I.

INTRODUCTION

The estimation and prediction of the kinematics

Tracking for Maneuvering of a dynamic object requires the use of a dynamic
model and discrete time data. A majority of target
Target Trajectories via the 3D trackers are based on a straight line maneuver
such as ®-¯ [1—3], ®-¯-° [4, 5] and linear Kalman
Circular Filter filters [6—8]. The model uncertainties and the
pilot-induced maneuvers are accounted for as
stochastic acceleration [9, 10]. Farooq et al. [11, 12]
present various maneuver compensation techniques.
The ambiguousness of the maneuver models has been
DIRK TENNE compensated by designing algorithms consisting of
TARUNRAJ SINGH multiple models [13—18]. On the contrary, complex
SUNY Buffalo nonlinear models have been developed to capture
the real object dynamics [19—21]. These models are
applied to special maneuvers such as the coordinated
turn [22, 23]. In general the coordinated turn
A circular prediction algorithm is proposed, which integrates models are a set of nonlinear, coupled equations,
the measured data into the filter and constrains the prediction to which are difficult to solve. The focus of the work
lie on a smooth curve modeled by an arc of a circle. The circular presented here is on the development of an algorithm
prediction is entirely defined in relation to three measurements in constraining the predicted state to a circular path.
three-dimensional space. It is therefore not necessary to calculate The unscented transformation developed by Julier
the center and the radius of the circle. To obtain the statistics and Uhlmann [24] is used in conjunction with the
of the circular prediction, the unscented transformation has circular target motion model derived here. The key
been utilized. The proposed hybrid filter combines the circular feature of the unscented transformation is that it
prediction and a constant velocity prediction by utilizing the
approximates the probability distribution rather than
approximating the nonlinear function by a Taylor
covariance intersection (CI). This combined prediction can
series. A review of the development and modifications
be updated with the subsequent measurement using a linear
of the original unscented transformation is presented
estimator. The proposed technique is compared with standard
in [25]. In addition, the proposed hybrid filter utilizes
filters and the interacting multiple model (IMM) approach on the covariance intersection (CI) algorithm developed
a benchmark trajectory which includes coordinated turns and by Uhlmann [26], [27] to combine the maneuvering
straight line maneuvers. and nonmaneuvering modes. The CI is a conservative
fusion technique if the cross-correlation is unknown.
Uhlmann arrived at the CI algorithm by showing
that the solution of a conservative fusion yielding a
positive definite error covariance, is a function of a
scalar weight w only.
A geometric approach of implementing a
circular turn has been introduced by Roecker and
McGillem [28]. It has been observed that aircrafts
normally maneuver on circular paths and a new polar
coordinate system has been implemented at the center
of the maneuver. Kawase et al. [29] designed an ®-¯
filter in conjunction with a two-dimensional circular
prediction algorithm. In [30] Kawase et al. combined
Manuscript received February 17, 2004; revised November 9, 2004;
released for publication April 22, 2005.
the circular prediction with a two-stage Kalman
filter, which has been used for maneuvering targets
IEEE Log No. T-AES/41/4/860800. by Alouani et al. [8]. The center-point approach
Refereeing of this contribution was handled by B. La Scala. (CPA) predicts in a polar coordinate system (R and
Authors’ addresses: D. Tenne, Center for Multisource and
Ã) whose origin is the center of the circle. The CPA
Information Fusion, SUNY Buffalo, Buffalo, NY; T. Singh, is not amenable for further stability, performance and
Dept. of Mechanical and Aerospace Engineering, SUNY uncertainty analysis, because of the complex center
Buffalo, 1009 Furnas Hall, Buffalo, NY 14260-4400, E-mail: coordinate calculation and discontinuities in the polar
(tsingh@[Link]). angle à between successive scans. This discontinuity
appears by switching from the previous circle to the
0018-9251/05/$17.00 °
c 2005 IEEE next circle as the radius and center change. Moreover,

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 4 OCTOBER 2005 1373
this algorithm has to distinguish between clockwise
and counterclockwise turns. Tenne and Singh [31]
presented a circular tracking algorithm in a local
coordinate system retaining a set of states to avoid
such discontinuities.
This paper develops a three-dimensional tracking
algorithm for maneuvering trajectories. Section II
introduces the planar circular tracking technique
in a local coordinate system, which is amenable to
extension to a three-dimensional tracking algorithm
as proposed in Section III. This section derives the
circular prediction plane on which the prediction is
performed using the Unscented transformation. The
Hybrid filter is introduced as a convex combination
of the circular prediction and a constant velocity
prediction. The paper concludes with an illustration
of the proposed technique on a benchmark trajectory Fig. 1. Properties of points lying on a circle.
consisting of coordinated turns and straight line (a) Basic properties. (b) Relative relationships.
maneuvers.

