Tracking Separating Targets with a Monopulse

Radar: Idealized Resolution

T. L. Ogle, W. D. Blair, G. C. Brown
Sensors and Electromagnetic Applications Laboratory,
Georgia Tech Research Institute,
Georgia Institute of Technology, Atlanta, GA 30332-0857

Abstract – The tracking of separating objects with two, and the AOA estimate indicated by the monopulse
a monopulse radar is a particularly challenging prob- ratio for that detection may be outside the angular ex-
lem in that the presence of merged measurements result tent of the two targets [1]. In other words, the mea-
in significant delays in the initiation of tracks on the surement is biased relative to the predicted measure-
new objects. Furthermore, the unknown inflation in ment of either target, and the variance of the measure-
the variance and the bias of the angle-of-arrival esti- ment is inflated relative to that for a single resolved
mates result in extra tracks and hinders track conti- target. Merged measurements often result in errors in
nuity. This paper presents an algorithm for tracking the track filtering and data association algorithms that
separating objects with a monopulse radar. The al- are severe enough to cause tracks to be lost and false
gorithm combines a Neyman-Pearson hypothesis test tracks started. The tracking of separating objects is a
for detecting merged measurements, a method of pars- particularly challenging problem in that the presence
ing merged measurements into two angle-of-arrival es- of merged measurements result in significant delays in
timates, and a track initiation procedure that takes ad- the initiation of tracks on new objects. Furthermore,
vantage of the parsed measurements. The new algo- the unrecognized inflation in the variance and the bias
rithm was implemented in a sophisticated computer of the AOA estimates result in the formation of extra
simulation environment to evaluate the performance tracks and hinders track continuity. Using an artifi-
improvements provided by the new algorithm. The out- cially low measurement variance when the presence of
put of the simulation environment is used to illustrate merged measurements is detected prevents the mea-
the challenges associated with tracking separating ob- surements from associating with existing tracks, which
jects with a monopulse radar. gives rise to undesirable extra tracks. The extra tracks
in effect steal measurements causing a delay in track
Keywords: Unresolved Objects, Target Track- initiation on the new object.
ing, Data Association, Radar Signal Processing,
Monopulse. In [2], Asseo studied the effects of rejecting measure-
ments that exceeded a threshold test that indicated
target multiplicity (i.e., merged measurements). Asseo
1 Introduction found that threshold tests using the magnitude of the
Monopulse is defined as the comparison of two quadrature or sum channels removed large errors in
squinted beams or simultaneous lobes in radar termi- the AOA estimates. This technique resulted in AOA
nology. It is commonly used in radar systems to de- estimates that tended to be closer to the centroid of
termine the angle-of-arrival (AOA) of a target. The the two targets, which was shown to improve aver-
AOA is estimated by comparing either the phase or age accuracy. While disregarding measurement data
the amplitudes of return signals from two beams that when merged measurements are detected was shown
are squinted about boresight, normally in azimuth or to improve tracking performance in [2], disregarding
elevation. A problem arises when the returns from two measurement data when merged measurements are de-
or more closely-spaced targets interfere, which means tected delays track initiation on new objects, which in
the return signals are not resolved in the frequency the case of separating ballistic missiles could be the
and time domains. In the case of merged measure- track on the re-entry vehicle (RV). The focus of this
ments, one detection occurs rather than the expected paper is minimizing the time at which the two objects

