Sei sulla pagina 1di 10

80 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 5, NO.

1, FEBRUARY 1997

Radar Tracking for Air Surveillance in a Stressful


Environment Using a Fuzzy-Gain Filter
Keith C. C. Chan, Member, IEEE, Vika Lee, and Henry Leung, Member, IEEE

Abstract—In this paper, we present a fuzzy-gain filter for target target dynamics. Many Kalman filter-based algorithms that are
tracking in a stressful environment where a target may accelerate developed to track maneuvering targets are, for example, of
at nonuniform rates and may also complete sharp turns within a such nature. They make assumptions about (process noise),
short time period. Furthermore, the target may be missing from
successive scans even during the turns, and its positions may be (system noise) and (measurement noise) [1], [5], [6], [11].
detected erroneously. The proposed tracker incorporates fuzzy Some of them also make assumptions of constant velocity
logic in a conventional - filter by the use of a set of fuzzy if-then and acceleration [2]–[6], [8], [10]. Some require that the
rules. Given the error and change of error in the last prediction, duration between scans (the data sampling interval) be small
these rules are used to determine the magnitude of and :
The proposed tracker has the advantage that it does not require
[2]–[3], [5]–[7], [9]–[10]. And some require the probability of
any assumption of statistical models of process and measurement detection be either unity or independent in successive scans
noise and of target dynamics. Furthermore, it does not need a [1]–[10].
maneuver detector even when tracking maneuvering targets. The It is, perhaps, because of the validity of these assumptions
performance of the fuzzy tracker is evaluated using real radar that many sophisticated model-based tracking algorithms have
tracking data generated from F-18 and other fighters, collected
jointly by the defense departments of Canada and the United not shown markedly improved performance over those based
States When compared against that of a conventional tracking on the simple fixed-gain - filter [12]. The - trackers are
algorithm based on a two-stage Kalman filter, its performance is still quite well received in the tracking community despite
found to be better both in terms of prediction accuracy and the the fact that they have their own limitations. For example, al-
ability to minimize the number of track losses. though the optimality regions are known, the exact magnitude
Index Terms—Fuzzy filter, fuzzy logic, fuzzy rules, fuzzy track- of and are often hard to determine. Different - trackers
ing. with different ’s and ’s may have to be used in response to
acceleration or deceleration or before, during, and after a turn.
I. INTRODUCTION Furthermore, to cope with maneuvering a maneuver detector is
often required in conjunction with an - tracker. In addition

T ARGET tracking is the prediction of the future trajectory


of an object based on its past states. How accurate
predictions can be made, therefore, depends on how accurate
to and therefore, there is also the need to decide when
the maneuver detector should be triggered.
These decisions can be very difficult in the case where the
a target’s past and present positions can be measured. Unfor-
targets being tracked have high dynamics (e.g., military aircraft
tunately, due to clutter and other limitations of the sensors
such as F-18’s). These targets may move undetected for several
and the signal processing algorithms, the positions of a target
successive scans when they make sharp turns. In such cases,
cannot be measured with perfect accuracy. The measurements
the time delay (typically of at least several scans) associated
can exhibit different random characteristics, and they can be
with the use of a maneuver detector will very likely result in
biased depending on a target’s angle of flight and its relative
track losses. By the time a target is detected to be making
position to the radar. Furthermore, the fact that a target such as
a turn, it would have already completed before the tracker is
a military aircraft may thrust, yaw, pitch, and roll, and that it
able to respond.
may deviate from a constant velocity trajectory due to aircraft
To overcome the problems associated with target tracking
control and turbulence, making target tracking all the more
in a stressful environment where a target may: i) accelerate
difficult.
To cope with these problems many methods have been or decelerate at nonuniform rates any time during its course,
developed. These methods are by and large model based. ii) perform sharp turns (i.e., high-alpha maneuvers) within a
They are usually developed under various assumptions about short time period (say, within 3–4 scans), iii) not be detected in
statistical models of process and measurement noise and about successive scans (i.e., missing observations) during the turns
or anywhere else along the track, iv) the target’s positions
inaccurately measured or detected to the extent of over 100%
Manuscript received August 15, 1995; revised April 15, 1996. deviation in the opposite direction from the real position, we
K. C. C. Chan and V. Lee are with the Department of Computing, The propose a tracking method based on a fuzzy-gain filter. This
Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong. method incorporates fuzzy logic in a conventional - filter by
H. Leung is with the Surface Radar Section, Radar Division, Defense
Research Establishment Ottawa, Ottawa, ON, K1A 0K2, Canada. the use of a set of fuzzy if-then rules. Given the position error
Publisher Item Identifier S 1063-6706(97)00040-4. and change-of-error in the last prediction, these rules are used
1063–6706/97$10.00  1997 IEEE
CHAN et al.: RADAR TRACKING FOR AIR SURVEILLANCE IN A STRESSFUL ENVIRONMENT 81