The unknown distance R14 in (3) is determined from

II. THE PLANAR CIRCULAR TRAJECTORY
4134 using the cosine rule
CONSTRAINT
2 sµ 3
¶2
To develop the algorithm, consider four points R34
lying on a circle as shown in Fig. 1(a). The four R14 = R13 4cos '2 § ¡ sin2 '2 5 : (4)
R13
points are connected to create two triangles 4123
and 4134, where the triangle sides are named by
In case of a prediction with constant angular velocity
the points which they connect, for example, R12 is
('2 = '1 and R34 = R23 ) it can be shown that the
the distance between points c and d. The fourth
smaller value of (4) yields the trivial solution of
point can be described relative to the points c to
R14 = R12 , and therefore the desired relationship for
e by a variety of angle and distance combinations.
the prediction is the positive radical. This equation
A convenient pair is the angle '2 and the distance
is also valid for the nonconstant angular velocity
R34 as indicated in Fig. 1(b). The desired prediction
case since the smaller value of R14 would place the
equations define the relationships between the fourth
predicted position in the vicinity of point d, which is
point, which is parameterized by R34 and '2 and
not the desirable solution.
the previous three points. To derive the prediction
From the triangle 4234 in Fig. 1(a), the variable
equation, consider the points d, e, and f relative
R34 can be calculated using the sine theorem
to point c, which is equivalent to introducing a
Cartesian coordinate system labeled u-v with its origin sin °1 sin '2
at point c. Observing Fig. 1(b), the distances are R34 = R23 = R (5)
sin °2 sin '1 23
defined in terms of ui and vi .
2 where the relationships °1 = '2 and °2 = '1 have been
R13 = u23 + v32 (1a)
applied, which can be easily observed from Fig. 1(a).
2
R14 = u24 + v42 (1b) The solution of R14 (4) contains the evaluation of a
2
square root, whose radicand can take negative values
R24 = (u4 ¡ u2 )2 + (v4 ¡ v2 )2 (1c) and can cause numerical instability. It is therefore,
2
R34 = (u4 ¡ u3 )2 + (v4 ¡ v3 )2 : (1d) desirable to derive an alternative form, which proves
advantageous in the prospective development. An
Equations (1c) and (1d) can be rewritten as alternative equation for the distance R14 can be
obtained by substituting (5) into (4) and rearranging
u2 u4 + v2 v4 = R12 R14 cos('1 + '2 ) (2a)
leading to
u3 u4 + v3 v4 = R13 R14 cos '2 (2b) q
sin '2 2 ¡ R 2 sin2 ' :
T R14 = R13 cos '2 + R23 1 (6)
and solving for the relative position [u4 v4 ] , yields sin '1 13
the desired relationship
· ¸ · ¸ · ¸ The square root expression can be modified by
u4 u2 v2 ¡1 R12 cos('1 + '2 ) observing Fig. 2, where the triangle 4123 has been
= R14 : (3)
v4 u3 v3 R13 cos '2 extended in the direction of the phasor R12 to build a

