Hc-culty ÜaÌÌlSlÍC NchlClc1taCkÍug
Piere Mnvielle
IEEE Aerospace and Electronic Systems Magazine
Volume 20 Number 8
August 2005
This magazine has not taditionally accepted contributions that
occupy more than a few fnished pages, or that contains materal that -at
frst glance - appears better suited for our other publications.
The accompanying aicle falls into that gap between our usual
articles for Systems and those for Transactns or one of our Tutorials.
Based on the extremely favorable reviews received, however,
we elected to brng this article to your attention Õ an "extra."
Are we giving you valuefor your time and efort?
Are we satising your needs?
Do you want more of this tpe of contribution?
Is the use of an insert efective in highlighting special contents?
Pleae let us know your thoughts about the depth of treatment of
content and this approach to te deliver of truly outstanding works that
]ust "do not/tt|eregular mold. "
Suggestions for topics and the authors are always welcome.
Evelyn Hr
Decades of Improvements in
Re-entry Ballistic Vehicle Tracking
Pere Mvele
Abstract- This paper deals with a missile defense
challenge, which has been studied extensively for decades
and still remains Ü topic of active research, the tracking of a
ballistic vehicle in its re-ent phase. With Anti-Ballistic
Missile or Anti-Tactical-Ballistic Missile goalsg most of
missile defenses need to track re-entry vehicles with a view
to locating them precisely and allowing low altitude
interception. The re-entry vehicle leaves a quiet exo­
atmospheric phase and a quasi-KepJeran motion to Ü endo­
atmospheric phase with large aerodynamic loads and a
sudden deceleration. The motion is then obviously non­
linear and frthenore both the extent and the evolution of
the drag are dificult to predict. Since the sixties, while more
and more efcient sensors, such as sophisticated phased­
aray radars, ¼CIC developed, the associated data processing
techniques have been improved, taking advantage of
computer performance increases. Although re-ent vehicle
tckig is undoubtedly a non-linear fltering problem, it
was frstly solved by rustic linear flters with fixed weight
and an assumed drag table. Then, Kalman flters were
exploited. Uncoupled and polynomial for a while, they are
now flly coupled and really non-linear. They are called
Extended Kalman filters, are based 1I te linearization
about the estimated state ad are the curent efcient and
classic solution to non-linear filtering. Moreover, they
perfor drag estimation which may be usefl to te
identifcation capabilities of the missile defense. I the
future, even more sophisticated techniques, such as the
promising and fashionable particle flteringg are likely to be
developed in re-ent vehicle tracking+
Anti-Ballistic Missile (ABM) or Anti-Tactical-Ballistic
Missile (ATBM) defenses may be confonted with
numerous challenging tacking problems, that depend on the
threat and on the defense layer architecture: the boost-phase
missile tracking for early waing or interception with laser
weapons, the multiple target tracking against Ml V s
Authors Curnt Ads
Sf n02 ¯ 33114 L Br ¯ Frc
(multiple independently targeted reentry vehicles) and
decoys, the multi-sensor tracking and fsion of infared
sensors, the latest crcial seconds of tracking by a hit-to-kill
interceptor, etc. Amongst these problems, the tracking of
ballistic vehicles in their re-entry phase has been stUdied
since 1960, following te work of Kalman and Bucy on
recursive linear filtering [I]. It still remains now a topic of
active research (as those few references [8,9,10,15,16,17,20]
of the last fve years testif it).
At the time of the sixties and the seventies, there was a
widespread interest in re-ent vehicle (RV) tacking [2], as
the Cold War urged the United States and the Soviet Union
to develop each their !¼Ü ABM defense in order to protect
themselves against the others. In 1975, the US Safeguard
system [3], which was descended fom U succession of
previous proj ects such as the Sentinel program and was
reduced following the 1972 ABM Treaty, became
operational and protected the Grand Forks site and the
Minutema ICBM (Inter Continental Ballistic Missile)
fields fom limited Russian or Chinese attacks.
Nevertheless, it was quickly said that the Safeguard system
could be easily overhelmed by the then new Russian
MIRVs. Also the system was completely closed i 1978 and
only its supporting rad remains now. At the same time, the
Russia elected to defend Moscow [3]; their ABM defense,
notably armed wit their nuclear interceptor Galosh, has
been upgraded for the past decades and still exists tody.
Meanwhile, A TBM defenses has spread with many efforts
on terinal interception and on RV tacking. For example,
Israel has developed the ATBM "Arrow" system that is able
to intercept Scud missiles; Arrow is based on an L-band
phased array radar and a missile that uses a fagmentation
warhead [4]. Furthennore, the curent U.S. National Missile
Defense (NM) project, frstly without any terminal layer
which can defend only Ü limited area, is to be merged with
ATBM projects and is likely to integrate terinal layers
which could specifcally defend stategic areas.
The paper focuses on techniques and algorithms which have
been used for RV tracking and, more precisely, for targets
which are not maneuvering . We will only bring up the radar
aspect even ifboth are closely linked. Since te beginning of
ABM defenses, phased array radars have been employed
and improved [4]. Their key advantages are their ability to
allow the beam from a large antenna to be rapidly steered
without mechanically moving the antenna, to maintain many
targets in tack, and to be easily hardened against nuclear
Maupt rive Novembr 11, 20; rv June 2, 20
0885-8985/0/ $17.0 ( 20 IE
blast. Moreover, vaouh radar techiqÌCs strengthened by
the use of digital processing, such as pulse compression,
MTl (moving-target indicator), and multimode, have
contributed to the advance of phased-aray radars. Fig 1-3
illustrate different applications of phased-aray ÎadÜrs. Fig I
represent5 "Cobra Judy", an U.S. shipboard radar built by
Rytheon and desiged for collection of data on foreig
ballistic missile tests and also for satellite tracking. Fig 2
represents the RuSsiÜn "Pill Box" radar, part of the Moscow
ABM system, ¾ith it four phased array providing a 3600
coverage. Fig 3 shows the transportable grounded-based
"Green Pine" radar, built by lAlELT A aÏd used within the
ATBM Israeli "Arrow" system.
Figure I- "Cobra Judy" (AN/SPQ-ll)
Figure 2 - "Pill Box" (Don-2NP)
Figure 3 -"Green Pine" (ELfM-2080)
Whether it is applied in the field of ballistic missile defense
or in the civilian feld of safety against the re-ent of
satellites and debris, RV tracking consists of locating and
following vehicles that leaves a quiet exo-atospheric phase
and a quasi-Keplerian motion to an endo-atmospheric phase
with large aerodynamic loadS and a sudden deceleration. RV
tacking is a nonlinear fltering problem: the motion is
obviously non-linear and both the extent and the evolution
of the dag are diffcult to predict. In addition to that
dynamic non-linearity, the measurement is stongly
nonlinear [2].
The paper deals with the evolution of the RV tracking
sofware techiques since the sixtiCs, fom the frst rustic
linear flters to the curent sophisticated foning methods.
The work is a summary and piecing together of previous
work; all inforation comes strictly fom the open literature
and the mentioned papers.
The oJgaiZation oíthe paper is as follows. SÕction 2 briefy
describes the problem of R V tracking. Section 3 intoduces
the optimal fltering applied to RV tracking. Then Section 4
descriles the evolution of the techniques over the past
decades. Section 5 give5 the concluding remarks.
2.1 Balistic Re-entr Vehicle Motion
The motion of the balliStic RV fons a complex dynamic
phenomenon [5,6]. The detenining factors are:
the vehicle characteristics which can deteriorate:
geomety, inertia (a5ym½et, unbalance), materials,
spin, etc.
the environment characteristics: local gravitational
feld, atospherc properties (pressure, density,.
temperature and derived quantities), hydometr, etc.
� """"'J�� ¯�
(a) tbe·maspbe·e ��
" 'V �
; :.

