I.

INTRODUCTION
The trajectory of a thrusting projectile can be
divided into two phases: thrusting and ballistic.
A Multiple IMM Estimation The target dynamics are substantially different and
characterized by different main forces in each phase.
Approach with Unbiased The resulting acceleration can be written in terms of
the thrust, drag, and gravity components [3] (the lift
Mixing for Thrusting Projectiles force is neglected in this work, since it is very small
in the scenarios considered) as
aF = aT + aD + aG : (1)
For the thrusting phase, we consider thrust, drag, and
TING YUAN
YAAKOV BAR-SHALOM, Fellow, IEEE gravity; for the ballistic phase, only gravity and drag.
PETER WILLETT, Fellow, IEEE For a trajectory with unknown drag coefficient and
E. MOZESON unknown thrust, it is difficult to clearly divide the
S. POLLAK trajectory into its phases. The interacting multiple
University of Connecticut
model (IMM) estimator with a thrust mode (TM) and
DAVID HARDIMAN a ballistic mode (BM) to match these two phases is a
AMRDEC
natural choice (it is assumed that no optical indication
of the plume is available, as in the real-data examples
considered).
Given a very short observation period of a
We present a procedure to estimate the state of
thrusting/ballistic projectile for the purpose of impact
thrusting/ballistic endoatmospheric projectiles for the end
point prediction (IPP), an accurate estimate of the
purpose of impact point prediction. The short observation time
target state (including its thrust and drag coefficient)
and the estimation ambiguity between drag and thrust in the is of the utmost importance. The IPP accuracy
dynamic model motivate the development of a multiple interacting depends mainly on the drag coefficient estimate
multiple model (MIMM) estimator with various drag coefficient since the prediction is done after the thrusting period
initializations. A simple unbiased IMM mixing procedure (useful is over, i.e., based on the BM. Note that the net
for quite general applications) is presented for state estimators acceleration (1) includes the (algebraic) sum of the
with unequal dimensions and applied for the thrusting and drag and thrust, thus causing an ambiguity in the
ballistic modes in the case considered. Results with real data are estimation when both are unknown. A better initial
given. estimate of the drag coefficient is conducive to a
more accurate estimate of the thrust, which then
results in a more accurate overall state estimate and
better IPP performance. The sensitivity of the system
performance to the drag coefficient estimate leads to
using a multiple IMM (MIMM) estimator to overcome
the above-mentioned ambiguity to more accurately
estimate the drag coefficient and thrust. The technique
involves evaluating the likelihood functions of the
IMM estimators with different initial drag coefficient
estimates to quantify how well the MIMM estimators
Manuscript received September 14, 2011; released for publication
fit the observation data [6].
December 12, 2011.
Another issue that arises is the mixing of the
IEEE Log No. T-AES/48/4/944203. mode-conditioned state estimates in an IMM estimator
Refereeing of this contribution was handled by L. Kaplan. when the modes used have different dimension state
vectors. The conventional approach in this case is to
This work supported by ARO W911NF-103-1-0369 and ONR
N00014-10-1-0029. augment with zeros the lower dimension state estimate
prior to the mixing [2]. However, this leads to a bias
Authors’ addresses: T. Yuan, Y. Bar-Shalom, P. Willett, E. Mozeson,
toward zero for the state components of the larger
and S. Pollak, Department of Electrical and Computer Engineering,
University of Connecticut, 371 Fairfield Rd., Storrs, CT state vector that are mixed with the extra components
06269-9005, E-mail: (p.willett@ieee.org); D. Hardiman, U.S. Army of the smaller state that are zero. A simple procedure
Aviation and Missile Research, Development, and Engineering to avoid this “biasing” is presented, together with a
Center (AMRDEC). suitable augmentation of the covariance of the smaller
state that yields an unbiased and consistent mixing.
0018-9251/12/$26.00 °
c 2012 IEEE This leads to a “direct” block mixing in contrast to

3250 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 48, NO. 4 OCTOBER 2012

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 the component-by-component mixing described in the Using the notation from [1] and omitting the time
recent work [4].1 index for simplicity, we assume an IMM estimator
Compared with a purely ballistic projectile [9] with two modes, say, mode 1 and mode 2. More
the IPP for thrusting projectiles is a significantly general cases can be handled in a similar manner. The
more challenging problem. When tracking a thrusting state of mode 1 has a higher dimension
projectile, there is an estimation ambiguity between · ¸
¢ xc
drag and thrust in the dynamic model: a sudden x1 = (2)
decrease in the thrust estimate causes a sudden xe
increase in the drag coefficient estimate and vice and the state of mode 2, which has lower dimension
versa. The unbiased mixing avoids the mixing of the (it consists of a subset of the components of x1 ), can
thrust estimate from the thrusting mode filter with be augmented as · ¸
zero from the BM filter (the latter assumes no thrust), 2¢
x
x = c (3)
which would bias the mixed thrust estimate towards 0
zero. This is the major difference between [9] and the
where xc denotes the common (c) components of
present work. The MIMM approach gives the best
states of the two modes and xe consists of the extra
choice of initial drag coefficient estimate, based on
(e) state components of mode 1 compared with that of
the characteristics of the thrusting projectile trajectory
mode 2.
observations.
The estimates of modes 1 and 2 are denoted as
The paper is organized in the following manner. · 1¸
We first give a brief overview of the standard IMM 1¢
x̂
x̂ = c1 (4)
estimator in Section II and the novel unbiased x̂e
mixing approach is introduced here to deal with
mixing the state estimates of modes with different with corresponding covariance matrix
· 1 ¸
dimensions. The dynamic and the measurement ¢ Pc Pce1
model and the corresponding discretized form are P1 = 1 (5)
Pec Pe1
presented in Section III. The IMM likelihood function
and · 2¸
used in selecting the best initial drag coefficient 2¢
x̂c
estimate from the MIMM estimator is presented x̂ = (6)
0
in Section IV. The IPP procedure is described in
Section V. The IMM estimator design and the filtering with corresponding covariance matrix
and prediction results for 30 real-data scenarios are · 2 ¸
¢ P 0
given in Section VI. Conclusions are presented in P2 = c (7)
0 0
Section VII.
respectively, where, according to the conventional
II. IMM ESTIMATOR WITH UNBIASED MIXING procedure [2], the lower dimension state is augmented
with zeros to make it of the same dimension as the
In order to estimate the target state during the
higher dimension state.
short observation interval as accurately as possible
We denote the mixing probabilities as
and then carry out the IPP, we use an approach
that relies on the IMM estimator with two different ¢
¹ijj = PfM i (k ¡ 1) j M j (k), Z k¡1 g, i, j 2 f1, 2g:
mode-matched filters for the TM and the BM.
Generally, in an IMM estimator with r modes that (8)
run in parallel, the state estimates at the beginning The standard mixing of the two estimates is
of a cycle are a mixture of estimates of the previous · 01 ¸ · 1 ¸ · 2¸
01 ¢
x̂c x̂c x̂c
cycle based on all the modes using the “mixing” x̂ = 01 = 1 ¹1j1 + ¹2j1
probabilities [1]. In the conventional IMM estimator, x̂e x̂e 0
" 1 #
mixing the estimates from different modes with x̂c ¹1j1 + x̂2c ¹2j1
unequal dimension state vectors causes a bias for the = (9)
x̂1e ¹1j1
extra components of the higher dimension state vector
[2]. It is necessary to introduce an unbiased approach and
· ¸ · ¸ · ¸
for the mixing, especially when the extra components ¢ x̂02
c x̂1c x̂2c
contain information of special interest, as in the case x̂02 = = ¹1j2 + ¹2j2
x̂02
e x̂1e 0
of high-thrust projectiles. " 1 #
x̂c ¹1j2 + x̂2c ¹2j2
= : (10)
1 The unbiased mixing work detailed here dates back to [10]. x̂1e ¹1j2
Our work on unbiased mixing and the work of [4] were carried
out independently, with the latter considering a more general Note that the last component of (5) is multiplied by
formulation. a factor ¹1j1 , less than one, while the last component

