EKF/UKF Maneuvering Target Tracking using

Coordinated Turn Models with Polar/Cartesian

Velocity
Michael Roth∗ , Gustaf Hendeby∗† , and Fredrik Gustafsson∗
∗ Dept.
Electrical Engineering, Linköping University, SE-581 83 Linköping, Sweden.
Email: {roth, hendeby, fredrik}@isy.liu.se
† Dept. of Sensor & EW Systems, Swedish Defence Research Agency (FOI), SE-581 11, Linköping, Sweden.

Email: gustaf.hendeby@foi.se

Abstract—Nonlinear Kalman filter adaptations such as ex- significant role in target tracking applications [1] where it can
tended Kalman filters (EKF) or unscented Kalman filters (UKF) alleviate observability issues [10].
provide approximate solutions to state estimation problems in We build our work upon the previous study [11] and show
nonlinear models. The algorithms utilize mean values and covari-
ance matrices to represent the probability densities in the other- how the choice of state coordinates and CT model variants
wise intractable Bayesian filtering equations. As a consequence, influences the tracking performance for target motion in the
their estimation performance can show significant dependence horizontal plane. In particular, we compare states that comprise
on the choice of state coordinates. The here considered problem either polar or Cartesian velocity. It is shown how different
of tracking maneuvering targets using coordinated turn (CT) assumptions on the process noise input during discretization
models is one practically relevant example: The velocity in the
target state can either be formulated in Cartesian or polar of the underlying CT differential equations yield further model
coordinates. We extend a previous study to a broader range variants for each velocity parametrization and how these can
of CT models that allow for changes in target speed and turn be used for filtering.
rate, and investigate UKF as well as EKF variants in terms of In terms of algorithms, we investigate the UKF as well
their performance and sensitivity to noise parameters. The results as the EKF variants that were covered in [11]. We provide
advocate for the use of polar CT models.
performance results for a range of CT Kalman filters and
I. I NTRODUCTION analyze their sensitivity with respect to the noise parameters
of the stochastic CT models. This is done via Monte Carlo
We investigate tracking of maneuvering targets under the (MC) simulations on a set of benchmark trajectories taken
assumption of coordinated turn (CT) target motion. The CT from [12], with emphasis on segments that correspond to target
models that we consider describe horizontal motion at nearly maneuvers. All filters are operated with range and bearing
constant speed along circle segments, as commonly encoun- measurements that are transformed to Cartesian positions and
tered in aviation, and appear in different flavors throughout the use individually optimized process noise parameters so as to
tracking and filtering literature [1–5]. Kalman filter variants facilitate a fair comparison of the model and filter choice.
that are based on these models are important building blocks The outline of the paper is as follows. The employed CT
of interacting multiple model (IMM) filters [2]. models are introduced in Section II. Kalman filter adaptations
Within the IMM framework, Kalman filters with CT models for nonlinear estimation problems are described in Section III.
are typically used to track the target during maneuvers [3, 6]. The conducted experiment and a discussion of obtained results
Our investigation concentrates on nonlinear CT variants that are given in Section IV. Conclusions are given in Section V.
allow for varying turn rate as well as varying target speed, in
contrast to their linear cousins with known turn rate [5]. This II. C OORDINATED T URN M ODELS
choice is backed by an analysis of real world aircraft data [7] The list of available motion models used for tracking is
that advocates for the varying speed and turn rate assumption. long. Extensive sources include the survey [5] or the dedicated
A potential benefit [3] of an IMM algorithm with these more chapters in [4]. In this section we describe the underlying
flexible CT models is a reduced number of required modes. continuous-time dynamics before deriving discrete time mod-
The nonlinear nature of the considered CT models requires els based on polar and Cartesian velocity representations.
the use of nonlinear Kalman filter adaptations such as extended Finally, we discuss the noise inputs of stochastic CT models.
Kalman filters [8] (EKF) or unscented Kalman filters [9]
(UKF). One consequence of the inherent approximations in A. From Particle Motion to Discrete-time CT Models
such algorithms is that the estimation performance depends Our focus is on coordinated turn models which have their
on the choice of the state coordinates. This observation dates origin in curvilinear particle motion [5]. Figure 1 illustrates the
back to the early days of Kalman filtering [8] and plays a involved variables. The target position in the horizontal plane
 x2  

where the indices denote sampling instants and w = a α
T
v1
v2
comprises the acceleration inputs. The sample time T en-
h ters the expression implicitly. For simplicity we treat w as