1149
are resolved (i.e., the track initiation time on the sec- tracking methodology using the two algorithms. Sim-
ond object). ulation results are given in Section 6, and concluding
Testing for the detection of unresolved objects has remarks are given in Section 7.
been extensively covered in the literature [3]. The test
used for this paper is a conditional generalized likeli- 2 Monopulse Radar Equations
hood ratio test (GLRT) for the detection of the pres-
ence of unresolved Rayleigh targets with a Neyman- and Statistics
Pearson algorithm. The Neyman-Pearson detection In an amplitude comparison monopulse radar sys-
algorithm uses both the in-phase and quadrature por- tem, a pulse of energy is transmitted directly at the
tions of the monopulse ratio and requires no a pri- predicted position of the target, and the target echo is
ori knowledge of the SNR or AOA of either target. received with two beams that are squinted relative to
A GLRT method for detecting merged measurements the predicted position of the target. (Note that two
was developed in [5] using a test statistic based on the beams are required for each angular coordinate.) Tra-
log likelihood ratio, and was shown to have the same ditionally, the AOA of a target is estimated with the
detection performance of the test used in this paper. in-phase part (i.e., the real part) of the monopulse ra-
Rather than computing a single estimate for the tio, which is formed by dividing the difference of the
AOA of the centroid of two unresolved targets, [3] pro- two received signals by their sum. The in-phase and
vides an approach that parses the single estimate for quadrature parts of the sum and difference signals for
two fully unresolved Rayleigh targets into two separate two unresolved Rayleigh targets can be written as
estimates about the centroid. Two objects are consid-
ered to be fully unresolved if the echoed pulses from sI = α1 cos φ1 + α2 cos φ2 + nSI (1)
both objects essentially overlap in time so that all of sQ = α1 sin φ1 + α2 sin φ2 + nSQ (2)
the echoed energy is captured in common samples of dI = α1 η1 cos φ1 + α2 η2 cos φ2 + ndI (3)
the output of the matched filter. This approach is used
dQ = α1 η1 sin φ1 + α2 η2 sin φ2 + ndQ (4)
in this paper to parse measurements once a merged
measurement detection threshold is exceeded. where αi is the Rayleigh distributed amplitude of the
The tracking of closely-spaced and possibly unre- received signal from target i, φi is the phase of the re-
solved targets can be viewed as three distinct cases. ceived signal from target i, ηi is the direction-of-arrival
The first case involves the tracking of targets that (DOA) for target i, and
are initially resolved into separate tracks and the tar-
gets become unresolved later in the scenario. The nSI ∼ N (0, σS2 ) nSQ ∼ N (0, σS2 ) (5)
second case involves targets that are initially unre- ndI ∼ N (0, σd2 ) ndQ ∼ N (0, σd2 ) (6)
solved, while the third case involves a single object
that splits into multiple objects. The third case is The notation N (x̄, σ 2 ) denotes a Gaussian distribu-
most relevant to tracking for ballistic missile defense, tion with mean x̄ and variance σ 2 . For this paper, the
which involves separations of attitude control module errors nSI , nSQ , ndI , and ndQ , are assumed to be inde-
(ACM)/RV from the booster tank and separation of pendent.
the RV from the ACM. The third case is the focus of Letting A and ψ denote the measured amplitude and
this paper. phase of the sum signal gives
For this paper, radar resolution is modeled as ide-
alized in that all of the energy for each target is de- sI = A cos ψ sQ = A sin ψ (7)
tected in a single resolution cell. Thus, two targets
are either completely resolved or the measurements Then the observed SNR will be defined as
are fully merged. This simplifies the problem because
A2
radar returns do not straddle range bins. If bin strad- <o = (8)
dling occurs, it is necessary to centroid measurements 2σS2
before the merged measurement detection algorithm.
Since α1 and α2 are Rayleigh distributed and φ1 and
The case of non-ideal resolution will be the focus of
φ2 are uniformly distributed on (−π, π], sI and sQ are
future work by the authors.
Gaussian random variables. Applying the transforma-
A brief background is given on monopulse radar tion of random variables in (7) to the PDF of sI and
equations and statistics in the next section. Sections 3 sQ gives the PDF of the observed SNR as
and 4 give an overview of the algorithm for detecting
merged measurements and the algorithm for parsing 1 h <o i
merged measurements respectively. Section 5 gives the f (<o |<R ) = exp − , <o ≥ 0 (9)
<R + 1 <R + 1

