I.

INTRODUCTION
Target tracking is a well-studied problem in
IMMPDAF for Radar the literature. Some specific problems of interest
in the single-target, single-sensor case are tracking
Management and Tracking maneuvering targets [ 11, tracking in the presence of
clutter [2], and electronic countermeasures (ECM)
Benchmark with ECM [12]. In addition to these tracking issues, to be
complete, a tracking system for a sophisticated
electronically steered antenna radar has to consider
T. KIRUBARAJAN, Student Member, IEEE radar scheduling, waveform selection and detection
Y. BAR-SHALOM, Fellow, IEEE
threshold selection [3]. Although many researchers
University of Connecticut have worked on these issues and many algorithms
W. D. BLAIR, Senior Member, IEEE
are available, there was no standard problem to
Georgia Tech. Res. Institute compare the performances of the various algorithms.
G. A. WATSON
Rectifying this, Blair, et al. developed the first
Naval Surface Warfare Center, Dahlgren benchmark problem [7], which considered only
the problem of tracking a maneuvering target and
pointing/scheduling of a phased array radar. Of all the
A framework is presented for controlling a phased array radar
algorithms considered for this problem, the interacting
for tracking highly maneuvering targets in the presence of false
multiple model (IMM) estimator yielded the best
alarms (FAs) and electronic countermeasures (ECMs). Algorithms performance [6, 11, 13, 231. The second benchmark
are presented for track formation and maintenance; adaptive problem [SI included false alarms (FAs) and ECM,
selection of target revisit interval, waveform and detection specifically, a standoff jammer (SOJ) and range gate
threshold; and neutralizing techniques for ECM, namely, against pull off (RGPO) as well as several possible radar
a standoff jammer (SOJ) and range gate pull off (RGPO). The waveforms (from which the resource allocator has to
interacting multiple model (IMM) estimator in combination select one at every revisit time). Preliminary results
with the probabilistic data association (PDA) technique is used for this problem showed that the IMM and multiple
for tracking. A constant false alarm rate (CFAR) approach hypothesis tracking (MHT) algorithms were the best
is used to adaptively select the detection threshold and radar solutions [5, 14, 221. The MHT algorithm, while 1-2
waveform, countering the effect of jammer-induced false
orders of magnitude costlier computationally than
the IMMPDAF (IMM estimator with probabilistic
measurements. The revisit interval is selected adaptively, based
data association filter (PDAF) modules [2]) for
on the predicted angular innovation standard deviations. This
the problem considered (5 hypotheses are needed,
trackerkadar-resource-allocator provides a complete solution to with a cost of 9 IMMs [21]), yielded comparable
the benchmark problem for target tracking and radar control results with the IMMPDAF. The benchmark problem
(presented in the companion paper). Simulation results show an of [SI was upgraded in [9] to require the radar
average sampling interval of about 2.5 s while maintaining a track resource allocator/manager to select the operating
loss less than the maximum allowed 4%. constant false alarm rate (CFAR) and include the
effects of the SOJ on the direction of arrival (DOA)
measurements; also the SOJ power was increased
to present a more challenging benchmark problem.
Manuscript received August 24, 1996; revised July 18, 1997 and While in [SI the primary performance criterion for
May 4, 1998. the tracking algorithm was minimization of radar
IEEE Log NO. T-AES13414101971. energy, the primary performance was changed in [9]
This work was supported under O W B M D O Grant
to minimization of a weighted combination of radar
NOOO14-91-J-1950, ONR Grant N00014-97-1-0502, AFOSR Grant time and energy. We present a framework based on
F49620-94-1-0150, and by funding from the In-House Laboratory the IMMPDAF for radar scheduling and control,
Independent Research (ILIR) Program at the Naval Surface Warfare target tracking and neutralizing techniques for ECM
Center Dahlgren Division in Dahlgren, VA. to solve the set of problems posed in [9].
Authors’ current addresses: T. Kirubarajan and Y. Bar-Shalom, The tracking algorithm is designed to control the
Electrical and Systems Engineering Department, University of beam pointing of the phased array radar adaptively
Connecticut, U157, 260 Glenbrook Rd., Storrs, CT 06269-2157; in order to maximize the revisit interval for the
W. D. Blair, GTRI-SEAL, 7220 Richardson Rd., Smyma, GA
30080; and G. A. Watson, Systems Research and Technology
target while meeting a maximum allowed track
Department, Naval Surface Warfare CenterIDahlgren Division, loss probability. Also, the SOJ has to be tracked (at
Dahlgren, VA 22448-5100. as low a rate as possible) so its interference with
target returns does not cause loss of the target track.
0018-9251/98/$10.00 @ 1998 IEEE The time and the pointing direction for the next

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 4 OCTOBER 1998 1115

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 radar dwell are determined by the radar resource discrete and continuous uncertainties. This is ideally
allocation algorithm. This also selects the type of suited for tracking maneuvering targets which exhibit
dwell, waveform for active dwells and the threshold discrete uncertainties in flight mode as well as
for measurement detection. The radar can carry out continuous ones. Here, the IMM is combined with
target track dwells (illuminating in the direction of the the probabilistic data association (PDA) technique,
predicted location of an existing track), search dwells which handles data association via weighting, to
(to detect and initialize tracks), and passive dwells to perform automatic track formation and maintenance.
track a jammer. Six target tracks are considered in this The PDA technique associates each validated
problem. The targets exhibit radar cross-section (RCS) measurement probabilistically to the estimated
fluctuations according to the Swerling type I11 model track. The probabilistic nature of the IMMPDAF
and perform maneuvers with lateral accelerations up allows the straightforward inclusion of neutralizing
to 7 g and longitudinal accelerations up to 2 g. Target techniques for ECM within the same framework. The
ranges vary from 20 km to 100 km and the elevation adaptive scheduling and detection techniques are also
angles vary between 2" and 80". The jammer stays described.
at ranges greater than 150 km and performs The organization of this paper is as follows:
accelerations up to 2 g. As an additional ECM, the In Section 11, tracking aspects of the IMMPDAF
target under track employs RGPO so as to pull the and its design are presented. Techniques for radar
tracker's range gate off the target by creating false scheduling/pointing and the IMM models used
returns. for tracking are also given. The ECM techniques
The tracking algorithm is required to maintain considered in the benchmark problem and our
tracks with a maximum track loss of 4%. A track neutralizing techniques are explained in Section 111.
is declared lost if the estimation error is greater Algorithms for SOJ power level estimation, target
than the two-way beamwidth in angles or 1.5 range RCS estimation, adaptive waveform selection,
gates in range. To compare the relative performances and adaptive detection threshold selection are also
of different algorithms, the average radar power presented. Simulation results for the six benchmark
per target, average radar time per second, average tracks are presented in Section IV.
sampling interval and the average rms error of the
position and velocity estimates are used. Here we
present the IMMPDAF solution for the second 11. TARGET TRACKING ALGORITHM
benchmark problem [SI consisting of the following:
algorithms for track formation and maintenance; In this section, the IMMPDAF technique for
adaptive selection of sampling interval, waveform and automatic track formation, maintenance, and
threshold; and neutralizing techniques for ECM. termination is presented. The coordinate selection for
The IMM estimator, combined with other data tracking, radar scheduling/pointing and the models
association techniques, has been shown to be an used for mode-matched filtering (the modules inside
effective technique for tracking maneuvering targets in the IMM estimator) are also discussed. These cover
the target tracking aspects of the solution to the
the presence of FAs based on real data [15, 241. Any
algorithm that handles this tracking problem needs to benchmark problem.
have the following:
1) the ability to identify (probabilistically) the A. Coordinate Selection
beginning and the end of maneuvers,
For target tracking in the track dwell mode of
2) the ability to discount FAs while associating
the radar, the number of detections at scan k (time
target-originated measurements correctly to the tk) is denoted by mk. The mth detection report c,(tk)
track. (m = 1,2,..., m k ) consists of a time stamp tk, range
The first requirement, known as maneuver r,, bearing b,, elevation e,, amplitude information
detection, involves swift switching of the effective (AI) p, given by the signal-to-noise ratio (SNR),
filter gain and/or state model [lo, 181. The next and the standard deviations of bearing and elevation
one, known as data association, involves assignment measurements, 0; and uk,respectively. Thus,
and/or weighting of multiple measurements. The
first requirement can be met with the approach of
random maneuver modeling where the unknown where the overbar indicates that this is in the spherical
input command is assumed to be a random process, coordinate system of the radar.
which is treated as a process noise with several The AI is used only to declare detections
intensities and structures [ 11. We adopt this strategy and select radar waveform for the next scan.
in the form of IMMPDAF [2] to solve the benchmark Since the use of AI, for example, as in [16],
problem. The IMM estimator accommodates both can be counterproductive in discounting RGPO