to determine the smoothing gains of and when making Equations (1)-(3) are used directly when an observation is
the next prediction. received on scan However, in the case where the probability
Based on the - filter, the proposed fuzzy tracker has, of detection is less than unity, there may be scans for which a
on one hand, the advantages which a typical - tracker target is not detected. Then, the smoothed position is set equal
has as well (i.e., it does not require any assumptions about to the prediction as follows:
noise models or target dynamics and is simple and relatively
fast). On the other hand, the fuzzy tracker does not suffer (7)
from the same limitations. It is capable of quickly adjusting (8)
the magnitude of and in response to changes in speed
and direction without a maneuver or an acceleration detector, This effectively amounts to setting The
and for this reason the proposed tracker is able to eliminate prediction for the next scan is computed as before.
any delay. Also, with fuzzy inference it can make better The predicted target position is simply a linear extrapolation
decisions by taking into consideration several different or even from the previous smoothed position with a constant velocity.
conflicting situations at the same time. As a result, it can The smoothing in an - filter is performed in each coor-
minimize the number of false alarms even if the data sampling dinate separately. In other words, (1)–(8) are also generalized
interval is relatively long, and a target is not detected during to the coordinate as follows:
the turns.
(9)
The performance of many model-based tracking methods
are often evaluated using simulated data. Their usefulness (10)
in real-tracking environment which usually depart from ideal
is, therefore, questionable. To better evaluate the practicality (11)
of the proposed fuzzy tracker, we used several sets of real- (12)
life radar tracking data collected by the joint effort of the (13)
defense departments of Canada and the United States. The
data sets consist of tracks generated by F-18 and other fighters. (14)
Using a two-stage Kalman filter as benchmark, we evaluated
(15)
the performance of the proposed fuzzy tracker. Experimental
results show that it is more robust in terms of its ability to (16)
prevent track losses and minimize prediction error.
where is the coordinate of the target’s observed
position at the th scan, is the coordinate of the target’s
II. THE - TRACKER predicted position at the th scan, is the coordinate of
The - tracker is probably the most extensively applied the smoothed target position at the th scan, and is the
fixed-coefficient filter. This filter can be defined by the fol- smoothed target velocity in the direction at the th scan.
lowing equations [12]: The range of smoothing gains and can be obtained from
the characteristic equation of the filter, i.e.,
(1) Applying the jury’s test yields the stability region
which provides admissible , values [12], [13]
(2)
and and (17)
(3)
While the stability region is known, it may not always be easy
where is the coordinate of the target’s observed for the exact and to be determined. If the coefficients are
position at the th scan, is the coordinate of the too small, the filters have good noise reduction ratio but can
target’s predicted position at the th scan, is the be too unresponsive to dynamic targets. If the coefficients are
coordinate of the smoothed target position at the th scan, too large, the filters have good following capabilities but can
is the smoothed target velocity in the direction at the be too sensitive to noise. In the case of a target moving in
th scan, is the radar scan time or the sampling interval, and a straight line with constant velocity in a relatively nonnoisy
are the fixed-coefficient filter parameters. The quantity environment, it may be easy for some fixed and to be
is normally defined to be unity, but in the case of missing determined; but in the case a target moving in a relatively noisy
observations, its value may be taken as the number of scans environment and its flight dynamics continuously changing,
since the last measurement. the coefficients cannot be fixed. Decisions as to what and
The usual initialization process for an - tracker is defined should be cannot be easily made. At some points during the
by course, they may need to be set relatively large, but at some
other points they need to be set relatively small. For example,
(4)
if a target deviates from a straight path and is maneuvering, the
(5) coefficients should be set relatively large to make sure that the
tracker follows the target closely. However, if the deviation
(6)
is due to measurement noise or detection error rather than the
82 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 5, NO. 1, FEBRUARY 1997

target maneuvering, and should be set relatively small so and


as to make sure that the tracker stays on the right track.
In other words, therefore, when a target is detected to deviate
from a straight line path, and can be set either large
or small depending on whether the deviation is due to noise If
or maneuvering. To avoid false alarms on one hand while
following a maneuvering target closely on the other, decisions
regarding the choice of and cannot be made based on If
binary logic. Decision boundaries, instead, have to be fuzzy so
that the chance of track losses can be reduced to a minimum by If
reducing the transient magnitude caused by an abrupt change
of decisions. (21)

