ARTICLE IN PRESS

Signal Processing 89 (2009) 2171–2177

Contents lists available at ScienceDirect

Signal Processing
journal homepage: www.elsevier.com/locate/sigpro

A novel interacting multiple model algorithm$

Qu HongQuana,, Pang LiPingb, Li ShaoHongb
a
College of Information Engineering, North China University of Technology, Beijing 100144, China
b
School of Electronics and Information Engineering, Beihang University, Beijing 100191, China

a r t i c l e in fo abstract

Article history: For maneuvering target tracking, the interacting multiple model (IMM) algorithm
Received 1 August 2008 employs a fixed model set. The performance of this algorithm depends on the model set
Received in revised form adopted. The result of using too many models is as bad as the case of too few models.
28 April 2009
Therefore, a variable structure IMM (VSIMM) was presented and applied to ground
Accepted 28 April 2009
target tracking. This algorithm improves performance and reduces computational load
Available online 6 May 2009
with using auxiliary information. But it is difficult to extend the VSIMM to other
Keywords: scenario (for example, aerial target), where there is not auxiliary information such as a
Interacting multiple model map. A novel interacting multiple model (Novel-IMM) algorithm was presented to solve
Variable structure IMM
the problem of model set adaptation without auxiliary information. The Novel-IMM
Model set adaptation
algorithm consists of N independent IMM filters operating in parallel, and each
Novel-IMM
independent IMM filter also consists of multiple sub-filters, which operate interactively.
In every time index, only one IMM output of a certain model set is used; but for a long
time, the algorithm will alternatively choose an output of the model set to be the
optimum final output. The Novel-IMM approach was illustrated in detail with an aerial
complex maneuvering target tracking example.
& 2009 Elsevier B.V. All rights reserved.

1. Introduction processing to process control [2,3]. The application of IMM

filters for detection and diagnosis of anticipated reaction
Computing the optimal state estimate for a jump wheel failures in the attitude control system has been
Markov system requires exponential complexity, and described and developed in Refs. [4,5]. The IMM method
hence, practical filtering algorithms are necessarily sub- for fault detection of railway vehicle from the measure-
optimal. In target tracking literatures, suboptimal multi- ment of the lateral acceleration of bogie and body and
ple model filtering algorithms, such as the interacting from rate of bogie has been described in Refs. [6,7].
multiple model (IMM) method, are widely used for state The IMM algorithm, which uses a fixed set of models,
estimation of such systems [1]. The IMM estimation has usually performs reasonably well for problems with a
received a great deal of attention in recent years due to its small model set. Many practical problems, however,
unique power and great success in handling problems involve more than just a small number of models [8].
with both structural and parametric uncertainties and/or When these algorithms are applied to solve target
changes, and in decomposing a complex problem into tracking problems, it happens that only using a small
simpler sub-problems, ranging from target tracking to number of models cannot get satisfying result. At any
fault detection and isolation, and from biomedical signal time, the target trajectory evolves according to one of a
finite number of predetermined model set. This requires
$
that the model set include as many models as necessary to
Supported by National Science Foundation of China (no. 50808007)
handle the varying target motion characteristics. But this
and the NCUT Young Major Research Foundation (no. 200802).
 Corresponding author. Tel.: +86 10 88801519. cannot improve the performance successfully. The result
E-mail address: [email protected] (H.Q. Qu). of using too many models is as bad as the case of too few

0165-1684/$ - see front matter & 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.sigpro.2009.04.033
 ARTICLE IN PRESS
2172 H.Q. Qu et al. / Signal Processing 89 (2009) 2171–2177