1116 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 4 OCTOBER 1998

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 measurements, which generally have higher SNR than Scan 1 (t = 0 s): As defined by the benchmark
target-originated measurements, AI is not utilized in problem, there is only one (target-originated, noisy)
the estimation process itself.' measurement. The position component of this
For target tracking, the measurements are measurement is used as the starting point for the
converted from spherical coordinates to Cartesian estimated track.
coordinates and then the IMMPDAF is used on these Scan 2 (t = 0.1 s): The beam is pointed at
converted measurements. This conversion avoids the the location of the first measurement. This yields,
use of extended Kalman filters and makes the problem possibly, more than one measurement and these
linear [ 171. The converted measurement report c,(t,) measurements are gated using the maximum possible
corresponding to Tm(tk)is given by [4]: velocity of the targets to avoid the formation of
impossible tracks. This validation region volume,
C,(tk) = [tk x, Ym z, Pm R,I (2) which is centered on the initial measurement, is given
where xm, ym, z,, and R, are the three position by
measurements in the Cartesian frame and their
covariance matrix, respectively. The converted values
are
x, = r, cos(b,) cos(e,) (3)
where 6, = 0.1 s is the sampling interval and imm,
y, = r, sin@,) cos(e,) (4)
ymax,and i , are the maximum speeds in the X ,
Y , and Z directions, respectively; Rk2,R;,, and R&
z, = r, sin(e,) ( 5 ) are the variances of position measurements in these
directions, obtained from the diagonal components
R, = T, . diag[(a;>', (a;)', (03~3 . Th (6) of (6). The maximum speed in each direction was
where ai is the standard deviation of range assumed to be 500 d s .
measurements at scan k and T, is the spherical-to- Using the measurement in the first scan and
Cartesian transformation matrix given by2 the measurement with the highest SNR in the
second scan, a track is formed using the two-point
cos(bm)cos(em) -r, sin(bm)cos(em) -rm cos(bm)sin(em) initialization technique [ 11. The track splitting used
sin(b,)cos(e,) rmcos(bm)cos(em) -rm sin(bm)sin(e,) in [2 and 141 was found unnecessary-the strongest
validated measurement was adequate. This technique
sin(em) 0 rmcos(e,)
yields the position and velocity estimates and the
(7) associated covariance matrices in all three coordinates.
Scan 3 (t = 0.2 s): The pointing direction for the
B. Track Formation radar is given by the predicted position at t = 0.2 s
using the estimates at scan 2. An IMMPDA filter with
In the presence of FAs, track formation is a crucial three models discussed in the sequel is initialized with
aspect. Incorrect track initiation will result in missed the estimates and covariance matrices obtained at the
detections leading to target loss. In [2] an automatic second scan. The acceleration component for the third
track formatioddeletion algorithm in the presence of order model is assumed zero with variance (u,,)~,
clutter was presented based on the IMM algorithm. where amax= 70 m/s2 is the maximum expected
Since in the present benchmark problem a noisy acceleration of the target.
measurement corresponding to the target of interest
is given in the first scan,3 forming new tracks for each From scan 3 on, the track is maintained using
unvalidated measurement, as suggested in [2] and as the IMMPDAF as described in Section IIC. In order
implemented in [14], at subsequent scans is expensive to maintain a high SNR for the target-originated
in terms of both radar energy and computational load. measurement during track formation, a high energy
In this implementation, track formation is simplified waveform is used, as discussed in Section 111. Also,
and handled as follows. at the second scan 3 dwells are requested by the
algorithm to ensure target detection. This simplified
approach cannot be used if the target-originated
'As in [16] one could use the AI but this would require a separate measurement is not given at the first scan. In that
model for the RGPO intensity which cannot be estimated in real case, the track formation technique in [2] can be used.
time due to its short duration and variability.
Immediate revisit with sampling interval 0.1 s is
2For the scenarios considered here this transformation is practically
unbiased and there is no need for the debiasing procedure of [17]. requested during track formation because the initial
3Assuming that this is a search pulse, i.e., without (monopulse) velocity of the target is not known (in the first scan
split-beam processing, the angular errors are uniformly distributed only the position is measured and there is no a priori
in the beam. velocity). This means that in the second scan, the

KIRUBARAJAN ET AL.: IMMPDAF FOR RADAR MANAGEMENT AND TRACKING BENCHMARK WITH ECM 1117

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 radar needs to be pointed at the first scan position, The corresponding values for a piecewise constant
assuming zero velocity. Waiting longer to obtain Wiener process acceleration (third-order) model are
the second measurement could result in the loss of
the target-originated measurement due to incorrect 1 6, 432 0 0 0 0 0
pointing. Also, in order to make the IMM mode 0 1 6 , O O 0 0 0 0 l
O
probabilities converge to the correct values as quickly 0 0 1 0 0 0 0 0
as possible, the target is revisited at a high rate.
0 0 0 1 6, 62/2 0 0 0 I
O
F(S,)= 0 0 0 0 1 6, 0 0 0
C. Track Maintenance
0 0 0 0 0 1 0 0 0
The state of the target at t k is x ( t , ) = 0 0 0 0 0 0 1 6, 62/2
[x(t,) x(t,> X ( l k ) y ( l k ) y<t,> Y(l,> z(tk) i(t,)
Z(lk)l’ where 0 0 0 0 0 0 0 1 6,
x(t,), y(t,), and z(tk) are the positions, jC(t,), y(t,), and - 0 0 0 0 0 0 0 0 1
i(t,) are the velocities and .i(t,,, y(t,), and Z(t,) are
the accelerations of the target in the corresponding
coordinates, respectively. The measurement vector
consists of the Cartesian position components at tk and
is denoted by z ( t k ) .
Assuming that the target motion is linear in the
Cartesian coordinate system, the state equation of the
target can be written as

X(tk) = JYG,)x(t,-,) + r(6,)v(~,-l) (9)

and the target-originated measurement is related to the
state according to

i 1
z(tk) = HX(tk)+ w(tk) (10) 1 0 0 0 0 0 0 0 0
where Sk = t, - tk-l. The white Gaussian noise H= 0 0 0 1 0 0 0 0 0 .
sequences v ( t k ) and w(tJ are independent and their 0 0 0 0 0 0 1 0 0
covariances are Q(6,) and R(tk), respectively.
For a piecewise constant white acceleration With the above matrices, the predicted state i ( t ; )
(second order) model, the matrices F(S,), r (S,), and at time t, is
H are given by %,-I = W&,-l) (13)
and the predicted measurement is
l S , O 0 0 0
~ 0 1 0 0 0 0 j i(t;) = HrZ(t,-) (14)

F(Sk) =
0 0 1 s , 0 0
0 0 0 1 0 0
I with associated innovation covariance

1) Probabilistic Data Association: During track

6, 0 0 (1 1) maintenance, each measurement at scan k is validated
against the established track. This is achieved by
0 632 0
r(&)= ’ setting up a validation region centered around the

[z(t,) - i(t,-)l’S(t,)-’[z(tk) - %tc>l5 Y (16)

1 1 :1:
where S(t,) is the expected covariance of the
innovation corresponding to the correct measurement

[:
and y = 16 (0.9989 probability mass [2]) is the gate
H = 0 0 1 0 0 0 . size. The appropriate covariance matrix to be used in
the above is discussed in the sequel.

1118 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34. NO. 4 OCTOBER 1998

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 The set of measurements validated for the track at
tk is
Z ( k ) = {zm(tk),m = 1,2,...,mk} (17)
and Po is the probability of detection of a
where mk is the number of measurements validated target-originated measurement. The probability that
and associated with the track. Also, the cumulative set a target-originated measurement, if detected, falls
of validated measurements up to and including scan within the validation gate is assumed to be unity.
k is denoted by Z k . All unvalidated measurements are Also, N[v;O , S ] denotes the normal probability
discarded. density function (pdf) with argument v, mean zero
With these mk validated measurements at tk one and covariance matrix S. The common validation
has the following mutually exclusive and exhaustive volume V(tk)is the union of the validation volumes
events: Vm(tk)used to validate the individual measurements
A associated with the target, given by
Em(tk) =
= ~flz’2KzlSm(tk)11/2 (26)

I
vm(tk)
{measurement zm(tk)is
from the target}, m = 1,...,mk .
where K, is the volume of the unit hypersphere of
{all measurements are false}, m =0 dimension n,, the dimension of the measurement z.
(18) For the three-dimensional position measurements
K, =47r/3 [ 2 ] .
Using the nonparametric version of the PDAF The updated state estimate is
[2] the validated measurements are associated
probabilistically to the track. The combined target
state estimate is obtained as
mk
r2(tk) = c@“(tk)2m(tk) (19) and the associated covariance matrix is
m=O

where Pm(tk) is the probability that the mth validated

measurement is correct and iim(tk)is the updated state
conditioned on that event. The conditionally updated
states are given by

im(tk) = % ( t i )+ wm(tk)v,(tk), m = 1,273 .. mk9

(20)
where Wm(tk)is the filter gain and vm(tk)= Zm(tk)-
% ( t i )is the innovation associated with the mth is the predicted state covariance and
validated measurement. The gain, which depends on
the measurement noise covariance, is .

