on the assumption that the angle of attack
reduces to is very small and can be neglected. They cannot be
1
®_ T = ®T ¯ = ®T2 : (41) expected to have good accuracy when the angle of
U attack is not small, not to mention the possibility of
On the other hand, ¯_ = ¯ 2 for a constant mass flow staging.
_ Thus, both (40) and (36)—(37) can be viewed
rate m. a) 9-state boost model: A booster dynamic
as counterparts of (41) for net accelerations. Note model in the ECI-CS, without the GT assumption, was
that constant T and constant U are equivalent for derived in [47]. From aT = ®T uT it follows that
a constant m._ The assumptions of constant m _ and
a_ T = ®_ T uT + −T £ aT (45)
constant U are justified for steady-state operation of
a rocket engine [45]. where −T is the angular velocity of uT . Assuming
The GT model (38) is nonlinear. For state and constant m_ and constant U, which implies T_ = 0 and
covariance propagation, numerical integration by the thus (41), it becomes
Runge-Kutta method and linearization were used in
[9]. Not only is this computationally inefficient, but ®T
a_ T = a + −T £ aT : (46)
it also inevitably injects possibly considerable errors, U T
especially those associated with linearization. In our Further, if the angle of attack is assumed constant,
opinion, a better alternative is the following. Note then the angular velocity −T can be obtained as
first that the continuous-time model (34)—(35) [or
equivalently, (36)—(37)] of ® and ¯ has the following 1
−T = v £ (aT + aG + aL ¡ aC ) (47)
exactly equivalent discrete-time model: kvr k2 r
®k ([47] has an oversight here) where vr = v ¡ − £ p is
®k+1 = (42)
1 ¡ ¯k T the Earth-relative velocity. Note that the coordinate
¯k acceleration aC due to the Earth’s rotation appears
¯k+1 = (43) because −T is in the ECI-CS, but vr is relative to
1 ¡ ¯k T
the atmosphere; thus, after neglecting aC , the lift aL ,
where T = tk+1 ¡ tk is the sampling interval. Thus, and the drag, the state-space model for a state vector
state prediction for ® and ¯ can be done precisely x = [p0 , v0 , a0T ]0 is
(under the stated assumptions) in discrete time
directly. Then the prediction of the kinematic p_ = v (48)
components can be done by solving the kinematic
v_ = aT + aG (49)
part of (38). This eliminates the need for numerical
integration to deal with the nonlinearity of ® and 1 1
a_ T = kaT kaT + [v £ (aT + aG )] £ aT : (50)
¯. Note that when noise is added, (36)—(37) and U kvr k2 r
(42)—(43) are not generally equivalent. Also, which
discrete-time value of ® and ¯ should be used for This 9-state model is more general than the
(39)? For a discussion of a similar situation, the GT models discussed above. Similar to the 7-state
reader is referred to Section VB of [64]. The exact GT model of (38) with (®, ¯) replaced by (40), it
discrete-time equivalent of (40) can be found for any presumes that the value of the exhaust velocity U is
given c to be available [47, 100]. Otherwise, a trivial solution is to
®k include U as an extra state variable with the model
®k+1 = : (44) _ = 0. Alternatively, (50) can be written in terms
1 ¡ ®k T=c U

LI & JILKOV: SURVEY OF MANEUVERING TARGET TRACKING, PART II 109

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
of ¯ through kaT k=U = ¯, and ¯ can be included in extension of the 9-state model (48)—(50) was proposed
the state vector, along with the model ¯_ = ¯ 2 . In the in [43] that can handle steering maneuvers with a
special case of a GT motion, (50) reduces to (41). variable angle of attack. It includes 3 additional state
This is the case because not only is (50) implied by variables for angular velocities. An extension of the
(i.e., derived from) (41) but it also implies (41): Under 8-state GT model (38) was reported in [9], which
aT = ®T uv , (50) is the same as (46) after neglecting incorporates into the state vector, as in [85], [86],
the lift and coordinate accelerations, and hence two additional angles (yaw and pitch) representing,
respectively, the horizontal and vertical orientations of
®_ T uv + ®T u_ v = a_ T the acceleration vector relative to the velocity vector.
® This, however, requires satisfactory modeling of the
= T aT + −T £ aT dynamics of the two angles. It is not clear whether
U
a “nearly constant” assumption (Wiener process
1 model) would be good enough for this purpose. Two
= ®2T uv + −v £ (®T uv )
U r
other extensions of the 8-state GT model (38) were
1 proposed in [52] that account for the spatial and
= ®2T uv + ®T (−v £ uv )
U r temporal correlation of the thrust, drag, and lift during
1 boost.
= ®2T uv + ®T u_ v These extended models, discussed next, are
U
formulated in an ENU-CS, where aG and aC of the
which implies (41). total acceleration a = aG ¡ aC + aT + aD + aL are
Assuming negligible lift and coordinate handled in the standard way (see Section IIC). The
accelerations, this 9-state general model with the thrust and aerodynamic forces are considered in a
GT constraint reduces to the 7-state GT model moving (relatively to the ENU-CS) VTC-CS, which
just mentioned with drag ignored (i.e., ® = ®T ). has its origin at the target mass center and axes
Indeed, with the GT constraint aT = ®T uv , (49) has defined by uv = vr =kvr k, ut = up £ uv , uc = uv £ ut
the same form as (39), and (50) is equivalent to with up = p=kpk. Modeling in such a “body-frame”
(40) without drag. This should come as no surprise. is more convenient than in the sensor (ENU) frame or
The thrust-specific quantities in the 9-state general in the absolute (ECI) frame. Note that this VTC-CS is
model correspond to those for the net acceleration slightly different from the VTC-CS defined in [17] for
in the 7-state GT model. This correspondence offers modeling maneuvering reentry (see Section IIIC1a)
a justification for the 7-state GT model. The GT and the one in [24].
constraint aT = ®T uv has three equations with one b) 10-state boost model: In the general case
unknown parameter. It renders the 9-state general (without the GT constraint), the net acceleration
model as having only 7 free-state variables. This is aN = aT + aD + aL is expressed and the corresponding
consistent with the GT model (38) with (®, ¯) replaced dynamic parameters are modeled both in the VTC-CS.
by (40), which has 7 states. Since the drag is collinear with uv and the lift
This constrained 9-state GT model was used in is orthogonal to uv , the net acceleration can be
[45] for target interception during a gravity turn. A decomposed in the VTC-CS as aN = [uv , ut , uc ]®,
modified version that ignores the Earth rotation in where ® = [®v , ®t , ®c ]0 with
(50) (i.e., replace vr with v) was used in [100] for Tv ¡ D Tt + Lt Tc + Lc
launch point estimation. ®v = , ®t = , ®v = :
m m m
The essential assumptions of the general model
(48)—(50) are constant T (or constant m and U) and (51)
constant angle of attack, as well as negligible lift, Here Tv , Tt , Tc are the components of the thrust force,
drag, and coordinate accelerations. It was argued D is the magnitude of the drag force (including the
in [47] that drag and coordinate accelerations are residual drag), and Lt , Lc are the components of
negligible relative to thrust and lift is neglected the lift force. Then, the total acceleration a in the
since angles of attack are usually small during ENU
ENU-CS is a = g1 + TVTC ® with unknown ®, where
boost flight. These assumptions are nonetheless ENU
g1 was given by (12) and TVTC = [uv , ut , uc ] is the
restrictive. For example, while lift is indeed small VTC-to-ENU coordinate transformation. Similar to
at zero angle of attack (during a GT) or at a very (34)—(35), under the assumptions of a constant net
high altitude (where the dynamic pressure is very force in the (moving) VTC-CS (i.e., constant Tv ¡ D,
low), they may become significant otherwise. In Tt + Lt , and Tt + Lc ) and m _ + ± = 0, differentiating (51)
general, at a non-zero angle of attack a boosting and ¯(t) yields a system of differential equations for
target is subject to considerable forces lateral to the ® and ¯ that is the same as (36)—(37) except now
velocity vector that cannot always be neglected. For ® = [®v , ®t , ®c ]0 . Thus the complete state-space model
this reason several boost models take into account for the 10-state vector x = [p0 , v 0 , ®0 , ¯]0 is [52] p_ = v,
ENU
lateral accelerations as well [36, 43, 9, 52]. An v_ = g1 + TVTC ®, and (36)—(37).