III. THE FUZZY-GAIN FILTER


and their values lie in Based on and ,
To define fuzzy decision boundaries for target tracking in a and can be defined as follows:
stressful environment, in this section we describe a fuzzy-gain
filter that is characterized by a set of fuzzy rules. Based on
the error and change of error in the last prediction, these rules
determine what magnitude and in an - filter should be If
set to when making the next prediction. Any change in target
dynamics or the tracking environment will, therefore, result in
changes in and If

A. System Variables If
The rules of the proposed fuzzy filter are expressed in
terms of two input variables and two output variables. The (22)
input variables and are defined in terms of the
prediction error and change of error at the th and
scan where
(18)
and If
(19)
and and If
The range of values for and
for are dependent on how
large the deviation can be between the measured and predicted If
positions of the target at the th scan. and
can theoretically be zero and and can (23)
be much greater than the maximum velocity of the target.
Since different targets have different velocity, the magnitude As a result, we define and to be
of and may vary from target to target. To
design a general fuzzy tracker which can be used with different
types of targets, we normalize the and components of (24)
and so that instead of and , we compute
and where
and

If
(25)

If
The universe of discourse and, hence, the range of values
which and each may take is in the interval [0,
If 1] regardless of the target types and their maximum velocity.
Values of and that fall into this range represent
(20) crisp values of and
CHAN et al.: RADAR TRACKING FOR AIR SURVEILLANCE IN A STRESSFUL ENVIRONMENT 83

Fig. 1. Membership functions for the fuzzy sets of E (k) and 1E (k):
0 0
Fig. 2. Membership functions for the fuzzy sets of and :
TABLE I
B. Membership Functions
FUZZY ASSOCIATIONS FOR (k)
The crisp values are mapped into some fuzzy sets defined in
the universe of discourse of and These fuzzy
sets are labeled in the linguistic terms of zero (ZE), small pos-
itive (SP), medium positive (MP), and large positive (LP). The
specific membership function for each fuzzy set determines the
meaning of these linguistic terms. These membership functions
are defined on the underlying universe of discourse of
and by the trapezoidal functions shown in Fig. 1. The TABLE II
grades of membership are assigned to different values of FUZZY ASSOCIATIONS FOR (k)
and based on these functions.
In determining the membership functions of each of these
fuzzy sets, we decided that the core of ZE be in [0.0, 0.1],
that of SP be in [0.3, 0.4], that of MP be in [0.6, 0.7] and that
of ZE be in [0.9, 1.0]. The intervals between the cores of the
fuzzy sets (which are all of equal width) are made boundaries
of them. Such choice of membership functions ensures that
the chance of each input data point falling into each fuzzy set universe of discourse in the sense that they are much faster
be roughly equal, and that the probability of each fuzzy rule in deciding what fuzzy values and should take on given
being fired is, as a result, more or less the same. specific values for and
While the antecedents of the fuzzy rules are made up of
and , their consequences are expressed in terms C. The Fuzzy Rules
of and where and are the coefficients of With the input and output variables defined above, the fuzzy
the proposed fuzzy - filter at the th scan. Even though, rules can be expressed as follows:
theoretically, and can take on a wide range of
IF AND
values, it is seldom necessary in practice. For the proposed
tracker, the universe of discourse for both coefficients are THEN AND
taken to be [0, 1]. where and are fuzzy sets defined in the universe
The fuzzy sets for both and are labeled in the of discourse of , and , respectively.
linguistic terms of ZE, SP, MP, LP, very large positive (VP), Since every such rule used by the proposed fuzzy tracker
and extremely large positive (EP). The specific membership can be considered as an implication that defines a fuzzy
function for each fuzzy set determines the meaning of the association, they are concisely written as a linguistic map or
linguistic terms and is defined on the underlying universe of fuzzy associative memory here, as shown in Tables I and II.
discourse of and by the triangular functions shown Each entry in the table associates input fuzzy sets with output
in Fig. 2. fuzzy sets and some proper and can be inferred for any
Unlike that of and the universe of discourse input in the universe of discourse.
is divided up into six instead of four fuzzy sets, and the core The rules (i.e., the table entries) were obtained by inter-
of these fuzzy sets are not equally spaced. The central region viewing some defense experts. Three interview sessions were
is more finely divided into four different fuzzy sets, SP, MP, conducted initially. The first two (which took about an hour
LP, and VP. This is because the region for optimum is each) involved two different individuals. Even though they
found more often in [0.3, 0.8] on one hand [2], [17], [18], were interviewed separately, the rules obtained were rather
and the movement of the tracker needs to be made relatively consistent. This is especially true with the more extreme cases
small to avoid false alarms on the other. Furthermore, when involving ZE’s and LP’s in the case of and ,
interviewing defense experts to obtain the fuzzy rules, they and ZE’s and EP’s in the case of and To resolve
seem to be more comfortable with such partitioning of the conflict, an extra session with both defense experts present
84 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 5, NO. 1, FEBRUARY 1997