1374 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 4 OCTOBER 2005
 III. THE THREE DIMENSIONAL EXTENSION

Commonly target position estimators function

in a three-dimensional environment. The radar
measurements, for example, are obtained in a
spherical coordinate system reporting the range,
Fig. 2. Right angle triangle extension to triangle 4123.
azimuth, and elevation, whereas the target’s position
is described in a global Cartesian coordinate system.
Kawase et al. [32] applied his two-dimensional CPA
right angle triangle with the hypotenuse R13 to a three-dimensional prediction, where the prediction
q is performed on a turning plane and its normal
2 ¡ R 2 sin2 ' = R cos ' ¡ R :
R23 vector is being estimated. The proposed curve-fitting
13 1 13 1 12
estimator constrains the prediction to a circular
Further simplification with the use of the sum of angle trajectory in a two-dimensional coordinate system.
theorem leads to This section addresses the fact that the three points
1 form a two-dimensional plane, which is oriented
R14 = [R13 sin('1 + '2 ) ¡ R12 sin '2 ] : (7) in three-dimensions. The circular estimation can be
sin '1
performed in this two-dimensional plane and later
Note that (7) is also valid for straight line maneuvers. back-transformed to the global three-dimensional
Assuming a constant speed (CS) angular turn (7) coordinate system.
reduces to
CS
R14 = 2R13 cos '2 ¡ R12 (8) A. Representing the Circular Prediction Plane
and in case of a straight line maneuver The circular prediction algorithm described in
CS the previous section is performed based on three
lim R14 = 2R13 ¡ R12 : (9) position measurements. These three locations define
'1 !0
two vectors ~r12 and ~r13 as shown in Fig. 1, which
For prospective development of a three- span a two-dimensional plane in three-dimensional
dimensional extension of the proposed algorithm, the space. The normal vector ~ns of the two-dimensional
prediction (3) can be interpreted as a vector quantity. plane can be obtained by evaluating the normalized
Define the vector ~r12 of length R12 pointing from cross product between ~r12 and ~r13 . The orientation of
point c towards point d, and vector ~r13 , respectively. the normal vector is positive when the three vectors
By evaluating the matrix inverse and noticing that constitute a right-handed coordinate system, which
the determinant can be expressed as u2 v3 ¡ u3 v2 = can also be determined by the right-hand screw rule
R12 R13 sin '1 , which is a direct result of the cross rotating ~r12 into the direction of ~r13 . The transition
product performed in a two-dimensional plane from the original Cartesian coordinate system to
the circular prediction plane can be achieved by a
k~r13 £~r12 k = R12 R13 sin '1 (10)
multi-step rotation of the coordinate system. The first
the prediction equation (3) can be rewritten as step consists of a rotation around the z-axis such that
· ¸ the projection of the normal vector onto the x-y plane
u4 cos('1 + '2 ) aligns with the x-axis. This rotation is performed by
= T(¼=2)~n13 R14
v4 sin '1 the three-dimensional transformation matrix G1 ('s ),
cos '2 where 's is the polar rotation angle of the normal
+ T(¡¼=2)~n12 R14 : (11) vector projected on to the x-y plane. Subsequent
sin '1
to the alignment with the x-axis, the second step
The following substitutions have been introduced. The rotates the coordinate system around the transformed
matrix T(®) is the rotational transformation matrix y-axis. This rotation is performed by the rotation
defined as matrix G2 (#s ), where #s is the direction cosine of
· ¸
cos ® sin ® the normal vector with respect to the z-axis. The
T(®) = : (12) combined coordinate transformation can be obtained
¡ sin ® cos ®
by constructing the rotation matrix G = G2 ('s )G1 ('s ):
The normalized vector ~n12 is the unit vector pointing 2 3
in the direction of ~r12 . The new prediction (11) cos 's cos #s sin 's cos #s ¡ sin #s
consists of a summation of the rotated and scaled unit 6 7
G = 4 ¡ sin 's cos 's 0 5 : (13)
vectors ~n12 and ~n13 in a two-dimensional plane. In a cos 's sin #s sin 's sin #s cos #s
three-dimensional environment the same rotation has
to be performed around the normal vector defined by Note that the rotation matrices Gi are orthogonal and
the two-dimensional plane. thus Gi¡1 = GiT .