☛
deterministic and introduce randomness once the discrete-
x1 time equations have been established. Schemes that discretize
related models under the assumption of random inputs are
Fig. 1. Illustration of state components. The target is depicted in the x1 – shown in [13] and [14], where the latter considers ω not as
x2 –plane. The attached velocity vector has components v1 and v2 in x1 and part of the state but as input.
x2 direction, respectively. v is the magnitude of the velocity vector and h the
heading angle.

B. Coordinated Turn with Polar Velocity

with respect to a fixed Cartesian coordinate frame is denoted
by x1 and x2 [m]; the velocity vector has components v1 and We begin with the discretization of (1) which is in general
v2 and magnitude v, all [m/s]; h [rad] denotes the heading not trivial. For a = 0 and α = 0, however, a compact solution
angle, given by the two argument arctangent atan2(v2 , v1 ). to the differential equations is readily available and given by
Furthermore, we define the turn rate ω = dh dt [rad/s] as the
time derivative of the heading angle. ωT ωT
x1 + 2v
 
We assume that the target is driven by linear and rotational ω sin( 2 ) cos(h + 2 )
 x2 + 2v sin( ωT ) sin(h + ωT ) 
acceleration inputs a [m/s2 ] and α [rad/s2 ], respectively. Here,  ω 2 2 
fp (x) =  v . (4)
a is also known as longitudinal acceleration that acts tangential  
 h + ωT 
to the target path, thus aligned with the Cartesian velocity
ω
vector. While a accounts for speed changes, α influences the
turn rate. The coupled differential equations
    The index p stands for “polar” and, naturally, fp (x) takes
ẋ1 v cos(h) states with polar velocity as input argument. Equation (4)
ẋ2   v sin(h) 
    is given in [11] but equivalent solutions can also be found
 v̇  =  a  (1)
    elsewhere [15].
 ḣ   ω 
If we assume non-zero a and α, the solutions to (1)
ω̇ α
become increasingly complicated, depending on the behavior
of a and α between sampling instants. A discussion of
T state equation for the target state x =
form a continuous-time
x1 x2 v h ω with velocity in polar coordinates. The such discretization problems, including suggested solutions, is
position entries in (1) depend nonlinearly on h, whereas the provided in [4]. Here, we settle for using (4) and concentrate
acceleration inputs a and α enter linearly. on the input part of (3), which amounts to finding G(x).
We investigate two schemes of which the first asserts that a
 If we change to aTstate with Cartesian velocity, x = and α act as impulses just before each sampling instant.
x1 x2 v1 v2 ω , the resulting differential equation
    Consequently, each input acts only on the state components
ẋ1 v1 that are directly influenced. The second approach assumes
ẋ2  
   v2 
 that a and α are constant between sampling instants, and is
 v̇1  = a cos(h) − v2 ω  (2) commonly known as zero-order-hold discretization. For the
   
 v̇2   a sin(h) + v1 ω  linear part (v, h, ω) of (1) the solution for constant w is
ω̇ α easily derived, but the solution for (x1 , x2 ) becomes very
complicated due to the time-varying h. Euler’s method, i.e.,
exhibits bilinear relations v1 ω and v2 ω. Furthermore, the asserting a constant h between sampling instants, yields a
assumption of mere longitudinal acceleration (apart from the simple approximate solution. The resulting G(x) of (3) are
turning acceleration) now yields a nonlinear dependency on given by
h = atan2(v2 , v1 ) in the velocity. The coordinate change
relating the two state vectors is obvious from the first two  
  T2
components in (1) and (2). 0 0 cos(h) 0
0  T22
The state differential equations (1) and (2) are both in the 0 sin(h)