110 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
This 10-state general model is a generalization D. Example of A Profile-Based Model
of the 8-state GT model (38): it takes into account
possible lateral forces during boost by lumping All models discussed so far are profile free. They
their effect into the net acceleration through (51). In describe the target motion by a state vector of a small
particular, if ®t = ®c = 0, it reduces to the 8-state GT dimension without recourse to additional information
model (38) but in the ENU-CS. In view of the strong of the target trajectory. Most of these models are
coupling between the thrust and aerodynamic forces in nonlinear and in general require numerical integration
the above model through the mass depletion function to propagate the trajectory. This is computationally
¯(t), as seen from (36)—(37), this model appears more difficult to implement in real time, especially when
sensible than augmenting the 8-state GT model with many targets are tracked simultaneously. Also,
two “independent” (uncoupled) variables (e.g., yaw they do not guarantee satisfactory accuracy. As an
and pitch as in [9], [85], [86]). To account for the alternative, a profile-dependent approach has been
modeling errors/model uncertainty, zero-mean white used traditionally in the BT tracking, where trajectory
noise should be added. profiles are obtained a priori (off-line) to propagate
c) 10-state stochastic boost model: If the the motion parameters in time [86, 25]. A trajectory
gravity and Earth rotation are accounted for and the profile (or template) typically consists of tabulated
thrust is aligned with the velocity, the lift is the only trajectory parameters [e.g., altitude, down-range and
possible significant lateral force in a GT motion. cross-range (in the orbital plane) position coordinates]
Being orthogonal to the velocity, the lift can be as functions of time since launch. If an accurate
represented as (see also [81]) profile is used, highly accurate estimates of the
current target states can be obtained. In practice, the
aL = 12 ´kvk2 (°t ut + °c uc )
exact profile of a foreign target is not known (nor
where °t and °c are the turn and climb lift parameters, necessarily in the set of available profiles) and is
respectively (similar to those in Section IIIC). Then, subject to target typing. Usually it is searched among
the net acceleration is a set of profiles, calculated in advance for the target
types of interest over a plausible range of trajectory
aN = aT + aD + aL parameters. This profile matching process has a
number of inherent issues and specific techniques
= ®uv + 21 ´kvk2 (°t ut + °c uc )
have been developed, which are beyond the scope of
ENU
= TVTC [®, 21 ´kvk2 °t , 12 ´kvk2 °c ]0 this survey.
Although much work has been done on
with ® = ®v = (T ¡ D)=m. Under the assumptions of profile-dependent modeling, relevant information
(34)—(35), ® and ¯ satisfy (36)—(37). In view of the available in the open literature is limited since
significant temporal correlation, it appears reasonable target profile information is usually classified. We
to assume that ° = [°t , °c ]0 can be approximated by a illustrate the idea via a simplified 2D model [98]
first-order Markov process (see, e.g., [23], [61]): (see also [67]) used in such a framework. More
detailed considerations of the approach and additional
°_ = ¡diag(·t , ·c )°(t) + w° (t) (52)
information can be found in [86], [25], [98], [9], [94],
where w° (t) denotes zero-mean white noise, and the [97], [78], and the references therein.
correlation time constants ·t , ·c are design parameters. The target trajectory is assumed (common in the
Thus, the corresponding 10-state stochastic model profile-based modeling) to remain approximately
with the state vector x = [p0 , v0 , ®, ¯, ° 0 ]0 is given by in a single (orbital) plane, defined by the position
p_ = v, vector p = [x, y, z]0 and velocity vector v = p_ in the
ECI-CS. The pair of vectors (p=kpk, h=khk) with
ENU
v_ = g1 + TVTC [®, 21 ´kvk2 °t , 12 ´kvk2 °c ]0
h = v ¡ (v0 p)p=kpk provides an orthonormal basis of
(36)—(37), and (52). this plane and determines the vertical and horizontal
Clearly, this model is another extension of the directions, respectively. The net acceleration vector
8-state GT model (38) that takes into account possible is then easily specified in the orbital plane as aN =
lateral (lift) forces during boost. Indeed, if °t = °c = 0 ®[(cos ')p=kpk + (sin ')h=khk] through its magnitude
(no lift), it coincides with (38) expressed in the ® and orientation–the pitch-over angle ' = 6 (aN , p)
ENU-CS. This 10-state stochastic model relies on from the vertical. After accounting for the gravity,
the modeling power of the first-order Markov model, such as given in (25), the motion model in the
which proved very efficient for numerous tracking ECI-CS is given by
· ¸
applications [23, 81, 61, 64]. ¹ p h
While these 10-state boost models proposed in p̈ = ®(t) cos '(t) ¡ + ®(t) sin '(t)
kpk2 kpk khk
[52] appear appealing, their effectiveness remains to
be tested. (53)