wm(tk)= P(ti)ff’[HP(tL)H’+ R,(tk)]-l

A
= P(ti)H’Sm(tk)-1 (21)

where Rm(tk)depends on the observed SNR for

measurement m [9].
The association event probabilities @,(tk) are given
by
Pm(lk) = P{Em(tk) I z!} (22) is the term analogous to the spread of the innovations
in the standard PDA [ 2 ] .Monopulse processing
results in different accuracies (standard deviations) for
different measurements within the same dwell. This
accounts for the difference in the above equations
from the standard PDA where the measurement
accuracies are assumed the same for all the validated
measurements.
To initialize the filter at k = 3 the following
estimates are used [ l ] :

KIRUBARAJAN ET AL.: IMMPDAF FOR RADAR MANAGEMENT AND TRACKING BENCHMARK WITH ECM 1119

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 The steps of the IMMPDAF are as follows [2].
Step 1 Mode Interaction or Mixing. The
mode-conditioned state estimate and the associated
covariances from the previous iteration are mixed
ri(t2)= (31) to obtain the initial condition for the mode-matched
filters.
The initial condition in cycle k for the PDAF
matched to the jth mode is computed using
r
iOj(tk-l = 'j('k-1 )pjlj (tk-l (35)
The associated covariance matrix is i=l

r ~ h n R,./S2

where h is the index corresponding to the validated where

measurement with the highest SNR in the second scan
and the superscripts x, y , and z denote the components
in the corresponding directions, respectively; the terms
R X ,RY, R Z ,R'y, RyZ,and Rxz are the elements of (6). - i , j = 1 , 2 ,...,r
If the two-point differencing results in a velocity )
component that exceeds the corresponding maximum
speed, it is replaced by that speed. Similarly, the
covariance terms corresponding to the velocity
components are upper bounded by the corresponding
maximum values. The accelerations are initialized as
zero with s.d. 7 g.
2) IMM Estimator Combined with PDA Technique:
In the IMM estimator it is assumed that at any time
the target trajectory evolves according to one of a
finite number of models, which differ in their noise
levels and/or structures [ 10, 11. By probabilistically
combining the estimates of the filters, typically Step 2 Mode-Conditioned Filtering. A PDAF is
Kalman, matched to these modes, an overall estimate used for each mode to calculate the mode-conditioned
is found. In the IMMPDAF the Kalman filter is state estimates and covariances. In addition, we
replaced with the PDA filter (given in Section IIC1 evaluate the likelihood function Rl(tk) of each mode
above for mode-conditioned filtering of the states) at tk using
which handles the data association.
Let r be the number of mode-matched filters used, 2 p [ z ( k ) I M(tk) = j , z k - ' l (38)
M ( t k )the index of the mode in effect in the interval
(tk-l,tk], and p.(t
! k ) be the probability that mode j
( j = 1,2,...,v) IS in effect in the above interval. Thus
pj(tk) = P{'(tk) =j I zk}. (33)
The mode transition probability is defined as
PI, = P{M(tk) = j I M(tk-1) = i}. (34)
where e j ( m ) is defined in (24) and b in (25). Note
The state estimates and their covariance matrix at that the likelihood function, as a pdf, has a physical
tk conditioned on the jth mode are denoted by ii,(tk) dimension which depends on mk. Since ratios of
and PJ (tk), respectively. these likelihood functions are to be calculated, they

1120 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 4 OCTOBER 1998

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 all have to have the same dimension, i.e., the same civilian air traffic surveillance, as demonstrated in [ 15
mk. Thus a common validation region (16) is vital and 241, this model is not necessary.
for all the models in the IMMPDAF. Typically the For a Wiener process acceleration model, the
“largest” innovation covariance matrix corresponding standard deviation ov3is chosen using
to “noisiest” model covers the others and, therefore,
ov3= min{P A, 6,amax} (44)
this can be used in (16) and (26).
Step 3 Mode Update. The mode probabilities are where A, is the maximum acceleration increment
updated based on the likelihood of each mode using per unit time (ierk), 6 is the sampling interval and
o < p < 1 [l].
For the targets under consideration, amax= 70 m/s2
and A, = 65 m/s3. Using these values, the process
noise standard deviations were taken as
Step 4 State Combination. The mode-conditioned
lJ”, = 3 m l s 2 (for nonmaneuvering intervals)
estimates and covariances are combined to find the
overall estimate ?(tk) and its covariance matrix P(tk) u = 3 5 m/s2 (for maneuvering intervals)
“2
are obtained as follows:
u
”3
= min{356,70} (for maneuver starthermination).
In addition to the process noise levels, the
elements of the Markov chain transition matrix
between the modes, defined in (341, are also design
I
parameters. Their selection depends on the sojourn
time in each motion mode. The transition probability
j=1
depends on the expected sojourn time via
(42)

3 ) Models in IMM Estimator: The selection of

the model structures and their parameters is one of the where T~ is the expected sojourn time of the ith mode,
critical aspects of the implementation of IMMPDAF. pii is the probability of transition from ith mode to the
Designing a good set of filters requires a priori same mode, and 6 is the sampling interval [ 11.
knowledge about the target motion, usually in the For the above models, pir, i = 1,2,3 are calculated
form of maximum accelerations and sojourn times using
in various motion modes [11. The tracks considered
in the benchmark problem span a wide variety of {
pii = min ui,max Zi,l - -
( :)} (46)
motion modes, from benign constant velocity motions where Zi = 0.1 and ui = 0.9 are the lower and upper
to maneuvers up to 7 g. To handle all possible motion limits, respectively, for the ith model transition
modes and to handle automatic track formation and probability. The expected sojourn times of 15, 4,
termination, the following models are used. and 2 s, are assumed for modes M’, M 2 , and M 3 ,
Benign Motion Model ( M 1): This second-order respectively.
model with low noise level (to be given later) has The selection of the off-diagonal elements of the
Markov transition matrix depends on the switching
a probability of target detection PD given by the
characteristics among the various modes and is done
expected SNR of the target, and corresponds to the
as follows:
nonmaneuvering intervals of the target trajectory. For
this model the process noise is, typically, taken to p12 = Oal(l -P11) P13 = 0.9(1 -p11)
model air turbulence. p21 =O.l(l-p22) p23=0‘9(1-p22)
Maneuver Model ( M 2 ) : This second-order
p31 = Oa3(l-P33) p32 = Oa7(l-p33)*
model with high noise level corresponds to on-going
maneuvers. For this white noise acceleration model The x, y , z components of target dynamics are
the process noise standard deviation ov2is obtained uncoupled and the same process noise is used in each
using coordinate. The decoupling is evident from (1 1) and
o v 2 = Q%ax (43) (12).

where amaxis the maximum acceleration in the D. Track Termination

corresponding modes and 0.5 < (Y 5 1 [l].
Maneuver Detection Model ( M 3 ) : This is a According to the benchmark problem, a track is
third-order (Wiener process acceleration) model with declared lost if the estimation error is greater than the
high level noise. For highly maneuvering targets two-way beamwidth in angles or 1.5 range gates in
like military attack aircraft, this model is useful for range. In addition to this problem-specific criterion,
detecting the onset and termination of maneuvers. For the IMMPDAF declares (on its own) track loss if the