YUAN, ET AL.: A MIMM ESTIMATION APPROACH WITH UNBIASED MIXING FOR THRUSTING PROJECTILES 3251

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 of (10) is ignored by the lower dimension (mode 2) Note that the common components after the
filter. Thus, there is an estimate bias for the extra unbiased mixing have the same covariance matrix
component(s) in the filter matched to mode 1 at the as that in the conventional mixing but the extra
start of each IMM cycle. component (thrust) variance stays unchanged after the
The remedy for this is to replace (4) in the mixing unbiased mixing.
(5) by the following modification (M) A generalized form for the unbiased mixing based
· 2¸ on the best linear unbiased estimator (BLUE) is
2 ¢
x̂c discussed in Appendix II.
x̂M = 1 (11)
x̂e
III. THE SYSTEM MODEL AND ITS DISCRETIZED
as if it had the same extra component estimate (thrust FORM
in our case) as mode 1.
For the problem considered, the higher dimension
Then the modified estimate mixing replaces (5) by
state, i.e., the state vector for the thrusting mode, is
x̂01
¢ 1 2 denoted as
M = x̂ ¹1j1 + x̂M ¹2j1
" 1 # · ¸
_
x(t) = [x(t) y(t) z(t) x(t) _
y(t) z_ (t) ®(t) ¿ (t)]0
x̂c ¹1j1 + x̂2c ¹2j1 ¢ x̂01
c
= = 01 (12) (21)
x̂1e x̂Me
where ®(t) is the drag coefficient and ¿ (t) is the thrust.
i.e., the extra component(s) x̂01 01 The time arguments are omitted where there is no
Me of x̂M are exactly the
1
same as those of x̂ and, consequently, there is no bias ambiguity.
after the modified mixing. It is known that the drag coefficient varies
significantly with the Mach number regime: subsonic,
Next, the modification from (5) to (11) should
transonic, and supersonic. This is accounted for by a
also be accounted for in the covariance resulting from
Mach number-dependent multiplier.
the mixing. The corresponding covariance matrix for
The dynamic model for this problem can be
the modified mixing will be (the details are shown in written as follows
Appendix I) 2 3 2 3 2 3 2 3
(· " 1 #) ẍ x_ x_ 0
1 1¸ P̃ P̃ 1 6 7 ¿ 6_7 6 7 6 7
Pc Pce c Mce 4 ÿ 5 = 4 y 5 + ®®m D 4 y_ 5 + g 4 0 5 + º̃1
PM01 = + ¹1j1 V
Pec1 Pe1 1
P̃Mec 1
P̃Me z̈ z_ z_ ¡1
½· 2 ¸ · 2 ¸¾
Pc 0 P̃c 0 (22)
+ + ¹2j1 : (13) and
0 Pe1 0 0
®_ = º̃2 (23)
The blocks of PM01 are obtained as
¿_ = º̃3 (24)
01 ¢
PMc =(Pc1 + P̃c1 )¹1j1 + (Pc2 + P̃c2 )¹2j1 (14) where

¢
² the first term on the right side of (22) is the specific
01
PMe = Pe1 ¹1j1 + Pe1 ¹2j1 = Pe1 (15) thrust in the x, y, and z directions. For the ballistic
phase, the thrust is zero;
¢
01
PMce =(Pce1 + P̃Mce
1
)¹1j1 (16) ² the second term is the drag part, which is related to
velocity and altitude;
where, according to equation [1, eq. (1.4.16-9)], the ² V is the magnitude of the velocity v = [x_ y_ z_ ]0 , i.e.,
spread-of-the-means terms are the speed (in m/s);
² ® is the drag coefficient (in m2 =kg) at subsonic
01 0
P̃c1 = (x̂1c ¡ x̂01 1
c )(x̂c ¡ x̂c ) (17) speed and ¿ is the thrust (in m2 =s);
01 0 ² ®m is the (dimensionless) Mach number-dependent
P̃c2 = (x̂2c ¡ x̂01 2
c )(x̂c ¡ x̂c ) : (18) drag coefficient multiplier, which is approximated
Note that, in view of (12), one has by a cubic spline curve shown in Fig. 1;
² D = ¡½(z)V=2, where ½(z) = ½0 e¡cz is the air density
1 ¢
=(x̂1c ¡ x̂01 1 01 0 1 (in kg=m3 ) at altitude z (in m) and c is the air
P̃Mce c )(x̂e ¡ x̂Me ) = 0, P̃Mec =0
density constant (in m¡1 ) [9];
(19) ² g is the standard acceleration due to gravity at
¢ sea level, assumed to be the same throughout the
1 01 0
P̃Me =(x̂1e ¡ x̂01 1
Me )(x̂e ¡ x̂Me ) = 0: (20) trajectory, with value 9:812 m/s2 ;
² º̃1 , º̃2 , and º̃3 are assumed to be continuous-time
The mixed estimate x̂02
c and its covariance matrix Pc
02
zero-mean white Gaussian noises. The drag
are obtained in the standard manner, using x̂c , Pc , x̂2c ,
1 1
coefficient and thrust acceleration are thus modeled
and Pc2 with the corresponding mixing probabilities. as Wiener processes with slow variation [1].

3252 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 48, NO. 4 OCTOBER 2012

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1. Cubic spline approximation of Mach number-dependent drag coefficient multiplier (for a “sharp nose” projectile [11]).

Combining the dynamic equations (22)—(24), we continuous-time white process noise

have the following compact form T2
x(k + 1) = x(k) + f[k, x(k)]T + A(k)f[k, x(k)] + º(k)
_ = f[x(t)] + º̃(t)
x(t) (25) 2
(30)
where
2 3 where A(k) is the Jacobian of (26) evaluated at x(k)
_
x(t)
6 7 and º(k) is the discretized continuous time process
6 _
y(t) 7 noise for the sampling interval T. Based on the
6 7
6 _
z (t) 7 assumption that ® is nearly constant and D is related
6 7
6 _ 7 to both z and V, the detailed form of A is given in
6 x(t) 7
6 ¿ _
+ ®(t)D(t)x(t) 7 Appendix III.
6 V(t) 7
6 7 The corresponding covariance matrix of the
6
f(x(t)) = 6 _
y(t) 7 (26)
_ 7 discretized process noise is
6 ¿ (t) + ®(t)D(t)y(t) 7
6 V(t) 7 2 2 T3
6 7 T2 3 3
6 7 I3 I3
6 ¿ (t) z_ (t) + ®(t)D(t)z_ (t) ¡ g 7 66 3 2 7 7
6 V(t) 7 64 2 5 qv 06£1 06£1 7
6 7 6 T 7
6 7 Q=6 7
4 0 5 6 2 I3 TI3 7 (31)
6 7
0 4 01£6 Tq® 0 5
and 01£6 0 Tq¿
º̃(t) = [º̃1 (t)0 º̃2 (t) º̃3 (t)]0 : (27)
where I3 is the 3 £ 3 identity matrix and the
The state vector equation is discretized by a continuous time process noise “intensities” qv , q®
second-order Taylor expansion [5] as and q¿ are the corresponding power spectral densities
(PSDs).
T2
_
x(t + T) = x(t) + x(t)T + ẍ(t) + HOT (28) The measurement model in Cartesian
2 coordinates is
where T is the sampling time interval and HOT 2 3
x(k) + w1 (k)
denotes the higher order terms that are neglected. Note 6 7¢
that z(k) = 4 y(k) + w2 (k) 5 = Hx(k) + w(k) (32)
_
dx(t) df[x(t)] @f[x(t)] z(k) + w3 (k)
ẍ(t) = = = _
x(t): (29)
dt dt @x where H = [I3 0] and the components of the
Using the discrete time notation x(k) = x(t)jt=kT measurement noise vector w(k) are independent
and x(k + 1) = x(t + T)jt=kT , we have the zero-mean white Gaussian noises with the same
following expression for (28) with the discretized standard deviation (SD) ¾p .