¢ _
I p. l o

¬ .¯

= I
,,' mes�spbe·e �
. ' ..
1 ! �
.. ! . y . ·c
� "I<ef �� ¹¦¸

¯ '
�� w`' �¯ 1#`` 3Ö � w� ¾
Atmospheric Volumlc d.slt (kgm3)
Figure 4 -Re-entry dynamics & atmospheric density
The atmospheric re-ent, ¾hich follows an eXo~
atmospherc phase (a) (ga,�-dominated with quasi­
Kepleria motion), is usually made up of two phases (Fig 4):
an accelerated phase (b) between 120 ad 60 k altitude
(gravitational forces still dominate atmospheric contact
forces) where the aerodynamic moment and velocity· vector
become in a line and where incidence fuctuations damp
down, a decelerated phase (c) under 60 k altitude
(aerodynamic drag is comparable, then largely dominates
gravitational forces) where incidence nominally converges.
During that last phase wit its considerable heat fu,
ablation might lead to inertial and aerodynamic asymmetries
whicl are harmfl to re-ent.
2.2 Radar Meaurements
Te phased aray radar scans the target with a vaing rate
(Fig 5). The radar is quite fee in the number of pulses
(monopulse or pulse tin) and in the design of each pulse
(e.g., modulation in phase, frequency). However the
tansmitted energy m¼st be enough to provide a corect
signal-to-noise ratio and detection.
When the radar manages to detect the target, a measurement
is generated (Fig 6). The extacted information consists of
the target agla direction (aimut AZ, elevation E), the
target range from radar d, and possibly, the Doppler
velocity (range rate). This measurement is imprecise with
possible bias and noises.
Scan tJ

.. `
/� .
Objective ,
Figure 5 ¬ Radar scanning of the re-entry vehicle
Measurement noise is usually assumed to be white and
Gaussian [5], with the following standard deviation:
(0 the radar resolution in the given dimension, SI the
signal-la-noise ratio).

Figure 6¬ Radar measurement imprecision
2.3 Re-entr Vehicle Tracking and Motion Characteristics
Generally speaking, te aims ofRV tacking are:
to follow the target and to create a persistent perception
of the target which is required by the situation
assessment process and by any action or action planning
that are going to lie within the duration,
to allow the engaging and guiding of interceptors or any
weapons in charge of the destruction or neutalization
ofthe target,
to determine the target impact point and te launching
point, enablig real-time 'counter-attack' (must be done
early during the trajector),
to contribute to the classifcation ad identifcation of
the target (e.g., lethal targetdecoy, type of the debris)
by exploiting cinematic fCatuÎes,
continue pointing at the target, producing
measurements, and maintaining tracking.
Thus, in spite of the dynamic non-linearity, tracking must
lead to a precise localization of the target and to d corect
prediction of this localization in the fture. Therefore, the
target position and the frst two derivatives, i.e. velocity and
acceleration, need to be estimated. In the deceleration phase
(3), aerodynamic drg is predominant and the acceleration
vector and te velocity vector are collinear. One parameter
may be enough to characterize the deceleration: the ballistic
coefcient p, which is introduced in the classic Allen
approximation [5,6), a re-entry model (1) with null
incidence and no lif, where it remains constant.