TENNE & SINGH: TRACKING FOR MANEUVERING TARGET TRAJECTORIES VIA THE 3D CIRCULAR FILTER 1375
B. Extended Circular Prediction Algorithm constrained to lie on a circle defined by the three
measurements.
Based on the development of the two-dimensional
circular prediction algorithm and the results of the IV. DEVELOPMENT OF THE HYBRID FILTER
circular prediction plane, this section describes the
circular prediction algorithm in three-dimensional The unscented transformation [24] is used in
space. The fact, that the circular prediction can conjunction with the circular target motion models
be described as a rotation of ~r12 and ~r13 can be as discussed in the previous section. In the following
exploited to extend the algorithm to include the third text, we describe the unscented transformation to
dimension. Equation (11) can be rewritten to include obtain the circular prediction and its covariance,
a zero z-component of the predicted position and the which is used in conjunction with the minimum mean
normalized vectors, where the rotation matrix T(®) square error estimator to update the prediction with
can be defined as the measurement.
2 3 The circular prediction on the three-dimensional
cos ® sin ® 0 prediction plane is described by (16), which depends
6 7 on the three position measurements. These are
T(®) = 4 ¡ sin ® cos ® 0 5 : (14)
generally obtained by a radar in spherical coordinates.
0 0 1
To obtain the statistics of the circular prediction, the
Substituting the normalized vectors ~n12 and ~n13 with three measurements are stacked to form an augmented
vectors described in the relative Cartesian coordinates vector of length n = 9:
denoted by a prime (11) can be written as x = [r1 Ãr1 #r1 r2 Ãr2 #r2 r3 Ãr3 #r3 ]T
2 3 2 0 3
u4 u3 (18)
6 7 6 0 7 R14 cos('1 + '2 ) consisting of the radar reports as range r, azimuth
4 v4 5 = T(¼=2)G 4 v3 5
R13 sin '1 Ãr , and elevation #r . Its covariance Px is a diagonal
0 w30 matrix with the repeated elements of the individual
2 0 3
u2 covariances ¾r2 , ¾Ã2 , and ¾#2 . The unscented
6 0 7 R14 cos '2 transformation can be applied to the augmented space,
+ T(¡¼=2)G 4 v2 5 (15)
R12 sin '1 where the so-called ¾ points are transformed to the
w20 Cartesian coordinate system and further propagated
where G is given by (13). The circular prediction with the circular prediction (16). The unscented
of (15) is described with respect to a Cartesian transformation selects a ¾-set consisting of 2n + 1
coordinate system where the x-y plane coincides points, which is the perturbation from the mean by
with the circular prediction plane. Including the back a scaled deviation. The deviations ¾i are defined as the
transformation into (15) involves the evaluation of columns of the matrix square root of Px [24], and the
the matrix product G = G¡1 T(¼=2)G. Note, since ¾-set is defined as
T(¡¼=2) = TT (¼=2) and G is orthogonal it can be ³0 = x̄ (19)
shown that the combined transformation matrix p
³i = ³0 + (n + ·)¾i for i = 1 : : : 2n (20)
G¡1 T(¡¼=2)G = G T . Therefore, the circular prediction
performed on the two-dimensional plane can be such that the ¾-set exhibits the same probabilistic
written with respect to the global Cartesian coordinate characteristics as the random variable x and · is a
system. By adding the coordinates of the first point free variable of the unscented filter, which in this
and substituting the relative coordinates with the particular application has been set to 2. The weights
global coordinates we obtain the circular prediction w0 = ·=(n + ·) and wi = 1=(2(n + ·)) have been
in global Cartesian coordinates selected to match the first four central moments of
2 3 2 3 2 3 2 3 x, which is true for all Gaussian random variables if ·
x4 x3 x2 x1
is selected to satisfy the constraint, n + · = 3.
6 7 6 7 6 7 6 7
4 y4 5 = G 4 y3 5 s1 + G T 4 y2 5 s2 + (I ¡ Gs1 ¡ G T s2 ) 4 y1 5 The ¾-set is transformed to the Cartesian
z4 z3 z2 z1 coordinate system and further propagated by the
circular prediction (16). This output set labeled ´i
(16) yields the statistics of the circular prediction (CP)
where the abbreviations: (mean xCP and covariance PCP ) by applying
R14 cos('1 + '2 ) R14 cos '2 2n
X
s1 = , s2 = (17)
R13 sin '1 R12 sin '1 xCP = ´¯ = wi ´i (21)
i=0
have been introduced. Performing the circular 2n
X
prediction with (16), we obtain a position on PCP = P´ = wi (´i ¡ ´)(´ ¯ T:
¯ i ¡ ´) (22)
the two-dimensional plane, which is furthermore i=0

1376 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 4 OCTOBER 2005
 Fig. 3. Unscented ¾-set illustrated on the circular prediction.

Fig. 3 illustrates the ¾-set and the propagation on reporting the range (r), azimuth (Ã), and elevation
an example configuration. The position uncertainties (#), which are related to the Cartesian coordinate
are symbolized by the 1-¾ covariance ellipsoids. system centered at the radar location by the standard
The standard unscented transformation creates 19 spherical to Cartesian coordinate transformation. With
circular trajectories, i.e., 19 possible combinations the aforementioned derivations it is of relative ease to
of the three uncertain measurements, to determine program the extended Kalman filter algorithm.
the statistics of the circular prediction. The small The hybrid filter consists of a convex combination
squares indicate these predictions, which result in the of the circular prediction and the predicted position
mean and covariance by evaluating the weighted sum. of the extended Kalman filter. These predictions are
The mean and covariance are illustrated by the large statistically correlated since they are conditioned
square and the prediction ellipsoid. on the same measurements. Therefore, the fusion
The predicted mean and covariances can now be is performed using the CI algorithm developed by
used in the minimum mean square error estimator to Uhlmann [34]. In contrast to the Kalman filter update,
update the prediction with the measurement, which the CI does not assume independent data to be fused,
has been converted to the Cartesian coordinates using thus yielding a conservative fusing algorithm. Assume
an unbiased transformation [33]. two random processes A and B, which are described
by their means (a, b) and covariances (Pa , Pb ). The
fused mean and covariance can be obtained as
A. The Hybrid Filter
Pc¡1 = wPa¡1 + (1 ¡ w)Pb¡1 (23)
The prediction algorithms constraining the target
motion on a smooth curve provide a reasonable Pc¡1 c = wPa¡1 a + (1 ¡ w)Pb¡1 b (24)
performance for circle-like trajectories. However, where w takes values in the range of [0, 1]. The
real target motions can be approximated by weight w can be interpreted as a tuning parameter of
piecewise curves which are circles and straight the CI. Its selection shapes the estimated covariance
lines. Furthermore, with the presence of noise, the either closer to the covariance of A (w ! 1) or to the
performance of a stand-alone circular filter would covariance of B (w ! 0). In the following section,
degrade. The proposed circular filter is therefore both random processes are equally weighted (w = 0:5).
integrated with traditional filters including straight line Upon arrival of a new measurement, the combined
target models. prediction can be updated using the linear estimator.
The constant velocity model in three-dimensions
includes the x, y, z position as well as the three
V. BENCHMARK TRAJECTORY IN THREE
velocities vx , vy , and vz , such that the state transition
DIMENSIONS
matrix can be constructed as shown in Appendix A.
This system is driven by white noise accelerations This section presents the performance
separated in each direction, and the process noise characteristics of the proposed hybrid filter compared
covariance can be calculated as shown in Appendix A, with the extended Kalman filter with a straight
where the inverse three-dimensional rotation matrix line constant velocity model and the interacting
of (13) has to be utilized. The radar measurements multiple model (IMM) algorithm with the modes
are obtained in the spherical coordinate system described in the subsequent text. The benchmark