YUAN, ET AL.: A MIMM ESTIMATION APPROACH WITH UNBIASED MIXING FOR THRUSTING PROJECTILES 3253

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 where ¤l (k) is the likelihood function of lth IMM
estimator, with r modes, at time k, given by [2]
r
¢X
¤l (k) = ¤il (k)¹il (k j k ¡ 1) (34)
i=1

where ¤il (k) is the likelihood function of mode i

of IMM estimator l at time k and ¹il (k j k ¡ 1) is
the predicted mode probability for mode i of IMM
estimator l. The mode likelihood is [1]

¤il (k) = p[z(k) j Mli (k), x̂0i 0i

l (k ¡ 1 j k ¡ 1), Pl (k ¡ 1 j k ¡ 1)]

= N [z(k); ẑil (k j k ¡ 1), Sli (k)] (35)

Fig. 2. MIMM approach.

where N [¢] is the Gaussian probability density
function, ẑil is the predicted measurement and Sli
is the innovation covariance in mode i of IMM
Assuming the position measurements are obtained
estimator l. The predicted mode probability can be
from intersection of line of sight (LOS) observations
written as
from two or more cameras (i.e., configuration III
multisensor data fusion [2]), the procedure of [7] can r
X
be used to obtain the exact position measurement ¹il (k j k ¡ 1) = PfMli (k) j Mlj (k ¡ 1), Z k¡1 g
noise covariance matrix. This requires the relative j=1

sensor-target geometry and the sensor accuracies (not £ PfMlj (k ¡ 1) j Z k¡1 g

available in the real-data examples considered in our
r
work). ¢X
= pji ¹jl (k ¡ 1) (36)
IV. THE MIMM APPROACH FOR SELECTION OF THE j=1
DRAG COEFFICIENT INITIAL ESTIMATE
where ¹jl (k ¡ 1) is the mode probability of mode j of
In the TM dynamic model, the drag parameter IMM estimator l at time k ¡ 1 and pji is the transition
(drag coefficient) and thrust parameter are separate probability from mode j to mode i over one time
state components as shown in (21). However, the
interval.
drag force and thrust force are acting simultaneously
Using (35) and (36), one can rewrite (34) as
(see (1)) and the IMM estimator has difficulty
( )
distinguishing between them if the initial uncertainty Xr X r
j
in the drag coefficient is large. A sudden decrease in ¤l (k) = N [z(k); ẑil (k j k ¡ 1), Sli (k)] pji ¹l (k ¡ 1) :
the drag coefficient estimate may trigger an increase j=1 j=1

in the thrust estimate. The sensitivity of the estimation (37)

to the initial drag coefficient estimate necessitates the
use of a MIMM estimator to overcome this “marginal
observability” problem [1].
The procedure starts by establishing a set of L
IMM estimators, each with an appropriate set of
modes (TM and BM in our case) to describe the
system behavior. Each IMM estimator is initialized
with a different value of the drag coefficient with
a suitable initial SD (generally, the initial SD is
taken equal to 25% of the initial estimate of the drag
coefficient). The filtering parameters are discussed in
Section VI.
In order to select the best initial drag coefficient
estimate, we need to determine the most likely one of
the L IMM estimators, during the observation period
(see Fig. 2). The likelihood function of IMM estimator
l for the time interval [k0 , K] is
K
Y (open-loop) EKF covariance equation. Then the 99%
¤kl 0 ,K = ¤l (k), l = 1, : : : , L
k=k0
Then
l¤ = arg max ¤kl 0 ,K = arg min[¡ ln ¤kl 0 ,K ] (38)
l l
selects the best IMM estimator. This yields the best
initial estimate of the drag coefficient.
V. IMPACT POINT PREDICTION
Based on the MIMM estimator, we can choose
at the end of the observation period the most
likely initial drag coefficient of the most probable
mode from the best IMM estimator (38). Then a
numerical IPP algorithm (we use the 4th-order
Runge-Kutta method [8]) is employed to predict
the trajectory down to its impact point. At the
same time the corresponding covariance is also
predicted to the impact point using a zero-gain
(33) chi-square probability region ellipse for the true
impact point, centered at the predicted impact point,

3254 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 48, NO. 4 OCTOBER 2012

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 is [1] tk is the sampling time and
· ¸
¡1
xip ¡ x̂pd yi = [ai0 ai1 ¢ ¢ ¢ ain ]0 , i = x, y, z: (43)
[xip ¡ x̂pd yip ¡ ŷpd ]Ppd = Â22 (99%)
xip ¡ ŷpd Assuming the projectile is moving with a nearly
(39) constant acceleration (NCA) (that is, n = 2) for the
where (xip yip ) is the true impact point, (x̂pd ŷpd ) is short time covered by N points, we get the estimate
the predicted impact point, Ppd is the corresponding for the parameter vector y (of dimension 3(n + 1) =
9) as
predicted covariance matrix, and Â22 (99%) denotes the 0 0
ŷ = [HyN (R N )¡1 HyN ]¡1 HyN (R N )¡1 zN (44)
99% point on the chi-square cumulative distribution
function with two degrees of freedom [1]. with the corresponding covariance matrix
We use as “warning zone” the region (39) 0
Py = [HyN (R N )¡1 HyN ]¡1 (45)
circumscribed by a circle (called the uncertainty
circle) where 32
· ¸ · ¸ z(1)
¸max 0 ¡1 xip ¡ x̂pd 6 . 7
[xip ¡ x̂pd yip ¡ ŷpd ] zN = 6 7
0 ¸max xip ¡ ŷpd 4 .. 5 (46)

= Â22 (99%) (40) z(N)

is the stacked vector of measurements (of dimension
where ¸max is the maximum eigenvalue of the
3N £ 1), 2 3
covariance matrix Ppd . The uncertainty circle (40) can Hy (1)
be used as the probability region for the true impact 6 . 7
point, centered at the predicted impact point. HyN = 6
4 .. 5
7 (47)

VI. REAL DATA RESULTS Hy (N)

The algorithm was evaluated on 30 cases of is the stacked measurement matrix (of dimension
field-collected data with the same sampling frequency 3N £ 9), and
30 Hz but different data lengths. 2 3
R(1) ¢ ¢ ¢ 0
6 . .. .. 7
A. MIMM Approach RN = 6
4 .. . 7
. 5 (48)
The MIMM estimator is chosen to consist of 0 ¢¢¢ R(N)
L = 4 IMM estimators with initial drag coefficient
ˆ
estimates ®(0): 0.18, 0.13, 0.065 and 0.035 (m2 =kg), is the block diagonal covariance matrix of
respectively. The initial SD of each drag coefficient measurement noise (of dimension 3N £ 3N) for the
estimate is 25% of the corresponding initial estimate fitting interval.
of the drag coefficient. Mapping the estimates (44) and the covariance
(45) to the time of the most recent measurement
B. N-Point Initialization and Measurement Noise [1], we get the initial estimates for position,
Variances velocity, and acceleration. Based on the initial drag
The initialization is crucial to the accuracy of the coefficient estimate and the acceleration estimates (last
state estimates. One-point or two-point initializations component of each direction, i.e., ai2 ), the thrust will
are practical approaches for many cases [1]; however, be initialized by calculating the acceleration in each
with a high sampling rate, as in the present case, direction (with the vertical one corrected by g).
an N-point initialization is feasible, reasonable, and Further, note that the goodness-of-fit error has the
of far preferable accuracy. The method is based following chi-square distribution [1]
on the polynomial fitting of a set of noisy scalar ¢
measurements [1]. JN¤ =[zN ¡ HpN ŷ]0 (R N )¡1 [zN ¡ HpN ŷ] » Â2Nnz ¡ny (49)
The N-point fitting of a polynomial of order n is
where nz = 3 is the dimension of z(i) and ny is the
done according to the model
dimension of y in our NCA model.
z(k) = Hy (k)y + w(k) The N-point initialization approach with N = 25 is
2 32 3 used (about 1 s data in our case) and the measurement
h(k)0 0 0 yx
¢6 76 7 noise SD is assumed time invariant with value 2 m
=4 0 h(k)0 0 5 4 yy 5 + w(k) (41) for each coordinate.2 Based on the initialization with
0 0 h(k) 0
yz the NCA model, the initial thrust estimate ¿ˆ (0) can
be calculated and the corresponding initial estimate’s
where k = 1, 2, : : : , N and SD can also be obtained. However, note that the NCA
· ¸
(t )n 0
h(k) = 1 tk ¢ ¢ ¢ k (42) 2 This
n! choice was based on evaluating (49).

