Kalman Tracking and Bayesian Detection For RADAR RFI Blanking
Kalman Tracking and Bayesian Detection For RADAR RFI Blanking
Kalman Tracking and Bayesian Detection For RADAR RFI Blanking
T
,
F =
_
_
1 0 T 0
0 1 0 T
0 0 1 0
0 0 0 1
_
_
,
G =
_
T
2
2
0 T 0
0
T
2
2
0 T
_
T
,
where x
n
is the state vector for a single aircraft, x
n
and
x
n
represent position and velocity respectively in the x
direction, F is the state transition matrix, and G is the
input distribution matrix.
The measurement model relates polar observations,
z
n
, to the state equation coordinates and includes ob-
servation measurement noise:
z
n
= h(x
n
) +v
n
, where (2)
h(x
n
) =
__
x
2
n
+ y
2
n
tan
1 yn
xn
_
,
v
n
= [v
r
(n), v
(n)]
T
,
r
n
and
n
are the measured range and bearing respec-
tively at snapshot n, with corresponding measurement
noise v
r
(n) and v
(n).
B. Kalman Prediction Equations
Given echo detections and associated track history up
to snapshot n, the rst step in a Kalman lter iteration is
to predict the next state vector and update the prediction
error covariance estimate as follows:
x(n + 1|n) = F x(n|n), (3)
P(n + 1|n) = FP(n|n)F +GQG, (4)
where
x(n|n) = ltered state estimate at t
n
given data through t
n
,
P(n|n) = ltered state error covariance,
x(n + 1|n) = predicted state estimate,
P(n + 1|n) = predicted state error covariance,
Q = process noise/acceleration covariance.
x(n|n) and P(n|n) are computed using lter equa-
tions, (6)(8), presented in Section II-C. Note that the
prediction point, ( x
n+1|n
, y
n+1|n
), is given by the rst
two elements of x(n + 1|n). We dene the elliptical
region, S
n+1
to be centered on this point and to have
radii r
x
and r
y
proportional to
_
P
1,1
(n + 1|n) and
_
P
2,2
(n + 1|n) respectively. Thus the larger the pre-
diction error variance, the larger S
n+1
grows to represent
our uncertainty as to where the next RADAR echo will
be detected.
Figure 3 illustrates this behavior. The plot shows track
evolution for real GBT data over ve snapshots for a
dense scene with multiple, overlapping aircraft tracks.
The ellipses show prediction regions, S
n+1
for each es-
tablished track. Note the variety of sizes, corresponding
to variations in track quality. For real-time processing,
these prediction regions would be blanked for the next
expected echoes, even before they are detected. However
for post processing, the prior information represented
by these ellipses can be used for a Bayesian combined
tracking-with-detection algorithm to improve sensitivity
to weak pulses.
Two successive associated pulse detections are re-
quired before a track can be initiated. x(0|0) is initialized
with the position of the second detection, and a two
sample velocity estimate computed from the position
difference between the detections.
To nd a practical initialization for P(0|0), we ran the
tracker on synthetic detection data which simulated the
aircraft motion seen in the real GBT data. After a large
number of Monte Carlo random trials, P(n|n) converged
on average to P(|) (2 10
3
)I. This value was
used to initialize P(0|0) when processing real data from
GBT.
In simulation, a
n
= [a
x
(n), a
y
(n)]
T
was generated
by lowpass ltering two mutually independent Gaussian
white noise time sequences (one each for a
x
(n) and
4
3 2.5 2 1.5 1 0.5 0 0.5 1 1.5 2
x 10
4
2
3
4
5
6
7
8
9
x 10
4
xy coordinate
x (meter)
y
(
m
e
t
e
r
)
detection
prediction
estimation
Fig. 3. An example of Kalman tracking performance for data
acquired at the GBT. Four aircraft tracks have been automatically
established and plotted, including a pair of crossing tracks. Data from
ve snapshots is shown with T 24s. The nal point plotted for each
track in the prediction point, ( x
n+1|n
, y
n+1|n
). Prediction regions,
Sn+1, shown by the dashed ellipses vary in size according to track
quality. Note that the center track has a large Sn+1 due to a missed
detection. Predictive real-time blanking is accomplished by excising
the prediction region data.