LI & JILKOV: SURVEY OF MANEUVERING TARGET TRACKING, PART II 111

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
where ®(t) and '(t) are known functions of time. In trajectory is pieced together using segments of the
such a case (given ® and ') the target state vector estimated trajectories by the independent elemental
consists of only position and velocity components: filters.
x = [p0 , v0 ]0 , and (53) yields a second-order nonlinear In the IMM architecture [10, 6], the
state-space model, which can be used to propagate the probabilistically weighted sums of estimates from all
target state by, e.g., numerical integration. elemental filters are always taken to be the (overall)
The knowledge of ®(t) and '(t) is a key state estimates. More importantly, elemental filters
assumption of this model, characteristic of the interact with one another by utilizing the most
profile-based modeling. Thus a thrust profile is recent estimates from every elemental filter. As such,
defined in [98] as a pair of curves P = f(®(t), '(t)) : the state estimates are in fact always a mixture of
t0 · t · tf g from the launch time t0 to the final state estimates based on all models, albeit with different
burnout time tf . A piecewise model for ®(t) and a weights.
linear model for '(t) were used in [98] to approximate A VSMM uses a time-varying set of models.
the actual profiles. Although not formally stated in [29], [28], its
This model is a simplified one. In a real-world MM-based tracking system actually uses a variable
application, other essential factors, such as the set of models (e.g., some models are deleted in real
Earth-rotational effects (see Section 5.6.1 of [9] and time), determined mostly by target altitude, and thus
[7]) and possible “out-of-plane” motions, should be has a variable structure.
accounted for. Another advantage of an MM-based tracker is that,
as reported in [28], it allows the inclusion of certain
V. INTEGRATION OF BT MOTION MODELS FOR “generic” (i.e., not phase specific) models for the
ENTIRE TRAJECTORY end-to-end BT motion to provide a “back-up” in case
other models in the working set do not work well,
Evidently, different models have been developed although these generic models may have degraded
for different regimes of a BT flight. In other words, accuracy compared with some phase-specific models.
no single model is adequate for most BT tracking It seems that a variable-structure IMM architecture
applications, although it was reported in [28] that the is particularly suitable for BT tracking. A more
first-order Markov booster models of Section IVB1b detailed description of these MM algorithms,
performed reasonably well over the entire BT including their pros and cons is beyond the scope
trajectories considered. Consequently there is a need of this survey. The reader is referred to the recent
to decide on the proper model(s) to use at any given surveys [66, 58, 68, 59].
time when tracking a BT. It should be emphasized that even when multiple
A natural idea is to use a decision logic–relying models are used, inclusion of at least one accurate
on, e.g., target altitude, and possibly other factors–to dynamics model for the flight phase in effect is still
determine the flight phase in effect at the time and use of vital importance for precision tracking.
the corresponding (single) model for tracking. Albeit
simple, it hinges on choosing the correct model, which VI. CONCLUDING REMARKS
is hard, particularly during a phase transition.
Alternatively and preferably, multiple models Numerous dynamic models of ballistic targets for
(MM) can be used simultaneously [6]. This approach tracking purposes have been developed. In essence,
was indeed taken and reported in [29], [92], [73], these models amount to various combinations of
[9], [28]. Three classes of MM architectures, models for three different classes of force–thrust,
corresponding to the three generations of MM aerodynamic forces, and gravity–at a variety of
algorithms [59, 66], have been proposed or accuracy levels in a suitable coordinate system.
used for BT tracking: autonomous MM (AMM) Table I summarizes the accelerations induced by these
[29, 28, 80, 79], interacting MM (IMM) [92, 73, 9, forces. Coast models consider gravity only; reentry
82, 22, 21], and variable-structure MM (VSMM) models rely on those of aerodynamic forces as well
[29, 28]. They all use a bank of elemental filters, each as gravity; and boost models account for all three
based on a particular model in the model set. forces, with thrust dominating. Table II summarizes
The multiple filters in the AMM architecture are the motion models proposed for all three flight phases.
completely independent during the tracking process. There are essentially three different gravity models
In [29], [28], the state estimate at any time is obtained used for BT tracking: the simplest constant-gravity
from the one based on the most probable model at model that assumes a flat, nonrotating Earth; the
the time, even though the state estimate of a standard inverse-square model based on a spherical Earth; and
AMM algorithm is a probabilistically weighted sum the sophisticated model with the J2 correction for an
of estimates from all elemental filters. As such, the ellipsoidal Earth. These models are more universally
model is automatically selected according to the applicable than models for other forces. The issue
computed model probabilities, and the estimated state here is not a lack of sufficiently accurate models, but

112 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
TABLE I
Accelerations on a Ballistic Target