(V is the target velocit magnitude, V the acceleration
magnitude and p the local atmospheric density)
In the real situation, the incidence is not null and p changes
while the incidence varies and the aerodynamic
characteristics, principally the drag component, move. The
obseration of P is evidently not direct and its estimation
precision relies on the estimation of bot the velocity and
the acceleration. In the exo-atmospheric phase, (1 ) suggests
the indetermination of J since the atmospheric density is
insignifcant and te acceleration negligible. The ballistic
coefcient fnds a meaning when the target penetrates the
dense layers of the atmosphere and when a variation in !
involves a variation in the deceleration. Then, the
deterination of P is essential to quantif the deceleration.
Its estimation depends on the quality and the rate of the
radar measurements. It also depends on the perforance of
tacking methods, i.e. the ability to put up with dynamic
non-linearity, to exploit and control measurements, to use
the kow ledge about environment, and to be content with a
priori information produced by upper sensors. Indeed,
trackng techniques need to integrate various information
with a suited modeling (Fig 7). That integration has evolved
while the techniques improved and the computer
perforances increased (cf. section 4).
& Modeling
A priori
• m ÜÜbÚ"meo'
.. M M...M Chon".ru""
Figure 7 ¬ RV tracking, information and modeling
RV tracking deals with the state deterination of a dynaic
system using noised measurements. Non-linearity may be
present both in the dynamics and the measurements. Applied
to RV tracking, non-linear filtering is described by the
dynamic model, the measurement model, and the a priori
knowledge (or initial distribution). The choice of each of
them determines the flter design, its estimation accuracy, its
numeric stability, and the required computation workload
Tis chdpter is a quick presentation of optimal fltering and
recursive Bayesian estimation for RV tracking with the
intention to be able to show i Section 4 how the successive
techniques position themselves in the general framework of
these non-linear non-Gaussian state-space models and how
they difer.
Dnamic State Model-The dynamic (system) model of the
taget is represented by a probÜbilistic description, i.e. a
non-linear, Markovian discrete-time stochastic process:
At time k, X is the (hi dden) statevectcr,
is the non­
linear (possibly time-varing) system transition fction, vk
is the model noise which is not necessar white, nor
Gaussian. We suppose herein that the target is not
maneuvering (no input interfering with dynamics).
The state vector describes the target state at time k. The
choice of the state space (and its dimension) is crucial. 1B
RV tacking, one naturally introduces the position and the
velocity of the target. Derivatives of higher order may be
introduced and are supposed to improve the precision when
the target position is extrapolated in the fture. However it
may be useless and even harmful to add certain parameters
to the state space when they can not be estimated. This
observability question is indeed ver general in estimation
techniques [7] and of course depnds on the measurements
and on the produced information. Concering RV tacking,
it seems difcult to introduce more than the deceleration,
quantifed by the ballistic coefcient, since the drag
component strongly fuctuates.
The choice of the dyamic model arises fom the knowledge
of the dynamic motion of the target, on its
reresentativeness about the motion characteristics and on
practical considerÜtions, such as computation time.
Likewise, the model noise includes the imprecision about
the motion coming from the approximate chosen model,
including for exaple uncertainties on the atmosphere state
and on the local gavitational field.
According to the Newton 's Law expressed in the non­
Galilean coordinate system of a ground observer, the
dynamic model can be generally described by the motion of
the center of gavity (G) {6]:
¿ -
A -
K = r à _·,
j¯,·:·,·i,·¤,·(·,¹ ·,)j
ad j
ate respectively the velocit ad te
acceleration ofG,
is the resultant on G of the aerodynamic forces,
is the grÖvitational acceleration which varies
according to the target position,
te expression in squÜre brackets combines the
centripetal acceleration and the Corialis ter.
depends on the local atmospheric density () ad its
predominant drag component is more precisely in proporion
to the dynamic pressure (q):
(V is te velocity magitude)
The description of the gavitational force comes within the
knowledge of the local gravitational feld. The model
complexity (spherical model; ellipsoidal model, local
model) is determined by the precision level which is
accessible to RV tracking. Te imperfect knowledge of the
atmosphere (difference between the real atmosphere profle
and the supposed one) also involves incertitude, and
possibly, bias or erors on diferent coinponents of the state
Measurement Model- Te measurement model is defined
by the following stochastic equation:
At time k, Zk is the measurement vector (dimension p),
hk is the (possibly time-varying) measurement fnction,
wk is the measurement noise (not necessary Gaussian but
conditionally independent given the process X ).
The radar measurement provides information about the
target range, the target angular directi on, and possibly the
Doppler velocity. For example, z
may be in dimension 3
(without Doppler velocity); Zk ¯ ¡¸ r d_' where
and r are the directional cosines which determine the
angular diection, d is the target range measurement. The
choice of the state coordinate system detennines the relation
between state and measurement, i.e. the degee of nOD­
linearity of hk . Notice that the system state is partially
observed (hk is not injective).
Prior Knowledge The prior knowledge (initial
distribution) is, at the frst time 0, the inforation on X that
the radar gets before the frst measurement. While
information about position and velocity can come fom
upper sensors, the information on the expected deceleratiop
can be provided by technical considerations. The
information may be both Ü1 estimdtor X of the state vector
and a quantifcation of the associated incertitude. In general,
this knowledge is given by the prior probability density
Optimal Filtering
Te rd scans the target and produces measurements at
diferent times (with a varying rate). The estimation problem
is a discrete time flterng problem which consists in
determiing the inforation state [7], meaning a fnction of
the available inforation (the sequence of noised
measurements) that completely summarizes the past of the
system and that estimates in the best way the system state in
the present and in the fture.
The sequence (Xk ¸¸¸can also be viewed as a Markov
process described by a transition probability
P(Xk E dXo, ... ,Xk_l) P(X
E «


k (Jk_1 ·1l
The measurements Zk' conditionally independent given the
(Jk )k'O'
are totally (in Ü probabilistic way)
described by the conditional densities
' Calling
measurements and Z
¯ ¸:,
the rndom sequence of
Z2 -¸|the associated
realization, and assuming the whiteness of both the two
noise sequences, the best inforation [7] that can be
produced wit Z
is the conditional density of ' given
'" Z
k |`I
The deterination of
is called optimal fltering. The
probability density fnction (PDF)
¸ is any distribution,
not reducible to the frst moments; the dimension of the
problem is generally infnite.
Recrsive Baesian ESlimalion- The optimal flter
estimation may be computed in a Bayesian recursive way,
and can be divided into two stages: a prediction stage and a
corection stage (cf. Fig 8).
Figure 8 ¯ Optimal fltering recursion
The prediction state is obtained by a convolution product
(0) of the model transition probability by the probability
density of te estimated state (a priori density) at the last
k-l- k-I =
k Q I
Zk-l- k-I
k_I -Z
The above equation is known as the Chapman�Kolmogorov
equation. Te transition probability
k can be expressed by
the following stochastic integal.
Q, Six Ix _ ·�····¹¹.

|·.!ª··�. '´'
k k-I "k-l
¡~l ¯ ´k-l
The correction stage is produced by Bayes rule with the
following relations between a priori and a posteriori

k=xk ',¸¸,·.