   2 
form ẋ = ϕ(x, a, α). Hence, we need to first discretize in order G1p = 
T 0, G2p (h) = T 0 (5)
 
0 2
to apply discrete-time filtering techniques later. This amounts 0 T 

 0 2
to (approximately) solving (1) and (2) for the state at time 0 T 0 T
instance k+1 given the state at time k, an initial value problem.
The results we obtain turn out to be of the form
for the impulse and constant interpretations of w, respectively.
xk+1 = f (xk ) + G(xk )wk , (3) Obviously, G1p is independent of the state.
C. Coordinated Turn with Cartesian Velocity III. EKF AND UKF FOR N ONLINEAR F ILTERING
The discrete-time version of (2) follows from (4) by con- In this section, an overview of different Kalman filter vari-
verting from polar to Cartesian velocity and is given by ants for nonlinear filtering is provided. For a more extensive

x1 + vω1 sin(ωT ) − vω2
1 − cos(ωT )

 treatment of filtering theory and the actual algorithms, the
x2 + v1 1 − cos(ωT ) + v2 sin(ωT ) reader is referred to [4] and the references below.
 ω ω 
fc (x) = 
 v1 cos(ωT ) − v2 sin(ωT ) .
 (6) A. Bayesian Filtering Theory
 v1 sin(ωT ) + v2 cos(ωT ) 
Filtering in nonlinear state-space models is a difficult prob-
ω
lem. The widely accepted Bayesian filtering equations [4]
Solving (2) under the assumption a = 0 and α = 0 yields
p(yk |xk )p(xk |y1:k−1 )
equivalent results. Often, (6) is written as product of an p(xk |y1:k ) = (8a)
ω-dependent matrix with the Cartesian state vector which p(y1:k )
Z
discloses a linear model for known ω [2, 5]. The corresponding p(xk |y1:k−1 ) = p(xk |xk−1 )p(xk−1 |y1:k−1 ) dxk−1 (8b)
input matrices in (3) follow from arguments similar to those Z
in the polar case and are given by p(y1:k ) = p(yk |xk )p(xk |y1:k−1 ) dxk (8c)
  T2 
0 0 22 cos(h) 0
 0 0  T sin(h) 0  are conceptually appealing but, due to the required marginal-
, G2c (h) =  T2 cos(h) 0  (7a) ization and conditioning operations, intractable for all but
   