were held. After about two hours of discussions (including of target dynamics. In this case, an “over prediction” of
one of the authors presenting the idea about a fuzzy-gain filter) the target position based on the current trend is made. If
some consensus was concluded. To ensure that the best set of the guess of flight direction is correct, the tracker should
fuzzy rules be obtained, both experts suggested testing the recover immediately.
rules against some simulated data. After many iterations of
trials and errors with a number of consultative sessions held IV. PERFORMANCE EVALUATION
with these experts, a set of fuzzy rules was finalized. These
To evaluate the performance of our fuzzy tracker, several
rules are able to trigger the best response to accommodate
sets of real data collected jointly by the defense departments
all flight patterns and dynamics including acceleration and
of Canada and the United States were used [14]. The data were
deceleration, maneuvering, missing detection, and noisy data
gathered in tracking experiments that took place at Canadian
points. The finalized set of fuzzy associations (i.e., the entries
Forces Base, Cold Lake, Alberta. Several F-18 fighter aircraft
in the above tables) exhibit the following characteristics.
flying in prescribed routes served as “raid” targets. Radar data
• The diagonal entries from the bottom left-hand corner in were recorded simultaneously from the AN/FPN-504 air traffic
both tables (except the one at the top right-hand corner) control (ATC) radar and the AN/FPS-508 air defense radar
are set to VP and ZE, respectively. These entries are nearby, using two multichannel analog tape recorders provided
mainly used to handle cases involving gentle maneuver. by the U.S. Air Force. The radar has a frequency of 1300
With VP and ZE, the tracker can move MHz. Its scanning rate is 10 s. The probability of detection
very close to a target without resetting the estimated for a target is roughly 0.8 and, hence, missed detection may
velocity. Since it takes relatively long to adapt to a new occur even for straight-line flight patterns. The beam width of
velocity model, modification of is minimized as much the radar is 2 . Therefore, the radar may not resolve closely
as possible to avoid false alarm. The value of can spaced targets when they are performing maneuvers, and some
remain more-or-less the same until the deviated point is targets will miss many measurements. The flying time was
confirmed to represent a real gentle maneuver or some about one half hour. The altitude was 20 000 ft and the speed
significant change of flight dynamics. was 500 knots. Rates of turn ranged from 1–6 g (i.e., the
• Entries in the top right-hand corner are usually used for full range). The identify friend or foe (IFF) system was used
track reinitiation after track losses. In such cases, setting to provide identity (ID) information about the targets. In the
both and to EP could greatly speed up the learning tracking study, the ID is used to distinguish between different
of the new target dynamics. fighters. Only the primary radar plot position information is
• Entries in the top left-hand corner are usually used to used for tracking.
deal with missing observations. By setting both and
to VP, the tracker can update the target dynamics after A. A Two-Stage Kalman Tracker
observations reappear.
• Other entries in the first column handle cases when the As benchmark, the performance of the fuzzy tracker is
last predictions are accurate. While should be kept large compared against a two-stage estimator which consists of two
so as to closely follow the target and prepare for the separate Kalman filters. The first of which is a steady-gain
tracking of any sudden maneuver, is set to ZE to avoid Kalman filter with reduced state model. It takes care of the
changing the predicted velocity by too much. This should case when a target moves in a rather straight line with a more
prevent false alarms resulting from noisy observations or less constant velocity. The second filter is a Kalman filter
from occurring. with maneuver model. It takes care of cases when a target
• The second and third entries in the first row are used to maneuvers or accelerates rapidly. The first filter is switched
handle smooth maneuvering. In such cases, the tracker on initially and, whenever maneuvering is detected; the second
needs to make a small change in velocity and a large filter is triggered. In more detail, these filters are characterized
change in position so as to closely follow the target. by the following equations:
• Entries in the second row are responsible for relatively
small changes of prediction error. This could happen as
a result of a noisy observation or a target beginning to (26)
maneuver. In the case of target maneuvering, the velocity
of the tracker does not need to be changed much (since (27)
maneuvering is only at its initial stage). In the other case (28)
when deviation is due to noisy observations, change in
velocity should be prevented. Hence, is set to ZE. where is the target state vector, is the measurement vector,
• The four entries at the bottom right region of the tables is the transition matrix, is the process noise
are used to handle stressful cases involving forward matrix at the th scan based on the th scan, is
acceleration, maneuver and/or noise. In these cases, the gains at the th scan, is the system noise covariance
and should be set neither too large nor too small. matrix, and is the measurement matrix with measurement
• Invocation of entries in the bottom right-hand corner may noise covariance, It should be noted that the symbols
indicate a potential track loss. This may be due to an ac- used to represent the target state vector and the measurement
cumulated effect of mistaken actions or a sudden change vector are very similar to those of the and coordinate of
CHAN et al.: RADAR TRACKING FOR AIR SURVEILLANCE IN A STRESSFUL ENVIRONMENT 85

