I.

INTRODUCTION

Considerable research has been undertaken

Interacting Multiple Model in the field of estimation theory in relation to the
multitarget-multisensor tracking (MTMST) problem.
Methods in Target Tracking: This is of interest in both military and civilian
applications. Ballistic Missile Defense and Airborne
A Survey Surveillance require idientification and tracking of
several hundred targets in real time. A typical scenario
includes maneuvering and nonmaneuvering targets
as well as environmental clutter and false alarms
produced by the sensor receiver noise and atmospheric
E. MAZOR disturbances. These problems occur in sophisticated
Technion, Israel Institute of Technology weapon delivery systems, satellite surveillance
A. AVERBUCH systems, air defense, oceanhattlefield surveillance, air
Tel Aviv University traffic control (ATC), and nonmilitary vehicle tracking
Y. BAR-SHALOM, Fellow, IEEE systems. In addition, the techniques have general
University of Connecticut application to statistical pattern recognition and
J. DAYAN robotics, etc. The main purpose of a tracking system
Technion, Israel Institute of Technology for traffic control or air defense is the estimation
of target trajectories in the controlled area and their
prediction into the near future. In a civilian ATC
system, the estimated trajectories (tracks) are used
The Interacting Multiple Model (IMM) estimator is a
to check the standard separation between pairs of
suboptimal hybrid filter that has been shown to be one of the
targets for maintenance of safety conditions (collision
most cost-effective hybrid state estimation schemes. The main
avoidance) and regularity of traffic flow. In a maritime
feature of this algorithm is its ability to estimate the state of a collision avoidance system, each ship monitors the
dynamic system with several behavior modes which can “switch” traffic situation in order to detect the possibility
from one to another. In particular, the IMM estimator can be a of colliding targets. In an air defense system, the
self-adjusting variable-bandwidth filter, which makes it natural for estimated trajectories are generally used to help
tracking maneuvering targets. The importance of this approach perform some of the following functions [511: threat
is that it is the best compromise available currently between identification, threat evaluation, weapon control and
complexity and performance: its computational requirements are assignment, calculation of the predicted position (for
nearly linear in the size of the problem (number of models) while fire or launch of missile), kill evaluation.
its performance is almost the same as that of an algorithm with Although the mathematical structure of the optimal
solution of the MTMST problem is well understood
quadratic complexity.
[7, 15, 181, the computational complexity of the
The objective of this work is to survey and put in perspective
optimal tracking algorithm (the Multiple Hypothesis
the existing IMM methods for target tracking problems. Special
Tracker (MHT)) limits its practical realization using
attention is given to the assumptions underlying each algorithm even the largest and fastest computers available in
and its applicability to various situations. the foreseeable future. Suboptimal implementations
of the MHT have many potential pitfalls.’ For this
reason, many different tracking algorithms have
been developed which sacrifice optimal performance
for the sake of compuiational feasibility. Intuition
Manuscript received August 15, 1995; revised May 24 and suggests that some suboptimal approximations to
November 7, 1996; January 3, 1997. the optimal algorithms result in little performance
IEEE Log NO. T-AES/34/1/00182. loss, but there exists no reliable quantitative analysis
of suboptimal performance that demonstrates that
The work of Y. Bar-Shalom was supported by ONR Grant
this is, in fact, the case:. The basic research problem
N00014-91-J-01950 and by AFOSR Grant F49620-95-1-0229,
is to find effective analytical techniques for the
Authors’ addresses: E. Mazor and J. Dayan, Dept. of Mechanical estimation problems. MTMST problems involve
Engineering, Technion, Israel Institute of Technology, Haifa
both the estimation of continuous-valued parameters
32000, Israel: A. Averbuch, School of Mathematical Sciences,
Tel Aviv University, Tel Aviv 69978, Israel: Y. Bar-Shalom, such as target positions, velocities and accelerations,
Dept. of Electrical and Systems Engineering, U157, University of
Connecticut, Storrs, CT 06269-2157, [email protected].
‘This was illustrated in [70] where an MHT (augmented with
maneuver hypotheses in addition to data association hypotheses)
0018-9251/98/$10.00 @ 1998 IEEE kept losing a test flight target.