1150
where <R is the SNR parameter of the Rayleigh signal where the <R1 term is the SNR of the target, and η1
given by is the DOA for that target.
Since yI is a conditional Gaussian random variable un-
E[α12 ] E[α22 ] der H0 or H1 , the ML estimate of ȳI , which is the
<R = + = <R1 + <R2 (10)
2σS2 2σS2 conditional mean of yI under H0 or H1 , is given for N
independent pulses by
where E[·] denotes the expected value and <R1 and
<R2 are the SNR for Rayleigh targets 1 and 2, re- hX
N i−1 X
N N
1 X
spectively. Since the relative RCS of the targets is as- ŷI = <on <ok yIk = <ok yIk
sumed to be known, let <R2 = λ<R1 for λ > 1. Then n=1
N YN
k=1 k=1
<R = (1 + λ)<R1 . For N subpulses at distinct fre- (18)
quencies (i.e., independent), the Maximum Likelihood where <ok and yIk denote the observed SNR and in-
(ML) estimate of <R is given by phase monopulse ratio for subpulse k and YN is given
by (11). Since the yIk are Gaussian random variables,
N ŷI is the minimum variance estimate of ȳI and it is a
b R = YN − 1, 1 X
< YN = <ok (11) Gaussian random variable with variance given by
N
k=1
"N #−1
where <ok denotes the observed SNR for subpulse p. p X p
b R is an unbiased, efficient estimator of <R with σŷ2I = <ok = (19)
Then < 2 2N YN
b R |<R ] = (<R + 1)2 /N. k=1
variance given by var[<
Denoting s = sI + jsQ and d = dI + jdQ , the in- The variance of the monopulse ratio of pulse k under
phase and quadrature parts of the monopulse ratio are hypothesis H0 is
given by
2 p̂
³d´ σ̂0k = (20)
sI dI + sQ dQ 2<ok
yI = Re = (12)
s s2I + s2Q
h σ2 ³ 1 ´ 2i
³d´ sI dQ − sQ dI d
p̂ = + 1+ ŷ (21)
yQ = Im = (13) σs2 b R1 I
<
s s2I + s2Q
The GLRT test statistic, TN , is defined as
where yI and yQ are conditionally independent, Gaus-
sian random variables with a common variance. The 0
TN = XN RN XN p̂−1 (22)
mean of yI is a “power” weighted average of the DOAs
of the two targets, while the mean of yQ is zero.
where
3 Detecting Unresolved Objects XN = [yI − ŷI · · · yIN − ŷI yQ1 · · · yQN ]0 (23)
In [6] a GLRT test is derived for the detection of
target multiplicity (i.e., merged measurments). The
RN = 2 diag[<o1 · · · <oN <o1 · · · <oN ] (24)
GLRT is a hypothesis test with the null hypothesis,
H0 , equal to the case of resolved objects, and H1 equal
to the case of unresolved measurements. The yIk and yQk are the in-phase and quadrature
The PDFs of yI and yQ for a single resolved target monopulse ratios for pulse k, and <ok is the observed
are given in [6] as SNR for pulse k. The TN is chi-square distributed with
³ < p ´ 2N − 1 degrees of freedom, so the detection rule, δ, for
R1
f (yI |H1 , <o , Φ) = N η1 , (14) merged measurements is given as
<R1 + 1 2<o
(
p b2
H0 , T N ≤ C
f (yQ |H0 , <o , Φ) = N (0, ) (15) δ= (25)
2<o b2
H1 , T N > C
where
h σ2 <R1 η12 i
p= d
+ . (16) where C b 2 represents the detection threshold of a chi-
σs2 <R1 + 1 square distributed random variable with 2N − 1 de-
and Φ denotes the parameter set grees of freedom and a given probability of detection.
Chi-square threshold values for several detection prob-
Φ = {<R1 , η1 , σS , σd }, (17) abilities and degrees of freedom are given in [7].

