Performance Evaluation of Interacting Multiple
Model using Nonlinear Filtering for Coastal Radar
Target Tracking System
Rika Sustika
Research Center for Informatics
Indonesian Institute of Sciences (LIPI)
Bandung, Indonesia
rika002@lipi.go.id
Abstract— This paper is examining the performance of
Interacting Multiple Model (IMM) using nonlinear filtering for
implementation on coastal radar target tracking system.
Nonlinear filtering algorithms that evaluated on this paper are
Converted Measurement Kalman Filter (CMKF) and Unscented
Kalman Filter (UKF). Many simulations on target tracking are
developed under assumption that noise characteristic is known
and there is no mismatch on noise modeling. In this paper, the
performance of IMM-CMKF and IMM-UKF was studied for
maneuvering and non maneuvering target dynamics with
condition that tracker (radar) doesn’t know process noise and
measurement noise characteristics. Parameters that used on
performance evaluation were execution time and percentage fit
error (PFE) of position estimation. Result from simulation
showed that execution time for IMM-CMKF is smaller than
IMM-UKF. Results also show that IMM-CMKF more robust
than IMM-UKF for condition where there is mismatch on noise
modeling.
Keywords- CMKF, filtering, IMM, radar, UKF, target tracking
I. INTRODUCTION
Primary objective of radar target tracking is to estimate
state trajectories of a moving target accurately by using noisy
measurement. There are many algorithms for radar target
tracking. Kalman filter is the most popular method in modern
target tracking systems because of its simplicity and
computational efficiency [1].
One of the issues in the design of target tracking system is
the choice of the target motion model. On the simplest
approximation, the model is assumed as the true dynamic
target and a single filter runs based on it. This approach has
several obvious flaws because the estimation does not take
into account a possible mismatch between the real target
dynamic and the filter model [2]. To solve this problem,
H.A.P Bloom introduced a safe adaptation or soft switching
method that known as Interacting Multiple Model (IMM) [3].
IMM use a bank of filter to estimate state variables of dynamic
system. Each filter used different model to characterize a
specific motion of a target, which makes it possible to describe
a whole motion. By using more than one model, the IMM
Joko Suryana
School of Electrical Engineering & Informatics
Institut Teknologi Bandung (ITB)
Bandung, Indonesia
algorithm is more capable to track targets with motion
uncertainty [2].
In tracking application, target motion is usually best
modeled using Cartesian coordinates [4]. When observation is
also in Cartesian coordinate, system can be modeled using
linear model. In this condition, IMM with Kalman filter can be
used to track the target. Unfortunately, on coastal radar system,
the target position measurements are commonly provided in
polar coordinate, in terms of range and azimuth with respect to
the radar location. The differences between tracking coordinate
and measurement coordinate make nonlinearity of the system
[5]. This matter makes Kalman filter can’t be used without
modification because Kalman filter only work on linear system
[7]. To solve this problem, there are some alternatives. The
simplest one is by converting the measurements to a Cartesian
frame, and then Kalman filter is used as filtering algorithm.
This method known as Converted Measurement Kalman Filter
(CMKF) [4]. Another solution is by using nonlinear filtering
based on Kalman Filter. The benchmark of nonlinear filtering
based on Kalman filter is Extended Kalman Filter (EKF) [8].
On 1995, Simon J. Julier and Jeffrey K. Uhlmann proposed
another algorithm that known as Unscented Kalman Filter
(UKF). In their paper they said that UKF is more accurate and
less difficult to implement than EKF as benchmark on
nonlinear filtering [6].
In this paper, we present the performance evaluation of
IMM method using two nonlinear filtering algorithms based on
Kalman filter, those are IMM-CMKF and IMM-UKF, for
implementation on coastal radar target tracking system.
Evaluation is done to know the performance of these
algorithms in condition when radar doesn’t know the real target
dynamic and real noise characteristic. Execution time is also
used as evaluation parameter to know the comparison of
computational efficiency of the algorithms.
II. MODEL AND IMM ALGORITHM
A. Basic Model
The tracking of the single target is based on the choice of a
model to describe the dynamic of a target. The simplest target
motion model is described in the Cartesian coordinate system
by linear discrete-time difference equation with additive noise nomalization factor. Mixed inputs (means and covariances) of
as [4] each filter are calculated as
(1) |
∑ (7)
|
where A is the state transition matrix (based on model of ∑
target dynamic), is state vector on time index k. The state (8)
vector ( ) consists of position and velocity or acceleration of 2) Filtering
the moving target on Cartesian coordinate, i.e, As described before on the introduction, for implementation
=[ ]T. is process noise that is assumed to be on coastal radar when tracking on Cartesian coordinate and
white and zero mean Gaussian with covariance Q. measurement on polar coordinate, filter that used on this IMM
method is nonlinear filtering. Two filtering method has been
The target is tracked by a ground based radar and provides evaluated, those are CMKF and UKF. These methods will be
measurement of range (r) and azimuth (θ). The measurement described latter.
model is given as [4]
(2) , , , ,
where is time index, is measurement, and is
measurement noise. For this case, when tracking is done on
Cartesian coordinate but measurement on polar coordinate,
measurement model can be given as [4] Interaction / Mixing
,
(3) ; ; ; ;
⁄ , , ,
where k is time index, is range of the target, is azimuth of
the target, and are target position on Cartesian coordinate,
and and are measurement noise on polar coordinate, Filter 1 Filter 2
that is assumed to be white and zero mean Gaussian with Λ Λ
covariance R.
, (4) , , , ,
where is range measurement standard deviation and is
azimuth measurement standard deviation.
Update Model Probability