KIRUBARAJAN ET AL.: IMMPDAF FOR RADAR MANAGEMENT AND TRACKING BENCHMARK WITH ECM 1121

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 track is not updated for 100 s. Alternatively, one can The predicted standard deviations in bearing and
include a “no target” model [2], which is useful for elevation are given by
automatic track termination, in the IMM mode set. In
a more general tracking problem, where the true target a;i = Y‘[O 1 01. S(t;). [O 1 01’ (49)
state is not known, the “no target” mode probability
or the track update interval would be the criterion for a; = J[O 0 11. S(@.[O 0 11’ (50)
track termination and the IMMPDAF would provide
a unified framework for track formation, maintenance respectively.
and termination. Step 7 The beamwidths at $ in bearing and
elevation, B; and Bi, respectively, are evaluated and
compared with the predicted standard deviation^.^
E. Adaptive Radar Scheduling If g f L5 B f / K b and a: 5 Bj/K,, where Kb and K,
Since the motion parameters of the target and the are integers, rI; is selected as the next sampling interval
jammer vary due to maneuvers, an adaptive sampling and tk is fixed at ti. Otherwise, the next (lower) T P l iS
policy is required-a short sampling interval during tried as the candidate sampling interval and steps 1-7
maneuvers and a long one during benign motions are are repeated.
desired. We use an improved version of the adaptive This technique allows a consistent selection of the
sampling selection scheme proposed in [ 111. The next sampling interval by being directly connected to the
sampling interval 6, = tk - tk-1 is selected (by the IMMPDAF estimator, its estimates and the associated
radar resource allocator/manager) from a predefined covariances. In the simulations, the sampling intervals
set of sampling intervals {2;}? where (q _< 2; 5 TN) are selected from the set A where
and the selection is based on the predicted values of
the radar angle innovation standard deviations relative A = {0.1,0.2,0.5,1,2,2.5,3,3.25,3.5}. (51)
to the radar beamwidth. The procedure is as follows. For the first 8 scans a sampling interval of 0.1 s
Step 1 The largest sampling interval is tried by is used to ensure accurate track f ~ r m a t i o nAlso,
.~
setting i = N,or equivalently (6, = TN). 0.1 s is used if no measurements are received
Step 2 The process noise variances for the IMM during a scan. In [ 141 a fixed value (1/6 of the
models and the Markov transition matrix are selected beamwidth at broadside) was used as the threshold
as discussed in Section IIC3 for the sampling interval for angle innovation standard deviations. In the
current implementation the threshold is also selected
7;. adaptively as 1/K of the beamwidth in the pointing
Step 3 The initial estimates and their covariances
for the IMM models are evaluated by mixing the direction as follows:
mode-conditioned estimates from the last update at
tk-l (Step 1 of the IMMPDAF in Section IIC2).
step 4 The predicted state at ti = tk-1 + 2;, its
covariance and the innovation covariance are evaluated where Zrgpo is a flag indicating RGPO activity to be
as in Section IIC 1. defined in Section 111.
Step 5 The predicted mode probabilities at ti are
evaluated by multiplying the Markov transition matrix Ill. NEUTRALIZING ECM (ECCM)
by the current mode probabilities.
Step 6 The combined innovation covariance S(t;) In this section we present the ECM techniques
and the combined predicted position = [xiy i zi] at ;; employed in the benchmark problem to deceive the
ti are found using the predicted mode probabilities. tracker and the neutralizing techniques implemented
The corresponding combined innovation covariance in our solution. As part of the derivation, adaptive
matrix s ( t i ) in the spherical coordinates of the radar is waveform and threshold selection schemes are also
found using - discussed. Also, the estimation of the RCS of the
S(@ = $2 . S(tL) . ($6)’ (47) target, which is used to determine the expected target
SNR, is presented.
where c,
the Cartesian-to-spherical transformation
matrix, is given by
A. Neutralizing the Standoff Jammer
Xi/G Y; 2; /r;
ql = The SOJ transmits broadband noise which can
-y’ k /SIk x;/4
“hide” the target return when the SOJ is in the
(-x;z;>/((r;>2$J (-Y;z;>/((r;>2s;> O I mainlobe of the radar beam (i.e., behind the target
(48)
and ri = + ( Y ; ) ~+ ( z : ) ~is the predicted range 4For the current problem, the radar’s range gate was large enough

and si = d m is the horizontal range at ti. so it imposed no practical constraint on the revisit time.
’Doubling the interval after every 3 measurements is also feasible.

1122 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 4 OCTOBER 1998

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 as seen from the radar) and causes reduced effective where n l ( k ) and ne@) are the receiver noises
target SNR (signal to total disturbance, Le., noise plus (independent standard Gaussian random variables) and
jammer) when it is in the sidelobe. In the benchmark the normalized antenna gain for the received signal E:
problem the jammer power is fixed at "io = 8 times is given by
the receiver noise power.6 When the jammer is in
the mainlobe, unless the radar waveform energy and ch = Q/(e;p&) +
detection threshold are increased, jammer-induced
false measurements will "hide" the target-originated + */(e;q, -e&) + */(-e;q, -e;q) (56)
measurement. To neutralize the effect of the jammer, and the normalized voltage gain pattern
its power level needs to be estimated, in addition to
its motion parameters, and the radar waveform and Qi(x,y)= V,(E;,2;4-x)Q(B:,i; + y). (57)
detection threshold should be selected such that a
CFAR is maintained given the predicted power and In (57), K(EL,;iG)and & ( $ , i k ) are the antenna
location of the SOJ with respect to the mainlobe. voltages in elevation and benaring, respectively, with
1) Estimation of Jammer State: The jammer the radar pointed at 2: and b: while the jammer is
maintains an oval holding pattern at an altitude of at E: and BL. Also, and etq are the squint angles
3.05 km at 168 m/s and remains about 150 km away in elevation and bearing, respectively, and $,; the
from the receiver. It is required to track the jammer phase of the voltage of SOJ, is uniformly distributed
from passive mode measurements received at the between 0 and 27r. A more detailed description of
radar. In passive mode, only one measurement is , the antenna voltage patterns can be found in [9, Sect.
received per scan and the measurement report Cj(ti) IIC]. Also, is given by
consists of a time stamp ti, bearing b:, elevation e;,
jammer-to-noise ratio (JNR) pi in dB and the standard
deviatjons of bearing and elevation measurements, $C
and a f , respectively. Thus, where nil@) and nj,(k) are the in-phase and
. . quadrature components of the broadband noise
Cj(tL)= [ t i bi e: p; crib crf]'. (53) (independent standard Gaussian random variables),
The jammer is tracked in spherical coordinates respectively, and Gs,(R) is the sensitivity time control7
(angles only) using the standard IMM estimator (STC) gain given by
with three mode-matched Kalman filters, namely
a second-order model with low process noise, a
second order model with high process noise, and
a third-order model with high process noise. The where R,, = 30 km. In passive mode Gstc(R)= 1.
jammer state at time t i , xj(i), consists of two angular The observed SNR8 of the ith range bin is given
positions (bearing and elevation) and velocities. The
by
third-order model includes accelerations as well. The
adaptive sampling technique given in Section IIE R: = $($k + sik) (60)
is employed for radar scheduling to obtain jammer and it can be shown that the return amplitude
measurements. The track is initialized using the
two-point initialization technique with the first two
d a
is Rayleigh distributed with parameterg

measurements. y0(Cjy + 1 [20].

2) Estimation of Jammer Power Level: The The estimated JNR is
objective is to estimate the jammer power level yo . N
in order to predict the effect of the jammer at the
next target revisit time and to select the waveform
and threshold to satisfy the CFAR condition and a
where N is the number of observations within the
minimum level of target Po. In passive mode, in which
dwell. Using
jammer measurements are obtained, the in-phase and
quadrature components of the sum voltage, slk and E(x2) = 2 2 (62)
sQk,respectively, are for a Rayleigh random variable (RV) with parameter
cr2, the mean of the JNR of the jammer measurement
SIk + nI(k)
= r,$: cos 4; (54)
SQk = r$; sin@: + n Q ( k ) (55) 7STC is used to prevent the detection of small objects at short range
~91.
6This is the power the jammer injects into each of the receiver 'Actually this is the signal-plus-noise to noise ratio.
channels (in-phase and quadrature, see [9]). Therefore, it degrades 'The pdf of a Rayleigh RV x with parameter u2 is p ( x ) = ' R ( x ; a 2 )=
the SNR by 2y0, i.e., IOlog,, 16 = 12 dB. ( x / 0 2 ) exp(-x2/202).

KIRUBARAJAN ET AL.: IMMPDAF FOR RADAR MANAGEMENT AND TRACKING BENCHMARK WITH ECM 1123

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 at t;, p i , can be shown to be (in dB) 4) Estimation of the Target Radar Cross Section:
The detection threshold maintains a CFAR. In order
iji = 1olog,o(ro(c{>2>. (63) to ensure target detection, an appropriate waveform
providing an expected target SNR above the threshold
An estimate of -yo at time t;, denoted by 70(tL),can is used. The waveform is selected (by the radar
be found from (63). Because of the widely varying resource allocator/manager, at every revisit time) to
nature of Rayleigh distributed RVs, such an estimate maintain a certain probability of target detection PD
is not reliable. To overcome this, we use an a filter to using an estimate of the mean target RCS.
find an estimate To(& as In active mode, in which the radar is used to track
the target, the in-phase and quadrature components of
the sum voltage of the target-originated measurement
are given by
where p; is the measured JNR at t! and clij = 0.8 is
the elter parameter. The a filter is initialized with
sIk = r:(CL)2cos& + r$icos4i + n I ( k ) (69)
?&) = 1. SQk = rL(xL)2sin$i + r;c; sin$: + nQ(k) (70)
The above estimation technique relies on
emulating the radar equations with the true jammer where n,(k) and nQ(k)are the in-phase and quadrature
state replaced by the estimated one. Generally, noise components, respectively, and
the radar model in the tracking system (or a close
approximation thereof) is known and, therefore,
this technique can be implemented in a real system.
Estimation results are given in Section IV. and ci is the RCS of the target at tk, T,” is the pulse
3) Adaptive Threshold Selection: From (58) it length of the waveform used at tk and R: is the target
follows that when the radar is in track mode, the range. Also, is similar to CL defined in (56) and
amplitudes of the false measurements are Rayleigh (57), but evaluated with the target at B: in bearing
distributed with parameter Y~G&?)(CL)~+ 1 where and EL in elevation (see [9]). The phase of the target
is as defined in (56), evaluated at time tk, the next return and the phase of the jammer return, qji and qji,
target revisit time, and with the radar pointed in the respectively, are independent and uniformly distributed
predicted direction of the target. Also, @L(x,y) is between 0 and 27r. The constant IC, is defined in [9].
modified as The RCS 0; is modeled as Swerling 111 typelo with
parameter cave. Using (60), (69), and (70), the mean
@ / ( x , Y ) = K(E{,i?i + x)&(BL,& +y ) (65) SNR” of the target-originated measurement, P k in dB,
where 2; and %
give the radar pointing directions for can be shown to be
target tracking.
For a detection threshold p k r dB at t k , the
probability of false alarms 4: is given by (72)
4: = P(R6 2 10(p../E 10)}
(66) Solving for 0; in (72), we can find an estimate of
the RCS. As in Section IIIA2, we use an a filter to
estimate the mean RCS, denoted by 6t.A recursive
estimate is obtained as
The detection threshold is selected according to the G; = at$-’ + (1 - a,)
following CFAR approach. The expected FA rate per
cell is fixed at Ffa= 2 x and for each scan k the 2( 10(ck/lo) - 1 - ~0(t~)G,,(~)(C~)2)(i~)4
X
threshold p:hr is given by (1 - €t)IC;T:Gstc(i:)(q4
+ l)lnFfJ)
P& = max(8 dB, lolog~o(-(”i,(~_,>G,,(R)(C:)2 (73)
(68) where j k is the “observed” target SNR in dB, R is
where +o(t~-l)is the estimate of the jammer power the center of the range gate at which the radar was
level at the previous scan for jammer tracking. The pointed, at = 0.6 is the filter parameter, $ is the
8 dB threshold, which is to be used in the absence estimated range of the target at tk and ct = 0.16 is the
of jamming (”/o = 0) yields the required CFAR of
1.82 x according to (67). ”The pdf of a random variable x of Swerling I11 type with
In the above, it is assumed that the estimate parameter u is f(x) = 4x/u2exp(-2x/o).
from time tiul is valid at time tk. This technique can “Actually
be used in the more general case where ^io is not this is the ratio between signal-plus-jammer-power-plus-noise and
constant, but changes slowly. noise.