G1c (h) =  T cos(h) 0

 T sin(h) 0 
 
 T sin(h) 0 
 the simplest cases. This has long been known [16]. Hence,
0 T considerable effort has been put into the development of
0 T
algorithms that approximate (8).
with impulse and constant w, respectively. Since we assume The most famous special case for which (8) can be solved
only longitudinal acceleration, G1c (h) and G2c (h) depend on h, comprises a linear state-space model with Gaussian noise
which is a function of v1 and v2 . Therefore, (7a) appears more and initial state. The solutions are given by a number of
nonlinear than (5). Alternatively, a third variant with individual recursive update equations that have become known as the
Cartesian acceleration inputs a1 and a2 can be formulated. Kalman filter (KF) [4, 8]. Of course we should note that
 T
With w = a1 a2 α we obtain the KF can be obtained from several different viewpoints
T2  that establish further notions of optimality for linear (not
2 0 0 necessarily Gaussian) problems [8].
 0 T2
 2 0
3
Gc =  T
 0 0 . (7b) B. Kalman Filter Adaptations for Nonlinear Filtering
 0 T 0 It is a natural idea to adapt the Kalman filter to nonlinear
0 0 T problems, and indeed this is where the algorithm made a first
Combining (6) and (7b) yields a widespread CT model variant considerable impact: The KF was employed to solve nonlinear
[2, 5]. Obviously, G3c does not depend on the state and navigation problems in the Apollo mission [17].
the resulting model allows acceleration in any direction, in The first nonlinear KF developed became known as the
contrast to the longitudinal only acceleration that we assumed extended Kalman filter (EKF), and is based on consecutive
before. In this respect, G3c bears a close resemblance to the linearization of the nonlinear dynamic and measurement equa-
input matrix of constant velocity models [4, 5]. tions about current state estimates. The linearization is typi-
cally carried out via Taylor series expansion but interpolation
D. Process Noise Assumptions schemes based on numerical differentiation also exist [18].
So far, we have discussed different deterministic CT models Going beyond linearization, the second-order EKF (EKF2)
that employ polar or Cartesian velocity. In order to obtain a retains the quadratic terms in the Taylor series. This attempt
stochastic CT model, we formulate a probabilistic description to approximate the system equations more accurately comes
of the input signals a and α. As indicated earlier, we pursue at the expense of a vastly increased computational complexity
a direct discrete-time approach [2] rather than discretizing if implemented without further care [19, 20].
random processes, which means that we treat a and α (whether A linear approximation cannot fully capture the behavior of
impulse or constant between sampling instants) as uncorrelated a nonlinear system. The unscented Kalman filter [9] (UKF) is
random variables with zero mean and standard deviations an algorithm based on deterministic sampling that addresses
σa [m/s2 ] and σα [rad/s2 ]. Furthermore, we treat the obtained this to some extent. Instead of approximating the state and
noise sequences ak and αk as white. The zero-mean input measurement functions of a nonlinear state space model, few
yields the typical nearly constant speed and nearly constant deterministic state realizations — called sigma points — are
turn rate behavior that is widely used in motion models for generated and propagated through the involved nonlinearities.
filtering. In fact, some authors prefer the name (nearly) con- The sample statistics of transformed sigma points are then
stant turn models to make a distinction to the more complex used in the KF equations. The main ingredient of the UKF
coordinated turns in aviation [5]. is a computation scheme for transforming mean values and
covariance matrices called unscented transformation (UT), Trajectory 1 Trajectory 2
which depends on a set of scalar parameters. It turns out that a
simple choice of UT parameters yields a numerical integration

x2 [km]
rule of degree 3, and gives rise to a UKF variant called the 20 −30
cubature Kalman filter [21] (CKF). The numerical integration
10
viewpoint in the CKF derivation facilitates accuracy statements −40
that are otherwise difficult to obtain for the UKF. A discussion 50 60 70 20 30 40
of this can be found in [20]. The fact that UKF variants Trajectory 3 Trajectory 4
are sampling based and do not involve Jacobian and Hessian 60 30
matrices, in contrast to EKF and EKF2, makes them simple 20
50

x2 [km]
to implement.
10
Relations between UKF and EKF2 have been established in 40
[22]. With the “Gaussian derivation” of the Kalman filter in 0
30
mind, a common framework for all of the above adaptations 30 40 50 60 10 20 30 40
can be formulated [20, 22]. After all, KF variants employ the Trajectory 5 Trajectory 6
same kind of measurement update equation that is linear in
the residual yk − ŷk|k−1 , the difference between predicted and 30
20

x2 [km]
measured output. The algorithms differ in the methods that 20 10
are used for computing mean values and covariance matrices
10 0
of nonlinearly transformed random variables. An extensive
survey of these moment computation schemes and their use 50 60 70 80 90 30 40 50 60 70
in Kalman filtering is given in [20]. x1 [km] x1 [km]