B. Interacting Multiple Model (IMM)

,
Interacting Multiple Model (IMM) algorithm is a solution
for target dynamic uncertainty or to handle a possible
mismatch between real target dynamic and filter model. The Combination
basic idea of IMM is assume a set of models as possible
candidates of the real dynamic target, run a bank of elemental
filters, each based on a unique model in the set, and generate ,
the overall estimates by a process based on the results of these
Figure 1. One cycle of IMM algorithm with 2 models
elemental filters. Figure 1 shows one cycle of interacting
multiple model algorithm that composed of two models. On
this paper, we used two filters with unique model on each
3) Update Model Probability
filter, one model used constant velocity (CV) and the other
In addition to mean and covariance, we compute the
used constant acceleration (CA) model.
likelihood of the measurement for each filter as
Λ v ; 0, S ( 9)
From Figure 1 we can say that basically IMM method
and probabilites of each model at time step k are calculated as
consists of four major steps: interaction or mixing, filtering,
update model probability, and combination. The equations for ∑ Λ (10)
each step are as follows [7]: Λ (11)
where c is normalizing factor.
1) Interaction
|
The mixing probabilities for each model are calculated as 4) Combination
∑ (5) Combination step is to compute state mean and covariance
| final, and computed with these equations
(6)
∑ (12)
Where is the probability of model Mi and is ∑ (13)
 III. FILTERING ALGORITHMS time update used equation on Kalman filter time update.
The explanation below is about two filtering algorithms that Unscented transform is used on measurement update because
have been evaluated on this research. nonlinearity is on measurement equation.

A. Converted Measurement Kalman Filter These equations below are measurement update on
Converted Measurement Kalman Filter is an alternative Unscented Kalman Filter algorithm [8].
approach for tracking in Cartesian coordinate using polar • Generation of sigma points
measurements. Polar coordinate measurement is transformed to χ | X | (27)
Cartesian coordinate systems, then conventional Kalman filter
χ | X n λ P| , i 1, … , n (28)
is applied. Kalman filter consists of two main steps, these are
time update and measurement update. χ | X n λ P| , i n 1, … ,2n
1) Time update [7]: (29)
• State matrix prediction • Map sigma points to measurement space
(14) y | h χ | , i=0,...,2n (30)
| |
• Covariance matrix prediction • Predict Z | , covariance S , and cross covariance of
| | (15) state dan measurement P .
Z | ∑ w y | (31)
T
2) Measurement update [7] S ∑ w y Z | y Z | R |
| |
• Innovation (residual) covariance (32)
| (16) P ∑ w χ X y Z T
(33)
| | | |
• Kalman gain update
• Filter gain
| (17)
K P S (34)
• State estimation update using last measurement • Update state dan covariance estimation
| | | (18) X | X | K Z Z | (35)
• Error covariance update (36)
| |
| | (19)