TENNE & SINGH: TRACKING FOR MANEUVERING TARGET TRAJECTORIES VIA THE 3D CIRCULAR FILTER 1377
 Fig. 4. Benchmark target maneuver with sample measurements.

Fig. 5. Consistency test of hybrid three-dimensional filter.

trajectory facilitates a combination of coordinated has been evaluated from the finite difference, whereas
turns and straight line maneuvers with constant, the initial position covariance is the measurement
zero, and increasing acceleration phases. The covariance Rk at time k converted into Cartesian
benchmark considers a target whose position is coordinates Rkc . The velocity state covariance is
sampled every T = 30 s as shown in Fig. 4. A sample initialized as (R2c + R3c )=T2 due to the finite difference.
of measurements is indicated by the markers £. The A consistency test as described in [35] has
target maneuver consists of three parts in the x-y been carried out to ensure proper operation of the
plane: hybrid filter. A filter is called consistent if its state
1) a circular trajectory with constant angular estimation is unbiased and the error covariance is
velocity, resulting in a constant acceleration of correctly approximated by the filter’s state covariance
0.0132 m/s2 , matrix. This test is performed off-line as Monte
2) a straight line maneuver with constant speed, Carlo simulation, where the ground truth is known.
3) an accelerating target on a circular trajectory, The hypothesis H0 postulates that the normalized
resulting in an increasing acceleration up to estimation error squared equals the dimension of
0.053 m/s2 , the state. Performing a Monte Carlo simulation the
where the target speed lies in the range of hypothesis H0 can be accepted if the test statistic lies
2.62 m/s—5.56 m/s. The targets elevation is simulated in the two-sided 95% probability concentration region.
as an up-and-down maneuver comparable to a Fig. 5 shows this region along with the average
sine function. The targets position is reported by normalized estimation error squared (NEES) based
a stationary radar located at the x = ¡250 m, y = on 100 Monte Carlo runs. Only a small part lies
200 m and z = 0 m as indicated by the circled outside the 95% region and the H0 hypothesis can be
position. The measurements are obtained in the radar accepted. Note that most filters fail to be consistent
coordinate system as range bearing whose standard during maneuvers that are outside of their respective
deviation in range is 10 m and 2 deg in the bearing models.
angles. A statistical comparison of the filters can be
All filters have been initialized with the first three achieved by a simple hypothesis test [35, pp. 76—77].
measurements and when necessary the velocity state Define the performance measure as the absolute

1378 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 4 OCTOBER 2005
 Fig. 6. Statistical test comparison of hypotheses H1 and H2 .

estimation error ²filter (k). Let us postulate the noise settings q2 = 0:01322 and q3 = 0:03312 , which
hypothesis that the error of filter 1 is larger than the correspond to the minimum acceleration and the
error of filter 2: average acceleration. The Markov chain transition
matrix between these models
H1 : ¢(k) = ²filter1 (k) ¡ ²filter2 (k) > 0: (25) 2 3
0:9 0:09 0:01
Based on Monte Carlo simulations the test statistic 6 7
pij = 4 0:15 0:7 0:15 5 (28)
¢̄(k) 0:01 0:15 0:84
¹(k) = (26)
¾¢ (k)
has been chosen to indicate a strong likelihood to
can be calculated, where the ¢̄ corresponds to the operate at the current model and to allow a smooth
mean of ¢(k) and ¾¢ (k) represents the deviation. If transition between the constant velocity and slow
¹ exceeds a threshold ¹0 then the hypothesis H1 is accelerating and fast accelerating models. Fig. 6 shows
accepted. The threshold is commonly defined by the the result of the statistical comparison as defined by
significance level of the null hypothesis. (26). The hybrid filter outperforms the IMM algorithm
The statistical test is performed to compare the during the constant acceleration phase, whereas the
performance of the straight line (SL) model to the IMM yields a better performance during the constant
combined straight line and circular prediction (hybrid) velocity phase. As the acceleration is increasing both
model. The hypothesis is postulated as filters perform well. The transition of the IMM to
the acceleration phase appears faster than the one
H1 : ¢(k) = ²SL (k) ¡ ²hybrid (k) > 0: (27) of the hybrid filter, which exhibits a smaller error at
The second hypothesis H2 compares the circular filter the mid-phase of the maneuvering circle path. Fig. 7
with the hybrid filter. Both test statistics are shown in compares the combined position rms error of the
Fig. 6. The hypothesis H1 compares the hybrid filter proposed hybrid and the IMM algorithm. Throughout
with the extended Kalman filter using a straight line the benchmark trajectory both filters perform likewise,
model, and it can be observed that the hybrid filter whereas the IMM algorithm exhibits a lower rms
performs better during the circular maneuver since the error during the constant velocity path and during
test statistic is above the threshold, whereas during the transition to the turning phase. The initial peak of
the straight line maneuver the test statistic falls below the IMM is the result of the finite difference velocity
the threshold. The comparison of the circular filter initialization.
and the hybrid filter yields better performance for the
circular filter during the turning maneuvers and the VI. REMARKS AND CONCLUSION
hybrid filter exhibits increased performance during the
straight line maneuver. In real applications the filters are confronted with
The third test H3 compares the IMM algorithm missed and false measurements. In case of missing
with the proposed hybrid filter. The three-mode measurements the proposed filter remains predicting
IMM consists of a constant velocity model and a on the previous circular path until a new measurement
Wiener process acceleration model with two process arrives. The missed measurements subsequently