YUAN, ET AL.: A MIMM ESTIMATION APPROACH WITH UNBIASED MIXING FOR THRUSTING PROJECTILES 3255

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 TABLE I and T = 1=30 s is the (fixed) sampling interval.
RMS Change Rate due to the Process Noise
Setting the MST as s1 = 2 s and s2 = 50 s for TM
and BM, respectively, we get the mode transition
Filter dv [(m/s)/s] d® [(m2 =kg)=s] d¿ [(m/s2 )=s]
probability matrix
IMM (TM) 7.5 ˆ
0:1®(0)/s 0:25¿ˆ (0)/s · ¸
IMM (BM) 4.3 ˆ
0:08®(0)/s N/A
0:9833 0:0167
¼= (58)
0:0007 0:9993
model is for thrust initialization only; each of these with initial mode probability vector [0:90 0:10].
IMM estimators, with different initial drag coefficient
estimates, starts from the beginning of measurements E. Real Data Results
with two-point initialization [1] for position and A sample estimated trajectory, indicating the
velocity. portions corresponding to the MIMM filtering and
to the IPP (including the uncertainty circle centered at
C. Selection of the Process Noise Intensities the IPP), is presented in Fig. 3. The estimated burnout
In the dynamic equations, we actually assume point (BoP) is also indicated.
a nearly constant velocity model for the kinematic The IPP performance of 30 real trajectories
component and a Wiener process with slow variation for short range, medium range, and long range are
for the drag coefficient and thrust. In order to satisfy presented Tables II—IV, respectively, and the following
this model’s assumptions, we need to choose small acronyms are used.
process noise “intensities” qv , q® , and q¿ in following
sense. # Trajectory ID, with indication of the
The process noise induced rms rate of change in projectile ranges S, M, L (short, medium,
velocity over a sampling interval T is long)
p Rprd Range prediction (m)
¢ qv T Ract Actual range (m) [This was the only
dv = [(m/s)/s]: (50)
T impact point information (ground truth)
For the drag coefficient the rate of change is when the data were recorded.]
p Rerr Range prediction error (m)
¢ q® T Rerr/R Ratio between range prediction error and
d® = [(m2 =kg)=s]: (51)
T actual range (prediction error percentage)
For the thrust the rate of change is Rad/R Ratio between the uncertainty radius and
p actual range (uncertainty circle percentage)
¢ q¿ T
d¿ = [(m/s2 )=s]: (52) A/B/C fPrediction error percentage less than
T 10%g/funcertainty circle percentage less
Then we have the following estimator design than 10%g/fboth A and Bg
procedure for selection of the process noise PSD M Major axis of uncertainty ellipse, also the
based on the corresponding rms rates of change radius of uncertainty circle (m)
m Minor axis of uncertainty ellipse (m)
qv = dv2 T (m2 =s3 ) (53)
DrgIni Initial drag coefficient estimate (m2 =kg)
q® = d®2 T (m4 =(kg2 ¢ s)) (54) DrgEnd Drag coefficient estimate at the end of the
observation interval (m2 =kg)
q¿ = d¿2 T (m2 =s5 ): (55)
ThrIni Initial thrust estimate (m=s2 )
The process noise intensities are chosen based on Vmax Maximum speed (velocity magnitude)
process noise induced rms change (in velocity/drag estimate (m=s) and the sampling point
coefficient/thrust) over an interval of 1 s shown in where it happened
Table I. BoP Burnout point–based on the best IMM’s
TM probability
D. IMM Mode Transition Probability Matrix TmP/TtP Total measurement points/total trajectory
The transition probability matrix for the two-mode (measurement+prediction) points (and
IMM is · ¸ percentage).
p11 1 ¡ p11
¼= (56)
1 ¡ p22 p22 The BoP is decided based on the best IMM’s
mode probabilities. The point from which the mode
where the elements of the matrix are obtained based
probability of TM stays lower than 0.03 for 10
on the mean sojourn time (MST) [1], s1 and s2 , for
consecutive sample points (i.e., BM dominates) is
the TM (i = 1) and the BM (i = 2), respectively,
recognized as the BoP. After the recognized BoP plus
T the 10 sample points for deciding the BoP, we stop
pii = 1 ¡ , i = 1, 2 (57)
si using TM and start using an additional BM, say BM1,

3256 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 48, NO. 4 OCTOBER 2012

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 Fig. 3. Sample estimated trajectory (trajectory 28).

TABLE II
IPP Performance for 5 Real Trajectories (Short Range)

# Rprd Ract Rerr Rerr/R Rad/R A/B/C M m DrgIni DrgEnd ThrIni Vmax BoP TmP/TtP

(S)06 7203 7541 ¡338 4.48 3.67 T/T/F 277 146 0.13 0.093 378 492(38) 41 150/883(17.0%)
(S)22 5092 5937 ¡845 14.23 4.6 F/T/F 273 227 0.13 0.194 274 394(39) 45 150/1235(12.1%)
(S)25 6650 6642 8 0.12 6.91 T/T/T 460 290 0.18 0.113 302 408(35) 40 139/1358(10.2%)
(S)29 6333 6747 ¡414 6.14 6.27 T/T/T 424 286 0.18 0.119 335 412(33) 39 150/1355(11.1%)
(S)30 4811 4884 ¡73 1.49 9.9 T/T/T 484 374 0.13 0.125 293 406(36) 40 117/1567(7.5%)

TABLE III
IPP Performance for 15 Real Trajectories (Medium Range)

# Rprd Ract Rerr Rerr/R Rad/R A/B/C M m DrgIni DrgEnd ThrIni Vmax BoP TmP/TtP

(M)01 6544 8932 ¡2388 26.74 4.43 F/T/F 396 245 0.13 0.118 325 458(41) 44 150/1233(12.2%)
(M)05 8729 9608 ¡879 9.15 4.46 T/T/F 429 286 0.065 0.088 479 552(27) 28 150/1347(11.1%)
(M)07 7934 9041 ¡1107 12.24 7.24 F/T/F 655 274 0.13 0.066 135 361(59) 62 149/1252(11.9%)
(M)08 8540 9041 ¡501 5.54 7.32 T/T/T 663 267 0.13 0.092 283 497(34) 37 101/1265(8.0%)
(M)09 8872 9020 ¡148 1.64 4.75 T/T/T 429 239 0.065 0.086 460 550(40) 43 138/1200(11.5%)
(M)10 8669 10067 ¡1398 13.89 5.12 F/T/F 516 242 0.065 0.091 342 515(39) 42 113/1203(9.4%)
(M)11 7829 9173 ¡1344 14.65 4.31 F/T/F 396 228 0.18 0.119 331 504(43) 47 138/1206(11.4%)
(M)12 8873 9023 ¡150 1.66 4.67 T/T/T 421 277 0.065 0.089 394 485(32) 35 150/1329(11.3%)
(M)13 7869 8133 ¡264 3.25 5.89 T/T/T 479 292 0.13 0.093 322 408(37) 43 149/1358(11.0%)
(M)14 9338 9338 0 0 5.09 T/T/T 476 305 0.065 0.08 413 522(33) 35 150/1399(10.7%)
(M)15 10336 8696 1640 18.86 8.36 F/T/F 728 337 0.13 0.065 364 503(38) 40 126/1453(8.7%)
(M)19 10412 9671 741 7.66 9.54 T/T/T 923 368 0.13 0.065 317 492(39) 43 110/1523(7.2%)
(M)20 8666 8258 408 4.94 5.77 T/T/T 477 316 0.065 0.1 394 521(34) 37 127/1448(8.8%)
(M)21 8150 8177 ¡27 0.33 5.47 T/T/T 448 257 0.13 0.083 310 438(41) 46 178/1261(14.1%)
(M)27 10070 9873 197 2 6.68 T/T/T 660 320 0.18 0.068 410 540(39) 41 150/1423(10.5%)