With the converted measurement Kalman filter, polar IV. SIMULATION SCENARIO
coordinate measurement first converted to Cartesian coordinate To simulate real condition that radar doesn’t know the
measurement using these equation. dynamic of target, we simulated two types of trajectory, those
cos (20) are non maneuvering and maneuvering trajectories. Non
sin (21) maneuvering trajectory is generated using CV (constant
Next, the measurement error matrix, R, needed to be adjusted velocity) model. Maneuvering trajectory is generated using CV
since data was measured as range and azimuth [8]. model, combined with CA (constant acceleration) model as
(22) maneuver dynamic. Sampling interval on simulation is 1
second and a total of 200 seconds. The target is tracked by a
(23) ground based radar with position on center coordinate (0,0).
Measurement data are generated by adding Gaussian noise to
when real trajectory. The values of parameters that are considered in
(24) the simulation can be presented on Table I.
(25) TABLE I
(26) SIMULATION PARAMETERS
Real and measurement trajectory generation parameters
B. Unscented Kalman Filter Parameter Value
Initial position [x y] [-1000 1000] (m)
Unscented Kalman Filter uses unscented transform to give Initial velocity [vx vy] [4 3] (m/s)
Gaussian approximation to the filtering solutions of nonlinear Psd of process noise (q_cv = q_ca) 0.1
filtering problem. In unscented transform, we deterministically Measurement noise:
chose a fixed number of sigma points, which capture the Range standard deviation (σr) 10 (m)
Azimuth standard deviation (σθ) 1o
desired moments (mean and covariance) of distribution of a
Time index of maneuvering 1-50 and 71-120 (s)
variable exactly. After that we propagate the sigma points
through the nonlinear function and estimate the moments of Filtering parameters
transformed variable from them. Parameter Value
Probabilities of switching model on IMM 0.95 0.05
0.05 0.95
For application on coastal radar, when tracking done on
Cartesian coordinate while measurement on polar coordinate, Prior probability of IMM [0.9 0.1]
 For initial estimation we used one point approach when the
first measurement data used as first estimation of and .
First estimation of velocity and on one point approach is
assigned as zero [8]. The performance is measured by the
execution time for each runs and the percentage fit error (PFE)
[9].

PFE 100 (37)

PFE 100 (38)

PFE PFE (39)

where and are target position estimation and and
are real target position. Every parameter is simulated
using 50 runs monte carlo simulation. For evaluation, in this Figure 3. Estimation trajectory 2 (maneuvering)
paper we say that performance is good enough for
implementation if PFE is under 3%. Execution time and percentage fit error as performance
parameter of the algorithms from this simulation can be seen
Simulations have done with three scenarios. First scenario
on Table II.
is when there is no mismatch on noise modeling. Its mean that
noise parameter that is used by filter is the same as real noise TABLE II
that was used on data trajectory and measurement generation as Filtering Result, 1st Scenario
can be seen on Table I. Second scenario is when there is
mismatch on noise modeling. Its mean that noise parameter Algorithm Execution time PFE (%)
that is used by filter is different with real noise that was used (second) Trajectory 1 Trajectory 2
IMM-CMKF 0.077 1.4541 1.6475
on real and measurement trajectory generation. Third scenario
IMM-UKF 0.285 1.3574 1.6312
is simulation when process noise and measurement noise is
variated. The third scenario is done to evaluated the robustness
of the algorithms when there is mismatch on noise modeling. Table II shows that execution time for CMKF algorithm is
smaller than UKF algorithm. Its mean that computation
V. SIMULATION RESULTS complexity of CMKF algorithm is simpler than UKF
st
A. 1 Scenario algorithm. In 50 runs of monte carlo simulation, percentage fit
error using IMM-CMKF is 1.4541% for trajectory 1 and
Table II and two figures below show the performance of IMM 1.6475% or trajectory 2. Percentage fit error using IMM-UKF
method using CMKF filtering (IMM-CMKF) compare with is 1.3574% for trajectory 1 and 1.6312% for trajectory 2.
IMM using UKF filtering (IMM-UKF), for condition when Results from all of the simulations on 1st scenario have PFE
there is no mismatch on noise modeling. under 3%. These results demonstrated that IMM-CMKF and
Sample for trajectory estimation as simulation result for 1st IMM-UKF have good performance when there is no mismatch
scenario can be seen on figure 2 and 3. Figure 2 is estimation on noise modeling. The good result also can be seen on figure
of non maneuvering trajectory and figure 3 is estimation of 2 and 3, when a target can be tracked by IMM-CMKF and
maneuvering trajectory. IMM-UKF with good performance. Estimation of trajectory
almost coincide with real trajectory.