models. As demonstrated in Ref. [9], using more models 2. The Novel-IMM algorithm
makes it no better. In fact, the performance will
deteriorate if too many models are used due to the 2.1. The IMM algorithm
excessive ‘‘competition’’ from the ‘‘unnecessary’’ (excess)
models. The dilemma is that more models have to be used Multiple model (MM) estimation is a powerful
to improve the accuracy, but the use of too many models approach to adaptive estimation. It is particularly good
would degrade the performance [10]. for systems subject to structural as well as parametric
Several modified IMM algorithms were proposed for changes. In this approach a model set is selected to
improvement of performance or computation efficiency in represent (or ‘‘cover’’) the possible system behavior
recent years. A selected filter IMM (SFIMM) algorithm patterns and the overall estimate is obtained by a certain
which uses a subset of filters with the specific subset combination of the estimates based on these models.
chosen using decision rules was presented to improve The multiple model approach is best described in terms
computation efficiency of maneuvering target tracking of stochastic hybrid systems.
[11]. A variable structure IMM (VSIMM) algorithm was The IMM algorithm consists of r interacting filters
presented to solve the dilemma where the model set not operating in parallel. In the IMM approach, at time k the
only differs across targets, but also varies with time for state estimate is computed under each possible current
a given target [2,3]. Not only does VSIMM inherit model using r filters, with each filter using a different
the effective cooperation strategies of the IMM and the combination of the previous model-conditioned estimates
superior output processing of the AMM (autonomous (mixed initial condition).
multiple model), but it also adapts to the outside world by Fig. 1 describes the IMM algorithm, which consists of r
producing new elemental filters if the existing ones are interacting filters operating in parallel, one cycle of the
not good enough and by eliminating those elemental algorithm consists of the following [23–25]:
filters that are harmful. This algorithm works successfully
in ground target tracking where the model set adapts
sequentially according to the target position and the road (1) Calculation of mixing probabilities:
network configuration [12–15]. 1
The major objective of the MM is to achieve best lk ðijjÞ ¼ p mi ; i; j ¼ 1; 2; . . . ; r
cj ij k
modeling accuracy with a minimum number of models.
X
r
Here VSIMM estimation has certain advantages. In a cj ¼ pij mik (1)
general setting, the problem of efficient model set design i¼1
for MM estimation is still open, Researches in the field of where lk(i|j) is mixing probability, [pij] is the model
VSIMM focus on the efficient model set design [16–18] transition probability matrix, mik is the model prob-
and model set adaptation algorithms [1,3,19,20], which ability, and cj is normalizing constant.
includes the decision for activating a candidate model as
well as terminating the model in effect. In general, those
decisions should consist of a set of complex rules based
on both a priori and a posteriori information about the
current system mode in effect, but there are not unified
theories and practical methods to guide us how to use
these rules. These theories about model adaptation
is much more challenging [21,22]. So the advantages in
all presented model set adaptation methods based on
the statistical hypothesis testing have been surprisingly
limited. In addition, it is very difficult to realize the
VSIMM without the auxiliary information.
A novel interacting multiple model (Novel-IMM)
algorithm has been presented in this paper to solve the
problem of model set adaptation without auxiliary
information. This method adopts independent parallel
model set method but not a serial model set adaptation
which is adopted in VSIMM. It consists of N independent
IMM filters operating in parallel, and each independent
IMM filter also consists of multiple sub-filters, which
operate interactively. In every time index, only one IMM
output of a certain model set is used; but for a long time,
the algorithm alternatively chooses an output of the
model set to be the optimum final output. The method
does not use the decision for model activation and
the decision for termination of the model in effect. The
computer simulations illustrate the Novel-IMM could
improve the performance of target tracking. Fig. 1. IMM algorithm (r ¼ 2).
 ARTICLE IN PRESS
H.Q. Qu et al. / Signal Processing 89 (2009) 2171–2177 2173

(2) Interaction: 2.2. The VSIMM algorithm

0j X
r
i
X^ kjk ¼ X^ kjk lk ðijjÞ; j ¼ 1; . . . ; r (2) Most of the work on IMM estimator has considered
i¼1 only fixed model sets, i.e. it is assumed that, at any time,
0j
where X^ kjk
is mixed initial condition for the filter the target trajectory evolves according to one of a finite
matched to model j. The covariance corresponding to number of predetermined model set. This requires that
the above is the model set include as many models as necessary to
handle the varying target motion characteristics.
X
r
i j i j
P 0j ¼ lk ðijjÞfP ikjk þ ðX^ kjk  X^ kjk ÞðX^ kjk  X^ kjk Þ0 g All the models in the estimator are kept throughout the
kjk
i¼1 entire tracking period and, when it is necessary for a large
j ¼ 1; :::; r (3) number of models, it brings extra computational load, still
more, it may degrade estimation accuracy. One remedy
1 for the above shortcomings is to vary the model set in
(3) Model-matched filtering: Compute X^ kþ1jkþ1 P1kþ1jkþ1
^ 2 2 the IMM estimator based on some criteria to yield better
X kþ1jkþ1 P kþ1jkþ1 using Kalman filter. The likelihood
estimation. This results in a variable structure IMM
function of each model is
(VSIMM) estimator where the model set not only differs
expððd
j
Þ2 =2Þ across targets, but also varies with time for a given target.
Ljkþ1 ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
kþ1
  ; j ¼ 1; . . . ; r (4) In [10], a general framework for variable structure hybrid
 