with a larger process noise setting to cover possibly of the BM. The mode transition probability matrix
higher uncertainties (projectiles may experience between BM and BM1 is
tumbling or high measurement uncertainties due to · ¸
0:7 0:3
change of the relative sensor-target geometry). The ¼= (59)
BM1 has the process noise induced rms change rate 0:3 0:7
in velocity as 2dv and the same other settings as those with initial mode probability vector [0:90 0:10].

YUAN, ET AL.: A MIMM ESTIMATION APPROACH WITH UNBIASED MIXING FOR THRUSTING PROJECTILES 3257

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 Fig. 4. Evolution of drag coefficient estimates in the MIMM (trajectory 12).

Fig. 5. Evolution of drag coefficient estimates in MIMM (trajectory 28).

TABLE IV
IPP Performance for 10 Real Trajectories (Long Range)

# Rprd Ract Rerr Rerr/R Rad/R A/B/C M m DrgIni DrgEnd ThrIni Vmax BoP TmP/TtP

(L)02 22765 20813 1952 9.38 8.89 T/T/F 1851 718 0.035 0.033 276 664(68) 73 139/2385(5.8%)
(L)03 14708 14585 123 0.84 4.43 T/T/T 646 299 0.065 0.045 390 670(48) 50 152/1398(10.9%)
(L)04 16849 13409 3440 25.65 10.05 F/F/F 1348 370 0.18 0.033 388 619(63) 69 144/1566(9.2%)
(L)16 16220 15740 480 3.05 5.16 T/T/T 813 326 0.035 0.036 342 669(51) 54 140/1458(9.6%)
(L)17 17136 16405 731 4.46 7.26 T/T/T 1191 461 0.035 0.044 350 616(50) 52 114/1805(6.3%)
(L)18 16474 17742 ¡1268 7.15 7.18 T/T/T 1274 455 0.065 0.05 393 688(46) 50 109/1804(6.0%)
(L)23 17494 17076 418 2.45 6.21 T/T/T 1061 453 0.035 0.039 331 671(31) 34 113/1768(6.4%)
(L)24 16031 16196 ¡165 1.02 6.32 T/T/T 1025 385 0.035 0.047 325 656(45) 50 113/1620(7.0%)
(L)26 16050 16871 ¡821 4.87 8.57 T/T/T 1447 416 0.13 0.046 329 616(63) 81 138/1705(8.1%)
(L)28 13624 13743 ¡119 0.87 3.46 T/T/T 476 235 0.035 0.041 345 631(36) 38 155/1198(12.9%)

3258 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 48, NO. 4 OCTOBER 2012

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 Fig. 6. Drag coefficient and thrust estimates in best IMM estimator (trajectory 12). Note that TM is discontinued at k = 45.

Fig. 7. Drag coefficient and thrust estimates in best IMM estimator (trajectory 28). Note that TM is discontinued at k = 48.

The evolution of the drag coefficient estimates in
the MIMM is shown in Figs. 4 and 5 (for trajectories
12 and 28, respectively). It can be seen that in the
MIMM estimator the drag coefficient estimates
of different IMM estimators converge toward the
end of the observation period. Based on (38), the
best IMM estimator (with the highest likelihood) is
selected.
The drag coefficient and thrust estimates in the
best IMM estimator selected using (38) are shown in
Figs. 6 and 7 (for trajectories 12 and 28, respectively).
The ambiguity between the drag coefficient and
thrust can be observed. A sudden decrease in the
drag coefficient estimate (as in Fig. 7 at time k = 17)
triggers an increase in the thrust estimate and vice
versa.
The estimation ambiguity between the drag
coefficient and thrust makes the unbiased mixing
strategy a key to correctly estimating the thrust
component in the IMM estimator. To illustrate this

YUAN, ET AL.: A MIMM ESTIMATION APPROACH WITH UNBIASED MIXING FOR THRUSTING PROJECTILES 3259

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 Fig. 8. Drag coefficient and thrust estimates (in best IMM estimator, trajectory 21) with conventional mixing.

Fig. 9. Drag coefficient and thrust estimates (in the best IMM estimator, trajectory 21) with unbiased mixing.

ambiguity in more detail, consider a comparison
between the IMM with the conventional (biased)
mixing and with the unbiased mixing. Figures 8 and 9
show the drag coefficient and thrust estimates obtained
with the conventional mixing and the unbiased mixing
for trajectory 21, respectively. Figure 8 shows that,
because of the biased estimates of the thrust, the drag
coefficient estimates try to go negative to compensate
for the thrust estimate that became practically zero
(the algorithm implementation forced them to stay
nonnegative). Note that in the IMM with conventional
mixing, the TM is discontinued prematurely at k = 30
(about 1 s after target acquisition) because the thrust
estimate is zero. This is “compensated” by a negative
drag coefficient estimate, which was, however, forced
to be nonnegative. In Fig. 9 it can be seen that TM
is discontinued at k = 56 (about 2 s after target
acquisition). This was based on the TM probability

3260 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 48, NO. 4 OCTOBER 2012

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 TABLE V
IPP Performance Evaluation Summary

Criterion Description % of Cases

Novel Unbiased Mixing MIMM Conventional Mixing Single IMM

A The prediction error percentage is less 23/30 (77%) 19/30 (63%)

than 10%
B The uncertainty circle percentage is less 29/30 (97%) 08/30 (27%)
than 10%
C True IP is in the uncertainty circle 20/30 (67%) 23/30 (77%)
centered at IPP
B&C IPP · 10% and true IP is in the 20/30 (67%) 04/30 (13%)
uncertainty circle centered at IPP

being less than 0.03 for 10 consecutive samples. The VII. CONCLUSIONS
drag coefficient estimates stabilize around 0:09 m2 =kg. Facing the challenges of very short time span of
Thus, the unbiased mixing can be seen to avoid the observations (about 10% of the whole trajectory)
tendency of the drag coefficient estimate to become on thrusting projectiles, the MIMM estimator with
negative to compensate for the thrust estimate that
different initial drag coefficient estimates is presented
becomes biased towards zero.
to alleviate the ambiguity in the estimation of the drag
The IPP performance is evaluated by three criteria
coefficient and thrust and to obtain more accurate
as shown in Table V. A single IMM with conventional
estimates of the drag coefficient and thrust. A novel
mixing using one initial drag coefficient estimate
unbiased mixing approach for an IMM estimator
0:1 m2 =kg with a large initial SD of 0:08 m2 =kg is
with modes that have unequal dimension state vector
employed as a performance comparison baseline. It
is introduced and is an important extension of the
can be seen that compared with a single IMM, the
conventional mixing in an IMM estimator. This gives
MIMM approach with the novel unbiased mixing
an unbiased estimate of the thrust component and
gives a significantly improved estimation performance,
leads to significant improvement for the final IPP
with smaller IPP errors and smaller uncertainties.
performance. The best IPP performance was achieved
If LOS measurement accuracy and the
sensor-target geometry would have been available, by employing the MIMM estimator with the novel
the use of the technique of [7] to obtain R(k) as a unbiased mixing.
time-varying covariance matrix would have allowed
APPENDIX I. COVARIANCE MATRIX OF HIGHER
overall better tracking and IPP performance. Another
DIMENSIONAL MODE IN UNBIASED MIXING
difficulty in working with the available real data
is the short span of the observation (5—15% of the The covariance matrix of mode 1 of the unbiased
trajectory) and the possibility of ill-conditioning of mixing is (M d (k), d = 1, 2, is the possible mode of
the LOS observations. mode-matched filter d at time k)