B. 2nd Scenario
In reality, we don’t know real noise characteristic, so there
is possibility that there is mismatch on noise modeling. Sample
for trajectory estimation as simulation result for 2nd scenario
when there is mismatch on noise modeling can be seen on
figure 4 and 5. Noise parameters that used on filtering step is
listed on Table III.
TABLE III
FILTERING PARAMETERS. 2ND SCENARIO
Filtering parameters Value
Psd of process noise (q_cv = q_ca) 0.01
Measurement noise:
Range standard deviation (σr) 10 m
Figure 2. Estimation trajectory 1 (non maneuvering) Azimuth standard deviation (σθ) 40
 Figure 4 is estimation of non maneuvering trajectory and modeling. Significant error on trajectory estimation on 2nd
figure 5 is estimation of maneuvering trajectory. scenario can also be seen on Figure 5, when target maneuvered.

C. 3rd Scenario
To see robustness of filtering method, we have simulated
the filtering algorithms with variation of process noise power
spectral density (q) and measurement noise standard deviation
(in this paper, only standard deviation of azimuth (θ) that be
variated). Table V below shows the filtering parameters that
used on 3rd scenario.
TABLE V
FILTERING PARAMETERS, 3RD SCENARIO
Filtering parameters (with noise variation)
Parameter Value
Psd of process noise (q_cv = q_ca) 0.01 until 1 with interval 0.01
Measurement noise:
Range standard deviation (σr) 10 (m)
Azimuth standard deviation (σθ) 0.1 o until 4 o with interval 0.1o
Figure 4. Estimation trajectory 1 (non maneuvering)

Simulation result can be seen on some figures below.

Figure 6 and 7 are simulation for trajectory 1 (non
maneuvering trajectory) and figure 8 and 9 are simulation for
trajectory 2 (maneuvering trajectory).

Figure 5. Estimation trajectory 2 (maneuvering)

Figure 6. PFE with variation of process noise on trajectory 1

Filtering results can be seen on Table IV below.
TABLE IV
FILTERING RESULT, 2ND SCENARIO
Algorithm Execution time PFE (%)
(second) Trajectory 1 Trajectory 2
IMM-CMKF 0.077 1.7877 2.9946
IMM-UKF 0.285 1.7413 3.0287

By comparing Table II and Table IV we can see that

percentage fit error (PFE) for 2nd scenario is bigger than PFE
on 1st scenario. Its mean that performance of target tracking is
lack when there is mismatch on noise modeling. From Table
IV we also can see that from 50 runs of monte carlo simulation,
percentage fit error of IMM-CMKF is 1.7877% for trajectory 1
and 2.9946% for trajectory 2. While for IMM-UKF, percentage
fit error is 1.7413% for trajectory 1 and 3.0287% for trajectory
2. This result demonstrated that performance of IMM-CMKF is
better than IMM-UKF when there is mismatch on noise Figure 7. PFE with variation of σθ on trajectory 1