ð2pÞM Sjkþ1  estimation based on a graph theoretic formulation was
proposed and applied to ground target tracking [12–15]. It
j
where ðdkþ1 Þ2 ¼ vjkþ1 ðSjkþ1 Þ1 ðvjkþ1 Þ0 , vjkþ1 , and Sjkþ1 are assumes that the total models can be covered by a number
the innovation vector and its variance matrix, and M is of model set, and a particular model set is running at any
the measurement dimension. given time determined by a hard decision. One cycle of the
(4) Model probability update: simple and practical algorithm is described in Fig. 2.
Model set adaptation is done by using additive
Ljkþ1 cj
mjkþ1 ¼ Pr j
; j ¼ 1; . . . ; r (5) auxiliary information. The targets under track are moving
j¼1 Lkþ1 cj along a constrained path, for example, a highway, with
varying obscuration due to changing terrain conditions. In
(5) Estimate and covariance combination: addition, the roads on which the targets travel can branch,
merge, or cross. Some of the targets can also move in an
X
r
j open field. Because of the varying terrain conditions, the
X^ kþ1jkþ1 ¼ X^ kþ1jkþ1 mjkþ1 ; j ¼ 1; . . . ; r (6)
j¼1
possible ways (models) in which a target trajectory can
evolve also vary. For example, at junctions, target motion
uncertainty increases—the target can move along any
of the roads meeting at the junction. Another example
X
r
j
P kþ1jkþ1 ¼ mjkþ1 fPjkþ1jkþ1 þ ½X^ kþ1jkþ1  X^ kþ1jkþ1  of terrain-conditioned motion is that of a target along a
j¼1 highway or a bridge, where the motion of the target is
j orthogonal to the road or the bridge is restricted. In this
 ½X^ kþ1jkþ1  X^ kþ1jkþ1 0 g (7)
case, the motion of a target is highly directional. Similarly,
Note that this combination is only for output purposes another constraint is imposed when the targets are
and it is not part of the algorithm recursions. It can be got moving in a file, where the target motion is unidirectional
from above equations that the performance will deterio- and the target velocity is such that ‘‘passing’’ is not
rate if too many models are used due to the excessive allowed.
competition from the unnecessary models. Kirubarajan presents a VS-IMM estimator, where filter
The IMM approach computes the state estimation that model set is adaptively modified depend on the terrain
accounts for each possible model using a suitable mixing topography (auxiliary information) [10]. For example, the
of the model-conditioned estimations depending on the added uncertainty at junctions is handled with model set
model probability. These algorithms are decision free
(no maneuver detection decision is needed) and undergo
a soft switching according to the latest updated mode
Model Set 1 Model Set 2 ... Model Set r
probabilities.
When target motion is complex, small number of
models is insufficient, worse still, using more models will
Model Set Selection (Auxiliary Information)
result in excessive competition from the unnecessary
models. In Ref. [9] the conclusion indicates that two
models is the best solution in normal cases when model Model Set n
transition matrix is effective. Considering model match,
three models are often used in IMM algorithms. One
model set cannot solve the diploma, so the multiple model General IMM
set algorithms are proposed and applied to complex target
motion tracking. Fig. 2. VSIMM algorithm.
 ARTICLE IN PRESS
2174 H.Q. Qu et al. / Signal Processing 89 (2009) 2171–2177