C. The Role of the State Coordinate System

Fig. 2. Trajectories and the segment in which the target is maneuvering
Kalman filter variants propagate mean values and covariance (marked in bold red).
matrices of filtered and predicted states throughout time.
Therefore, a widespread interpretation is that the densities
in (8) are assumed to be Gaussian. This is a very restrictive as- to accentuate the behavior during turns, only segments of the
sumption and cannot account for phenomena like multimodal- trajectories with pronounced acceleration are included in the
ity or skewness. As a consequence of the limited flexibility, evaluation. These sections are marked in Figure 2.
the choice of state coordinates can have a significant influence For each trajectory, range and bearing measurements are
on the estimation performance of any nonlinear Kalman filter generated by a noisy sensor in the origin, with white Gaussian
variant. noise with standard deviation σθ = 0.1 [◦ ] and σr = 100 [m]
for bearing and range, respectively. The measurements and
IV. E XPERIMENT AND D ISCUSSION their variances are then converted to Cartesian coordinates to
In this section different filter and CT model combinations yield position measurements with time-varying measurement
are assessed on a maneuvering target tracking problem. covariance matrices. The transformation is performed using
both the Gauss approximation and the UT. Both methods
A. Experiment Setup produce comparable results for the chosen σθ and σr .
Monte Carlo simulations are used to evaluate the different The tracking performance is evaluated by computing the
filters and state-space representations. For this purpose, the six root mean square error (RMSE) for the true target position
benchmark target trajectories of [12] are used which represent compared to the predicted state position. We average first
different maneuvering aircraft targets, ranging from a slow over MC realizations and second over time steps. Using the
cargo aircraft, over smaller maneuverable Learjet-type aircraft, predicted state position instead of the filtered position ensures
to very agile and fast aircraft. We illustrate the horizontal that we evaluate the full state and the model assumption, not
trajectories in Figure 2 and refer the reader to [12] for a only the position estimate which would be predominant using
more comprehensive description and further plots, including the filtered position. As mentioned above, only sections of
acceleration magnitude. the trajectories with acceleration are included in the RMSE
The CT models evaluated here cover only two-dimensional computations. Hence the performance is only evaluated for
CT motion and therefore only the target motion in the hor- the highlighted segments in Figure 2.
izontal plane is considered. Motivation for such a problem The specific filters that are compared are EKF, EKF2, and
separation is given in [3, 7]. Furthermore, all the trajectories UKF for all CT models of Section II. Per algorithm type
contain considerable constant velocity segments for which the this amounts to two polar and three Cartesian models that
standard KF is known to work very well, whereas CT models are tested. For each batch of measurement realizations, the
are better suited to handle the turns. Several solutions exist to filters are run with different noise settings. Picking the best
utilize this difference in behavior, the IMM is one. Therefore, parameters for all combinations facilitates a fair comparison.
The magnitude of the process noise in each algorithm is varied 200
using a logarithmic scale with a few samples per decade, with

avg. position RMSE [m]

the actual measurement noise that was used to generate the 1: 142.0
measurements. These simulations also indicate how sensitive 150 2: 138.0
the filter-model combination is to the process noise parameters. 3: 155.2
The process noise covariance is computed according to 4: 103.6
100 5: 159.4
Qk = G(xk ) diag(σa2 , σα2 )G(xk )T (9) 6: 153.0

for all of the filter-model combinations except for those that

employ G3c of (7b). There, a slightly altered expression is used. 50
0.5 1 1.5 2 2.5 3 3.5
To set a baseline for the experiments, a linear Kalman log(σa)
filter with constant velocity (CV) model is applied along all
nonlinear filters. Again, the position RMSE is computed for Fig. 3. Average position RMSE [m] of a linear Kalman filter with CV model
the several process noise settings so as to find the best possible for all trajectories. Only segments in which the target maneuvers are taken
into account.
performance on the maneuver segments of the trajectories.
B. Simulation Results and Discussion 97.7
0 200
In order to give the reader some qualitative performance
insights we illustrate selected results of the best algorithms −0.5
in each category instead of showing results for all model and −1 150
filter combinations. We begin by first assessing the linear KF