a
x
(n)). A lter cutoff of f
c
=
1
100T
produced smooth
simulated aircraft turning maneuvers which were consis-
tent with tracks seen in the real GBT data. The process
covariance is modeled as a constant matrix of the form
Q = E{a
n
a
T
n
} =
_
2
x
2
xy
2
xy
2
y
_
. (5)
Values of
2
x
=
2
y
= 12.0 and
2
xy
=
2
yx
= 0 were
used in the results described below and were established
by qualitatively matching simulated tracks with the real
GBT data.
C. Kalman Update (Filter) Equations
When a new snapshot of detections from CLEAN is
available, the Kalman update step completes the iteration
begun with equations (3) and (4) as follows:
x(n + 1|n + 1) = x(n + 1|n) +K(n + 1)
_
z(n + 1) h
_
x(n + 1|n)
_
_
, (6)
P(n + 1|n + 1) =
_
I K(n + 1)H(n + 1)
_
P(n + 1|n), (7)
K(n + 1) = P(n + 1|n)H
T
(n + 1)
_
H(n + 1)P(n + 1|n)H
T
(n + 1) +R
1
,(8)
where
H(n + 1)
=
_
h
x
_
x= x(n+1|n)
(9)
=
_
1
_
x
2
(n + 1|n) + y
2
(n + 1|n)
_
_
x(n + 1|n) y(n + 1|n) 0 0
y(n + 1|n) x(n + 1|n) 0 0
_
,
and where K(n + 1) is the Kalman gain matrix. Range
and bearing measurement noise are assumed to be mutu-
ally independent, so the measurement error covariance,
R, has form
R =
_
2
r
0
0
2
_
. (10)
2
r
is proportional to receiver noise variance and the
square of transmit pulse length.
2
is proportional noise
and the square of transmit antenna angular rotation
rate divided by transmit pulse repetition rate. We have
estimated these parameters empirically for the GBT data
and treat them as constants in the Kalman update.
After computing (6), (7), and (8), index n is incre-
mented to complete the iteration which started with
prediction equations (3) and (4).
D. Track Management
In practical multiple target automatic tracking applica-
tions it is necessary to deal with a number of ambiguities
when interpreting the pulse detection data. We have
developed a set of rulebased procedures (more fully
described in [14]) to address the following issues.
1) Track Creation and Association. For each new
snapshot the CLEAN algorithm produces a set
of detections. Each of these must be classied
as being a newly detected aircraft for which a
track must be created, or as belonging to an
existing track. Detections which lie within a xed
distance, d
a
, from an existing tracks prediction
point, ( x
n+1|n
, y
n+1|n
), are associated with that
track. To avoid ambiguities, detections which sat-
isfy this criterion for two or more distinct tracks
are assigned to the track whose prediction point is
closest.
Two successive associated detections are required
to start a track. New detections which are not
within a distance d
a
of any existing prediction
point are designated candidate starting points. The
track is created if in the succeeding snapshot a
detection within distance d
n
of the candidate point
is found which is not associated with any existing
track. Since no velocity information is available
5
from a single candidate point, it is necessary that
d
n
>> d
a
. In practice d
n
is set to be the maximum
distance an aircraft can travel in T seconds.
2) Missed Detections. If in a given snapshot no new
detection is associated with a particular track, it
is assumed that the aircraft is still present, but
that the detection was missed due to random
variation in echo amplitude. Consider missing k
successive detections for a given track. In this
case, Kalman prediction equations (1), (3), and
(4) from prior snapshot, t
nk
, are recomputed
as a multistep prediction by replacing T with
kT. This produces the desired kstep prediction
point ( x
n+1|nk
, y
n+1|nk
) but the prediction error
covariance, P(n+1|nk) increases, and the size
of S
n+1
grows as compared to a normal single step
prediction.