1124 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 4 OCTOBER 1998

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 rescaling factor that accounts for radar mispointing threshold is not less than a certain “excess factor”
(obtained assuming an average mispointing of 116 of pe. The list of available waveforms is given in [9,
the radar beamwidth). Also, E; and C: are evaluated Sect. IIA]. The waveform index wk is selected
at the estimates. The filter is initialized with 3; = 1 according to
(the unit for Go is m2). w, = min i (75)
Since the index of the target-originated i :&(i)>pkr+iJe
measurement is not known, the “observed” target where &(i) is the predicted SNR in dB provided by
SNR is evaluated as a weighted sum of all the
measurements in the validation region as follows:
the ith waveform at the predicted location, is dhr
the detection threshold found in Section IIIA3 and
r-1 mL pe = 6 dB is a constant used to make sure that the
(74) predicted SNR is above the threshold by this excess.
i = l j=1 The predicted SNR ;k(i) is given by
where &(tk) is the ith mode probability given in (40),
pij is the jth association event probability in model i
of the IMM estimator as defined in (23) and p j is the
SNR of the jth measurement.
For a model in the IMM estimator, the sum
of the weights used to find the “observed” target
(76)
SNR is generally less than unity and the value of Po
(probability that none of the measurements originated where C i and EL are evaluated at the corresponding
from the target) will affect the “observed” SNR, estimates. This is similar to (72) except that it uses the
which decreases with increasing Po. One advantage estimated RCS and estimated jammer power.
of this approach is that the state estimation results are
directly used in the energy calculation, both intuitively
and quantitatively-when the state estimator does not B. Neutralizing Range Gate Pull Off
have much confidence in the received measurements, The purpose of RGPO is to pull the radar range
indicated by a high Po, the “observed” SNR drops, gate off the target by repeating an amplified version
which results in a low RCS estimate and, possibly, of the radar pulse with delay-by emitting a false
a higher waveform in the next scan. This acts as a return. If a tracking algorithm follows the false RGPO
compensating feature when the target measurement measurement, instead of the target, it will eventually
is not detected or lies outside the validation gate. lose the target. The false RGPO measurements
5 ) Adaptive Waveform Selection: The waveform are emitted by the target such that they appear to
for the next scan is selected using the estimates of move away from the true target position (in range
jammer power and target RCS, as well as the chosen only and away from the radar) with either linear
detection threshold. One of two major criteria of or quadratic motion. When RGPO is active, two
the benchmark problem is to track the targets with strong signals, one from the target and the other
minimum radar energy and radar time subject to from RGPO, are received at the radar, in addition to
the 4% track loss constraint. These are the “costs” any false measurements due to the noise. In order to
associated with an algorithm, which are to be discount the false RGPO measurements, the tracking
minimized. algorithm has to recognize that RGPO is active and
To maintain a high detection probability and give less “weight” to the false measurement, whose
a low false alarm rate by compensating for the index is not known. We present a weighting scheme,
jammer-originated noise, a high energy waveform is where the discriminating feature is the range of the
required. On the other hand, high energy waveforms, measurements. A more detailed description of RGPO
which use longer pulses, have low measurement is presented in [9, Sect. 1111.
accuracies (high standard deviations),” which is
In the linear and quadratic cases, the range Rkf‘ of
undesirable in tracking highly maneuvering targets.
the RGPO measurement at time tk is given by
The object is to balance the conflicting requirements
of “cost,” detection probability, jammer compensation, (77
and measurement accuracy.
The waveform at tk is found such that the ratio and
1
of the return signal power to the chosen detection R,f’ = RL + 2apO(tk- to)’ (78

‘*Because, beyond some point, higher energy can be obtained only

respectively. The rate of pull-off, in the linear case, is
with longer pulses. The angular accuracies stay about the same if denoted by vpo and the acceleration in the quadratic
the signal to jammer-plus-noise ratio is the same as the SNR in the case by apo.The RGPO is activated by the target at
absence of jamming. time to, which is not known to the tracking algorithm.

KIRUBARAJAN ET AL.: IMMPDAF FOR RADAR MANAGEMENT AND TRACKING BENCHMARK WITH ECM 1125

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 Thus, the “intended” range of the RGPO is larger than measurement, whose range information, unlike that
the true target range. When two “strong” returns are of the RGPO measurement, does not have any special
received, which usually (but not always) indicates characteristics, could receive favorable weighting. One
RGPO, the one at the longer range cannot simply drawback of the use of the “anti-RGPO discriminating
be discarded as the RGPO measurement because the function” is that it is counter-effective against Range
closer one could be a FA and the farther one the true Gate Pull In (RGPI), where the target under track
measurement. attempts to place a false return consistently in front
One approach to tackle the RGPO is to form a new of the true measurement. However, RGPI is probably
track when RGPO is detected and maintain two tracks not practical against an asynchronous sensor.
until RGPO stops. We derive an alternative approach It is also important to maintain target
in the same spirit of the IMMPDAF where multiple detection throughout the duration of RGPO. If the
measurements are combined with different weights to target-originated measurement is not detected when
find an estimate. The steps are as follows. RGPO is active, the range gate is pulled off easily.
In order to avoid this, the excess SNR pe (w.r.t. the
Step 1 At scan k, if two measurements with SNR threshold) is increased to pe = 12 dB.
above prgpo= 12 dB are received, RGPO is declared 1) Track Reacquisition in Presence of Range
and RGPO indicator Zrgpo is set to 1. Otherwise, Gate Pull Ofs Failure to detect target-originated
‘rgpo = 0. measurements when RGPO is active, especially at the
Step 2 When Zrgpo = 1 at tk, the association event beginning of an RGPO epoch, could result in more
probabilities of PDA, defined in (23), are rescaled as missed detections and, eventually, in track loss. A
tracking algorithm must be “intelligent” enough to
realize that target-originated measurements are being
missed and to reacquire the target when track loss is
D
m=O (79) anticipated.
b+ e(j) When there are no measurements in the validation
where b and e(m) are as defined in (24) and (25), gate for L number of scans between tk-L and tk-1, the
respectively. The weighting function W(m) of the mth following track reacquisition scheme is employed.
measurement is given by 1) An immediate revisit with sampling
interval 0.1 s (tk = tk-l + 0.1) in search dwell
mode is requested. In search mode, multiple
measurements with different ranges in the pointed
where direction are returned. In this case no monopulse
w(m>= exp((rk - r m > / r ~ > (81) processing is done, Le., the “measured angles”
are the beam-pointing angles. Since the beam
is an “anti-RGPO discriminating function” based on encompasses more than a validation gate, this
the range of the measurements, and rm is the range facilitates reacquisition. Also, a suitable waveform
of the mth measurement, Fk is the average range is selected, as in Section IIIA5, with the detection
of all the measurements at tk and rc is the standard threshold pkr = 10 dB and pe = 6 dB.
deviation of range measurements, which is used as a 2) The position components of the received
normalizing constant. measurements are weighted with w(m),defined in
The effect of this approach is that the association (81), to find the position of the reacquired track.
probabilities of the measurements at higher ranges 3) An immediate revisit (tk+l = tk + 0.1) in track
are reduced while that of the closest one is increased. mode is requested at the position of the reacquired
However, the sum of the association probabilities track for the next scan and the track is maintained as
PJ that the target-originated measurement before in Section IIC.
is detected and validated remains constant at
(cy:, e ( j ) ) / ( b+ e(1)). Another aspect is that
The use of a higher threshold for search mode
at tk eliminates false measurements and only
for the mth measurement the weighting depends on
e(m)w(m); this achieves a “compromise” between target-originated and RGPO measurements are
IMMPDA filter’s confidence in the mth measurement, received at the radar. The range penalty function
represented by the likelihood e(m), and the penalty w(m) discounts any RGPO measurements which
function w(m) based on the range measurement. are farther from the radar than the target-originated
measurements. In the simulations L = 3.
An alternative approach would be to declare that
RGPO is active when two measurements with high IV. SIMULATION RESULTS
association probabilities are received and then to
use the procedure in Step 2 to discount the RGPO. In this section the simulation results obtained
The disadvantage of this approach is that a false using the algorithms described in Sections I1 and