¢ 01 0
PM01 = E((x ¡ x̂01 1
M )(x ¡ x̂M ) j M (k), Z
k¡1
)
01 0
= E[(x ¡ x̂1 + x̂1 ¡ x̂01 1 1 1 1
M )(x ¡ x̂ + x̂ ¡ x̂M ) j M (k ¡ 1), M (k), Z
k¡1
]¹1j1
01 0
+ E[(x ¡ x̂2M + x̂2M ¡ x̂01 2 2 2 1
M )(x ¡ x̂M + x̂M ¡ x̂M ) j M (k ¡ 1), M (k), Z
k¡1
]¹2j1
= fE[(x ¡ x̂1 )(x ¡ x̂1 )0 j M 1 (k ¡ 1), M 1 (k), Z k¡1 ] + (x̂1 ¡ x̂01 1 01 0
M )(x̂ ¡ x̂M ) g¹1j1

fE[(x ¡ x̂2M )(x ¡ x̂2M )0 j M 2 (k ¡ 1), M 1 (k), Z k¡1 ] + (x̂2M ¡ x̂01 2 01 0

M )(x̂M ¡ x̂M ) g¹2j1
(· ¸ μ· 1 ¸ · 01 ¸¶ μ· 1 ¸ · 01 ¸¶0 )
Pc1 Pce1 x̂c x̂c x̂c x̂c
= 1 1
+ 1
¡ 01 1
¡ 01 ¹1j1
Pec Pe x̂e x̂Me x̂e x̂Me
( "μ· ¸ · ¸¶ μ· ¸ · ¸¶0 # μ· ¸ · ¸¶ μ· 2 ¸ · 01 ¸¶0 )
xc x̂2c xc x̂2c x̂2c x̂01
c x̂c x̂c
+ E ¡ 1 ¡ 1 + 1
¡ 01 1
¡ 01 ¹2j1
xe x̂e xe x̂e x̂e x̂Me x̂e x̂Me
(· ¸ " 1 1
#) ½· 2 ¸ · 2 ¸¾
¢ Pc1 Pce1 P̃c P̃Mce Pc Pce21 P̃c 0
= + ¹1j1 + + ¹2j1 (60)
Pec1 Pe1 1
P̃Mec 1
P̃Me Pec12 Pe1 0 0

YUAN, ET AL.: A MIMM ESTIMATION APPROACH WITH UNBIASED MIXING FOR THRUSTING PROJECTILES 3261

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 where Pce21 = 0 and Pec12 = 0, since the estimates and for the mode 2 filter it is
of different components from different modes are
assumed to be uncorrelated. Then (13) is obtained. x̂02 = x̂1 ¹1j2 + x̂2 ¹2j2 (68)
2 3 2 3
APPENDIX II. BLUE FOR UNBIASED MIXING Ge1 Ga1 0 0 0 0
6 7 6 7
Let us consider a generalized form for the =6 1
4 Gb Gc1 07 6 2
5 x̂¹1j2 + 4 0 Gc Gb2 7
5 x̂¹2j2
unbiased mixing based on the BLUE. Assuming a 0 0 0 0 Ga2 Ge2
“base” state vector as
2 13 (69)
xe 2 3
¢6 7 (Ge1 x̂1e + Ga1 x̂c )¹1j2
x = 4 xc 5 (61) 6 7
x2e =6 1 1 1 2 2 2 7
4 (Gb x̂e + Gc x̂c )¹1j2 + (Gb x̂e + Gc x̂c )¹2j2 5
and the state vectors of the IMM mode 1 and mode 2 Ge2 x̂2e + (Ga2 x̂c )¹2j2
as, respectively, (70)
2 1 3 2 1 1 3 2 3
Ge Ga1 0 Ge xe + Ga1 xc x̂02
¤
6 7 6 7 ¢ 6 02 7
x1 = 4 Gb1 Gc1 0 5 x = 4 Gb1 x1e + Gc1 xc 5 (62) =6 7
4 x̂c 5 : (71)
0 0 0 0 x̂02
e
2 3 2 3
0 0 0 0 Note that the last n components of (66) are ignored by
2 6 7 6 2 2 7 the mode 1 filter while the last n components of (70)
x = 4 0 Gc2 Gb2 5 x = 4 Gb xe + Gc2 xc 5 (63)
are multiplied by a factor ¹2j2 less than one; the first
0 Ga2 Ge2 Ge2 x2e + Ga2 xc
m components of (70) are ignored by the mode 2 filter
where, without loss of generality, the diagonal while the first m components of (66) are multiplied
non-zero matrices Ge1 , Ge2 , and Gci , i = 1, 2 are m £ m, by a factor ¹1j1 less than one. This would lead to
n £ n, and p £ p invertible matrices, respectively.3 estimation bias for the components x1e and x2e in each
Then the state of mode 1 is a linear combination of IMM cycle.
certain components of the state of interest x; the state The remedy is replacing estimates of mode 1 and
of mode 2 is a linear combination of certain another mode 2 as follows
components of the state of interest x. At the same time 2 1 1 3
Ge x̂e + Ga1 x̂c
each mode may have certain extra components that the 6 7
other does not have (x1e or x2e ) and also shares certain x̂1M = 6 1 1 1 7
4 Gb x̂e + Gc x̂c 5 (72)
common components both have (xc ). Ge2 x̂2e + Ga2 x̂c
The standard mixing of the estimates for the mode
1 filter is and 2 3
Ge1 x̂1e + Ga1 x̂c
x̂01 = x̂1 ¹1j1 + x̂2 ¹2j1 (64) 6 7
x̂2M = 6 2 2 2 7
4 Gb x̂e + Gc x̂c 5 : (73)
2 1 3 2 3
Ge Ga1 0 0 0 0
6 1 7 6 7 Ge2 x̂2e + Ga2 x̂c
= 4 Gb Gc 0 5 x̂¹1j1 + 4 0 Gc2
1
Gb2 5 x̂¹2j1
Then the modified estimate mixing for the mode 1
0 0 0 0 Ga2 Ge2 filter and the mode 2 filter are
(65)
2 3 x̂01 1 2
M = x̂M ¹1j1 + x̂M ¹2j1 (74)
(Ge1 x̂1e + Ga1 x̂c )¹1j1
6 7 2 3
= 4 (Gb1 x̂1e + Gc1 x̂c )¹1j1 + (Gb2 x̂2e + Gc2 x̂c )¹2j1 5 (Ge1 x̂1e + Ga1 x̂c )¹1j1 + (Ge1 x̂1e + Ga1 x̂c )¹2j1
6 7
(Ge2 x̂2e + Ga2 x̂c )¹2j1 =6 1 1 1 2 2 2 7
4 (Gb x̂e + Gc x̂c )¹1j1 + (Gb x̂e + Gc x̂c )¹2j1 5
(66)
2 3 (Ge2 x̂2e + Ga2 x̂c )¹1j1 + (Ge2 x̂2e + Ga2 x̂c )¹2j1
x̂01
e
¢6 7 (75)
= 4 x̂01
c 5 (67)
2 3
x̂01 Ge1 x̂1e + Ga1 x̂c
¤
6 1 1 7
=6 1 2 2 2 7
4 (Gb x̂e + Gc x̂c )¹1j1 + (Gb x̂e + Gc x̂c )¹2j1 5
3 The invertibility of the matrices guarantee the 1:1 mapping

between the local state vector and the corresponding components

Ge2 x̂2e + Ga2 x̂c
of the “base” state vector. (76)