TENNE & SINGH: TRACKING FOR MANEUVERING TARGET TRAJECTORIES VIA THE 3D CIRCULAR FILTER 1379
 Fig. 7. Combined position rms error comparison.

have to be replaced by the corresponding prediction. APPENDIX A. CONSTANT VELOCITY MODEL

The limits on how long this filter can predict on the
A common model used for target tracking,
previous circle and when the target is declared to be
assumes a target moving on a straight line with a
lost have to be identified. False radar reports could
constant velocity, where the system is driven by white
be handled by the circular filter using a least squares noise acceleration. The linear state space is comprised
approach to identify the circular path based on the of the Cartesian position and its corresponding
history of a window of measurements. velocities yielding the state transition matrix:
In this work, the fact that a target trajectory has · n£n ¸
I TIn£n
to be smooth even if the acceleration is impulsive © = n£n (29)
has been exploited to develop a new technique for 0 In£n
target tracking. The proposed algorithm integrates where In£n is the identity matrix of dimension
the measured data into the filter and constrains the n = 3 and 0n£n corresponds to the null matrix. The
prediction to lie on a smooth curve, modeled, for measurements are again assumed to be the radar
example by the arc of a circle. This filter recalculates reports in polar coordinates. Subsequently, the
the equations of the circle at every time instant and extended Kalman filter equations have to be applied.
thus, in the presence of reasonably fast sampling This model is driven by white noise acceleration,
approximates any smooth trajectory via multiple arcs which can be best described in the target’s coordinate
of circles. frame, where the Cartesian abscissa is aligned with the
The development of the circular filter is illustrated heading direction. The accelerations are divided into
in two-dimensional space and subsequently extended the parallel acceleration ap , the normal acceleration an
to three-dimensional space using orthogonal acting on the target in perpendicular direction, and the
binormal acceleration ab . The variances are defined as
transformations. For the determination of the
¾p2 , ¾n2 , and ¾b2 , respectively. The covariance matrix
covariance of the estimate, the proposed algorithm
described in the Cartesian system can be obtained
relies on the unscented transformation, a novel using the inverse three-dimensional rotation matrix
approach to approximate the statistics of a random of (13)
variable subject to a nonlinear transformation. The use Scart = GT diag[¾p2 ¾n2 ¾b2 ]G: (30)
of the unscented transformation enables us to calculate
the statistics of the circular prediction algorithm. The process noise Qd is recalculated for every time
The hybrid filter consists of a combination of the step. Due to the nonlinear state space system, the
orientation of the covariances depends on the heading
straight line prediction and the circular prediction. It
angle #. Assuming a discretized white noise the
has been shown on a benchmark maneuver that the
process noise is defined as [6]:
hybrid filter outperforms standard filters, such as the
extended Kalman filter using a straight line model. Qd =
In addition, a closed-form solution to the circular ½Z T Z T ¾
prediction algorithm has been derived, which has been E ©(T ¡ ¿1 )Bwc (¿1 )w(¿2 )Tc B T ©(T ¡ ¿2 )T d¿1 d¿2
0 0
implemented in conjunction with the extended Kalman
filter. Statistical tests have been performed to validate (31)
the performance increase using the hybrid filter. where © is the Jacobian of the state space, B is
The proposed filter in two-dimensional or the input matrix, and wc is the continuous white
three-dimensional can complement the bank of filters noise vector with the covariance Efwc (¿1 )w(¿2 )Tc g =
in an IMM filter [35] or the variable-dimension Qc ±(¿2 ¡ ¿1 )
filter developed by Bar-Shalom and Birmiwal [14]. 2 T3 T2 3
Scart Scart
Maneuver detectors have been applied to switch 6 3 2 7
Qd = 4 5: (32)
between different target models describing various T 2
maneuvers of the target. Scart TScart
2