1151
4 Parsing Merged Measure- 5 Tracking Methodology
ments The merged measurement detection and measure-
ment parsing algorithms are placed at the output of
The PDFs of yI and yQ under H1 , for two unresolved
the signal processor. Tracker feedback is used to asso-
Rayleigh targets, are given as
³< η + < η ciate detections to existing tracks in range only using a
R1 1 R2 2 q ´ 2D assignment algorithm. If a detection associates to
f (yI |H1 , <o , Φ) = N , (26)
<R1 + <R2 + 1 2<o an existing track, then a GLRT test is performed for
q target multiplicity. A buffer of the test statistic, TN ,
f (yQ |H1 , <o , Φ) = N (0, ) (27)
2<o in (25) is maintained for all detections that associate
where to the same track. There are N buffers for N tracks.
h σ2 <R1 η12 + <R2 η22 + <R1 <R2 (η1 − η2 )2 i If a given associated detection exceeds the GLRT test,
q = 2d + , then a second GLRT test is performed using the sum
σS <R1 + <R2 + 1
(28) of the last five test statistic values and the appropriate
and Φ denotes the parameter set chi-square threshold for the increased degrees of free-
dom due to the sum. Merged measurements are parsed
Φ = {<R1 , <R2 , η1 , η2 , σS , σd }. (29) only when both tests detect target multiplicity. Mea-
The <<1 and <<2 terms are the SNRs of the two tar- surement parsing is performed with (32) and (33). The
gets, while η1 and η2 are the DOAs for the targets. new detections are then passed on to the local tracker
The larger errors caused by the combined SNR of the with the other detections returned from that dwell.
unresolved targets occur because the targets produce The local tracker uses a 2D assignment algorithm to
a value of <o that is small relative to <<1 <<2 (<<1 + assign measurements to tracks. Measurements are then
<<2 + 1)−1 . processed with an Interacting Multiple Model (IMM)
Since the yIk are Gaussian random variables, ŷI is filter developed for ballistic missile tracking, and the
the minimum variance estimate of ȳI and it is a Gaus- respective tracks are then updated. If measurements
sian random variable with variance given by fail to associate with any existing tracks, then a new
"N #−1 premature track is initiated. The premature track is
q X q classified as ‘tentative’ upon the second measurement
σŷ2I = <ok = (30)
2 2N YN associating to that track (i.e., upon the second mea-
k=1
surement update). If the local tracker fails to update
The DOA estimator for two unresolved targets is a track within five scheduled dwells in different radar
derived in [3], where the first step in parsing a merged resource periods, then the track is dropped. The ten-
measurement is to calculate an estimate of q, given in tative track is not promoted to a ‘firm’ status until
(28), with eleven radar resource periods have passed. Once the
X 0 <N XN
q̂ = N (31) track achieves a ‘firm’ or ‘tentative’ status, the track
2N − 1 is reported for metrics to be scored on that track.
where XN is defined in (23), RN is defined in (24), and
N is the number of subpulses. Then, the DOAs, η, for 6 Simulation Results
the two unresolved targets are given by
s The simulations included a ground X-band radar
<R2 q̃ and a ballistic missile with a 600 km range as shown
η̂1 = ŷI + (32)
ˆR
<R1 < in Figure 1. The missile’s full stack separated into an
s RV and a booster tank at 96.1 seconds. The radar
<R1 q̃ cross section (RCS) of the missile’s full stack, RV, and
η̂2 = ŷI − (33)
ˆR
<R2 < booster tank are all equal (i.e., constant RCS). The
radar’s broadside is pointed directly at the separation
where
 event, and the sensor resolution is modeled as ideal-
σd2

 0, 2
 q̂ ≥ σs2 ized.
q̃ = 4ηbw <R1 <R2 ˆ 4ηbw2
<R1 <R2 ˆ σd2 Simulations were run with no merged measurement
 (<R1 +<R2 )2 <R q̂ ≥ (<R1 +<R2 )2 <R + σs2

 σ2
detection or measurement parsing (i.e., a baseline
q̂ − σd2 , otherwise case). Figure 2 shows the GLRT test statistic for track
s
(34) 1 on Monte Carlo run 1 for this case. The chi-square
ˆ R is given in (11), and ηbw denotes the one-way, half
< threshold used in all cases is based on a 0.01 proba-
power point on the antenna gain pattern of the sum bility of false alarm. The plot shows test exceedances
channel. ηbw is approximated as k2m , where km is the just after separation, and that the second track is ini-
monopulse error slope. tiated at 127.9 seconds. Therefore, it takes more than