TABLE III
THE MEAN SQUARE PREDICTION ERROR FOR BOTH FUZZY AND KALMAN TRACKERS

TABLE IV
THE NUMBER OF TRACK LOSES FOR BOTH FUZZY AND KALMAN TRACKERS

a measured or predicted position used in Sections II and III, the two-stage Kalman tracker was evaluated based on i) the
but they are not related. They are used here mainly because number of track losses (which is an indication of the overall
they have been in widespread use in the field. reliability and robustness of the trackers) and ii) the mean-
For the tracking data, the average velocity of the targets square prediction error (which is an indication of the trackers’
being tracked is 0.2 nautical miles per second and the sampling ability to closely follow the targets). How soon the trackers
period is 10 s. The average displacement is therefore were able to recover from incorrect actions and how the initial
nautical miles. Initially, and (the velocity) settings of the filter parameters affect the performance of the
are set to zero. For the reduced-state model, the parameters trackers were also investigated. The results are presented and
for (26)–(28) are, therefore, set as follows: discussed in the rest of this section.
In Tables III and IV, we show the mean-square prediction
errors and the number of track losses, respectively, for both
the fuzzy and the Kalman trackers as we vary the gate size
For the maneuver model proposed in [11] to be able to track
from 0.10 to 0.15 to 0.22, and then to 0.30 nautical miles.
targets capable of maximum acceleration of 6 g, has to be
As shown in these tables, the two-stage Kalman tracker did
set to 2 g, and has to be set to 50 s. As a result, we have
not perform as well as the fuzzy tracker, both in terms of its
the following parameter settings:
ability to prevent track losses and to closely follow the targets.
The differences in performance are particularly significant as
we reduce the gate size. For the fuzzy tracker, the number
of track losses increased from 6 to 9 to 18, and then to 57
as the gate size is reduced from 0.30 to 0.22 to 0.15, and
then to 0.10, respectively. However, for the Kalman tracker
the reduction of gate size has a much larger impact on its
performance. The number of track losses increased from 27 to
63 to 112, and then to 148 as the gate size is reduced from 0.30
to 0.22 to 0.15, and then to 0.10, respectively. The difference
between the two trackers in the increase of the mean-square
prediction error is also rather significant. In the case of the
B. Results and Discussions Kalman tracker, there is a large 16.3% increase as the gate
Using the real data sets described at the beginning of this size is reduced. However, in the case of the proposed fuzzy
section, the performance of the proposed fuzzy tracker and tracker, the increase is a mere 4.3%. From Table III, one could
86 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 5, NO. 1, FEBRUARY 1997

Fig. 3. Performance of fuzzy tracker in a very stressful environment. Fig. 4. Performance of Kalman tracker in a very stressful environment.