TABLE II
Motion Models of a Ballistic Target

LI & JILKOV: SURVEY OF MANEUVERING TARGET TRACKING, PART II 113

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
a choice of an appropriate model for the problem at
hand. This choice depends mainly upon the intended
coverage of the motion regime(s) and the affordable
sophistication level. As a result, coast models require
the most sophisticated, while relatively simpler ones
can be used for boosters and reentry vehicles.
Aerodynamic forces include mainly drag and
lift. The lift is responsible for maneuvers of reentry
vehicles and, along with its induced drag, is basically
not predicable in the case of a foreign BT beyond the
fact that the lift is normal to the velocity. Therefore,
stochastic modeling techniques play an important role
here. The state of the art here is at the level of the
Singer model for aircraft maneuvers. In our opinion,
many other stochastic models, particularly those of an
adaptive nature, described in Section IV of [64] have
potential here and deserve investigation. The zero-lift Fig. 2. Coordinate systems.
drag is relatively more predicable and its models rely
heavily on aerodynamics. An improvement here calls North pole N. Its fundamental plane OxI yI coincides
for the adoption of a better yet affordable atmospheric with the Earth’s equatorial plane.
model for pressure, temperature, air density, and so The Earth-centered (Earth) fixed (ECF, ECEF, or
on. ECR)-CS OxF yF zF also has its origin at the Earth
The thrust is propellant specific, and an accurate center O, its axis zF ´ zI , and fundamental plane
model that is generic enough is hard to come by. OxF yF coincident with the equatorial plane. Its axes
Available booster models either treat it entirely as OxF and OyF , however, rotate with the Earth around
a random process with a stochastic model or (but the Earth’s spin axis OzF ´ OzI as OxF points to the
not both) impose simplifying constraints on it based prime meridian. See [77] for more details.
on rocket dynamics. We believe a better approach The ENU-CS OS xS yS zS has its origin OS at some
is to impose simplifying constraints and model the point on the Earth surface or above it (usually at the
deviation from the constraints as a random process. location of a sensor). Its up-axis OS zS is normal to
In view of the similarity between the widely used GT the Earth’s reference ellipsoid,9 usually defined by
models for boosters and the CT models for aircraft the geodetic latitude Á. The axes OS xS and OS yS are
maneuver, attention should be paid to the advances tangential to the Earth reference ellipsoid with yS
in the latter, particularly in terms of kinematic pointing North and xS East.
constraints, for their applicability to the former. Another CS (not depicted in Fig. 2), commonly
The Coriolis and centrifugal forces arising from used in BT tracking, is the radar face (RF)-CS10
the rotation of the Earth exist only in a noninertial [23, 70]. It can be defined from the local radar
frame. Its effects should be accounted for usually only ENU-CS by two angles of rotation [70]. For a phased
when the BT under track is traveling over a large area array radar, the x and y axes of the RF-CS lie on the
or a long time period, such as the coast flight. radar face, with x axis along the intersection of the
Compared with dynamics models of other powered radar face with the local horizontal plane, and z is
vehicles, BT models rely more on the aerodynamics along its normal (boresight)
and ballistics, and less on stochastic modeling p direction. Such a radar
measures the range r = x2 + y2 + z2 and the direction
techniques. While efforts for better modeling based on cosines u = x=r and v = y=r. This nonorthogonal
dynamics should continue, in our opinion, powerful coordinate system (r, u, v) is often referred to as the
stochastic modeling techniques deserve more attention RUV-CS.
in the future. ¡!
In Fig. 2, points P and OS (i.e., vectors ½ = OP
¡¡!
and r = OOS ) denote target and sensor positions,
APPENDIX. COORDINATE SYSTEMS
respectively, in the ECI-CS or ECF-CS. Throughout
The CS commonly used in BT tracking are this paper, the target position and velocity vectors
illustrated in Fig. 2. More details can be found in e.g. are p = [x, y, z]0 and v = p_ = [x_ , y_ , z_ ]0 , respectively,
[7], [70], [26], [8], [9], [75], [77]. in whatever coordinate system is considered. For
The ECI-CS OxI yI zI is fixed in an inertial
space (i.e., fixed relative to the “fixed stars”). It 9 Note
¡¡! that the local vertical direction differs from the radial axis
is a right-handed system with the origin O at the OOS . They coincide if a spherical Earth model is used.
Earth center, axis OxI pointing in the vernal equinox 10 Note that it differs from the so-called radar reference CS, which is

direction, axis OzI pointing in the direction of the actually just an ENU-CS at radar site [23].