1152
 5

4.5

Launch
4
Object resolved
Ground X−band radar
at 127.9 s.
3.5

# of Measurements
X
3
RV separates from
booster tank at 96.1 s. 2.5

2
600 km Ballistic
Missile Trajectory
1.5

0.5
Impact

0
50 65 80 96.1 110 127.9 140 155 170 185 200

Time (s)

Figure 1: Scenario geometry used for simulations. Figure 3: Measurement Counts without Measurement
Parsing.

300

1
250

0.8
200

Completeness achieved
TN Test Statistic

at 130 s.
Tracker Completeness

0.6
150

RV separates from Second track initiated

at 127.9 s. 0.4
100 booster tank at 96.1 s.

0.2
50

0
0
50 65 80 96.1 110 127.9 140 155 170 185 200
50 65 80 96.1 110 125 140 155 170 185 200
Time (s)
Time (s)

Figure 2: GLRT Test Statistic for Az and El without Figure 4: Tracker Completeness Metric without Mea-
Measurement Parsing. surement Parsing.

30 seconds to resolve the objects in this case. Figure 4 shows the tracker completeness metric for 100
A count of the measurements at the output of the Monte Carlo experiments with no merged measure-
signal processor sent to the tracker as a function of ment detection or measurement parsing. As expected,
time is shown in Figure 3. The plot shows that from the average tracker completeness metric drops to 0.5
96.1 seconds to approximately 116 seconds the RV and just after separation, and the tracker does not regain
the booster tank are unresolved. From approximately full completeness of 1 until 130 seconds.
116 seconds to 133 seconds the objects are oscillat- The redundant track ratio is defined as the ratio of
ing between resolved and unresolved, and beyond 133 total number of tracks to the total number of valid re-
seconds the objects are fully resolved. Measurement portable objects. Figure 5 shows the redundant track
counts greater than 2 indicate the presence of a false ratio averaged over 100 Monte Carlo runs for the case
alarm. with no measurement parsing (i.e., the baseline). Re-
Tracker completeness is defined as the ratio of the dundant tracks are present due to the oscillation be-
total number of tracks held at a given time to the to- tween resolved and unresolved measurements.
tal number of valid reportable objects at that time. The GLRT test statistic for track 1 of Monte Carlo
In other words, a tracker completeness of 1 is perfect, run 1 for the case with merged measurement detec-
because we have one track for every trackable object. tion and measurement parsing is shown in Figure 6.

1153
 1.05

1.04

1.03 5
Redundant Track Ratio

4.5

1.02
4
Second track initiated
at 101.5 s.
3.5

1.01

# of Measurements
3

2.5
1

2
50 65 80 96.1 110 125 140 155 170 185 200
Time (s) 1.5

Figure 5: Redundant Track Ratio without Measure-

0.5
ment Parsing.
0
50 65 80 96.1 110 125 140 155 170 185 200

Time (s)
300
T/P AZ
T/P EL
Chi Square Figure 7: Measurement Counts with Measurement
250
Parsing.

200
TN Test Statistic

150
Second track initiated
at 101.5 s.
RV separates from
booster tank at 96.1 s.
100

50

0
50 65 80 96.1 110 125 140 155 170 185 200

Time (s)

Figure 6: GLRT Test Statistic for Az and El with Mea-

surement Parsing. 0.8

Completeness achieved
at 119 s.
Tracker Completeness

0.6

The second track is initiated at 101.5 seconds. The 0.4

plots show that the GLRT test is exceeded well after
the second track is initiated. Again, this is due to the
0.2
oscillation between resolved and unresolved measure-
ments.
0
The measurement counts in Figure 7, show that from
96.1 seconds to 101.5 seconds the objects are unre- 50 65 80 96.1 110 125 140 155 170 185 200
Time (s)
solved, from 101.5 seconds to 123 seconds the mea-
surement count oscillates between resolved and unre-
Figure 8: Tracker Completeness Metric with Measure-
solved objects, and beyond 123 seconds the objects are
ment Parsing.
resolved. In this case, it is shown that measurement
parsing provides resolved measurements to the tracker
earlier than the baseline case, but it also provides more
false alarm measurements to the tracker after the ob-
jects are fully resolved.
The tracker completeness metric for several Monte