3) Track Dropping. A track which has no associated
detections in three successive snapshots is termi-
nated.
4) Track Splitting. If two or more new detections are
associated with a single track then the track is split
into separate tracks for each new detection. These
split tracks have a common history for t t
n1
but for t t
n
are computed as distinct tracks. This
scenario arises when aircraft paths cross or when
a new aircraft detection occurs close to an existing
track prediction point.
III. A BAYESIAN COMBINED KALMAN TRACKER
AND DETECTOR
In a conventional RADAR detector, all range-bearing
bins are assumed to be equally likely to contain an
echo. Detections are made when the magnitudesquared
matched lter output exceeds a predetermined constant
threshold, .
1
Thresholds are set to yield a specied
probability of false alarm (PFA). For a xed PFA, the
probability of detection (PD) is a function of the receiver
design and signal statistics, such as signal to noise ratio.
In the context of the Kalman tracker, one need not
assume all bins have the same probability of detection.
The track histories provide prior information which
indicates a higher probability of echoes being detected in
prediction regions, S
n+1
. We propose a Bayesian detec-
tion scheme where a spatially dependent prior probability
density function for the presence of an echo, p(x, y),
is computed using the S
n+1
ellipses to designate areas
1
The widely used constant false alarm rate (CFAR) detector scales
a constant, c, by a local estimate of noise and clutter standard
deviation in order to form a variable threshold, (r, ). The Bayesian
detection process described here can still be applied to c to achieve
the claimed effect.
Fig. 4. The threshold is determined by the prior distribution p(x, y)
of the presence of pulse. 0 is the constant threshold outside the
elliptical regions, Sn+1. (x, y) is at a local minimum corresponding
to each prediction point. The two concavities represent decreased
threshold according to the prior probability inside two different sized
Sn+1.
of increased density. With this approach it is possible
to increase the overall PD without an increase in PFA.
A detailed theoretical development of this detector is
found in [14] and in a forthcoming paper. Here it is
simply noted in summary that the effect of the Bayesian
detector is to make the detection threshold, (x, y)
spatially varying, with local minima at the prediction
point centroids of the S
n+1
regions, as illustrated in
Figure 4. Section IV presents a comparison of spectral
density estimates with conventional and the proposed
Bayesian detectors. The proposed scheme produces less
RADAR pulse bias in the spectral estimate.
IV. EXPERIMENTAL RESULTS
A set of real data recorded at the GBT for a 10 MHz
wide band around 1292 MHz was used to test the echo
detection algorithm, tracking, and blanking performance.
A sequence of les representing a continuous block
of 10 minutes of data (50 RADAR antenna rotations)
was recorded in January, 2003. By using the tracker
information, the new Bayesian algorithm is able to detect
some weaker echoes which are miss-detected by the
conventional method. This was accomplished without
increasing the false alarm rate signicantly. All blanking
was implemented by zero-stufng [1], that is, placing
zeros into the time samples where RADAR interference
is detected.
The ability to excise RADAR interference was eval-
uated for three different blanking techniques: 1) sim-
ple time window blanking, 2) combined time window
and detected pulse blanking using conventional pulse
detection, and 3) combined time window and detected
pulse blanking using the proposed Bayesian detector.
The Kalman tracker was used for both detected pulse
blanking methods.
Figure 5 shows the unblanked spectrum accumulated
over a 1.2 second window which contained terrain echo
clutter and aircraft echoes. Note the dominant spectral
peak around 5.5 MHz caused by terrain echoes. The
6
0 1 2 3 4 5 6 7 8 9 10
0
2
4
6
8
10
12
Frequency in MHz
R
e
la
t
iv
e
P
o
w
e
r
unblanked
time window blanking ONLY
Fig. 5. The unblanked and time window blanked spectra integrated
over 1.2 seconds while the RADAR beam was sweep overhead at the
GBT.
spectrum with simple time window blanking is also
shown, and at this scale RFI appears to have been
eliminated.