3262 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 48, NO. 4 OCTOBER 2012

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 and where “*” implies the elements ignored by mode 2
filter no matter what they are; the spread-of-the-means
x̂02
M = x̂1M ¹1j2 + x̂2M ¹2j2 (77)
2 3 terms are
(Ge1 x̂1e + Ga1 x̂c )¹1j2 + (Ge1 x̂1e + Ga1 x̂c )¹2j2 02 0
6 7 P̃c1,02 = (x̂1c ¡ x̂02 1
c )(x̂c ¡ x̂c ) (86)
=6 1 1 1 2 2 2 7
4 (Gb x̂e + Gc x̂c )¹1j2 + (Gb x̂e + Gc x̂c )¹2j2 5 02 0
P̃c2,02 = (x̂2c ¡ x̂02 2
c )(x̂c ¡ x̂c ) : (87)
(Ge2 x̂2e + Ga2 x̂c )¹1j2 + (Ge2 x̂2e + Ga2 x̂c )¹2j2
(78) In most situations we are dealing with the
2 3 following simple case (with x as the state of interest):
Ge1 x̂1e + Ga1 x̂c
2 3 2 13
6 7 Im 0 0 xe
= 4 (Gb1 x̂1e + Gc1 x̂c )¹1j2 + (Gb2 x̂2e + Gc2 x̂c )¹2j2 5 :
1 6 7 6 7
x = 4 0 Ip 0 5 x = 4 xc 5 (88)
Ge2 x̂2e + Ga2 x̂c
0 0 0 0
(79) 2 3 2 3
0 0 0 0
Given the augmented covariance matrices of (62) 2 6 7 6 7
and (63) as x = 4 0 Ip 0 5 x = 4 xc 5 (89)
2 1 3
Pe Pec1 0 0 0 In x2e
¢6 7
P 1 = 4 Pce1 Pc1 0 5 (80)
where Ir , r = m, n, p, are the r £ r identity matrices.
0 0 0 Further setting In in (89) as an all-zero matrix and
and 2 3 eliminating the corresponding rows and columns, we
0 0 0 get the special case discussed in Section II with the
¢6 7
P 2 = 4 0 Pc2 Pce2 5 (81) state of interest as
· 1¸
0 Pec2 Pe2 xe
x= :
the corresponding covariance matrices for the xc
unbiased mixing can be obtained accordingly as
follows. APPENDIX III. THE JACOBIAN MATRIX FOR THE
For mode 1 filter the corresponding augmented THRUSTING MODE
covariance matrix is
82 1 1 3 2 39 The Jacobian A for the TM, under the assumption
> Pe Pec ¤
<
0 0 ¤ >
=
6 7 6 7 that ® (under subsonic speed) and ¿ are constant, can
PM01 = 4 Pce1 Pc1 ¤ 5 + 4 0 P̃c1,01 ¤ 5 ¹1j1 be shown as (the time arguments are omitted)
>
: >
;
¤ ¤ ¤ ¤ ¤ ¤ 2 3
82 1 3 2 39 0 0 0 1 0 0 0 0
P 0 ¤ 0 0 ¤ > 60 0 0
>
< e
6 7 6 7
= 6 0 1 0 0 0 7
7
+ 4 0 Pc2 ¤ 5 + 4 0 P̃c2,01 ¤ 5 ¹2j1 6 7
> > 60 0 0 0 0 1 0 0 7
: ; 6 7
¤ ¤ ¤ ¤ ¤ ¤ 6 7
6 0 0 a43 a44 a45 a46 a47 a48 7
A=6 6 7 (90)
(82) 7
6 0 0 a53 a54 a55 a56 a57 a58 7
where “*” implies the elements ignored by mode 1 6 7
60 0 a 7
filter no matter what they are; the spread-of-the-means 6 63 a64 a65 a66 a67 a68 7
6 7
terms are 40 0 0 0 0 0 0 0 5
01 0
P̃c1,01 = (x̂1c ¡ x̂01 1
c )(x̂c ¡ x̂c ) (83) 0 0 0 0 0 0 0 0
01 0
P̃c2,01 = (x̂2c ¡ x̂01 2
c )(x̂c ¡ x̂c ) : (84) where, note that D depends on both altitude z and
velocity v,
Similarly, for mode 2 filter, the corresponding
augmented covariance matrix is @D
a43 = ®®m x_ (91)
82 3 2 39 @z
> ¤ ¤ ¤ ¤ ¤ ¤ >
< =
6 7 6 7 V2 ¡ x_ 2 @D
PM02 = 4 ¤ Pc1 0 5 + 4 ¤ P̃c1,02 0 5 ¹1j2 a44 = ¿ + ®®m D + ®®m x_ (92)
>
: >
; V3 @ x_
¤ 0 Pe2 ¤ 0 0
82 3 2 39 ¡y_ @D
¤ ¤ ¤ ¤ ¤ ¤ > a45 = ¿ x_ 3 + ®®m x_ (93)
>
< = V @ y_
6 7 6 7
+ 4 ¤ Pc2 Pce2 5 + 4 ¤ P̃c2,02 0 5 ¹2j2 ¡z_ @D
>
: >
; a46 = ¿ x_ + ®®m x_ (94)
¤ Pec2 Pe2 ¤ 0 0 V3 @ z_
(85) a47 = Dx_ (95)

YUAN, ET AL.: A MIMM ESTIMATION APPROACH WITH UNBIASED MIXING FOR THRUSTING PROJECTILES 3263

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 x_ REFERENCES
a48 = (96)
V [1] Bar-Shalom, Y., Li, X. R., and Kirubarajan, T.
@D Estimation with Applications to Tracking and Navigation:
a53 = ®®m y_ (97) Algorithms and Software for Information Extraction.
@z Hoboken, NJ: Wiley, 2001.
¡x_ @D [2] Bar-Shalom, Y., Willett, P. K., and Tian, X.
a54 = ¿ y_ 3 + ®®m y_ (98) Tracking and Data Fusion.
V @ x_
Storrs, CT: YBS Publishing, 2011.
2 2
V ¡ y_ @D [3] Li, X. R. and Jilkov, V. P.
a55 = ¿ + ®®m D + ®®m y_ (99) Survey of maneuvering target tracking. Part II: Motion
V3 @ y_
models of ballistic and space targets.
¡z_ @D IEEE Transactions on Aerospace and Electronic Systems,
a56 = ¿ y_ 3 + ®®m y_ (100) 46, 1 (Feb. 2010), 96—119.
V @ z_
[4] Lopez, R., Danés, P., and Royer, F.
a57 = D y _ (101) Extending the IMM filter to heterogeneous-order state
space models.
y_
a58 = (102) Proceedings of the 49th IEEE Conference on Decision and
V Control, Atlanta GA, Dec. 2010, pp. 7369—7374.
@D [5] Mehra, R. K.
a63 = ®®m z_ (103) A comparison of several nonlinear filters for reentry
@z vehicle tracking.
¡x_ @D IEEE Transactions on Automatic Control, AC-16, 4 (Aug.
a64 = ¿ z_ 3 + ®®m z_ (104) 1971), 307—319.
V @ x_
[6] Osborne, III, R. W., Bar-Shalom, Y., and Kirubarajan, T.
¡y_ @D Radar measurement noise variance estimation with several
a65 = ¿ z_ 3 + ®®m z_ (105)
V @ y_ targets of opportunity.
IEEE Transactions on Aerospace and Electronic Systems,
V2 ¡ z_ 2 @D 44, 3 (July 2008), 985—995.
a66 = ¿ + ®®m D + ®®m z_ (106)
V3 @ z_ [7] Osborne, III, R. W. and Bar-Shalom, Y.
Statistical efficiency of composite measurements from
a67 = D z_ (107) LOS observation.
z_ Proceedings of SPIE, Signal Processing, Sensor Fusion
a68 = (108) and Target Recognition, vol. 8050-07, Orlando FL, Apr.
V 2011.
with [8] Press, W. H., et al.
Numerical Recipes in FORTRAN: The Art of Scientific
@D c½(z)V Computing (2nd ed.).
= = ¡cD (109)
@z 2 New York: Cambridge, 1992.
@D ½(z)x_ x_ [9] Ravindra, V., Bar-Shalom, Y., and Willett, P. K.
=¡ =D 2 (110) Projectile identification and impact point prediction.
@ x_ 2V V IEEE Transactions on Aerospace and Electronic Systems,
@D ½(z)y_ y_ 46, 4 (Oct. 2010), 2004—2021.
=¡ =D 2 (111) [10] Yuan, T., et al.
@ y_ 2V V Point prediction for short range thrusting projectiles.
@D ½(z)z_ z_ Proceedings of SPIE, Signal and Data Processing of Small
=¡ =D 2: (112) Targets, vol. 7698-55, Orlando FL, Apr. 2010.
@ z_ 2V V
[11] http://www.braeunig.us/space/cd.htm.
The Jacobian for the BM can be easily obtained
by setting ¿ to zero in the above expressions and
eliminating the last row and column.