114 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
¡¡! ¡! [11] Brookner, E.
example, p = OS P in the ENU-CS and p = OP in the Tracking and Kalman Filtering Made Easy.
ECI-CS. Note that the velocity in the ECF-CS can be New York: Wiley, 1989.
expressed in the ECI-CS as follows: [12] Cardillo, G. P., Mrstik, A. V., and Plambeck, T.
A track filter for reentry objects with uncertain drag.
x_ F = x_ I + ! yI , y_ F = y_ I ¡ ! xI , z_ F = z_ I IEEE Transactions on Aerospace and Electronic Systems,
35, 2 (Apr. 1999), 395—409.
(54) [13] Chadwick, W. R., and Zarchan, P.
where ! is the Earth rotation rate. Interception of spiraling ballistic missiles.
The choice of a CS is a complex issue, depending In Proceedings of the 1995 American Control Conference,
on numerous factors and related to many elements of Seattle, WA, June 1995, 4476—4483.
a tracking system [9, 75]. For more information, the [14] Chan, Y. T., Hu, A. G. C., and Plant, J. B.
A Kalman filter based tracking scheme with input
reader is referred to [62].
estimation.
IEEE Transactions on Aerospace and Electronic Systems,
AES-15, 2 (Mar. 1979), 237—244.
ACKNOWLEDGMENT
[15] Chang, C. B.
Ballistic trajectory estimation with angle only
The authors would like to thank Yaakov measurements.
Bar-Shalom and Paul Easthope for their comments IEEE Transactions on Automatic Control, AC-25 (June
on the conference version [61] of this paper. 1980), 474—480.
[16] Chang, C. B., and Tabaczynski, J.
Application of state estimation to target tracking.
REFERENCES
IEEE Transactions on Automatic Control, AC-29, 2 (Feb.
[1] Ali, A., and Sartuk, K. 1984), 98—109.
Development of a satellite borne tactical ballistic missile [17] Chang, C. B., Whiting, R. H., and Athans, M.
trajectory prediction tool. On the state and parameter estimation for maneuvering
In Proceedings of the 2003 International Conference reentry vehicles.
on Recent Advances in Space Technologies, Nov. 2003, IEEE Transactions Automatic Control, AC-22, 2 (Feb.
601—604. 1977), 99—105.
[2] Athans, M., Whiting, R. H., and Gruber, M. [18] Chen, G., Wang, J. R., and Shieh, L. S.
A suboptimal estimation algorithm with probabilistic Interval Kalman filtering.
editing for false measurements with applications to target IEEE Transactions on Aerospace and Electronic Systems,
tracking with wake phenomena. 33, 1 (Jan. 1997), 250—259.
IEEE Transactions on Automatic Control, AC-22 (June [19] Chen, H., Bar-Shalom, Y., Pattipati, K. R., and
1977), 372—384. Kirubarajan, T.
[3] Athans, M., Wishner, R. P., and Bertolini, A. MDL approach for multiple low observable track
Suboptimal state estimation for continuous-time nonlinear initiation.
systems from discrete noisy measurements. IEEE Transactions on Aerospace and Electronic Systems,
IEEE Transactions on Automatic Control, AC-13 (Oct. 39, 3 (July 2003), 862—882.
1968), 504—514.
[20] Chen, Y., Wen, C., Xu, J-X., and Sun, M.
[4] Bar-Shalom, Y., and Li, X. R.
High-order iterative learning identification of projectile’s
Estimation and Tracking: Principles, Techniques, and
aerodynamic drag coefficient curve from radar measured
Software.
velocity data.
Boston: Artech House, 1993. Reprinted by YBS
IEEE Transactions on Control Systems Technology, 6, 4
Publishing, 1998.
(July 1998), 563—570.
[5] Bar-Shalom, Y., and Li, X. R.
Multitarget-Multisensor Tracking: Principles and [21] Conover, S., Kerce, J. C., Brown, G., Ehrman, L., and
Techniques. Hardiman, D.
Storrs, CT: YBS Publishing, 1995. Impact point prediction of small ballistic munitions with
an interacting multiple model estimator.
[6] Bar-Shalom, Y., Li, X. R., and Kirubarajan, T.
In Proceedings of SPIE Conference on Acquisition,
Estimation with Applications to Tracking and Navigation:
Tracking, Pointing, and Laser Systems Technologies XXI,
Theory, Algorithms, and Software.
vol. 6569, Orlando, FL, Apr. 2007.
New York: Wiley, 2001.
[7] Bate, R. R., Mueller, D. D., and White, J. E. Fundamentals [22] Cooperman, R. L.
of Astrodynamics. Tactical ballistic missile tracking using the interacting
Mineola, NY: Dover, 1971. multiple model algorithm.
[8] Blackman, S. S. In Proceedings of the 2002 International Conference on
Multiple Target Tracking with Radar Applications. Information Fusion, Annapolis, MD, July 2002, 824—831.
Norwood, MA: Artech House, 1986. [23] Costa, P. J.
[9] Blackman, S. S., and Popoli, R. F. Adaptive model architecture and extended Kalman-Bucy
Design and Analysis of Modern Tracking Systems. filters.
Norwood, MA: Artech House, 1999. IEEE Transactions on Aerospace and Electronic Systems,
[10] Blom, H. A. P., and Bar-Shalom, Y. 30, 2 (Apr. 1994), 525—533.
The interacting multiple model algorithm for systems [24] Crosson, E. L., Romine, J. B., and Willner, D.
with Markovian switching coefficients. Boost-phase acceleration estimation.
IEEE Transactions on Automatic Control, 33, 8 (Aug. In Proceedings of the IEEE International Radar
1988), 780—783. Conference, 2000, 210—214.