which represents motion along the possible roads One of outputs of the IMM filter of the N model sets is
(hard decision using additive auxiliary information). used to match target motion, but the chosen model set
These additional model set is replaced by the primitive would vary with time. Therefore, it is very important to
model set after the target passes through the junction. set up a method to choose the optimum output in these
At each scan, the structure of the estimator for every independent IMM filters.
target is individually modified based on the known To determine the optimum final output, the Novel-IMM
topography. This enables the estimator to handle the algorithm performs a continuous MSPT (model set prob-
variation in the possible motion modes across targets as ability test) for N hypotheses (H1: model set 1 is chosen to
well as with time for each target. match target motion, H2: model set 2 is chosen, y, HN:
The VSIMM algorithm eliminates the need for carrying model set N is chosen), and only one IMM output with its
all the possible models throughout the entire tracking corresponding model set is chosen as the optimum output
period, significantly improving performance. But the or final estimation. So model set adaptation is done by
model set adaptation is very difficult to extend to aerial using MSPT without any auxiliary information.
target without auxiliary information. Independence between different IMM filters can solve
the excessive competition from the unnecessary models.
2.3. The Novel-IMM algorithm The MSPT algorithm helps to choose the model set which
matches the target motion best. When target motion
Although much effort has been made and several ways changed, the MSPT algorithm will choose another model
have been tried, the IMM and VSIMM are not very set accordingly. For each IMM, the model set does not
successful for complex motion track without prior change with time.
information. The objective of this paper is to develop a The flow chart for one cycle of the Novel-IMM logic is
better method to solve the dilemma about how many shown in Fig. 3. We consider two model sets.
models to be used and how to interact with them. A recursion of the Novel-IMM algorithm consists of
The IMM algorithm consists of r interacting filters four fundamental steps:
operating in parallel. In the IMM approach, at time k the
state estimation is computed under each possible current (1) Parallel independent IMM: X^ kþ1jkþ1 ðiÞ, Pkþ1jkþ1 ðiÞ, mkþ1 ðiÞ,
model using r filters, with each filter using a different and Lkþ1 ðiÞ are computed for each model set using
1 1
combination of the previous model-conditioned estimates general IMM (see Fig. 1). X^ kjk ð1Þ and P^ kjk ð1Þ are target
(mixed initial condition). state and its covariance estimation of model 1 in
2 2
The Novel-IMM algorithm consists of N independent model set 1 at k time index, while X^ kjk ð1Þ and P^ kjk ð1Þ
IMM filters operating in parallel, and these multiple IMM are target state and its covariance estimation of model
filters are independent. But each independent IMM filter 2 in model set 1. PT(1) and mk(1) are transition matrix
also consists of multiple sub-filters, which operate and model probability vector of model set 1.
1 1 2 2
interactively. At time k, the state estimation is computed X^ kjk ð2Þ,P^ kjk ð2Þ,X^ kjk ð2Þ,P^ kjk ð2Þ,P T ð2Þ, and mk(2) are corre-
with all possible models in the N model sets, and in these sponding terms of model set 2. PT(1) and PT(2) are as
possible compute, every IMM filter uses their own same as Pij in Eq. (1). mk(1) and mk(2) are model
independent initial condition without being combined. probability.

Fig. 3. Novel-IMM algorithm (two model sets and two models in each model set).
 ARTICLE IN PRESS
H.Q. Qu et al. / Signal Processing 89 (2009) 2171–2177 2175

(2) The likelihood function value of every model set: multiple model set algorithm and its model set is adaptive
X
r not using any additional auxiliary information.
C kþ1 ðnÞ ¼ Ljkþ1 ðnÞmjk ðnÞ; n ¼ 1; . . . ; N (8)
j¼1
3. Simulations