The noralizing denominator, which generates non­
linearity, can be developed Ü.

, ·
,¸,=g • Ix
"l- k-I ·dk (11)
¯^ k k g ¯º
Analytical solutions to tose rClations are only realizble in
a few simple situations. Thus, in the Linear Gaussian
Hypotheses¸ the two stages come to the Kalman flter. I
the non-linear general situation, numerical suboptimal
algorithms need to be used.
In the R V tacking problem, that obviously does not flfl
te Gaussian Hypotheses, only approxiate methods may
be used. Those methods have evolved since their beginnings
as it is described in the �ext chapter.
Estimators- With the knowledge of te posterior density
_ ,it is possible to compute estimates with respect to any
criterion. For example, the minimum mean-squae eror
(MMSE) estimate is the conditional mean of Jk :
while the maximum a posteriori (MA) estimate is the
maximum of ]¸ (it may be more suited when the posterior
density _ is multi-modal). Furthermore, a measure of
aÛcuraLy of a state estimate (e.g., covariance) may also be
obtined fom ]¸ and confdence regions may be
A attCpt of chronology about the main successive RV
trÖcking techniques is represented in Fig 9, with frst their
development, then thei practical use and fnally their
decline. The lists of numbers in square brackets make up
references connected to the implementation of each
technique for R V tackingµ Afer the fxed�gain flters,
which were the frst rustic flter of the 1950s and te
beginning of the 1960s, Kla flters were exploitÖd« They
were quite quickly replaced by derived fors called
Extended Kalman filters (EKF) which, sice the seventies,
have been the classic solution to non-linear fltering¡ Iterated
Extended Kalman flters (lEKF), close to EKF, were also
Linear Gaussian Hypotheses: Gauss-Markovian process +whit Gaussian
noises (mutally independent) ¬ Gaussian dprior pf [8J
used. Among the curent emerging methods, the Unscented
Kalman flters (UK) and, above all, particle flters (PF)
may replace EKF in the medium or long ter.
Find-Gain Filtrs
Kalman Filtrs
EKF, IEKF, et.
UF, PF, 0M.
Figure 9 -RV tracking chronology
I this chapter, the principles of the successive techniques
are frstly prsented. Then, an analysis of the evolution is
developed, including a comparison of the perforance and
of the complexity.
Various appellations have been given to the fxed-gain
flters, among which fxed-weight filters, steady-state flters
and Wiener flters in their fequency for. The most famous
implementations of fxed-gain flters, suitable for noisy
kinematic models, are the a-1 flters and the a-14 flters,
Only the prnciples will be presented next; refer to [7] for a
complete description of fxed-gain flters.
Optimal filter approximations- Those fIed-gain flters
require the stong assumption that the system is linear and
time-invariant, meaning both the state and the measurement
equations (2) et (5) are linear and time-constant
(stationarity). Under suitable conditions, the state estimation
covariance (the estimation of the inceritude on the
estimated state) will converge to a steady-state value. The
corection stage may then be computed using a fxed
multiplying factor that weight the importance attributed to
te current measurement. In R V tacking, these required
assumptions are not valid at all. Rougly speag, it could
be gsped as if this wrong assumption was compensated,
concering the state model, by a high model noise.
Choosing a piecewise constnt Wiener acceleration model
[7,IOJ, the state model (2) is represented by the third order
'¡,, ¯ t ', ¬+¸ (I2)
1 T
wit t ¯ 0 1 T
0 0 1
I RV tacking, the main alteratives for choosing the state
coordinate system are a Cartesian coordinate system, such as
an Earth-centered fixed
(ECF) coordinate system, and a
"sensor" coordinate system, such as a range, azimuth and
elevation (RAE) coordinate system or a range, directional
cosines coordinate system. Whatever the coordinate system,
" the state vector ought to integrate components on position
(P) (RAE for example), on velocity (V) and on acceleration
(A) required by the third order model (12).
Decoupling te state, each state vector component in each
coordinate is propagated using the polynomial expansion
(12). Then, the prediction stage (8) consists in (for the rge
component R ofa RE coordinate system):
where í_
_ and í_¸¡
¡ are respectively the estimated range at
time k and the predicted range at time k+ 1. Working in the
RAE state coordinate system, the measurement model (5) is
straight and the correction stage (10) may be computed by
the following equations (for the range estimation):
r -¨(·.
is the measurement eror and where K
K k and K i ÜÜ precomputed filter weights [10].
Implementation principles- Fig 10 sums up the above
description of a fxed-gain fltering recursion applied to RV
tracking. The state vector dimension is 9, with 3 position
components (), 3 velocit components (V) ad 3
acceleration components (A), The initialization consists in
the state vector Xo (given by U had-over).
As it was mentioned before, the prediction or propagation of
the state estimator Xk1k is uncoupled and polynomial. More,
this prediction stage may use an a priori drg table in order
to take a drag deceleration into account. This drag table,
where the ballistic coefcient P is a function of the altitude,
is precomputed and estimated for an a priori trajectory [10]
which must corespond more or less to the expected
traj ector .
The radar process, including actions such as pointing and
signal processing (detection, SI evaluation, etc.), uses the
information of the predicted state x
and will fmally
x¡ k
produce a measurement. The correction stage consults a data
base containing the fxed weights that are a fnction of the
altitude. The weights are precomputed by a supervised
training, based on the a priory trajectory [10]. At the end of
the corection stage (14), the update estimated state i
k~l k+¡
is provided.
¯ ECF coordinate system: (0 XT YT ZT) cartesian and non-Galilean
coordinate system, 0: Ear center, Zj along Earth axis of rotation, X[ and
YT rotate with the Earh.
hx�d wei ghlS
Figure Iã Fixed-Gain ftering recursion
Remark- The computational and memory requirements of
the fxed-gain flters are known to be ver modest.
Moreover, they were implementable as early as the 1950s
with the then amplifers and time-invariant network
elements, such as resistors and capacitors [19].
However, their efciency and their accuracy are mediocre,
still worse when the measurement conditions Ü difcult
(low rates, misdetections) or when the tajectory and its dag
evolution is quite different fom the a priori trajetory [10].
In that case, they can encounter instabilities and
divergences. Implicitly based on a high model noise which
compensates a poor model process, fxed-gain flters for R V
tracking are not able to reach a precise level of estimation.
Furthermore, they are not adapted to te initial tansient
period when they should give more weight to the
measurements than to the weak previous inforation. Fixed�
gain flters are probably more suitable for the exo­
atmospheric phase where their assumptions are pretty well
Finally fxed-gain flters applied to RV tracking are
completely incapable of estimating the deceleration and te
ballistic coefcient evolution of the target, as they need t
use this evolution as an Ü priori hypothesis, both for the state
prediction and the achieving of the correction weight K.
Thus, they can not provide this discrimination information.
Variants- Several variants and improvements may be
added to' R V tracking fxed-gain flters. Among them, the
propagation equations can be coupled or the precomputed
weights can be refined with a growing memory mechanism
[8). This system avoids the maladjustment of the initial
transient period by the temporary use of a zero-process noise
flter [7].
4.2 Kalmanfilters
Kalman flters, known a Kalman-Bucy filters when they are
used in continuous time problems, frst appeared at the
beginning of the sixties [I] and were a major breakthough
for the linear estimation in dynamic systems. Refer to (7] for
a complete description of Kalman flters,
Optimal flter approximations- The Linear Gaussian
Hypotheses, described i chapter 3, is required for this flter,
The PDF ], that we tr to detenine, is Gaussian and is
completely described by the mean ik ¯ t¸x,¡z·¯ z' }
and the covariance matrix Pk Cov(
ì Pk
represents the (estimated) incertitude on the estimated $lBIß.
In RV tracking, the state process is not linear. As for the
above fixed-gain flters, a straight process model, such as a
piecewise constant Wiener acceleration model, may be used
with Û adapted model noise. About the measurement
model, the description of te state vector in a "sensor"
coordinate system involves that (5) is linear:
H,Xk+Wk |l°)
0 0
.. .J
1 0 and
is the model noise
0 1
(additive, zero mean and white) and Rk tI5 covariance.
The Kalman flter recursion is close to the fxed-gain one,
with the propagation of the covariance matrix � (16-18)
and a dynamically calculated weigt K (17) .
Pk+l1k ·¯· �,. '¢k' + V·
H ' ó -I
K hi
�+llk' k+I'
Pk+llk+1 Pk+l1k -Kk+1 ¯,.
wlcr¤ ó,,, -J
.,,Pk+l1k • H
+1 +Rk+1
implementation princi/es- Fig II sums up the above
description of a Kalman fltering recursion applied to RV
Te state vector dimension is 9, with 3 position components
(P), 3 velocity components (V), and 3 acceleration
components (A). The initialization step consists i both an
estimated state .0 and the associated incertitude Po (given
by a hand-over).
The prediction stage and the corection stage propagate at
once the estimated state and the covariance matx. Te
ballistic coefcient may be computed [9], from the
estimation of the velocity ad the acceleration, and using te
relation (I) (adding if the need aises a gavitational ter).
Remark-The computational and memor requirement of
Kalman flters are slightly more important ta the fxed­
gain filters. As a matter of fact, Kalman flters wit al their
matx calculations needed the integated digitl circuit
modules which were developed from the beginning of the
sixties [19].
«..s-.,s .s- -......·. .-Ûf.i--l p .. -· ·.l.-.!c+, I!S
cst-..i-- .. ·..s- --.., ¸-; c-....- .s-.- ·..-.- a|.-..
..- --. .-.. .- .-.-,·... IB .s·.· ....- --s-. ,.-.-..
,-«|-+,- .--.. .-- -..|..... --·--, --..-- Tus, .s..·
-s.i-8ti-u CÍ!hc 06Cc!cI8lICD.--,----. .. ¬.., ...-..-s -,
.s· --..·----. --..·. «..s-.. .-, .---.s.-,
Figure i I - Kalman fltering recursion
Variants ÏIHcIcu! Q-ss:-l· ..·..-..
.-M,.cd0c.ÛMQ.-s 0I .-..-+ -·+-·±.·s
,·-.-.., .·- ,-...-|- .-s .-a.--.-
.--,......-.| .-.+
· s Extended-Kalmanfilters
¸ : .; ...-
-·+·. --+-|
--..-., .s-
ÎXlcBCc0~ÄÌÇ8u t..-·· (EK) , derived from Kalan
a..... .. . ..--,..-.| ..,-...s- .+.,.-: .- ---..-..·
·,..--. WDctc .s- .-,.----....-- -. .s- .,..-.. a|..· ..
--.., [8BU .- av -....-, -.,, .-.....-.· a·.-· .- ¸·, .-·
. .--,.-.- s-....,..-- -. -·.--s-+·..-.- a|.-..
Optimal filter approximationstss +..-..., ...-. QÌ8Cc IB
.s- HN !t8CKlDg Î-.lit, -..-.-, .s- B0B¡Iucðly -. .s-
+,-.-... .-s .s· --...·.---. -,....--. ». -..|.-·, --.s
.s- --s·. ,·-.-.. --..- .-+ lhc Ïe..Mr-¯cBt --..- a
.ss..... .-.- --.- .-s «-..· rs- -..- ....-,..-- .. .s--
!h8llhc PDF ]¸ will Icm8lB --·- -· .cS. Gaussian 8D0WIÌ¡
Dc Q