1380 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 4 OCTOBER 2005
REFERENCES [16] Bar-Shalom, Y., Chang, K., and Blom, H.
Tracking a maneuvering target using input estimation vs.
[1] Sklansky, J. the interacting multiple model algorithm.
Optimizing the dynamic parameter of a track-while-scan IEEE Transactions on Aerospace and Electronic Systems,
system. 25 (Mar. 1989), 296—300.
RCA Laboratories, Princton, NJ, June 1957. [17] Jilkov, V., Angelova, D., and Semerdjiev, T.
[2] Benedict, T. R., and Bordner, G. W. Design and comparison of mode-set adaptive IMM
Synthesis of an optimal set of radar track-while-scan algorithms for maneuvering target tracking.
smoothing equations. IEEE Transactions on Aerospace and Electronic Systems,
IRE Transactions on Automatic Control, AC-1 (July 1962). 35, 1 (Jan. 1999), 343—350.
[3] Kalata, P. R. [18] Kirubarajan, T., and Bar-Shalom, Y.
® ¡ ¯ target tracking systems: A survey. Kalman filter versus imm estimator: When do we need
In Proceedings of the American Control Conference, 1992. the latter?
[4] Simpson, H. R. IEEE Transactions on Aerospace and Electronic Systems,
Performance measures and optimization condition for a 39, 4 (Oct. 2003), 1452—1457.
third order sampled-data tracker.
[19] Bishop, R., and Antoulas, A. C.
IEEE Transactions on Automatic Control, AC-12 (June
Non-linear approach to the aircraft tracking problem.
1962).
Journal of Guidance, Control and Dynamics, 17, 5 (Sept.
[5] Kalata, P. R. 1994), 1124—1130.
The tracking index: A generalized parameter for ® ¡ ¯
and ® ¡ ¯ ¡ ° target trackers. [20] Blackman, S., and Popoli, R.
IEEE Transactions on Aerospace and Electronic Systems, Design Analysis of Modern Tracking Systems.
AES-20, 2 (Mar. 1984). Norwood, MA: Artech House, 1999.
[6] Bar-Shalom, Y., and Fortmann, T. E. [21] Bullock, T., and Sangsuk-Iam, S.
Tracking and Data Association. Maneuver detection and tracking with a nonlinear target
New York: Academic Press, 1988. model.
[7] Lawton, J. A., Jesionowski, R. J., and Zarchan, P. In Proceedings of the Conference on Decision and Control,
Comparison of four filtering options for radar tracking vol. 1, Las Vegas, NV, Dec. 1984, 1122—1126.
problem. [22] von Mises, R.
Journal of Guidance, Control and Dynamics, 21, 4 (July Theory of Flight.
1998). New York: McGraw-Hill, 1945.
[8] Alouani, A., Xia, P., Rice, T., and Blair, W. [23] Nabaa, N., and Bishop, R. H.
A two-stage Kalman estimator for tracking maneuvering Validation and comparison of coordinated turn aircraft
targets. maneuver models.
In Proceedings of the IEEE International Conference on IEEE Transactions on Aerospace and Electronic Systems,
Systems, Man, and Cybernetics, vol. 2, 1991. 36 (Jan. 2000), 250—259.
[9] Singer, R. A. [24] Julier, S. J., Uhlmann, J. K., and Durrant-Whyte, H. F.
Estimating optimal tracking filter performance for New approach for filtering nonlinear systems.
manned maneuvering targets. In Proceedings of the American Control Conference, vol. 3,
IEEE Transactions on Aerospace and Electronic Systems, 1995, 1628—1632.
AES-5 (Nov. 1970).
[25] Julier, S., and Uhlmann, J.
[10] Berg, R. F. Unscented filtering and nonlinear estimation.
Estimation and prediction for maneuvering target Proceedings of the IEEE, 92, 3 (Mar. 2004), 401—422.
trajectories.
[26] Uhlmann, J. K.
IEEE Transactions on Automatic Control, AC-28 (Mar.
Dynamic map building and localization: New theoretical
1983), 294—304.
foundations.
[11] Farooq, M., and Bruder, S.
Ph.D. dissertation, University of Oxford, Trinity Term
Information type filters for tracking a maneuvering target.
1995.
IEEE Transactions on Aerospace and Electronic Systems,
26, 3 (May 1990), 441—454. [27] Uhlmann, J.
Covariance consistency methods for fault-tolerant
[12] Farooq, M., Bruder, S., Quach, T., and Lim, S.
distributed data fusion.
Adaptive filtering techniques for manoeuvering targets.
Information Fusion, 4 (2003), 201—215.
In Proceedings of the 34th Midwest Symposium on Circuits
and Systems, vol. 1, May 1991, 31—34. [28] Roecker, J. A., and McGillem, C. D.
[13] Magill, D. Target tracking in maneuver-centered coordinates.
Optimal adaptive estimation of sampled stochastic IEEE Transactions on Aerospace and Electronic Systems,
processes. 25, 6 (Nov. 1989), 836—843.
IEEE Transactions on Automatic Control, AC-10 (1965), [29] Kawase, T., Tsarunosono, H., Ehara, N., and Sasase, I.
434—439. An adaptive-gain alpha-beta tracker combined with
[14] Bar-Shalom, Y., and Birmiwal, K. circular prediction for maneuvering target tracking.
Variable dimension filter for maneuvering target tracking. In Proceedings of IEEE Speech and Image Technologies
IEEE Transactions on Aerospace and Electronic Systems, For Computing and Telecommunications, vol. 2, 1997.
AES-18, 5 (Sept. 1982), 621—629. [30] Kawase, T., Tsarunosono, H., Ehara, N., and Sasase, I.
[15] Blom, H., and Bar-Shalom, Y. Two-stage Kalman estimator using advanced circular
The intercating multiple model algorithm for systems predictions for maneuvering target tracking.
with markovian switching coefficients. In Proceedings of the IEEE International Conference on
IEEE Transactions on Automatic Control, 33 (Aug. 1988), Acoustics, Speech, and Signal Processing, vol. 4, May
780—783. 1998.