3264 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 48, NO. 4 OCTOBER 2012

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 Ting Yuan received the B.S. and M.S. degrees both in electrical and engineering
from University of Electronic Science and Technology of China. He is currently
pursuing the Ph.D. degree in electrical and computer engineering, University of
Connecticut, Storrs.
His research interests include statistical signal processing and multi-sensor
track/information fusion.

YUAN, ET AL.: A MIMM ESTIMATION APPROACH WITH UNBIASED MIXING FOR THRUSTING PROJECTILES 3265

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 Yaakov Bar-Shalom (S’63–M’66–SM’80–F’84) was born on May 11, 1941.
He received the B.S. and M.S. degrees from the Technion, Israel Institute of
Technology, in 1963 and 1967 and the Ph.D. degree from Princeton University,
Princeton, NJ, in 1970, all in electrical engineering.
From 1970 to 1976 he was with Systems Control, Inc., Palo Alto, California.
Currently he is Board of Trustees Distinguished Professor in the Department
of Electrical and Computer Engineering, the Marianne E. Klewin Professor in
Engineering, and the director of the ESP (Estimation and Signal Processing) Lab
at the University of Connecticut.
His current research interests are in estimation theory, target tracking and data
fusion. He has published over 400 papers and book chapters in these areas and
in stochastic adaptive control. He coauthored the monograph Tracking and Data
Association (Academic Press, 1988), the graduate texts Estimation and Tracking:
Principles, Techniques and Software (Artech House, 1993), Estimation with
Applications to Tracking and Navigation: Algorithms and Software for Information
Extraction (Wiley, 2001), the advanced graduate texts, Multitarget-Multisensor
Tracking: Principles and Techniques (YBS Publishing, 1995), Tracking and Data
Fusion (YBS Publishing, 2011), and edited the books Multitarget-Multisensor
Tracking: Applications and Advances (Artech House, Vol. I, 1990; Vol. II, 1992;
Vol. III, 2000).
He has been elected Fellow of IEEE for “contributions to the theory of
stochastic systems and of multi-target tracking.” He has been consulting to
numerous companies and government agencies, and originated the series of
Multitarget-Multisensor Tracking short courses offered via UCLA Extension, at
Government Laboratories, private companies and overseas.
During 1976 and 1977 he served as Associate Editor of the IEEE Transactions
on Automatic Control and from 1978 to 1981 as Associate Editor of Automatica.
He was Program Chairman of the 1982 American Control Conference, General
Chairman of the 1985 ACC, and Cochairman of the 1989 IEEE International
Conference on Control and Applications. During 1983—1987 he served as
Chairman of the Conference Activities Board of the IEEE Control Systems
Society and during 1987—1989 was a Member of the Board of Governors of
the IEEE CSS. He was a Member of the Board of Directors of the International
Society of Information Fusion (1999—2004) and served as General Chairman
of FUSION 2000, President of ISIF in 2000 and 2002 and Vice President for
publications in 2004—2011.
In 1987 he received the IEEE CSS Distinguished Member Award. Since
1995 he is a Distinguished Lecturer of the IEEE AESS and has given numerous
keynote addresses at major national and international conferences. He is
corecipient of the M. Barry Carlton Award for the best paper in the IEEE
Transactions on Aerospace and Electronic Systems in 1995 and 2000 and recipient
of the 1998 University of Connecticut AAUP Excellence Award for Research. In
2002 he received the J. Mignona Data Fusion Award from the DoD JDL Data
Fusion Group. He is a Member of the Connecticut Academy of Science and
Engineering. In 2008 he was awarded the IEEE Dennis J. Picard Medal for Radar
Technologies and Applications.

3266 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 48, NO. 4 OCTOBER 2012

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.
 Peter Willett (F’03) received his B.A.Sc. (engineering science) from the
University of Toronto in 1982, and his Ph.D. degree from Princeton University
in 1986.
He has been a faculty member at the University of Connecticut since
1986, and since 1998 has been a professor. His primary areas of research
have been statistical signal processing, detection, machine learning, data
fusion, and tracking. He has also interests in and has published in the areas of
change/abnormality detection, optical pattern recognition, communications, and
industrial/security condition monitoring.
Dr. Willett is Editor-in-Chief of IEEE Transactions on Aerospace and
Electronic Systems, and until recently was associate editor for three active
journals–IEEE Transactions on Aerospace and Electronic Systems (for Data
Fusion and Target Tracking) and IEEE Transactions on Systems, Man, and
Cybernetics, parts A and B. He is also associate editor for the IEEE AES
Magazine, associate editor for ISIF’s electronic Journal of Advances in Information
Fusion, is a member of the editorial board of IEEE’s Signal Processing Magazine
and was first editor of the AES Magazine’s periodic Tutorial issues. He was a
member of the IEEE AESS Board of Governors 2003—2009. He was General
Cochair (with Stefano Coraluppi) for the 2006 ISIF/IEEE Fusion Conference
in Florence, Italy and for the 2008 ISIF/IEEE Fusion Conference in Cologne,
Germany, Program Cochair (with Eugene Santos) for the 2003 IEEE Conference
on Systems, Man & Cybernetics in Washington DC, and Program Cochair (with
Pramod Varshney) for the 1999 Fusion Conference in Sunnyvale.

David F. Hardiman obtained a B.S. in electrical engineering from University of

Alabama in Huntsville, M.S. in electrical engineering from Georgia Institute of
Technology, and an MBA from Vanderbilt University.
He is an Electrical & Computer Engineer, U.S. Army Aviation and
Missile Research, Development, and Engineering Center (AMRDEC), and
has over 23 years of technical and management experience in engineering,
research, and development. His areas of expertise include project management;
hardware/software design and development for real-time embedded computer
systems; digital signal processing; image processing; algorithm optimization; and
estimation/detection theory.

Mozeson and Pollack biographies are not available.

YUAN, ET AL.: A MIMM ESTIMATION APPROACH WITH UNBIASED MIXING FOR THRUSTING PROJECTILES 3267

Authorized licensed use limited to: INSTITUTO MILITAR DE ENGENHARIA. Downloaded on November 20,2023 at 17:27:38 UTC from IEEE Xplore. Restrictions apply.