JI0I8!Ic¡ s·....-.+ -, .s- --.-
X _ � � ¸.k ¬ Z ' ¦ .-s !hc ...-....-s .-.....-.·
-.e.· r, • a.......,, -·- ·-.-... .- . .-·-. -·,.-..-- CÍ
!hc ÛDl (o. .-.--s -·+-·, -· .s- ---..--.· -,....--. CÍ.s-
+,-÷.. .-s .s- --...·----. «..s .s- .-.--..-- -c
--...-.-, ..- ....- .-s --...·--.-. ,·-s....--. s-· . s·..
ut0ct t¬ lh8l .n`o.-. .-. .--,.....-- of the i..--..-
¸·¸ .. .s- ...-.. -...-..- -c .-- ....- .-s .s·
a .'1'
1η ÌÜ
.--,.....-- -. !hc i..--..-
X' +ll'
!Dc ,·-:...-s
Implementation principles- t., .: ..-. ., .s- .--.-
s-....,..-- -. . tm t...·.-, .-..·.-- .,,..-s .- av
......-, .-..-+...-, . .-.-. .--«.-s,- CÍ!Dc VN ^chiÛ.C
--...- --- -.- ..- . ....- .-·.-. -. ..---..-- · ¸: | . . s;
', ¯ ¸· Y Ý V +