TENNE & SINGH: TRACKING FOR MANEUVERING TARGET TRAJECTORIES VIA THE 3D CIRCULAR FILTER 1381
[31] Tenne, D., and Singh, T. [34] Uhlmann, J., Julier, S., Kamgar-Parsi, B., Lanzagorta, M.,
Circular prediction algorithms–Hybrid filters. and Shyu, H-J.
In Proceedings of the American Control Conference, vol. 1, NASA Mars Rover: A testbed for evaluating applications
May 2002, 172—177. of covariance intersection.
[32] Kawase, T., Tsarunosono, H., Ehara, N., and Sasase, I. In Proceedings of the 1999 Unmanned Ground Vehicle
An adaptive-gain alpha-beta tracker combined with Technology Conference, vol. 3693, Naval Research Lab.,
three-dimensional circular prediction using estimation of Washington, D.C. Also in Proceedings of SPIE, 1999,
the plane state. 140—149.
In Proceedings of the IEEE International Conference on [35] Bar-Shalom, Y., and Li, X-R.
Acoustics, Speech, and Signal Processing, vol. 5, May Estimation and Tracking: Principles, Techniques and
1999. Software.
[33] Lerro, D., and Bar-Shalom, Y. Storrs, CT: YBS Publishing, 1998.
Tracking with debiased consistent converted
measurements versus EKF.
IEEE Transactions on Aerospace and Electronic Systems,
29, 3 (July 1993).

Dirk Tenne received his M.S. and Ph.D. degrees from the Department of
Mechanical and Aerospace Engineering, State University of New York at Buffalo
in 1998 and 2004, respectively. He received the Diplom-Ingenieur from the
Technische Universität Darmstadt, Germany in 1999.
Since 2004 he is a postdoctoral fellow at the Center for Multisource and
Information Fusion, SUNY Buffalo. His research includes target tracking and
statistically robust controller design.
Dr. Tenne has published papers at the Fusion Conference, the American
Control Conference, and the American Institute of Aeronautics and Astronautics
journals.

Tarunraj Singh received his B.E, M.E, and Ph.D degrees in mechanical
engineering from Bangalore University, Indian Institute of Science, and the
University of Waterloo, respectively.
He was a postdoctoral fellow in the Aerospace Engineering Dept. of Texas
A & M University prior to starting his tenure at the University at Buffalo in
1993, where he is currently a Professor in the Department of Mechanical and
Aerospace Engineering. He was a von Humboldt fellow and spent his sabbatical
at the Technische Universität Darmstadt in Germany and at the IBM Almaden
Research center in 2000—2001. He was a NASA Summer Faculty Fellow at
the Goddard Space Flight Center in 2003. His research is supported by the
National Science Foundation, AFOSR, NSA, Office of Naval Research and
various industries including MOOG Inc. Praxair and Delphi Thermal Systems.
His research interests are in robust vibration control, optimal control, nonlinear
estimation and intelligent transportation.
Dr. Singh has published over 100 refereed journal and conference papers
and has presented over 30 invited seminars at various universities and research
laboratories.

1382 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 41, NO. 4 OCTOBER 2005