see that different noise level had been used for testing of the
robustness of both the fuzzy and the Kalman tracker.
It should also be noted from Table IV, that the deterioration
in the performance of the Kalman tracker was quite drastic
in some cases. For example, it lost track of target 130A only
once—when the gate size was 0.30. However, when the gate
size was reduced to 0.22, it lost track of 130A ten times. Close
examination of the data reveals that the main reason for this
sudden jump in errors is that the Kalman tracker is rather slow
to recover from incorrect actions. The track formed by 130A
contains a gentle maneuver. When the gate size is relatively
large, the Kalman tracker was able to keep track of the target,
but when we reduce the gate size, the Kalman tracker lost
track of it right at the beginning of the maneuver. Once it
lost the target, the Kalman tracker took a rather long time to
recover from its mistakes since Kalman filter require a rather
long learning period before it enters a steady state (i.e., stable
and reliable prediction). On the contrary, the fuzzy tracker was Fig. 5. Performance of fuzzy tracker in a rather nonstressful environment.
able to recover from incorrect decisions much more quickly
and this explains the rather large difference in the number of at least several scans) required for a maneuver detector to
track losses. respond to it. Since the target made a 360 turn in an interval
For better understanding of the differences in performance of six scans, a delay of two scans could mean an angular
of the two trackers, we have extracted some sessions of the deviation of about 120 Thus, it is not surprising for the two-
tracks from the real data sets for illustration and discussion stage Kalman tracker to “overshoot” when predicting the next
here. Figs. 3–10 show the actual and predicted positions made target positions resulting, therefore, in track losses.
by both the fuzzy and the two-stage Kalman tracker when Figs. 5 and 6 illustrate the performance of the fuzzy and the
tracking targets in four different stressful environment each of two-stage Kalman tracker in a different stressful environment.
which exhibiting different characteristics. A target that is moving in a nearly straight line with constant
The first of them involves a target rapidly accelerating and velocity is being tracked. There is almost no missing obser-
making sharp turns. The target was missing from some scans vations. In this relatively nonstressful environment, it should
and it was also misrecognized several times during its flight. be noted that the Kalman tracker predicts a rather zigzag path.
Figs. 3 and 4 illustrate, respectively, the predictions made by It is quite sensitive to noise even though the noise level is at
the proposed fuzzy tracker and the two-stage Kalman tracker. a minimal.
As shown in the figure, the Kalman tracker suffers from some Figs. 7 and 8 illustrate the ability of the fuzzy and the two-
chain effect of mistaken actions, and this results in the amount stage Kalman tracker to closely follow a target that is rapidly
of deviation between the actual and the predicted positions accelerating along a rather straight line (i.e., accelerating but
accumulating scan by scan. This slowness in response to nonmaneuvering). The environment in which the tracking
sudden changes in flight dynamics is due to the time (usually experiment is performed is rather nonstressful in the sense
CHAN et al.: RADAR TRACKING FOR AIR SURVEILLANCE IN A STRESSFUL ENVIRONMENT 87

Fig. 6. Performance of Kalman tracker in a rather nonstressful environment. Fig. 8. Performance of Kalman tracker when the target accelerates rapidly.

Fig. 7. Performance of fuzzy tracker when the target accelerates rapidly. Fig. 9. Performance of fuzzy tracker in the presence of missing observations.

that with the exception of the target accelerating rapidly there It should be noted that the likelihood of a target missing
are no missing observation and the noise level is relatively from successive scans during a maneuver is quite high in real
low. As shown in the figure, in such a tracking environment, life. This is especially true with high dynamic aircraft. Due to
the Kalman tracker made more mistakes leading to more track the angle of flight and their ability to make sharp turns, they
losses. are usually much more difficult to detect during maneuvering.
Finally, Figs. 9 and 10 illustrate the problems associated It should also be noted that the fuzzy tracker uses Mamdani’s
with target tracking in the presence of missing observations. interpolation method based on center of area.
The environment in which the tracking task was performed Other than the two-stage Kalman filter, the performance of
had a relatively low noise level. The target that was being the fuzzy tracker had also been evaluated by comparing it
tracked did not maneuver or rapidly accelerate. As shown in with an - tracker with set to 0.5 and set to 0.167 which
the figures, the Kalman tracker did not perform well in this had been proved to be the best of fixed gains by Cohen [2],
case. The maneuver detector which it employed was triggered [17], who based them on Benedict and Bordner’s formula for
by a noisy data point and the Kalman tracker mistook it as optimum gain [18]. The result of the performance analysis are
a sign for target maneuvering. It was “tricked” to trace off discussed in [19].
a nonexistent maneuvering path. Quite unlike the Kalman
tracker, the fuzzy tracker did not have such a problem. This
is due mainly to its ability to take into consideration several V. CONCLUSION
possibilities at the same time (as more than one rule can be In this article, we demonstrated the potential for using
fired simultaneously). fuzzy logic in the tracking of targets in stressful environment
88 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 5, NO. 1, FEBRUARY 1997