1154
 1.05
References
[1] S.M. Sherman, Monopulse Principles and Tech-
1.04
niques, Artech House, Inc., Dedham, MA, 1984.

1.03
[2] S. J. Asseo, “Effect of Monopulse Signal Thresh-
Redundant Track Ratio

olding on Tracking Multiple Targets,” IEEE

Trans. on Aero. Elect. Sys., pp. 504-509, Nov.
1.02
1974.

1.01 [3] W. D. Blair and M. Brandt-Pearce, “Monopulse

DOA Estimation for Two Unresolved Rayleigh
1
Targets,” IEEE Trans. on Aerospace and Elec-
tronic Systems, Vol. 37, No. 2, pp. 452-469, Apr.
0 20 40 60 80 100
Time (s)
120 140 160 180 200
2001.
[4] W. D. Blair and M. Brandt-Pearce, “Unresolved
Figure 9: Redundant Track Ratio with Measurement
Rayleigh target detection using monopulse mea-
Parsing.
surements,” IEEE Trans. on Aerospace and Elec-
tronic Systems, Vol. 34, No. 2, pp. 543-551, Apr.
1998.
Carlo experiments is shown in Figure 8 for this case. [5] A. Sinha, T. Kirubarajan, and Y. Bar-Shalom,
When averaged over 100 Monte Carlo experiments, the “Maximum Likelihoo Angle Extractor for Two
tracker regains completeness quicker than the baseline Closely Spaced Targets,” IEEE Trans. on
case, with a completeness of 0.99 regained by 119 sec- Aerospace and Electronic Systems, Vol. 38, No.
onds. In Figure 9, the redundant track ratio for the 1, pp. 183-203, Jan. 2002
case with measurement parsing shows a significant in-
crease in redundant tracks, which could be reduced by [6] W. D. Blair, “Monopulse Processing for Tracking
delaying the tentative status on the track. Unresolved Targets,” Naval Surface Warfare Cen-
ter, Dahlgren Division, Dahlgren, VA 22448-5000,
September 1997.
7 Conclusions
This paper presents a method for tracking separating [7] Y. Bar-Shalom and X.R. Li, Multitarget-
targets with a monopulse radar under idealized reso- Multisensor Tracking: Principles and Techniques,
lution. The algorithm presented combines a Neyman- YBS Publishers, Box U-157, Storrs, CT, 06296-
Pearson hypothesis test for detecting merged measure- 3157, 1995.
ments, a method of parsing merged measurements into [8] X. Zhang, P. Willett, and Y. Bar-Shalom,
two AOA estimates, and a track initiation procedure “Monopulse Radar Detection and Localization of
that takes advantage of the parsed measurements. The Multiple Targets via Joint Multiple-Bin Process-
tracking of separating objects is a particulary challeng- ing,” Proceedings of the 2003 IEEE Radar Con-
ing problem in that the presence of merged measure- ference, pp. 232-237, May 2003.
ments results in significant delays in the initiation of
tracks on the new objects, which in the case of separat-
ing ballistic missiles could be the track on the re-entry
vehicle (RV). The results from simulations show that
over 100 Monte Carlo experiments, an earlier track ini-
tiation time for the second object is achieved. In this
paper, the average track initiation time on the sec-
ond object was reduced by 11 seconds. However, the
algorithms do result in a higher false alarm rate af-
ter the objects are fully resolved, which in this case
caused redundant tracks. Mitigating the the redun-
dant tracks and considering non-ideal resolution using
the technique of [8] will be the focus of future work by
the authors.

1155