where (x, , ., ..· ...,-. ,-....-- .--.·.-.... . - -cr ,·,.
vy, vJ ..- ..·,-. vc¡CCI!y .--·s.-..-. - -c- .-s p .. .--
-....... .-·+..--. .s- .-..........-- -t .s· ....- .-..-·
·,·....|, .. ,..-- -, .- . ,·.-- c......- inceritde -- .s-
-........ .-·a..--. �.

Figure 1 2 ¬ EKF ftering recursion
PCCU!0IBg .- !hc »-«.-- . ,.« ·, -.,·-..-+ i ..- non­
c....-.- tc- WI!h --., . +.., !cH,
x = v _ y = Vy z = vz

V = C + Q
x + 2Qv -çJ-
X X y

Vy ¯ C¸ + O¯y - 2Ov¸-çJ-÷

Vz =
«--·- C¸ ,C
.·- .-- _8V;l8!:On8Ì Ío..c .--,----..
V .· .s· .-|-..., -Ügitsc q .. .-- s,-.-.. ,·-...·- ad
n .. .-- t...- .-....-- ·..-
·s- prediction ...,- CCuSl8lS CÍBumcIl68ÌÌy .-.-,....-,, -,
.- ODE solver such .. the Range-Kutta method, .s- abve
flly coupled ballistic equations ( 1 9). It also consists in
propagating the covariance matrix Pk by (16)
wii c ¯ ( _¸ . About ilie ballistic coeficient �, its
dynamics are considered as almost constant [2, 1 0, 1 5] i ie
dynamic model (p Cvolution is gene|lly slow for a non-
maneuvering vehicle). That does not keep the esiimatet J
from varying during ie tacking, due to the adde 1 process
noise (� proess is actually an independent-increment
Wiener process).
The coÏÖÜtion stage is quite similar to the Kalman one,
apar te use of the Jacobian ¸_¸
Remark-EK are much more greedier in calculations and
memor tha the forer ones. However, that can strongly
differ according to the choice of the coupleduncoupled
equations, on the integraton method, and on may
optimizations which can be introduced in all the steps.
RV tracking EK, that uses knowledge on the local
gravitational feld and on the atmospheric densit, leads to a
more accurate estimation of the state vector. The ballistic
coefcient p is estimated [2, 1 0, 1 5] without the noisy
behavior of the Kalman metod. Moreover. it ofers Ü
estimation of the incertitude around the estimated /l
Nevertheless, unstability and divergence may occur when
the initializatioÜ is to imprecise, when the measurement
conditions are degaded and when, more generally, the local
approximations are not checked.
Many methods, more or less sophisticated, have been
srudied to improve RV tracking EKF [I I], mostly on the
propagation of the covariance matrix. Among them, the
choice of the coordinate system seems to be the more
sensitive. According to [2, 1 1], a "sensor" coordinate system
(yet not ver practical to express te dynamics (3)), like te
RAE, would be the best option. Measurement equations
remain linear and the process model holds all te
nonlinearity. Moreover, it allows a decoupling of the eror
covariance matrix and reduces it ill-conditioning. Tus, an
EK in Û "sensor" coordinate system would be as efcient
as an lEKF [2], a more complex derived for described
straight afer.
4.4 Iterated Etended-Kalman flters
The efects oflinearization erors in the EK can be reduced
by relinearizing the measurement equation (5) around te
updated state rater t Ielying only on the predicted state.
Thus, Iterated Extended-Kalman flters (IEK) consist of a
few added iterations (the number is decided either a priori or
based on a convergence criterion) of the EKF correction
. h
¸ ¸
stage, computmg eac hme the Jaco Ia O at the new
updated state. This technique may improve R V tracking
when obviously the measurement eqUation is nonlinear, i. e.
when the state coordinate system is Cartesian [2]. Refer to
[7] for a complete description of Iterated Extended-Kalman
Variants- Several other techniques, close to FK anc
lEK, have been suggested for RV tacking. For example,
[2] describes a single-stage iteration filter which is more or
less an one iteÏation lEK' ,with, what is more, a
relinearization of the dynamics aroUnd a smoothed (meaning
backward predicted) estimate, In the other hand, Interacting
Multiple Model (IMM) techniques can be usefl a they are
able to switch between a bark of flteÎs, each of them
adapted to a diferent phase (boost, exo-atmospheric, endo­
atmospheric) of the ballistic miss�le [20].
4.5 Unscented Kalman filters
Unscented Kalman flters (UKF) are an emerging method,
closely related to EK. I oIdCr to capture the nonlinearity,
their basic idea is to propagate the matrix covariance, not
only by asi�¿Û local approximation at the estimated state
and a (ew ters of a Taylor series expansion, but by
prpagating a judiLioQs number of deterministically chosen
points (coherent with the state uncertainty) through the
nonlinear dynamics model and by using those predicted
points a a approximation of the predicted covariance
matTx. Refer to [1 7] for a complete description of U.
[ 1 6, 1 7] illustrate the application of UKF to simplifed
versions of R V tracking.
4.6 Particle flters
Paricle flters (PF), also called Sequential Monte Carlo
methods, emerged towards the end of te 1980s. They are
simulation based metods and include several closely
related algorithms known under the names of Bootstrap
flters, Condensation, Monte Carlo flters, interacting
particle approximations, etc. These stochastic algoritms
can also be looked upon as genetic algorithms, with
alterate phases of crossover, selection and mutation [14].
Refer to [ 13, 14] for a complete descrption of paicle flters.
Optimal flter approximations-The main idea of PF is the
approximation of the PDF ]
by a set or a population of
random samples (or particles) {Xk(i): i ¯ 1, . . . ,Nd, where NE
is the number of samples. The associated empirical