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 1 JANUARY 1998 103

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 and the testing of discrete hypotheses, such as the a powerful approach resulting in parallel processing
association of measurement data with targets. Thus, algorithms and robust solutions [72, 76, 921. The case
it is natural to pose the problem of MTMST as a when the transition probability matrix is unknown
hybrid state estimation problem, that is, estimation deserves attention and has been considered by several
for a partially observed stochastic process with authors for some restricted versions of the general
discrete- and continuous-valued states. The hybrid problem [77, 931. The case where jumps in the system
state viewpoint [9, 38, 49, 85, 971 provides both a structure are modeled by a semi-Markov chain was
natural formulation for many types of surveillance first discussed in [791.
and tracking problems and a powerful framework for A unified tutorial presentation on the topic of MM
deriving theoretically optimal and practical suboptimal estimation and other related techniques from a target
tracking algorithms. Hybrid systems are characterized tracking oriented perspective can be found in [14, 15,
by the following. 181. Algorithms for hybrid state estimation in general
consist of a bank of model-matched filter modules
1) State (consisting of kinematic and, possibly,
and some algorithm to organize the cooperation
feature components) that evolves according to a
between the individual filters. The modules can be
stochastic difference (or differential) equation model,
Kalman filters (KF) or other filters as discussed
2) Model that is governed by a discrete stochastic
below. Compared with other MM approaches the main
process: it is one of a finite number of possible
advantage of hybrid system-based solutions is that
models (each corresponds to a behavior mode), that
they are not subject to the so-called “obliviousness to
undergo jumps (switches) from one model (behavior
detection” difficulty: the a priori jump probabilities
mode) to another according to a set of transition
used in the regime dynamics make hybrid algorithms
probabilities (the dynamic Multiple Model (MM)
alert to regime changes whereas classical solutions are
approach).
biased toward the no-maneuver hypothesis after long
The value of hybrid models for tracking algorithms quiescent periods [34]. One of the significant schemes
is that the occurrence of target maneuvers can be is the so-called Generalized Pseudo-Bayesian (GPB)
explicitly included in the kinematic equations through method [ 1, 40, 931 and the other is the Interacting
regime jumps [33, 731. The MM adaptive estimation Multiple Model (IMM) algorithm [28, 301. The
approach is based on the fact that the behavior of a general structure of these algorithms consists of a
target cannot be characterized at all times by a single bank of filters for the state cooperating with a filter
model, but a finite number of models can adequately for the parameters. A GPB algorithm of order k
describe its behavior in different regimes. Several (GPBk) needs Nf” filters in its bank [93], where Nf
approaches have been proposed to perform the state is the number of models in the bank and k is the
estimation of a system together with identification number of sampling periods over which mode jumps
of each model (out of a finite set). Although static are considered. The IMM algorithm is conceptually
(i.e., nonswitching) MM estimation algorithms have similar to GPB2 and performs nearly as well as the
been known since the mid 1960s, practical algorithms GPB2 method, but requires only Nf filters to operate
for the switching case have only recently become in parallel, i.e., it has notably less computations,
available. A unified treatment and a brief survey of namely, the cost of GPBl [30]. The IMM estimator,
the various suboptimal state estimation and system originally proposed by [28], is a suboptimal hybrid
structure detection algorithms were presented in filter that was shown to achieve an excellent
[93] for discrete systems with abruptly changing compromise between performance and complexity.
structure where the changes are modeled by a finite Its complexity is nearly linear (with a small quadratic
state Markov chain. The algorithms considered component) in the size of the problem (number of
here assume that the transition probability matrix models) while its performance is almost the same as
governing the Markov chain is known. The proposed that of the GPB2. Another advantage (shared with the
approach assumes that the actual (true) system is one other MM estimators) is its modularity. It can be set
of finite number of possible models and that the a up using KF or extended KF (EKF) as its building
priori probability of each model is known. Modeling blocks to account for nonlinearities in the motion
with Markov jump parameters may be thought of equation (e.g., for coordinated turns) and in the
as an extension of the original static MM approach measurement equation for range-azimuthal-elevation
(or partitioning approach). In the former, Tugnait observations, or probabilistic data association filters
[93J allows switchings among the hypothesized (PDAF) and joint PDAF (JPDAF), based on KF
models unlike the static MM approach where a or EKF, if data association (due to false alarms or
candidate model is assumed to describe the true neighboring targets) is a significant problem.
system behavior for all time. In several applications, One of the first applications of the IMM
switching parameter modeling is more realistic is the work of [33] for the European ATC
[85] because the underlying true system does have (EUROCONTROL), based on the assumption that the
time-varying characteristics. Note that partitioning is governing system equations describe a Markov chain