I126 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 4 OCTOBER 1998

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 TABLE I
Performance of IMMPDAF in Presence of FA, RGPO, SOJ

I11 are presented. The energy and radar time costs, C, corresponds to the case where the radar energy is
as defined in [9], computational requirements and more critical and the second one, C,, corresponds to
root-mean-square errors (RMSE) are given. The the case where the radar time is more critical. The
simultaneous operation of the various algorithms radar energy is proportional to the pulsewidth while
presented in this paper is illustrated: the estimation radar time is inversely proportional to the sampling
of RCS and jammer power, the adaptive selection of interval. Thus the selection of the weighting parameter
sampling interval, waveform and detection threshold pi is driven by the more important cost.
and the switching of mode probabilities. The secondary measure of performance is the
computational complexity of the tracking algorithm.
A. Performance Measures This is evaluated as the number of floating point
operations per second (FLOPS). Since track
The tracking algorithm is tested on the following initialization is both a critical and a computationally
six benchmark tracks. l3 expensive phase, we also present the computational
requirement for track initialization separately. Other
Target 1. A large military cargo aircraft with auxiliary performance measures are prediction errors
maneuvers up to 3 g. (RMSE) in position and velocity. More details on
Target 2. A Learjet or commercial aircraft which performance measures and target configurations can
is smaller and more maneuverable than target 1 with be found in [9].
maneuvers up to 4 g. In Table I, the performance measures and their
Target 3. A high speed medium bomber with averages for the IMMPDAF in the presence of FA,
maneuvers up to 4 g. RGPO, and SOJ are given. The averages are obtained
Target 4. Another medium bomber with good by adding the corresponding performance metrics of
maneuverability up to 6 g. the six targets (with those of target 1 added twice)
Targets 5 and 6. Fightedattack aircraft with very and dividing the sum by 7. In the table, the maneuver
high maneuverability up to 7 g. density is the percentage of the total time that the
The primary measures (evaluation criteria) of target acceleration exceeds 0.5 g. The average flop
a tracking algorithm, as defined in the benchmark count per second was obtained by dividing the total
problem, are number of floating point operations by the target track
length. This is the computational requirement for
ci= Eave+ /pave
i = 1,2 (82) target and jammer tracking, neutralizing techniques
for ECM and adaptive parameter selection for the
y, = io3 p2 = io5 (83)
estimator, i.e., it excludes the computational load for
where
- EaVeis the average radar energy per second, radar emulation.
T,, is the average radar time per second and pi is a The average costs C, and C, are 8.7 and 45.6,
weighting parameter. The average radar energy per respectively. The factors affecting the costs of
second is the sum of the energy of the track dwells individual tracks are the RCS of the target, its range,
requested by the tracking algorithm divided by the maneuver density and magnitude, jammer’s effect
duration of the trajectory. The average radar time per and the effect of ECM. For example, target 5 is
second is the sum of the radar time of the track dwells the “toughest” target with the lowest RCS, highest
divided by the duration of the trajectory [9]. Each maneuver density and at a high range (> 80 km).
dwell requires s of radar time. The first cost Also, the SOJ enters the mainlobe of the radar beam
while tracking target 5. In contrast, target 4 is the
13The tracking algorithm does not know the type of the target under “easiest” one at the shortest range with the lowest
track; the parameters are selected to handle any target. maneuver density. The RCS is higher and the effects

KIRUBARAJAN ET AL.: IMMPDAF FOR RADAR MANAGEMENT AND TRACKING BENCHMARK WITH ECM 1127

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 TABLE I1 IMMPDAF is 62.5 kFLOP/radar cycle. Due to the
Computational Load of IMMPDAF in Presence of FA, RGPO, high sampling rate (10 Hz) during track formation,
SOJ
the computational load is at its peak then. In Table I1
I Target I PeakLoad I AverageLoad I the peak computational requirement, together with
No. (kFLOPS) (kFLOPS) the average one for comparison, is given. Although
1 205.9 22.2 the peak requirement is ten times higher than the
2 251.8 24.3
3 224.8 24.6
average one, it is only 1/300 of the currently available
4 209.0 24.3 computational power, for example, of a Pentium
5 290.4 27.1 processor running at 133 MHz.
6 212.5 24.6 With the revisit time calculations taking about
Average 232.4 24.5 the same amount of computation as a cycle of the
IMMPDAF, but running at half the rate of the Kalman
filter (which runs at constant rate), the IMMPDAF
of RGPO and SOJ are less than those in target 5. with adaptive revisit time is about 10 times costlier
The effect of these parameters is evident from the computationally, than a Kalman filter. Due to its
difference in the costs of target 4 and target 5. ability to save radar resources, which are much
The average sampling interval requeited by more expensive than computational resources, the
the IMMPDAF is 2.48 s, which represents a 40% IMMPDAF is a viable alternative to the Kalman filter,
increase (Le., average sampling rate lowered by 70%) which is the standard “workhorse” in many current
compared with the preliminary results14 reported tracking systems.
in [5 and 141. The selection of sampling interval In Table I11 the performance of the IMMPDAF
parameters is based on a “worst-case” design which in the presence of FA and SOJ is presented. The
can handle a wide variety of targets. Performance effect of RGPO can be seen by comparing Table I
can be improved by optimizing the parameters to with Table 111. The RMSE of position estimates are
a subclass of trajectories, for example, cargo or 20-30% higher in the presence of RGPO because of
commercial aircraft. Increasing maneuver density and the increased uncertainty. In spite of the apparently
maximum acceleration reduce the sampling interval, increased prediction errors, the IMMPDAF continues
which in turn increases the costs. The reduction in to maintain the tracks satisfying the requirements
the average sampling interval of target 5 is about of the benchmark problem. This robustness of the
15% compared with the sampling interval of target 1, IMMPDAF to the effects of RGPO is due to the
which has the lowest maximum acceleration. weighting scheme introduced in Section IIIB and the
The average flopI5 requirement is 25 kFLOPS, track reacquisition scheme of Section IIIBA (without
which can be compared with the flop rate of 78 such counter-RGPO techniques the track loss is about
MFLOPS of a Pentium processor running at 133 MHz 35%, which is far above the acceptable track loss of
[ 191. Thus, the real-time implementation of the 4%). These techniques neutralize the RGPO without
complete tracking system is possible. With the average forming new tracks unlike the MHT algorithm, where
revisit interval of 2.5 s, the flop requirement of the track splitting is used to handle RGPO [5].The
presence of RGPO also reduces the average sampling
interval by 8% and increases the cost functions C, and
14An exact comparison is not possible because the current
benchmark problem is significantly more complex than the one
C, by 7%. This is due to the extra energy requested
presented in [8]. by the algorithm when RGPO is declared. The most
15The flop count is obtained using the built-in MATLAB function significant change in the sampling interval occurs
flops. Note that these counts are rather pessimistic; the actual flop in target 5 , where the 9% increase in the sampling
requirement would be considerably lower. interval is reflected in reduction of C,.

TABLE 111
Performance of IMMPDAF in Presence of FA and SOJ

1128 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 4 OCTOBER 1998

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 TABLE IV
Performance of IMMPDAF in Presence of FA and RGPO

-,
9 I 30
I

-m 7
-m
$6
8

1
0 20 40 60 80 100 120 140 160 180 200
-10' J
ElapsedTime (s) 35 40 45 50 55 60 65 70
X position (km)
Fig. 1. Estimated jammer power level (true T, = 8).
Fig. 2. Trajectory of target 6 (I-initial position, F-final position).