log(σα)
with a CV model, then EKF and UKF with Cartesian CT −1.5
models, and finally EKF and UKF with polar CT models. The −2 100
presented RMSE values were computed from the predicted
−2.5
position estimates on 10 measurement realizations.
The UKF variants showed almost no effect to different −3 50
values of the UT parameters. Therefore, the most basic version −1 0 1 2
log(σa)
of a UKF with zero central weight has been used. This is in
fact the CKF of [21].
Fig. 4. RMSE as function of noise standard deviation log(σa ) and log(σα )
Figure 3 shows the prediction results that were obtained for EKF2 with Cartesian CT model on Trajectory 1. Only segments in which
with the CV KF. Illustrated is the average position RMSE the target maneuvers are taken into account.
against the logarithm log(σa ) (base 10) of the acceleration
noise standard deviation σa that was employed in the KF.
The performance on Trajectory 4 is better than on the other Trajectory 5 on which the performance is equally bad. A closer
trajectories, which can be explained by a longer segment look at Trajectory 5 reveals that it contains both segments
with only minor maneuvering that is well suited to the CV of very little maneuvering (constant low acceleration) and
assumption. Furthermore, we include the best RMSE value heavy maneuvering and is as such difficult to handle with
on each trajectory in the legend. A sensible more advanced a single filter. Changing from Cartesian EKF to EKF2 does
algorithm must beat these numbers in order to be worth the not yield significantly different performance but has another
extra effort. interesting effect. The obtained RMSE values become less
Next, we turn to CT models with Cartesian velocity. These sensitive to variations in σa . This can be seen by comparing
require the user to select the two parameters σa and σα . Figures 4 and 5 that were obtained with the same model.
Figure 4 serves as example for all upcoming illustrations and A similar observation has been made in [6] where EKF2
displays the average position RMSE as contour plot of log(σa ) is said to improve robustness of a Cartesian velocity CT
and log(σα ). Dark red areas correspond to parameter choices EKF. Apart from this desirable effect, EKF2 significantly
that produced an RMSE over 200 m. The color scale is valid increases the computation time. Owing to the sparsity of the
for the other plots of the same kind. The shown results are for required Hessian matrices; however, the computations could
an EKF2 using a CT model with state-dependent input matrix be significantly reduced by utilizing the problem structure.
G2c (h) of (7a) on Trajectory 1. The best RMSE in the plot A comparison of Cartesian UKF variants yields the results
title shows that the algorithm clearly outperforms the CV KF. of Figure 6 which are comparable to the EKF in Figure 5 on
A cross marks the optimal parameter choices. Trajectories 3–6 and slightly worse on Trajectories 1 and 2.
We begin with an analysis of EKF and compare the three Again, G2c (h) of (7a) gives the best results, but also using the
covariance matrices of (7a) and (7b). It turns out that G2c (h) simple G3c of (7b) performs well. Similar to EKF2, the UKF
of (7a) provides better results than G1c (h) and G3c . The seems to be less sensitive to changes in σa than the EKF.
corresponding RMSE plots are given in Figure 5. Apparently, We next turn to the CT models with polar velocity. The
we can do better than with a CV KF. One exception is matrices G1p and G2p (h) in (5) give rise to two variants. For
 94.7 98.9 111.6 106.3
0 0
−1 −1
log(σα)

log(σα)
−2 −2
−3 −3
130.4 107.7 132.2 114.5
0 0
−1 −1
log(σα)

log(σα)
−2 −2
−3 −3
159.1 134.0 154.4 134.2
0 0
−1 −1
log(σα)

log(σα)
−2 −2
−3 −3
−1 0 1 2 −1 0 1 2 −1 0 1 2 −1 0 1 2
log(σa) log(σa) log(σa) log(σa)

Fig. 5. RMSE on all trajectories for the best EKF with Cartesian CT model. Fig. 6. RMSE on all trajectories for the best UKF with Cartesian CT model.
Color bar given in Figure 4. Arrangement of trajectories similar to Figure 2. Color bar given in Figure 4. Arrangement of trajectories similar to Figure 2.