(3) Normalized likelihood function value of every model set: The new approach is illustrated in detail with two
The weight probability of model set kk+1(n) is updated examples of complex aerial maneuvering target tracking.
from Ck+1(n) and kk(n): The sensor sampling period T is 1 s.
C ðnÞkk ðnÞ The trajectory 1 is a target flying in the (x,y) plane,
kkþ1 ðnÞ ¼ PN kþ1 ; n ¼ 1; . . . ; N (9) starting with an initial position [10 km, 40 km]0 and an
n¼1 C kþ1 ðnÞkk ðnÞ initial velocity [300 m/s, 0 m/s]0 , Fig. 4(a) shows trajectory
1 that executes a 5-motion sequences (CV–CA–CV–
CT–CV):
(4) Estimate and covariance: According to MSPT, the
output of IMM, whose sequential likelihood function
value is maximum, is chosen as target state output. So (1) CV motion in 30 s;
the final output, X^ kþ1jkþ1 and P kþ1jkjþ1 , are computed as (2) CA motion in 30 s, its acceleration is (10 m/s2,
follows: 10 m/s2);
(3) CV motion in 30 s;
kkþ1 ðnm Þ ¼ max fkkþ1 ðnÞg (10) (4) coordinated turn motion in 8 s, its radius is 1.5 km, its
n¼1;...;N
acceleration is 60 m/s2; and
(5) CV motion in 30 s.
X^ kþ1jkþ1 ¼ X^ kþ1jkþ1 ðnm Þ (11)
The trajectory 2 is a target flying in the (x,y) plane,
starting with an initial position [10 km, 40 km]0 and an
P kþ1jkjþ1 ¼ P kþ1jkþ1 ðnm Þ (12) initial velocity [0 m/s, 300 m/s]0 , Fig. 4(b) shows trajec-
From the above interpretation, we insure the indepen- tory 2 that executes a 9-motion sequences (CV–CA–CV–
dence of all IMM filters, so the deterioration of perfor- CA–CV–CT–CV–CT–CV):
mance due to the excessive competition from the
unnecessary models is solved. While in every IMM filter, (1) CV motion in 50 s;
filter results of every model in every model set are (2) CA motion in 15 s, its acceleration is (20 m/s2,
interacted, so the good performance of general IMM is 20 m/s2);
preserved. Therefore, the Novel-IMM algorithm not only (3) CV motion in 50 s;
inherits the merit of IMM method, but also overcomes the (4) CA motion in 10 s, its acceleration is (30 m/s2,
weakness of IMM. 30 m/s2);
The VSIMM is a series of multiple model set adaptation (5) CV motion in 50 s;
algorithm and its model set selection depends on the (6) coordinated turn motion in 31 s, its radius is 1.5 km, its
auxiliary information. While the Novel-IMM is a parallel acceleration is 60 m/s2 (trajectory 2);

45 45
40
40 1 CV
35
1 CV
2 CA 30
35 2 CA 3 CV 4 CA
25
Y (km)
Y (km)

8 CT
3 CV
20
30 15 7 CV
9 CV
5 CV
10
5 CV 4 CT
25 5
0 6 CT

20 -5
5 10 15 20 25 0 10 20 30 40
X (km) X (km)

Fig. 4. Target trajectory: (a) trajectory 1 and (b) trajectory 2.

 ARTICLE IN PRESS
2176 H.Q. Qu et al. / Signal Processing 89 (2009) 2171–2177

(7) CV motion in 50 s; probabilities of model set CVCA2 and CVCT2, respectively.

(8) coordinated turn motion in 31 s, its radius is 1.5 km, its The transitions between the model sets are quite clear.
acceleration is 60 m/s2 (trajectory 2); and During the CA portion, the probability of model set CVCA2
(9) CV motion in 50 s. is much higher than CVCT2, but during the coordinated
turn motion portion, the probability of CVCT2 is much
higher than CVCA2. The RMSE of IMM and Novel-IMM
The Novel-IMM algorithm includes two model sets
are given in Fig. 6. During the CA maneuvering, two
(CVCA2 and CVCT2).
algorithms get the same performance, but during the
The first model set CVCA2 consists of three models: CV
coordinated turn motion, Novel-IMM algorithm has better
model, CA model with large process noise covariance [26]
performance than IMM.
and CA model with small process noise covariance [26].
In addition, for a ground target track with the map
The second model set CVCT2 consists of three models:
auxiliary information, the Novel-IMM algorithm can also
CV model, clockwise CT (coordinated turn) model with
use the map auxiliary information for the hard decision,
known turn rate [27], counter clockwise CT model with
now only one model set is used for Novel-IMM algorithm
known turn rate [27].
at a given time. In this case, Novel-IMM algorithm equals
The standard covariance of models’ process noise is
to VSIMM algorithm. Therefore, Novel-IMM algorithm can
sv ¼ 2 m/s, while the standard deviation of measurement
be used in more situations with or without auxiliary
noise is sx ¼ sy ¼ 50 m. The matrix of the probabilities
information. But VSIMM algorithm cannot be used in
should have a large diagonal element, and it is as follows:
above simulation case without any prior auxiliary in-
2 3
0:95 0:025 0:025 formation.
6 0:025 0:95 0:025 7 To validate the performance of the method, the RMSE
P¼4 5
of VSIMM and Novel-IMM are given in Fig. 7. The
0:025 0:025 0:95
trajectory of the target is as shown in Fig. 4(b). In VSIMM
The IMM algorithm includes 1 model set CVCA2 as the method the maneuvering onsets are known as the prior
first model set in the Novel-IMM algorithms. auxiliary condition while there is not prior information
The average model set probabilities for the Novel-IMM in Novel-IMM method. Performance of the Novel-IMM
are given in Fig. 5, where CVCA2 and CVCT2 denote the is nearly same as VSIMM, but VSIMM could only be used