LI & JILKOV: SURVEY OF MANEUVERING TARGET TRACKING, PART II 115

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
[25] Danis, N. J. [40] Gray, J. E., and Smith-Carroll, A. S.
Space-based tactical ballistic missile launch parameter On the usage of derived quantities in tracking: Energy
estimation. and other estimators.
IEEE Transactions on Aerospace and Electronic Systems, In Proceedings of the 2002 IEEE Aerospace Conference,
29, 2 (Apr. 1993), 412—424. Mar. 2002, 4-1571—4-1580.
[26] Daum, F. E., and Fitzgerald, R. J. [41] Hardiman, D. F., Kerce, J. C., and Brown, G. C.
Decoupled Kalman filters for phased array radar tracking. Nonlinear estimation techniques for impact point
IEEE Transactions on Automatic Control, AC-28 (Mar. prediction of ballistic targets.
1983), 269—282. In Proceedings of 2006 SPIE Conference on Signal and
Data Processing of Small Targets, vol. 6236, Orlando, FL,
[27] Davies, M. (Ed.)
2006, 62360C.
The Standard Handbook for Aeronautical and Astronautical
Engineers. [42] Hernandez, M. L., Ristic, B., Farina, A., and Timmoneri, L.
Columbus, OH: McGraw-Hill, 2003. A comparison of two Cramer-Rao bounds for non-linear
filtering with Pd < 1.
[28] Easthope, P. F. IEEE Transactions on Signal Processing, 52, 9 (Sept.
Using TOTS for more accurate and responsive 2004), 2361—2370.
multi-sensor, end-to-end ballistic missile tracking.
[43] Hough, M., Baum, F., and Huang, J.
In Proceedings of SPIE Conference on Signal and Data
Nonlinear recursive estimation of boost trajectories,
Processing of Small Targets 2000, vol. 4048, Apr. 2000.
including batch initialization and burnout estimation.
[29] Easthope, P. F., and Heyes, N. W. Presented at the AIAA/AAS Astrodynamics Specialist
Multiple-model target-oriented tracking system. Conference and Exhibit, Providence, RI, Aug. 16—19,
In Proceedings of SPIE Conference on Signal and Data 2004.
Processing of Small Targets 1994, vol. 2235, Apr. 1994. [44] Hough, M. E.
[30] Farina, A., Benvenuti, D., and Ristic, B. Improved performance of recursive tracking filters using
Estimation accuracy of a landing point of a ballistic batch initialization and process noise adaptation.
target. AIAA Journal of Guidance, Control, and Dynamics, 22, 5
In Proceedings of the 2002 International Conference on (1999), 675—681.
Information Fusion, Annapolis, MD, July 2002, 2—9. [45] Hough, M. E.
[31] Farina, A., Del Gaudio, M. G., D’Elia, U., Immediata, S., Optimal guidance and nonlinear estimation for
Ortenzi, L., Timmoneri, L., and Toma, M. F. interception of accelerating targets.
Detection and tracking of ballistic target. AIAA Journal of Guidance, Control, and Dynamics, 18, 5
In Proceedings of the 2004 IEEE International Radar (1995), 958—968.
Conference, Apr. 2004, 450—456. [46] Hough, M. E.
Optimal guidance and nonlinear estimation for
[32] Farina, A., Ristic, B., and Benvenuti, D.
interception of decelerating targets.
Tracking a ballistic target: Comparison of several filters.
AIAA Journal of Guidance, Control, and Dynamics, 18, 2
IEEE Transactions on Aerospace and Electronic Systems,
(1995), 316—324.
38, 3 (2002), 1916—1924.
[47] Hough, M. E.
[33] Farina, A., Ristic, B., and Timmoneri, L. Nonlinear recursive filter for boost trajectories.
Cramer-Rao bound for nonlinear filtering with Pd < 1 and AIAA Journal of Guidance, Control, and Dynamics, 24, 5
its application to target tracking. (2001), 991—997.
IEEE Transactions of Signal Processing, 50, 8 (Aug.
[48] Hutchins, R. G., and Pace, P. E.
2002), 1916—1924.
Studies in trajectory tracking and launch point
[34] Fitzgerald, R. J. determination for ballistic missile defense.
On reentry vehicle tracking in various coordinate systems. In Proceedings of 2006 SPIE Conference on Signal and
IEEE Transactions on Automatic Control, AC-19 (July Data Processing of Small Targets, vol. 6236, Orlando, FL,
1974), 581—582. 2006, 62360Y.
[35] Foster, G. J., Petruzzo, J. J., and Phan, T. N. [49] Hutchins, R. G., and San Jose, A.
Track filtering of boosting targets. IMM tracking of a theater ballistic missile during boost
In Proceedings of the 35th Southeastern Symposium on phase.
System Theory, Mar. 2003, 450—454. In Proceedings of 1998 SPIE Conference on Signal and
Data Processing of Small Targets, vol. 3373, 1998,
[36] Fowler, J. L.
528—531.
Boost and post-boost target state estimation with
angles-only measurements in a dynamic spherical [50] Hutchins, R. G., and San Jose, A.
coordinate system. Trajectory tracking and backfitting techniques against
In Proceedings of the AAS/AIAA Astrodynamics theater ballistic missiles.
Conference, Aug. 19—21, 1991, 41—54. In Proceedings of 1999 SPIE Conference on Signal and
Data Processing of Small Targets, vol. 3809, 1999,
[37] Gallais, P. 522—526.
Atmospheric Re-Entry Vehicle Mechanics.
[51] Jazwinski, A. H.
New York: Springer, 2007.
Stochastic Processes and Filtering Theory.
[38] Gelb, A. (Ed.) New York: Academic Press, 1970.
Applied Optimal Estimation. [52] Jilkov, V. P., Li, X. R., and Ru, J.
Cambridge, MA: MIT Press, 1974. Modeling ballistic target motion during boost for
[39] Gorecki, F. D., and Piehler, M. J. tracking.
Comparative results for a special class of robust nonlinear In Proceedings of 2007 SPIE Conference on Signal and
tracking algorithms. Data Processing of Small Targets, vol. 6699, San Diego,
Journal of Guidance, 12, 3 (May—June 1989), 289—296. CA, Aug. 2007, 669909-1—669909-12.