[4] Y. Bar-Shalom and K. Birmiwal, “Variable dimension filter for ma-


neuvering target tracking,” IEEE Trans. Aerosp. Electron. Syst., vol.
AES-18, pp. 621–629, Sept. 1982.
[5] T. C. Wang and P. K. Varshney, “A tracking algorithm for maneuvering
targets,” IEEE Trans. Aerosp. Electron. Syst., vol. 29, pp. 910–924, July
1993.
[6] D. J. Mook and I. M. Shyu, “Nonlinear aircraft tracking filter utilizing
control variable estimation,” Amer. Inst. Aeronaut. Astronaut., vol. 15,
pp. 228–237, Feb. 1992.
[7] K. Rokhsaz and J. E. Steck, “Use of neural networks in control of high-
alpha maneuvers,” J. Guidance, Contr., Dynamics, vol. 16, no. 5, pp.
934–939, 1993.
[8] Y. T. Chan and F. Couture, “Maneuver detection and track correction by
input estimation,” Inst. Elec. Eng. Proc., vol. 140, pp. 21–28, Feb. 1993.
[9] W. D. Blair, J. E. Gray, and M. D. Boyd, “Design analysis for the
two-stage , , estimator,” IEEE Proc., vol. 2, pp. 1050–1054, Apr.
1991.
[10] M. Munu, I. Harrison, and M. S. Woolfson, “Comparison of the Kalman
and - filters for the tracking of targets using phased array radar,” in
Proc. IEE Conf. Radar ’92, Brighton, U.K., Oct. 1992, no. 365, pp.
196–199.
[11] R. A. Singer, “Estimating optimal tracking filter performance for
manned maneuvering targets,” IEEE Trans. Aerosp. Electron. Syst.,
vol. AES-5, pp. 473–483, July 1970.
Fig. 10. Performance of Kalman tracker in the presence of missing obser- [12] S. S. Blackman, Multiple Target Tracking with Radar Applications.
vations. Norwood, MA: Artech House, 1986.
[13] P. R. Kalata, “A-B target tracking systems: A survey,” in Proc. Amer.
Contr. Conf., Chicago, IL, June 1992, pp. 832–836.
characterized not only by high noise levels, but also by [14] M. Gregoire, “Raid tracking trails (RATT): Digital plot data base,” Rep.
1026, Defense Res. Establishment Ottawa, ON, Canada, 1990.
targets accelerating at nonuniform rates, making sharp turns, [15] K. C. C. Chan, V. Lee, and H. Leung, “Automatic generation of fuzzy
or missing from successive scans even during maneuvering. rules using a natural genetic algorithm,” in Proc. SPIE Conf. Applicat.
By incorporating fuzzy logic in conventional - filters, the Fuzzy Logic Technol. III, Orlando, FL, Apr. 8–12, 1996.
[16] V. Lee, K. C. C. Chan, and H. Leung, “Adaptive system noise covariance
coefficients are allowed to be modified dynamically based for performance enhancement of kalman filter based algorithms,” in
on the conclusions of a set of fuzzy rules. These rules Proc. SPIE Conf. Acquisition, Tracking, Pointing, X, Orlando, FL, Apr.
determine and according to the magnitudes of the last 8–12, 1996.
[17] S. A. Cohen, “Adaptive variable update rate algorithm for tracking
prediction error and change-of-error. To a certain extent they targets with a phase array-radar,” Inst. Elec. Eng. Proc., Pt. F, vol. 133,
reflect the likelihood of, say, the tracker encountering a noisy pp. 277–280, June 1986.
[18] T. R. Benedict and G. W. Bordner, “Synthesis of an optimal set of radar
measurement, a target moving in a straight path, or a target track-while-scan smoothing equations,” IRE Trans., vol. 7, pp. 27–32,
maneuvering. Whenever more than one possible situation July 1962.
needs to be considered simultaneously, more than one rule is [19] K. C. C. Chan, V. Lee, and H. Leung, “Robust target tracking using a
fuzzy filter,” in Proc. IEEE Int. Conf. Syst., Man, Cybern,, Vancouver,
fired. The degrees at which these rules are fired are combined BC, Oct. 22–25, 1995.
quantitatively to determine the final values of and
The effectiveness of the fuzzy tracker was evaluated using
some real data sets representing a wide range of different
stressful environment. The performance in terms of prediction
error and the number of lost-tracks is compared against that
of a two-stage Kalman filter. The results show that the fuzzy
tracker has a very good degree of fault tolerance and a better Keith C. C. Chan (M’94) received the B.Math.
recovery rate. While these results are encouraging, more efforts (computer science and statistics), M.A.Sc., and
are required to study the robustness of the fuzzy tracker with Ph.D. degrees (systems design engineering), in
1984, 1985, and 1989, respectively, all from the
respect to changes in the membership functions and the rules. University of Waterloo, ON, Canada.
Also, as is typically the case with other rule-based approaches, From 1984 to 1989, he worked as a Research
whether or not this approach can be generalized to work with Assistant in the Pattern Analysis Machine Intel-
ligence Laboratory, University of Waterloo, ON,
all kinds of targets in all types of environment depends on Canada. Soon after graduation, he joined the IBM
whether or not some automated approaches to finding the best Canada Laboratory, Toronto, where he was involved
rules and membership functions can be developed [15], [16]. in software development projects in the Image
Systems Center and the Application Development Technology Center. He
joined the Department of Electrical and Computer Engineering at Ryerson
REFERENCES Polytechnic University, Toronto, ON, Canada, as an Associate Professor in
1993. Since 1994, he has been with the Department of Computing, The Hong
[1] C. B. Chang and J. A. Tabaczynski, “Application of state estimation to Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong, where he
target tracking,” IEEE Trans. Automat. Contr., vol. AC-29, pp. 98–109, is currently Associate Professor and Director of the Software Technology
Feb. 1984. Facilities Center. His research interest is in computational intelligence, data
[2] M. Munu, I. Harrison, D. Wilkin, and M. S. Woolfson, “Comparison of mining, pattern recognition, and software engineering.
adaptive target-tracking algorithms for phased-array radar,” Inst. Elec. Dr. Chan was an adjunct faculty member of the Department of Systems
Eng. Proc., Pt. F., vol. 139, no. 5, pp. 336–342, Oct. 1992. Design Engineering, University of Waterloo, ON, Canada, the Department
[3] W. D. Blair, “Fixed-gain two-stage estimators for tracking maneuvering of Electrical and Computer Engineering, Ryerson Polytechnic University,
targets,” IEEE Trans. Aerosp. Electron. Syst., vol. 29, pp. 1004–1014, Totonto, and the Department of Electrical Engineering, The University of
July 1993. Western Ontario, London, ON, Canada.
CHAN et al.: RADAR TRACKING FOR AIR SURVEILLANCE IN A STRESSFUL ENVIRONMENT 89