1
Weight Probability

1
Weight Probability

CVCA2 CVCA2
CVCT2 CVCT2

0.5 0.5

0 0
30 40 50 60 70 80 90 100 110 120 130 0 50 100 150 200 250 300 350
Time (s) Time (s)

Fig. 5. Model set weight probability: (a) trajectory 1 and (b) trajectory 2.

150 100
IMM IMM
RMSEx (m)

RMSEy (m)

Novel-IMM Novel-IMM
100
50
50

0 0
30 40 50 60 70 80 90 100 110 120 130 0 50 100 150 200 250 300 350
Time (s) Time (s)
150 80
IMM IMM
RMSEVx (m/s)

RMSEVy (m/s)

Novel-IMM
Novel-IMM 60
100
40
50 20
0 0
30 40 50 60 70 80 90 100 110 120 130 0 50 100 150 200 250 300 350
Time (s) Time (s)

Fig. 6. Position RMSE for IMM and Novel-IMM: (a) trajectory 1 and (b) trajectory 2.
 ARTICLE IN PRESS
H.Q. Qu et al. / Signal Processing 89 (2009) 2171–2177 2177

80 60
Novel-IMM Novel-IMM

RMSEVy (m/s)
60 VSIMM VSIMM
RMSEy (m)

40
40
20
20
0 0
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Time (s) Time (s)

Fig. 7. Position RMSE for VSIMM and Novel-IMM (trajectory 2).