However, at the expanded scale of Figure 6, a dif-
ference is seen between spectral estimates using the
new Bayesian detection as compared with conventional
detection. In this particular data window there were no
strong aircraft echoes so there is little difference between
time window blanking and conventional detected pulse
blanking. There are three weak aircraft echoes which
were detected with the new algorithm, and the resulting
spectrum shows reduced bias near 5.5 MHz correspond-
ing to the RADAR pulse center frequency.
Detected pulse blanking using Kalman lter tracking
techniques has been shown to reduce RFI bias due to
RADAR pulses in GBT observations. Also, the new
Bayesian combined tracking and detection algorithm has
been shown to improve blanking of weak aircraft echo
pulses.
REFERENCES
[1] Q. Zhang, Y. Zheng, S. Wilson, J. Fisher, and R. Bradley,
Combating pulsed radar interference in radio astronomy, The
Astronomical Journal, vol. 126, pp. 15881594, 2003.
[2] J. Fisher, Summary of r data samples at green bank, National
Radio Astronomy Observatory, Green Bank Observatory, Tech.
Rep., 2001.
[3] , Analysis of radar data from february 6, 2001, National
Radio Astronomy Observatory, Green Bank Observatory, Tech.
Rep., 2001.
[4] S. Ellingson and G. Hampson, Mitigation of radar interfernece
in l-band radio astronomy, The Astrophysical Journal Suple-
ment Series, vol. 147, pp. 167176, 2003.
[5] , R and asynchronous pulse blanking in the 1230-1375
mhz band at arecibo, Ohio State University, Tech. Rep., 2002.
[6] P. Ravier and R. Weber, Robustness in r detection for time-
blanking, LESI and ESPEO, University of Orlans, France,
Tech. Rep., 2000.
0 1 2 3 4 5 6 7 8 9 10
0.4
0.41
0.42
0.43
0.44
0.45
0.46
0.47
Frequency in MHz
R
e
la
t
iv
e
P
o
w
e
r
time window + conventional pulse detected blanking
time window + improved pulse detected blanking
Fig. 6. The spectra with simple time window blanking combined
with detected pulse blanking (using new and conventional echo
detection algorithms, respectively).
[7] A. Leshem and A.-J. v.d. Veen, Introduction to interference
mitigation techniques in radio astronomy, in Perspectives in
Radio Astronomy, Technologies for Large Antenna Arrays,
A. Smolders and M. van Haarlem, Eds. Dwingeloo, The
Netherlands: NFRA, 1999.
[8] A. Boonstra, A. Leshem, A.-J. van der Veen, A. Kokkeler, and
G. Schoonderbeek, The effect of blanking of tdma interference
on radio-astronomical observations: experimental results, in
Proc. of the IEEE International Conf. on Acoust,, Speech, and
Signal Processing, vol. 6, 2000, pp. 35463549.
[9] A. Leshem, A.-J. v.d. Veen, and E. Deprettere, Detection and
blanking of gsm interference in radio-astronomical observa-
tions, in Proceedings of the 2nd IEEE Workshop on Signal
Processing Advances in Wireless Communications, 1999, pp.
374377.
[10] K.V.Ramachandra, Kalman Filtering Techniques for Radar
Tracking. Marcel Dekker, 2000.
[11] P. Zarchan and H. Musoff, Fundamentals of kalman ltering:
A practical approach, Progressin Astronautics and Aeronautics,
vol. 190, pp. 257291, 2000.
[12] B. Mahafza, Radar Systems Analysis and Design Using MAT-
LAB. Chapman and Hall/CRC Press, 2000.
[13] J. H ogbom, Aperture synthesis with a nonregular distribution
of interferometer baselines, Astronomy and Astrophysics Sup-
plement, vol. 15, p. 417, 1974.
[14] W. Dong, Time blanking for gbt data with radar r, Masters
thesis, Brigham Young University, 2004.