The difference in the average computational

requirement in presence of RGPO depends the estimate converges after about 10 samples (15 s
predominantly on the sampling interval. The effect of elapsed time) to within 20% of the correct value of
the extra weighting step (the anti-RGPO discriminant) 8. The RMSE over the entire period is 0.25. Since no
is not significant. It is noted that target 5 requires an assumption about the variability of ^/o is made, this
additional 20% flop rate in the presence of RGPO. estimation scheme can be used to estimate a (slowly)
This can be compared with the 17% increase in the time-varying jammer power level as well. Only
sampling interval. The same effect can be observed in knowledge of the dependence of the measured jammer
other targets as well. power on "lo is required as discussed in Section IIIA2.
The performance of the IMMPDAF in the presence Fig. 2 shows the trajectory of target 6. The
of FA and RGPO is presented in Table IV. Comparing target maintains constant speed and course for 30 s
Tables I and IV, we can note the 35-40% reduction in before making a 7 g turn at t = 30 s. The new course
the average FLOPS in the absence of SOJ. This can is maintained for another 30 s and a 6 g turn is
be considered as the computational requirement for performed while the throttle is reduced and the aircraft
tracking the SOJ. is nosed over in order to rapidly decrease altitude.
This turn-and-dive maneuver is accompanied by
RGPO and the SOJ happens to be in the mainlobe
B. Operation of Algorithms of the radar at the same time to maximize the tracking
difficulty. After 30 s another 6 g turn at full throttle
An accurate estimation of jammer power level and is performed. Finally, constant velocity is maintained
target RCS and a good adaptive selection of sampling after another 7 g turn at t = 150 s [9].
interval, waveform, and detection threshold ensure the Figs. 3 and 4 illustrate the operation of the
correct operation of the estimator. In this section these algorithms on Target 6. In Fig. 3 the average of the
aspects are illustrated and commented upon. RCS estimates, detection threshold and the selected
In Fig. 1, the estimated jammer power level +o is waveform are shown. The RCS estimates can be
shown for one run of scenario 6. It can be seen that compared with the true value of 1.9 m2. The adaptive

KIRUBARAJAN ET AL.: IMMPDAF FOR RADAR MANAGEMENT AND TRACKING BENCHMARK WITH ECM 1129

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 8 1.5
11
0 20 40 60 80 100 120 140 160 180
I
200
!: I-
g-10

% -15 t
“0 20 40 60 80 100 120 140 160 180 200

-40

-45
I 15 20 25 30 35 40 45 50
0 20 40 60 80 100 120 140 160 180 200 X position (km)
Time (s)
Fig. 5. Trajectory of target 2 (I-initial position, F-final position.
Fig. 3 . RCS (- - - - true, -estimated), detection threshold (in
dB), waveform (average index between 1 and 8) of target 6.
estimator and the sampling intervals requested by
the estimator are shown. The mode probabilities
100, 1
“follow” the changes in the target maneuver-the
IMM estimator becomes an “adaptive-bandwidth’’
filter. This property makes the IMM estimator superior
to the standard Kalman filter, which, in the nearest
neighbor and strongest neighbor forms [2], was found
1
to be ineffective for this benchmark problem [5,
s2
141. The reduction in the mode probability of mode
0.5
m M ’ between 15-20 s illustrates the effectiveness
P of the RGPO; initially the RGPO is effective by
“0 20 40 60 80 100 120 140 160 180 200 causing a “false-alarm’’ in maneuver detection. The
filter recovers from this initial error because of the
anti-RGPO discrimination function that is used to
discount the RGPO measurements. The same effect
is observed between 102-105 s, when RGPO is active
“0 20 40 60 80 100 120 140 160 180 200 in combination with a maneuver onset (turn and dive)
Time (s)
and presence in the mainlobe of the SOJ. The revisit
Fig. 4. Acceleration magnitude, mode probabilities (in bands
interval is larger when the target is not maneuvering
separated by lines as follows: M’I-lMZI- - - / M 3 ) and sampling
interval (in s) of target 6. because the estimator has more confidence it its
predictions indicated by low innovation standard
deviations. During maneuvers, when predictions are
nature of the detection threshold selection scheme less accurate, the revisit rate is increased.
can be seen from the figure-between 69-76 s The highly adaptive nature (across target classes)
and 98-128 s the SOJ is in the mainlobe and the of the algorithms presented in this work for tracking
corresponding threshold values are increased (up to and radar control can be seen from their performance
20dB) to compensate for the predicted effect of the when tracking target 2, which with its benign motion
SOJ. Similarly, the adaptive change in the waveform characteristics contrasts with target 6. Also, the SOJ
corresponding to the selected detection threshold does not enter the mainlobe. Fig. 5 shows the target
can be noticed. The waveforms are indexed from 1 trajectory where a mild 2.5 g turn is followed by a 4 g
to 8; with the index increasing by one, the power is turn.
about 2 dB higher (see [9] for details). The first 6 Figs. 6 and 7 present the simulation results. The
waveforms are of high range resolution while the detection threshold remains unchanged except at scans
last two (7 and 8) have a lower resolution (larger where target was reacquired using a higher threshold
resolution cell due to longer pulses). Typically, the as described in Section IIIAC. This is consistent
high-power low-resolution waveforms were used only with the fact that the SOJ does not enter the radar
in the presence of the SOJ in the mainlobe to avoid mainlobe. Reacquisition took place during maneuvers
loss of target detection. when RGPO was also active. The waveform energy
In Fig. 4 the magnitude of the true acceleration requested by the algorithm decreases with time as the
of the target, the mode probabilities in the IMM target approaches the sensor-the estimator becomes

1130 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 4 OCTOBER 1998

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 for neutralizing the effects of two types of ECM,
namely, an SOJ and RGPO, were presented.
Algorithms for estimating the RCS of the target and
the unknown jammer power level and schemes for the
adaptive selection of the radar waveform, detection
threshold, and sampling interval (revisit time) were
v
c
0 also given.
;a
c The IMM estimator was used in combination with
the PDA technique to track 6 benchmark targets,
150
'0 50 100
which evolve in widely different motion modes. A
8, I
single set of algorithms was presented with a built-in
adaptive variable-bandwidth feature to handle all
possible motion modes and the considered ECM
combinations.
-0 50 100 150
Time (s) Simulation results were presented demonstrating
Fig. 6. RCS (- - - - true, __ estimated), detection threshold (in the usefulness of the algorithms. An average sampling
dB), and waveform (average index between 1 and 8) of target 2.) interval of about 2.5 s is possible. Computational
requirements are such that the entire radar control
and estimation can be done in real-time using existing
computational power. The costs, in terms of radar
time and energy, of using these algorithms were
presented. Estimation results for an SOJ, which was
0 50 100
tracked using angle only measurements, were also
presented.
Some performance improvements are possible by
tuning the parameters, especially the decision regions
for waveform, detection threshold and sampling
0 50 100 150 interval selection. For example, when tracking cargo
aircraft, which maneuver substantially less than attack
aircraft, the predicted angle error threshold used to
select the next revisit time from the set (51) can be
I increased. Also the maximum revisit time can be
extended (see [6]).Another possible improvement is
Time (s)
the use of amplitude information in the estimation
Fig. 7. Acceleration magnitude, mode probabilities (in bands process itself, however, this has to consider the fact
separated by lines as follows: M'-/M2/- - - / M 3 ) and sampling
that generally RGPO measurements yield higher SNR
interval (in s) of target 2.
than the target-originated ones.
The major contribution of this work is that
confident about target detection. Mode probabilities, it provides the first unified framework for target
as shown in Fig. 7, change adaptively according to tracking and radar management by linking these
the maneuvers of the target. When the RGPO is active two aspects directly. The estimation results are used
between 12-17 s, the mode probability of mode M' quantitatively to select the operating conditions of
in the estimator and the sampling interval drop before the radar. For example, the confidence measures of
the estimator discounts the false RGPO measurements. the filter in the measurements and estimates, via the
These results demonstrate the capabilities of the likelihoods and association probabilities, are used
tracking and radar management scheme presented in to select the waveform and detection threshold at
this work for the benchmark problem with widely the next scan, which is also scheduled adaptively
varying targets (from benign civil aircraft to highly based on the latest performance of the estimation
maneuvering military attack aircraft) in the presence filter. Theoretical as well as intuitive explanations
of ECM and FA. for this connection between radar control and target
tracking were also presented. These algorithms can
V. CONCLUSIONS be used, without human intervention and in real-time,
for tracking a single target in the presence of FA and
We derived and presented a set of algorithms ECM and for radar resource allocationhanagement.
for radar management (pointing, scheduling, and Additional issues to be addressed in the future
waveform selection) and tracking highly maneuvering include: multiplehnresolved targets, multipath, RGPI,
targets in the presence of FA and ECM. Techniques multiple simultaneous RGPOs, and multiple sensors.