116 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
[53] Kameda, H., Tsujimichi, S., and Kosuge, Y. [65] Li, X. R., and Jilkov, V. P.
Target tracking for maneuvering reentry vehicles using A survey of maneuvering target tracking–Approximation
multiple maneuvering models. techniques for nonlinear filtering.
In Proceedings of the 36th SICE (Society of Instrument In Proceedings of 2004 SPIE Conference on Signal and
and Control Engineers) Annual Conference, Japan, 1997, Data Processing of Small Targets, vol. 5428, Orlando, FL,
1031—1036. Apr. 2004.
[54] Keil, K-H. [66] Li, X. R., and Jilkov, V. P.
Generic case study: Evaluation of early warning satellites Survey of maneuvering target tracking. Part V:
cueing radars against TBM. Multiple-model methods.
In Proceedings of 1999 SPIE Conference on Signal and IEEE Transactions on Aerospace and Electronic Systems,
Data Processing of Small Targets, vol. 3809, Denver, CO, 41, 4 (Oct. 2005), 1255—1321.
July 1999, 297—307. [67] Li, Y., Kirubarajan, T., Bar-Shalom, Y., and
[55] Kerr, T. H. Yeddanapudi, M.
Streamlining measurement iteration for EKF target Trajectory and launch point estimation for ballistic
tracking. missiles from boost phase LOS measurements.
IEEE Transactions of Aerospace and Electronic Systems, In Proceedings of the 1999 IEEE Aerospace Conference,
27, 2 (Mar. 1991), 408—421. Snowmass at Aspen, CO, Mar. 1999, 425—442.
[56] Lawton, J. A., Jesionowski, R. J., and Zarchan, P. [68] Mazor, E., Averbuch, A., Bar-Shalom, Y., and Dayan, J.
Comparison of four filtering options for a radar tracking Interacting multiple model methods in target tracking: A
problem. survey.
AIAA Journal of Guidance, Control, and Dynamics, 21, 4 IEEE Transactions on Aerospace and Electronic Systems,
(July—Aug. 1998), 618—623. 34, 1 (1998), 103—123.
[69] McCoy, R. L.
[57] Lee, S-C., and Liu, C-Y.
Modern Exterior Ballistics: The Launch and Flight
Trajectory estimation of reentry vehicle by use of on-line
Dynamics of Symmetric Projectiles.
input estimator.
Atglen, PA: Schiffer, 1999.
Journal of Guidance, Control, and Dynamics, 22, 6 (1999).
[70] Mehra, R. K.
[58] Li, X. R. A comparison of several nonlinear filters for reentry
Hybrid estimation techniques. vehicle tracking.
In C. T. Leondes (Ed.), Control and Dynamic Systems: IEEE Transactions on Automatic Control, AC-16 (Aug.
Advances in Theory and Applications, vol. 76, New York: 1971), 307—319.
Academic Press, 1996, 213—287.
[71] Mehra, R. K.
[59] Li, X. R. Approaches to adaptive filtering.
Engineer’s guide to variable-structure multiple-model IEEE Transactions on Automatic Control, AC-17, 5 (Oct.
estimation for tracking. 1972), 693—698.
In Y. Bar-Shalom and W. D. Blair (Eds.),
[72] Miele, A.
Multitarget-Multisensor Tracking: Applications and
Flight Mechanics: Theory of Flight Paths, vol. 1. Reading,
Advances, vol. III., Boston, MA: Artech House, 2000, ch.
MA: Addison-Wesley, 1962.
10, 499—567.
[73] Miller, M., Drummond, O., and Perrella, A.
[60] Li, X. R., and Bar-Shalom, Y. Multiple-model filters for boost-to-coast transition of
A recursive multiple model approach to noise theater ballistic missiles.
identification. In Proceedings of SPIE Conference on Signal and Data
IEEE Transactions of Aerospace and Electronic Systems, Processing of Small Targets, 1998, SPIE, vol. 3373, 1998,
30, 3 (July 1994), 671—684. 355—376.
[61] Li, X. R., and Jilkov, V. P. [74] Minvielle, P.
A survey of maneuvering target tracking–Part II: Tracking a ballistic re-entry vehicle with a sequential
Ballistic target models. Monte-Carlo filter.
In Proceedings of 2001 SPIE Conference on Signal and In Proceedings of the 2002 IEEE Aerospace Conference,
Data Processing of Small Targets, vol. 4473, San Diego, Mar. 2002, 4-1773—4-1787.
CA, July—Aug. 2001, 559—581. [75] Moore, J. R., and Blair, W. D.
[62] Li, X. R., and Jilkov, V. P. Practical aspects of multisensor tracking.
A survey of maneuvering target tracking–Part III: In Y. Bar-Shalom and W. D. Blair (Eds.),
Measurement models. Multitarget-Multisensor Tracking: Applications and
In Proceedings of 2001 SPIE Conference on Signal and Advances, vol. III, Norwood, MA: Artech House, 2000,
Data Processing of Small Targets, vol. 4473, San Diego, 1—76.
CA, July—Aug. 2001, 423—446. [76] Morrison, N.
[63] Li, X. R., and Jilkov, V. P. Introduction to Sequential Smoothing and Prediction.
A survey of maneuvering target tracking–Part IV: New York: McGraw-Hill, 1969.
Decision-based methods. [77] NIMA
In Proceedings of 2002 SPIE Conference on Signal and Department of Defense World Geodetic System 1984.
Data Processing of Small Targets, vol. 4728, Orlando, FL, Technical Report NIMA TR8350.2, National Imagery and
Apr. 2002, 511—534. Mapping Agency, 2000.
[64] Li, X. R., and Jilkov, V. P. [78] Petruzzo, J. J., and Foster, G. J.
A survey of maneuvering targets tracking. Part I: Combined tracking and threat typing algorithm for
Dynamic models. boosting missiles.
IEEE Transactions on Aerospace and Electronic Systems, In Proceedings of 2003 SPIE Conference on Signal and
39, 4 (Oct. 2003), 1333—1364. Data Processing of Small Targets, vol. 5204, 2003, 86—96.