under the assumption that the true model sequence is [8] X.R. Li, Multiple-model estimation with variable structure: some
known while Novel-IMM could be used without the true theoretical considerations, in: Proceedings of the Third Conference
on Decision and Control, 1994, pp. 1199–1204.
model sequence. [9] L. Bloomer, J.E. Gray, Are more models better? The effect of the
Indeed, the Novel-IMM algorithm must cost more model transition matrix on the IMM filter, in: Proceedings of
computational resources than IMM for real time tracking the 34th Southeastern Symposium on System Theory, IEEE, 2002,
pp. 20–25.
system, but this problem can be solved easily with the [10] X.R. Li, Y. Bar-Shalom, Multiple-model estimation with variable
development of high speed DSP and computer. structure, IEEE Trans. Autom. Control 41 (9) (1996) 478–493.
[11] H.-J. Lin, D.P. Atherton, An investigation of the SFIMM algorithm
for tracking manoeuvring targets, in: Proceedings of the 32nd
Conference on Decision and Control, San Antonlo, TX, 1993,
pp. 930–935.
4. Conclusion [12] B. Paneutier, K. Benameur, U. Nimier, M. Rombaut. VSIMM using
road map information for a ground target tracking, in: Seventh
International Conference on Information fusion, 2005, pp. 24–31.
A Novel-IMM algorithm was presented which leads to
[13] T. Kirubarajan, Y. Bar-Shalom, K.R. Pattipati, I. Kadar, B. Abrams,
a systematic treatment of model set adaptation without E. Eadan, Tracking ground target with road constraint using an IMM
additional auxiliary information. The Novel-IMM algo- estimator, in: Proceedings of IEEE Aerospace Conference, (5), 1998,
pp. 5–12.
rithm consists of N independent IMM filters operating in
[14] T. Kirubarajan, Y. Bar-Shalom, K.R. Pattipati, Topography-based VS-
parallel, and these multiple IMM filters are independent. IMM estimator for large-scale ground target tracking, in: IEE
The Novel-IMM uses MSPT algorithm to choose the Colloquium Target Tracking: algorithms and Applications, London,
outputing model set, which matches the target motion UK, 1999, pp. 11/1–11/4.
[15] T. Kirubarajan, Y. Bar-Shalom, K.R. Pattipati, I. Kadar, Ground target
better, according to the target motion. For each IMM, the tracking with variable structure IMM estimator, IEEE Trans. Aero-
model set does not change with time. The Novel-IMM space Electron. Syst. 36 (1) (2000) 26–46.
algorithm solves the deterioration of performance due to [16] X.R. Li, Model set design for multiple-model estimation—part I, in:
Proceedings of the 2002 International Conference on Information
the excessive competition from the unnecessary models, Fusion, Annapolis, MD, 2002, pp. 26–33.
and inherits the merit of the IMM method and overcomes [17] X.R. Li, Z.-L. Zhao, P. Zhang, C. He, Model set design for multiple-
weakness of the IMM. It is a new thought for the study model estimation—part II: examples, in: Proceedings of the 2002
International Conference on Information Fusion, Annapolis, MD,
and design of the MM estimation algorithms. 2002, pp. 1347–1354.
[18] X.R. Li, Z.-L. Zhao, X.-B. Li, General model set design methods for
References multiple-model approach, IEEE Trans. Autom. Control 50 (9) (2005)
1260–1276.
[19] X.R. Li, V.P. Jilkov, R.U. Jifeng, Multiple-model estimation with
[1] L.A. Johnston, V. Krishnamurthy, An improvement to the interacting variable structure—part VI: expected-mode augmentation, IEEE
multiple model (IMM) algorithm, IEEE Trans. Signal Process. 49 (12) Trans. Aerospace Electron. Syst. 41 (3) (2005) 853–867.
(2001) 2909–2923. [20] X.R. Li, Y. Zhang, Multiple-model estimation with variable structure.
[2] X.R. Li, A survey of maneuvering target tracking—part II: model set Part V: likely-model set algorithm, IEEE Trans. Aerospace Electron.
adaptation, IEEE Trans. Autom. Control 45 (11) (2000) 2047–2060. Syst. 36 (2) (2000) 448–466.
[3] X.R. Li, X.R. Zhi, Y. Zhang, A survey of maneuvering target [21] X.R. Li, V.P. Jilkov, Survey of maneuvering target tracking. Part V:
tracking—part III: model-group switching algorithm, IEEE Trans. multiple-model methods, IEEE Trans. Aerospace Electron. Syst. 41
Aerospace Electron. Syst. 35 (1) (1999) 225–241. (4) (2005) 1255–1321.
[4] N. Tudoroiu, K. Khorasani, Fault detection and diagnosis for [22] X.R. Li, V.P. Jilkov, A survey of maneuvering target tracking—part IV:
satellite’s attitude control system (ACS) using an interactive multi- decision-based methods, signal and data processing of small
ple model (IMM) approach, in: Proceedings of the 2005 IEEE targets, in: Proceedings of SPIE, vol. 4728, 2002, pp. 511–534.
Conference on Control Applications, Toronto, Canada, August 28–31, [23] H.A.P. Blom, Y. Bar-Shalom, The interacting multiple model
2005, pp. 1287–1292. algorithm for systems with Markovian switching coefficients, IEEE
[5] N. Tudoroiu, K. Khorasani, Satellite fault diagnosis using a bank of Trans. Autom. Control 33 (8) (1988) 780–783.
interacting Kalman filters, IEEE Trans. Aerospace Electron. Syst. 43 [24] S. Blackman, R. Popoli, Design and Analysis of Modern Tracking
(4) (2007) 1334–1350. Systems, Artech House, Boston, 1999.
[6] Y. Hayashi, H. Tsunashima, Y. Marumo, Detection of railway vehicles [25] Y.B. Shalom, X.R. Li, Estimation and Tracking: Principles, Techni-
using multiple model approach, in: SICE-ICASE International Joint ques, and Software, Artech House, Boston, 1993.
Conference 2006, Bexco, Busan, Korea October 18–21, 2006, pp. [26] X.R. Li, V.P. Jilkov, Survey of maneuvering target tracking. Part I:
2812–2817. dynamic models, IEEE Trans. Aerospace Electron. Syst. 39 (4) (2003)
[7] Y. Hayashi, H. Tsunashima, Y. Marumo, Fault detection of railway 1333–1363.
vehicle suspensions using multiple model approach, in: SICE [27] G.A. Watson, W.D. Blair, IMM algorithm for tracking that maneuver
Annual Conference 2007, Kagawa University, Japan, September through coordinated turns, in: SPIE Signal and Data Processing of
17–20, 2007, pp. 1539–1543. Small Targets, vol. 1698, 1992, pp. 236–247.