(0 is the Dirac
N£ i�Ï
delta fnction); it is supposed to tend to
¿ Ü NE tends to
infnity. Then, the integral calculation of te prediction and
the Bayesian correction stage are coMputed in a Monte
Carlo way, propagating te paricles by the stochastic model
process, stengthening their ¾eight according to their
likelihood towards the current measurement, sampling, and
resapling if necessary, etc.
I the prediction stage, each sample of {Xk.l(i): i ¯ 1, . . . , NE}
(reresenting ¹¸

¸... ¸ k-J at time k-l) is propagated
independently (independent realizations of noise Vk_l ) with
the dynamic model:
¦·,(i) /
1, . . . , N
,is then representative of the PDF
CF· l l
/_¸¸¸..�¸.. of the optimal flter equation (8).
I the correction stage, the likelihood of each sample x,(/)
is evaluated, relating to the new measurement q. The
noralized (importance) weight of each sample becomes:


(21 )
The Bayesian correction ( 1 O) i s then realized, with the
Bootstrap flter, by resampling NE samples fom
¦·,(i) i¯ 1, µ . e , N
¦ according to the importance
weights q _ . The most probable samples are selected
preferentially and it can be demonstated that the set of
random samples �, (i) i
1, « e e , N
¦is representative of
the a posteriori PDF

(1 4).
¬ Z =Z
In this famework, both the process model and the
measurement do not need to be linear and the approximation
is not simply local as it used to be for EKF. Any process ad
measurement noises (for example multi-modal distrbutions)
can be taken into account.
Implementation principJes- Fig 13 sums up the above
description of a paricle fltering recursion applied to RV
tracking. Using a state vector of dimension 7 described in
the Cartesian coordinate system ECF [15], the sample
components (position, veloity, and ballistic coefcient) are
¸· y Z Vx Vy Vz ¡|
The initialization consists of a set of random samples (a few
tens of thousands for example) that represents the
incertitude on the state (given by a hand-over). The
prediction stage (20) propagates each sample, applying the
process noise (mutation effect) and integating the flly
coupled ballistic equations (1 9) with an ODE solver. The
corection stage (21 ) selects the most probable samples by
resampling techniques, Multinomial resamplng being the
simplest but not the most effcient method [ 1 3] .
Various state estimators can be computed, such Ü the

MMSE estimator (approximated by ),or the
NE Ì=l
MAP estimator, more suited when [¿ is multi-modal.
Remark-The computational and memor requirements of
paricle flters are undoubtedly much higher than EK ones.
However, as for EK, that can strongly depend on various
choices and many optimizations can be introduced. Among
them, some techniques are now emerging that manage to
reduce the number of propagated particles while maintaining
the estimation accuracy of ]¿ [ 1 3]. An other way to
accelerate PF methods could be to use parallel processing.
Particle flters could be specially usefl when the tracking
conditions are unusual, i.e. for example when the
CF-1 2
initialization is very imprecise (EK may diverge as the
local EK approximation is too wrong) or when the
measurements stop being quasi-Gaussian and when several
modes appear [ 1 5]. More generally, SMC methods can
improve RV tracking for they give a larger framework to
integrate imprecise inforation about dynamics,
measurement, and knowledge.
Figure 13 ¯ Particle fltering recursion
However, degeneracy problems (meaning the number of
selected samples rapidly collapsing to a few values, and
even a single value) may occur if one does not take care.
Variants-Many techniques are being developed to improve
PF and to solve degeneracy problems. Among them, the
addition of a sampling Markov Chain Monte Carlo
(MCMC) step [ 1 3] or the massive resor of EK or U for
each particle in order to guide its propagation wit the latest
Other new methods are cur ently tested for non-linear
fltering. For example, grd-based numerical approaches are
developed to solve the stochastic diferential equations of
the optimal flterng recursion [ 18].
4.6Analysis ofthe evolution
The improvements of the R V tracking data processing
techniques have been contin].lous since the 1 960s. More and
more accurate, they have became more robust as [ 10]
testifed it for an EKF implementation at the millimeter
wave (MMW) radar of Kwajalein (Marshall Islands). They
are more and more able to get out of diffcult tracking
conditions, such as data loss periods [10] or unexpected RV
dynamics [ 15]. Afer a few decades, they are now capable of
modeling and integrating, inside their behavior, knowledge
about the RV motion, and also capable of taking the
uncerainties (dynamics, environment, measurement) into
account in te closest way. By this very fact, they no longer
require tuning or an a priori table, either before or during
tracks [ 1 0] . Finally, they can potentially perform a precise
estim�tion of the ballistic coefcient and thus may
contibute to te classifcation or identifcation of the target.
These improvements have been mainly accomplished tanks
to the great advances made in the digital computer
technology in the last fouf decades. The well-known
checked Moore' s Law of Fig 14 illustates it, with its
exponential inÜÏease of the number of tansistors on a chip
and the processor speed in MIPs (million instructions per
second). Te successive Intel processors are located on the
Number of
transistors •
, � � Pe�tiu�IV �
1 0 million- . ¢ ~ �=+ =- !- w « - ! - . w r . + • , � .

g ¿ g .. ª
• 500
+ w 25
1 million
1 00000
i i i i

.j ... w . = = « - i - ' ':' _ . - ' r' " . ..