EKF, the results show little preference for either of the two, close for Trajectories 3 to 6, and the RMSE is less sensitive
so we present the results for G2p (h) in Figure 7. These show to σa for Trajectories 1 and 2 (i.e., for the least agile targets).
another improvement over the EKF with Cartesian CT model.
The regions of low RMSE on Trajectories 1 and 2 appear V. C ONCLUDING R EMARKS
larger than for all Cartesian CT filters. The filter is robust to
deviations from the optimal σa and σα . We have shown a range of coordinated turn (CT) models
Changing to EKF2 does not yield any significant improve- using either Cartesian or polar velocity and how to use them in
ments which might be explained by the fact that the lower part a Kalman filtering framework for maneuvering target tracking.
of (4) is actually linear. The corresponding Hessian matrices The results of the conducted simulation study are in favor
are zero. A similar statement can be made for UKF, as shown of polar velocity. This confirms the results of the previous
in Figure 8 for G2p (h). study [11] and extends it to the case of varying target speed.
For polar CT models, the performance in terms of position
C. Summary RMSE of the predicted state appears to be comparable for
We made the following observations. For the considered EKF and UKF. As the UKF does not require the derivation and
trajectory segments, the use of nonlinear Kalman filters with implementation of Jacobians it might be more straightforward
coordinated turn models yields an improvement over a linear to implement. The RMSE provided by the Cartesian velocity
Kalman filter with constant velocity model. EKF and UKF turned out slightly worse. Interestingly, the
For Cartesian velocity UKF and EKF2 can reduce the sensitivity of the RMSE with respect to the noise parameters
sensitivity to the process noise parameters. Still EKF shows was decreased by using EKF2 and UKF in the Cartesian case.
the best quantitative performance. In terms of the model, a This, in addition to the simpler implementation and lower
state dependent input matrix G2c (h) from (7a) appears best if computational cost of UKF over EKF2 results in a recom-
a Cartesian velocity representation is used. mendation for UKF if Cartesian CT models are preferred.
Using polar velocity representation instead can yield a Our results indicate that for each CT filter there is one
further decrease in RMSE. Here, the choice of input matrix (5) set of noise parameters that yields good performance on all
is less significant. UKF and EKF provide comparable accuracy. investigated trajectories, although the maneuver characteristics
The RMSE contour plots in Figures 4 to 8 reveal that are quite different from trajectory to trajectory. Furthermore,
each individual algorithm can be run with the same parameter we have shown that the use of such flexible CT filters yields
settings for all trajectories without much increase in RMSE. improved results (on maneuvering targets) when compared to
The optimal σα are close for all trajectories, the optimal σa are a well tuned linear KF with constant velocity model. These
 84.4 90.9 96.4 94.2
0 0
−1 −1
log(σα)

log(σα)
−2 −2
−3 −3
131.3 103.2 129.3 112.9
0 0
−1 −1
log(σα)

log(σα)
−2 −2
−3 −3
147.0 125.6 146.3 129.7
0 0
−1 −1
log(σα)

log(σα)
−2 −2
−3 −3
−1 0 1 2 −1 0 1 2 −1 0 1 2 −1 0 1 2
log(σa) log(σa) log(σa) log(σa)

Fig. 7. RMSE on all trajectories for the best EKF with polar CT model. Fig. 8. RMSE on all trajectories for the best UKF with polar CT model.
Color bar given in Figure 4. Arrangement of trajectories similar to Figure 2. Color bar given in Figure 4. Arrangement of trajectories similar to Figure 2.