KIRUBARAJAN ET AL.: IMMPDAF FOR RADAR MANAGEMENT AND TRACKING BENCHMARK WITH ECM 1131

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 REFERENCES [13] Kalata, P. (1984)
An a-/3 target tracking approach to the benchmark
Bar-Shalom, Y., and Li, X. R. (1993) tracking problem.
Estimation and Tracking: Principles, Techniques and In Proceedings of the American Control Conference,
Software. Baltimore, MD, June, 1994.
Dedham, MA: Artech House, 1993. Kirubarajan, T., Bar-Shalom, Y., and Daeipour, E. (1995)
Reprinted by YBS Publishing, 1998. Adaptive beam pointing control of a phased array radar in
Bar-Shalom, Y., and Li, X. R. (1995) the presence of ECM and false alarms using IMMPDAF.
Multitarget-MultisensorTracking: Principles and In Proceedings of the American Control Conference,
Techniques. Seattle, WA, June, 1995.
Storrs, CT: YBS Publishing, 1995. Kirubarajan, T., Yeddanapudi, M., Bar-Shalom, Y., and
Bierson, G. (1990) Pattipati, K. R. (1996)
Optimal Radar Trucking Systems. Comparison of IMMPDA and IMM-assignment
New York: Wiley, 1990. algorithms on real air traffic surveillance data.
Blackman, S. S. (1986) In Proceedings of the SPIE Conference on Signal and Data
Multiple-Target Tracking with Radar Applications. Processing of Small Targets, Orlando, FL, Apr. 1996.
Dedham, MA: Artech House, 1986. Lerro, D., and Bar-Shalom, Y. (1993)
Blackman, S. S., Bush, M., Demos, G., and Popoli, R. Interacting multiple model tracking with target amplitude
(1995) feature.
IMM/MHT tracking and data association for benchmark IEEE Transactions on Aerospace and Electronic Systems,
tracking problem. 29, 2 (Apr. 1993), 494-509.
In Proceedings of the American Control Conference, Lerro, D., and Bar-Shalom, Y. (1993)
Seattle, WA, June, 1995. Tracking with debiased consistent converted measurement
Blair, W. D., and Watson, G. A. (1994) versus the EKF.
IMM algorithm for solution to benchmark problem for IEEE Transactions on Aerospace and Electronic Systems,
tracking maneuvering targets. 29, 3 (July 1993), 1015-1022.
In Proceedings of SPIE Acquisition, Tracking and Pointing Li, X. R., and Bar-Shalom, Y. (1993)
Conference,Orlando, FL, Apr. 1994. Design of an interacting multiple model algorithm for air
Blair, W. D., Watson, G. A., and Hoffman, S. A. (1994) traffic control tracking.
Benchmark problem for beam pointing control of phased IEEE Transactions on Control Systems Technology, 1, 3
array radar against maneuvering targets. (Sept. 1993), 186-194.
In Proceedings of the American Control Conference, Messmar, H. P. (1995)
Baltimore, MD, June, 1994. The Indispensable Pentium Book.
Blair, W. D., Watson, G. A., Hoffman, S. A., and Gentry, Reading, MA: Addison-Wesley, 1995.
G. L. (1995) Papoulis, A. (1991)
Benchmark problem for beam pointing control of phased ProbabiMy, Random Variables and Stochastic Processes.
array radar against maneuvering targets in the presence of New York: McGraw-Hill, 1991.
ECM and false alarms. Blackman, S. S. (1998)
In Proceedings of the American Control Conference, Personal communication, 1998.
Seattle, WA, June, 1995. Popoli, R. E, Blackman, S. S., and Busch, M. T. (1996)
Blair, W. D., Watson, G. A,, Kirubarajan, T., and Application of multiple hypothesis tracking to agile beam
Bar-Shalom, Y. (1998) radar tracking.
Benchmark for radar resource allocation and tracking in In Proceedings of Signal and Data Processing of Small
the presence of ECM. Targets 1996,2759 (Apr. 1996), 418428.
IEEE Transactions on Aerospace and Electronic Systems, Sastry, C. R., Slocumb, B. J., West, P. D., Kamen, E. W.,
this issue, 1097-1 114. and Stalford, H. L. (1994)
Blom, H. A. P. (1984) Tracking a maneuvering target using jump filters.
A sophisticated tracking algorithm for ATC surveillance In Proceedings of the American Control Conference,
data. Baltimore, MD, June, 1994.
In Proceedings of the International Radar Conference, Yeddanapudi, M., Bar-Shalom, Y., and Pattipati, K. R.
Paris, May, 1984. (1995)
Daeipour, E., Bar-Shalom, Y., and Li, X. R. (1994) IMM estimation for multitarget-multisensor air traffic
Adaptive beam pointing control of phased array radar surveillance.
using an IMM estimator. IEEE Proceedings, 85, 1 (Feb. 1997), 80-94.
In Proceedings of the American Control Conference,
Baltimore, MD, June, 1994.
Daum, E E. (1996)
Book review-Multitarget-multisensor tracking: principles
and techniques.
IEEE Systems Magazine, 11, 7 (July 1996), 39-42.

1132 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 4 OCTOBER 1998

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 Thiagalingam Kirubarajan (S’95) was born in Sri Lanka in 1969. He received
his B.A. degree in electrical and information engineering from Cambridge
University, England, in 1991, M.S. and Ph.D. degrees from the University of
Connecticut in 1995 and 1998, respectively.
While in England he worked for the Central Electricity Research Laboratories,
Leatherhead, Surrey, as a research assistant. From 1991 to 1993 he was an
assistant lecturer in Electrical and Computer Engineering at the University of
Peradeniya, Sri Lanka. During 1993-1998 he was a graduate studenthesearch
assistant at the University of Connecticut, Storrs. His research interests are in
estimation and target tracking. Currently he is Research Assistant Professor of
Electrical and Systems Engineering at the University of Connecticut.

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
in 1970, all in electrical engineering.
From 1970 to 1976 he was with Systems Control, Inc., Palo Alto, CA.
Currently he is Professor of Electrical and Systems Engineering and Director of
the ESP Lab (Estimation and Signal Processing) at the University of Connecticut.
His research interests are in estimation theory and stochastic adaptive control and
has published over 220 papers in these areas. In view of the causality principle
between the given name of a person (in this case, “(he) will track”, in the modern
version of the original language of the Bible) and he profession of this person,
his interests have focused on tracking. His other interests are stochastic control of
vertical airfoils and of pairs of inclined foot supports on crystals. He coauthored
the monograph Tracking and Data Association (Academic Press, 1988), the
graduate text Estimation and Tracking: Principles, Techniques and Software (Artech
House, 1993), the text Multitarget-Multisensor Tracking: Principles and Techniques
(YBS Publishing, 1995), and edited the books Multitarget-Multisensor Tracking:
Applications and Advances (Artech House, Vol. I, 1990; Vol. 11, 1992). He has
been elected Fellow of IEEE for “contributions to the theory of stochastic systems
and of multitarget tracking.” He has been consulting to numerous companies, and
originated the series of Muhitarget-Muhisensor Tracking short courses offered via
UCLA Extension, at Government Laboratories, private companies and overseas.
He has also developed the commercially available interactive software packages
MULTIDATTMfor automatic track formation and tracking of maneuvering
or splitting targets in clutter, PASSDATm for data association from multiple
passive sensors, BEARDATTMfor target localization from bearing and frequency
measurements in clutter, IMDATTMfor image segmentation and target centroid
target and FUSEDATTMfor fusion of possibly heterogeneous multisensor data for
tracking.
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 Co-Chairman 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. In 1987 he received the IEEE CSS Distinguished Member Award.
Since 1995 he is a Distinguished Lecturer of the IEEE AESS. He is co-recipient
of the M. Barry Carlton Award for the best paper in the IEEE Transactions on
Aerospace and Electronic Systems in 1995.

KIRUBARAJAN ET AL.: IMMPDAF FOR RADAR MANAGEMENT AND TRACKING BENCHMARK WITH ECM 1133

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.
 William Dale Blair (S’83-M’85-SM’96) received the B.S. and M.S. degrees
in electrical engineering from Tennessee Technological University in 1985 and
1987. He received the Ph.D. degree in electrical engineering from the University
of Virginia in January 1998.
While a graduate research assistant, Dr. Blair performed robotic controls
research for the Center of Excellence of Manufacturing Research and Technology
Utilization at Tennessee Tech. In 1987 he joined the Naval System Division of
FMC Corporation in Dahlgren, VA as an electrical engineer where his work
involved the development and evaluation of new algorithms for weapons control.
In 1990, Dr. Blair joined the Naval Surface Warfare Center, Dahlgren Division
(NSWCDD) in Dahlgren, VA. At NSWCDD he originated two benchmark
problems for target tracking and radar resource allocation, which served as
themes for invited sessions at the 1994 and 1995 American Control Conferences.
He also led a project that demonstrated through a real-time tracking experiment
that modem tracking algorithms can be utilized to reduce the radar time and
energy required by a phased array radar to support surveillance tracking.
Dr. Blair’s research interests include radar signal processing and control,
resource allocation for multifunction radars, multisensor resource allocation,
tracking maneuvering targets, and multisensor integration and data fusion. Dr.
Blair’s research has been reported in over eighty publications. He has received
technical awards that include an NSWCDD Technical Excellence Award in
1991 for his contributions to the development of innovative techniques for target
trajectory estimation and prediction and the NSWCDD Independent Exploratory
Development Excellence Award in 1993. He served as a lecturer in 1993, 1994,
and 1998 for the UCLA Extension Program in three short courses related to
advanced topics in target tracking. Dr. Blair is the Deputy Editor-in-Chief
and Editor for Radar Systems for the IEEE Transactions on Aerospace and
Electronic Systems. He regularly serves as a reviewer for the ZEEE Transactions
on Automatic Control, American Control Conference, and IEEE Conference
on Decision and Control. He was also co-organizer for invited sessions at the
American Control Conferences in 1993, 1994, and 1995.
Mr. Blair is a member of the IEEE Control Systems, Information Theory, and
Education Societies; Tau Beta Pi, Eta Kappa Nu, and Phi Kappa Phi.

G. A. Watson Photograph and biography not available.

1134 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 4 OCTOBER 1998

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on December 10, 2009 at 11:35 from IEEE Xplore. Restrictions apply.