LI & JILKOV: SURVEY OF MANEUVERING TARGET TRACKING, PART II 117

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
[79] Ravindra, V. C., Bar-Shalom, Y., and Willett, P. [93] Tewari, A.
Projectile identification and impact point prediction. Atmospheric and Space Slight Dynamics: Modeling and
IEEE Transactions on Aerospace and Electronic Systems, to Simulation with MATLAB and Simulink.
be published. Boston: Birkhuser, 2006.
[80] Ravindra, V. C., Bar-Shalom, Y., and Willett, P. [94] Tsai, M. J., and Rogal, F. A.
Impact point prediction and projectile identification. Angle-only tracking and prediction of boost vehicle
In Proceedings of 2007 SPIE Conference on Signal and position.
Data Processing of Small Targets, vol. 6699, San Diego, In Proceedings of SPIE on Signal and Data Processing for
CA, Aug. 2007, 66991A-1—66991A-12. Small Targets, vol. 1481, 1991.
[81] Regan, F. J., and Anandakrishnan, S. M. [95] Vinh, N. X.
Dynamics of Atmospheric Re-Entry. Hypersonic and Planetary Entry Mechanics.
AIAA, New York, 1993. Ann Arbor, MI: The University of Michigan Press, 1980.
[82] Register, A. H., Blair, W. D., and Brown, G. [96] Wishner, R. P., Larson, R. E., and Athans, M.
Improved radar revisit time control to observe ballistic Status of radar tracking algorithms.
missile mode transitions. In Proceedings of the Symposium of Nonlinear Estimation,
In Proceedings of the 34th Southeastern Simposium on San Diego, CA, Sept. 1970.
System Theory, Huntsville, AL, Mar. 2002, 349—353. [97] Woods, E., and Queeney, T.
[83] Ristic, B., Arulampalam, S., and Gordon, N. Multisensor detection and tracking of tactical ballistic
Beyond the Kalman Filter. Particle Filters for Tracking missiles using knowledge-based state estimation.
Applications. In Proceedings of SPIE Signal Processing, Sensor Fusion,
Norwood, MA: Artech House, 2004. and Target Recognition III, vol. 2232, 1994.
[84] Ristic, B., Farina, A., Benvenuti, D., and Arulampalam, M. [98] Yeddanapudi, M., and Bar-Shalom, Y.
Performance bounds and comparison of nonlinear filters Trajectory prediction for missiles based on boost-phase
for tracking a ballistic object on re-entry. LOS measurements.
IEE Proceedings on Radar Sonar Navigation, 150, 2 (Apr. In Proceedings of 1997 SPIE Conference on Signal and
2003), 65—70. Data Processing of Small Targets, vol. 3163, July 1997,
[85] Roecker, J. A. 316—328.
Track monitoring when tracking with multiple 2-D [99] Yeddanapudi, M., Bar-Shalom, Y., Pattipati, Y., and Deb, S.
passive sensors. Ballistic missile track initiation from satellite
IEEE Transactions on Aerospace and Electronic Systems, observations.
27, 6 (1991), 872—875. IEEE Transactions on Aerospace and Electronic Systems,
[86] Rudd, J. G., March, R. A., and Roecker, J. A. 31, 3 (July 1995), 1054—1071.
Surveillance and tracking of ballistic missile launches. [100] Van Zandt, J. R.
IBM Journal of Research and Development, 38, 2 (Mar. Boost phase tracking with an unscented filter.
1994), 195—216. In Proceedings of 2002 SPIE Conference on Signal and
[87] Saulson, B. G., and Chang, K. C. Data Processing of Small Targets, vol. 4728, Orlando, FL,
Nonlinear estimation comparison for ballistic missile Apr. 2002, 263—274.
tracking. [101] Archan, P.
Optical Engineering, 43 (2004), 1424—1438. Tracking and intercepting spiraling ballistic missiles.
[88] Sioris, G. M., Chen, G., and Wang, J. In Proceedings of the 2000 IEEE Position Location and
Tracking an incoming ballistic missile using an extended Navigation Symposium, San Diego, CA, Mar. 2000,
interval Kalman filter. 277—284.
IEEE Transactions on Aerospace and Electronic Systems, [102] Archan, P.
33, 1 (Jan. 1997), 232—240. Tactical and Strategic Missile Guidance, vol. 199 of
[89] Sivananthan, S., Kirubarajan, T., and Bar-Shalom, Y. Progress in Astronautics and Aeronautics (4th ed.).
Radar power multiplier for acquisition of low observables AIAA, Reston, VA:, 2002.
using an ESA radar. [103] Zhao, Z-L., Chen, H., Chen, G., Kwan, C-M., and Li, X. R.
IEEE Transactions on Aerospace and Electronic Systems, Comparision of several ballistic target tracking filters.
37, 2 (Apr. 2001), 401—418. Presented at the American Control Conference 2006, June
[90] Sorenson, H. W. 14—16, 2006.
Kalman filtering techniques. [104] Zhao, Z-L., Chen, H., Chen, G., Kwan, C-M., and Li, X. R.
In C. T. Leondes (Ed.), Advances in Control Systems IMM-LMMSE filtering algorithm for ballistic target
Theory and Applications, vol. 3, New York: Academic tracking with unknown ballistic coefficient.
Press, 1966, 219—292. Also in [91], 90—126. In Proceedings of 2006 SPIE Conference on Signal and
[91] Sorenson, H. W. (Ed.) Data Processing of Small Targets, vol. 6236, Apr. 2006,
Kalman Filtering: theory and Application. 62360K.
Los Alamitos, CA: IEEE Press, 1985
[92] Sviestins, E.
Multi-radar tracking for theater misisile defence.
In Proceedings of 1995 SPIE Conference on Signal and
Data Processing of Small Targets, vol. 2561, San Diego,
CA, July 1995, 384—394.

118 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 46, NO. 1 JANUARY 2010

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.
X. Rong Li (S’90–M’92–SM’95–F’04) 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 Technology Research Center.
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; information fusion; and target tracking, detection, and
classification, performance evaluation, statistical inference, stochastic systems,
and electric power.
Dr. Li has authored or coauthored four books, seven book chapters, and
more than 200 journal and conference proceedings papers. He has served the
International Society of Information Fusion as 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 (1995—1996) and as editor (1996—2003); 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 North America, Europe, Asia, and Australia. 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 involved in research and development of radar
tracking systems. From 1989 to 1999 he was a Research Fellow with the Institute
of Parallel Processing, Bulgarian Academy of Sciences, Sofia, where he worked
as a key researcher in numerous academic and industry projects (including
international) in the areas of estimation, target tracking, sensor data fusion, and
parallel processing. In 1999 he joined the Department of Electrical Engineering,
University of New Orleans, where he is currently an associate professor.
Dr. Jilkov is author/coauthor of over 80 journal articles and conference papers.
His current research interests include stochastic systems, nonlinear filtering,
applied decision and estimation, target tracking, information fusion, and parallel
processing.

LI & JILKOV: SURVEY OF MANEUVERING TARGET TRACKING, PART II 119

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:53 UTC from IE Xplore. Restricon aply.

You might also like