insights advocate very much for the use of the investigated CT

filters in IMM tracking algorithms. [7] N. Nabaa and R. Bishop, “Validation and comparison of
coordinated turn aircraft maneuver models,” IEEE Trans.
ACKNOWLEDGMENT Aerosp. Electron. Syst., vol. 36, no. 1, pp. 250–259, 2000.
The work was supported by the project Cooperative Local- [8] B. D. Anderson and J. B. Moore, Optimal Filtering.
ization (CoopLoc) funded by Swedish Foundation for Strategic Prentice Hall, Jun. 1979.
Research (SSF) and the Swedish strategic research center [9] S. J. Julier and J. K. Uhlmann, “Unscented filtering and
Security Link. nonlinear estimation,” Proc. IEEE, vol. 92, no. 3, pp.
401– 422, Mar. 2004.
R EFERENCES [10] M. Mallick, M. Morelande, L. Mihaylova, S. Arulam-
[1] S. Blackman and R. Popoli, Design and Analysis of palam, and Y. Yan, “Comparison of angle-only filtering
Modern Tracking Systems. Artech House, Aug. 1999. algorithms in 3D using Cartesian and modified spherical
[2] Y. Bar-Shalom, X. R. Li, and T. Kirubarajan, Estimation coordinates,” in 2012 15th International Conference on
with Applications to Tracking and Navigation. Wiley- Information Fusion (FUSION), Jul. 2012, pp. 1392 –
Interscience, Jun. 2001. 1399.
[3] Y. Bar-Shalom, P. K. Willett, and X. Tian, Tracking and [11] F. Gustafsson and A. Isaksson, “Best choice of coordinate
Data Fusion: A Handbook of Algorithms. Storrs, CT: system for tracking coordinated turns,” in Proceedings
YBS Publishing, 2011. of the 35th IEEE Conference on Decision and Control,
[4] F. Gustafsson, Statistical Sensor Fusion. Studentlitter- 1996, vol. 3, Dec. 1996, pp. 3145 –3150.
atur AB, Mar. 2010. [12] W. Blair, G. Watson, T. Kirubarajan, and Y. Bar-Shalom,
[5] X. R. Li and V. Jilkov, “Survey of maneuvering target “Benchmark for radar allocation and tracking in ECM,”
tracking. Part I. Dynamic models,” IEEE Trans. Aerosp. IEEE Trans. Aerosp. Electron. Syst., vol. 34, no. 4, pp.
Electron. Syst., vol. 39, no. 4, pp. 1333 – 1364, Oct. 1097–1114, Oct. 1998.
2003. [13] M. Morelande and N. Gordon, “Target tracking through
[6] P. Vacher, I. Barret, and M. Gauvrit, “Design on a a coordinated turn,” in IEEE International Conference
tracking algorithm for an advanced ATC system,” in on Acoustics, Speech, and Signal Processing, 2005. Pro-
Multitarget-Multisensor Tracking: Applications and Ad- ceedings. (ICASSP ’05), 2005, pp. iv/21–iv/24.
vances. Norwood, MA: Artech House, 1992, vol. 2, pp. [14] P.-A. Kountouriotis and S. Maskell, “Maneuvering tar-
1–28. get tracking using an unbiased nearly constant heading
 model,” in 2012 15th International Conference on Infor-
mation Fusion (FUSION), Jul. 2012, pp. 2249–2255.
[15] J. L. Gertz, “Multisensor surveillance for improved air-
craft tracking,” MIT Lincoln Laboratory Journal, vol. 2,
no. 3, pp. 381–396, 1989.
[16] A. H. Jazwinski, Stochastic Processes and Filtering
Theory. Academic Press, Mar. 1970.
[17] M. Grewal and A. Andrews, “Applications of Kalman
filtering in aerospace 1960 to the present,” IEEE Control
Syst. Mag., vol. 30, no. 3, pp. 69–78, 2010.
[18] M. Nørgaard, N. K. Poulsen, and O. Ravn, “New de-
velopments in state estimation for nonlinear systems,”
Automatica, vol. 36, no. 11, pp. 1627–1638, Nov. 2000.
[19] M. Roth and F. Gustafsson, “An efficient implementation
of the second order extended Kalman filter,” in Proceed-
ings of the 14th International Conference on Information
Fusion (FUSION), Jul. 2011.
[20] M. Roth, “Kalman filters for nonlinear systems and
heavy-tailed noise,” Licentiate thesis, Linköping Univer-
sity, Linköping, Sweden, Sep. 2013. [Online]. Available:
http://urn.kb.se/resolve?urn=urn:nbn:se:liu:diva-97544
[21] I. Arasaratnam and S. Haykin, “Cubature Kalman filters,”
IEEE Trans. Autom. Control, vol. 54, no. 6, pp. 1254–
1269, Jun. 2009.
[22] F. Gustafsson and G. Hendeby, “Some relations between
Extended and unscented Kalman filters,” IEEE Trans.
Signal Process., vol. 60, no. 2, pp. 545 –555, Feb. 2012.