Vika Lee received the B.A. degree in computing Henry Leung (S’88–M’90) received the B.Math.
from the Hong Kong Polytechnic University, Hung degree in applied mathematics from the University
Hom, Kowloon, Hong Kong, in 1994, (fuzzy logic of Waterloo, ON, Canada, in 1984, the M.Sc. degree
and target tracking). He is currently an M.Phil. in mathematics from the University of Toronto,
candidate in the Department of Computing, The ON, Canada, in 1985, and the M.Eng. and Ph.D.
Hong Kong Polytechnic University, Hung Hom, degrees in engineering physics and electrical engi-
Kowloon, Hong Kong. neering from the McMaster University, Hamilton,
He was involved in various research projects in ON, Canada, in 1986 and 1991, respectively.
geographical information systems as an undergrad- In 1990, he was a Research Engineer at the
uate student and had co-authored a paper in that Communications Research Laboratory, the McMas-
area. In his spare time, he works as a freelance ter University, Hamilton, ON, Canada. In 1991, he
programmer. His research interests include fuzzy logic, genetic algorithms, joined the Defense Research Establishment Ottawa, ON, Canada, as a defense
target tracking, simulation, the traveling salesman problem, classification, and scientist. His research interests include adaptive signal processing, neural
georgraphical information systems. network, nonlinear dynamics, and data fusion.

Potrebbero piacerti anche