Figure 8. PFE with variation of process noise on trajectory 2
Figure 9. PFE with variation of σθ on trajectory 2
Figure 6 and 7 show that on tracking non maneuvering
target, error on noise modeling doesn’t make significant
influence to filtering performance. IMM-UKF has better
performance than IMM-CMKF but not too superior. The two
algorithms have good performance, with percentage fit error
under 3%.
Figure 8 and 9 show, when there are a maneuver on target
dynamic, mismatch on noise modeling (process noise or
measurement noise) make significant influence to filtering
performance, especially on IMM-UKF. From 50 runs monte
carlo simulation for each parameter, there are some
divergences on filtering using IMM-UKF. It makes mean of
percentage fit error from monte carlo simulation using IMM-
UKF greater than 3%. In this condition, CMKF give better
performance. Estimation error (PFE) from filtering using IMM-
CMKF lower than estimation error using IMM-UKF. From this
simulation we can say that IMM-CMKF is more robust than
UKF algorithm for condition when there is mismatch on noise
modeling.
VI. CONCLUSION
IMM-CMKF and IMM-UKF give good performance to
track a target with or without maneuvering, when there is no
mismatch on noise modeling. UKF algorithm is a little better
than CMKF algorithm in this condition, but execution time of
UKF is more than CMKF. When there is error on noise
modeling, IMM with CMKF algorithm has better performance
than IMM with UKF algorithm. IMM-UKF still has good
performance when no maneuver on target dynamic but the
performance is bad when there is maneuver on target dynamic.
It can be concluded that IMM-CMKF more efficient on
computation time and more robust than IMM-UKF for
implementation on coastal radar target tracking system.
REFERENCES
[1] K.Rameshbabu, J.Swarnadurga, et.al.,”Target Tracking System Using
Kalman Filter”, International Journal of Advanced Engineering
Research and Studies”, Vol II, pp. 90-94, Oct-Dec 2012.
[2] X. Rong Li, & V. P. Jilkov, “Survey of Maneuvering Target Tracking
Part V: Multiple-Model Methods”, IEEE Transactions On Aerospace
And Electronic Systems, 41(4), 1255-1321, Oct 2005.
[3] H.A.P Bloom, “An Efficient Filter for Abruptly Changing System”.
Proceedings of the 23rd IEEE Conference on Decision and Control, Las
Vegas, NV, Dec. 1984, 656-658.
[4] Sudesh K. Kashyap, G.Girija, and J.R Raol,”Error Model Converted
Measurement and Error Model Modified Extended Kalman Filters for
Target Tracking”, Defence Science Journal, vol 56, No 5, 2006.
[5] S. J. Julier, J.K. Uhlman, “A Consistent, Debiased Method for
Converting Between Polar and Cartesian Coordinate System”, In The
Proceedings Of AeroSense: The 11th International Symposium On
Aerospce/Defense Sensing, Simulation and Controls, 10-121, 1997.
[6] S. J. Julier, J. K. Uhlmann, ” A New Extension of the Kalman Filter to
Nonlinear Systems”, In The Proceedings Of AeroSense: The 11th
International Symposium On Aerospce/Defense Sensing Simulation and
Controls, 182-193, 1997.
[7] J. Hartikainen, A. Solin,A, S. Sarkka, “Optimal Filtering with Kalman
Filters and Smoothers: A Manual for the Matlab Toolbok EKF/UKF”,
Department of Biomedical Engineering and Computational Science,
Aalto University School of Science, Finland, Version 1.3, August, 2011,
http://becs.aalto.fi/en/research/bayes/ekfukf/documentation.pdf,
accessed on 10 Juni 2012.
[8] B. Balaji & Z. Ding, “A Performance Comparison of Nonlinear Filtering
Techniques Based on Recorded Radar Datasets”, Proceedings of SPIE,
7445(1), 2009.
[9] VPS Naidu, Girija G & Shanthakumar N, “Three Model IMM-EKF for
Tracking Targets Executing Evasive Maneuvers”, Paper no. AIAA2007-
1204, AIAA Aerospace Meeting and Exhibition, 8-11 Jan, 2007.