I i 6 i
. - .!_ . _ . j • J,t�l + . , _ .�. _ . _ � _ .
l 6
i t l
. - I
i i i i
. - 0. 1 1
• • ��
1 000
--=. l- +---+--..0. 01
1 975 1 980 1 985 1 990 1995 2000 Year
Figure 14 - Moore' s Law
increase M processor speed
o memor capacit
1 960 1 975 1990 2005 Year
Figure 15 * Processor global improvement
and RV tracking requirements
Fig 1 5 represents the relative exponential increase of both
the processor and the memory capacity since 1 960 (this
rough evaluation is based on the civilian computer evolution
but is likely to apply more or less to more confdential
computers). Furthenore, domains of computationa (and
memory) requirements are shown for each RV tracking
technique. This evaluation comes partly fom references
[8,15] and partly, we must admit it, fom a rough guess. The
inceritude is due to the req1iImeMt variation between
various implementations of a RV tacking technique (e.g.,
coupleddecoupled, polynomialfODE integration, number of
propagated particles for a particle flter). Of course, and it
could not have been diferent, all those techniques are
consistent with the computer technology of their time.
Fixed-gain flters were implementable with the then rustic
analogic and vacuum tubes technology, Kalman flter
needed te digital technology of the 1 960s ( I 9] and, since
the 19708, Extended-Kalman filters have been riding on te
wave of integated circuit and microprocessor development.
The particle flters have been in gestation for a long time
[ 1 3] and are now emerging while the computer
peronnances aford it. SMC methods are doubtÛsb now
more attainable than the fxed-gain flters weÎe at the
beginning of the sixties.
Some of the RV tracking techniques, that have followed one
another since the sixties, have been presented. Frm the frst
simple rustic linear flters to the curent Monte-Carlo
simulation based methods, the performances, meaning the
accuracy and the robustness, have naturally raised and the
classification abilities have been developed. Those
techniques have benefted fom the considerable increase of
computer capabilities.
More generally, the RV tracking techiques and the
associated perforances must be considered jointly with the
radar system (or any other sensor), within a defense system
and agaist a expected threat.
( 1] R.E. Kalman and RS. Bucy, ¹ New result in linear
fltering and prediction theor I , 1. Baic Eng. ¿ vol. 83,
1 961 , pp. 95- 1 08
[2] R.K. Mehra, ² A Comparison of Several Nonlinear
Filters for Reentr Vehicle Tracking », IEEE Transactions
on Automatic Control, Vol. AC- 1 6, N° 4, AugÓst 1971
[3J Col. D. Smit, U A Brief Histor of "Missiles " and
Ballitic Missile Defense», CentÔÏ for DefeÜsC Information,
http://wwv. ·.cdi.orgostspots/issuebrief/ch2/
[4J "Detection & Ranging - Radar in the 2dh Centu
IEEE Aerospace and Electronic Systems Magazie, Jubilee
Issue, October 2000
[5] FJ. Regan and S.M. Anandakshan, Dnamics of
Atmospheric Re-entr, AIAA, Education series, 1 993
[6J P. Galais, Mecanique d vol, CEACEST A, 1998
[7] Y. BܬShalom and X.R Li, Estimation and Tracking -
Principles, Technique and Software, Artech Houe, 1 993
[8] lAo Lawton, RJ. Jesionowski ad P. Zarchan,
"Comparison of Four Filtering Otions for a Radar
Tracking Problem", Joual of Guidance, Control and
Dynamics, Vol. nO 4, July-August 1 998
[9] R.J. Jesionowski and P. Zarchan, "Comparion of
Filtering Options For Ballistic Coeficient Estimation ",
Nasa report, 1 998
[ 1 0] G.P. Cardillo, A. V. Mrstick and T. Plambeck, "A Track
Filter for Reentr Objetcs with Uncertain Drag", IEEE
Transactions on Aerospace and Electronic Systems, Vol. 35,
n° 2, Aprl, 1999
[ 1 1 ] F.E. Daum and R J. Fitzgerald, « Decoup/ed Kalman
Filters for Phased Arra Radar Tracking }, IEEE
Transactions on Automatic CoÏtrol¡ Vol. AC-28, N" 3,
March 1983
CF- 1 3
[12] PJ. Costa and W.H. Moore, "Extended Kalman-Buc
Filters for Radar Tracking and ldentiication", Proceedings
of the 1 991 IEEE National Radar Conference, IEEE
Aerospace and Electronic Systems Society, 1 991 , pp. 1 27-
1 3 1
[ 1 3] A. Doucet, N. de Freitas and N. Gordon, Sequential
Monte-Carlo Method in Practice, Springer, 2001
[14] N.J. Gordon, DJ. Salmond ad A.F.M. Smit, "Novel
approach to non-linearlnon-Gaussian Baesian state
estimation," lEE Proceedings, Part F, Vol 140, nOl, 107-
JJ 3, April, 1 993
[ 1 5] P. Minvielle, « Tracking a Ballitic Re-entry Vehicle
with a Sequential Monte-Carlo Filter », 2002 IEEE
Aerospace Conference, March 9-1 6, 2002
[ 1 6] A. Fana, B. Ristic and D. Benvenuti, V Tracking a
Ballistic Target: Comparison ofSeveral Nonlinear Filters »,
IEEE Transactions on Aerospace and Electonic Systems,
Vol. 38, July 2002
[ 1 7] SJ. Julier ad J.K. Uhlmann, « A Ne Extension o/the
Kalman Fi/ter to Nonlinear Systems I³¡ The Proceedings of
AeroSense: the 1 1 Interational Symposium on
AerospacelDefense SÔÛsing Simulation and Controls,
Orlando, FL, 1997
[ 1 8] S. Challa and Y. Bar-Shalom, « Nonlinear Filter
Design using Foker-Planck-Kolmogorov Probabilit
Densit Evolutions», IEEE Transactions on Aerospace and
Electronic Systems, Vol. 36, N° 1 , Jaua 2000
[ 19] B. D. O. Anderson and J.B. Moore, "Optimal Filtering,"
Ed. T. Kailatb, Prentice Hall, 1 979
(20) R. L. Coopea, "Tactical Ballistic Missile Tracking
using the lnteracting Mltile Model AlgOrithm", Fif
Inteational Conference on Infonnation Fusion, 8- 1 1 July,

Sign up to vote on this title
UsefulNot useful