104 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 1 JANUARY 1998

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 process. The state of this chain represents the mode of Qjand Rj, respectively. It is assumed that v and w,
the target flight during a particular time interval. This uncorrelated with x(O), are mutually uncorrelated2 and
mode is assumed to be among the possible modes that f , g, and h are known. Also x ( 0 ) is assumed to be
of the set of all modal states at all times. The set of Gaussian with approprlate mean and covariance. All
modes consists, for example, of a constant velocity vectors and matrices are assumed to have appropriate
model and several maneuver models. The process is dimensions. Note that (2) implies the availability
characterized by its a priori transition probabilities of noisy, base state information only and that the
[ 141, which constitute the transition matrix. These mode information is embedded (hidden) in the
are design parameters for the algorithm. In the measurement sequence. This description covers the
framework of the Hadamard project of the French case of nonstationary noises by modeling the noise
Civil Administration, advanced tracking methods were sequences as jumping from one set of parameters to
investigated in [94]. The project dealt with the aircraft another one.
motion modeling and showed that stringent speed In a fixed-structure (or fixed mode-set) hybrid
estimation can be obtained only with MM algorithms. system, a set of modes must be selected in advance
The IMM estimation algorithm has triggered a Denoting by Mf this fixed set of Nf modes (this set
large variety of related approaches to target tracking. might be smaller than J4, or, simply M, is not really
The goal of this survey paper is to provide an known), the system (1)--(3) is then approximated as
overview of these important developments. The
organization is as follows. In Section I1 we describe x(k + 1 ) = fj[k,xCk)I f gj[k,x(k),vj(k)I
the problem of hybrid state estimation which estimates v jeM;" 4)
the base state and the modal state based on the
z ( k ) = hj[k,X(k)l + W j ( k ) 'd j E Mf (5)
measurement sequence. In Section I11 we give detailed
description of the IMM algorithm and its varieties. with Mf "smaller" than M, or just because M, is not
Comparison of the main MM algorithms is given in perfectly known, and jumps between the system
Section IV. Parallel implementation is discussed in modes are modeled by switching from one model to
Section V. Nonsimulation performance prediction another, governed by thie same law as given in (3).
techniques for IMM are the topic of Section VI. The Here, Nf is the cardinality of the set Mf,i.e., the
application of an IMM estimator for a phased array possible number of system modes, and subscript j
radar is reviewed in Section VII. Finally, using IMM denoted quantities pertaining to model (mode) m j .
algorithms for target tracking with glint noise and A jump linear fixed-structure hybrid system with
noise identification is discussed in Section VIII. mode transition modeled by a semi-Markov process,
such as the state-dependent Markov chain, can be
described by the equatilons [29, 31-33, 37-39, 781
II. PROBLEM FORMULATION
x(k + 1 ) = ~ j ( k ) x ( : k+) rj(k)vj(k)
A general description for a hybrid system with
additive noise is [65] 'd j EMf (6)

x ( k + 1) = f [ k , x ( k ) , m ( k+ 1)l z ( k ) = H;(k)x(k) + W j ( k ) 'd j E Mf (7)

+ g [ k , x ( k ) , m ( k + l),v[k,m(k + 1)ll ( 1 ) where rj is the process noise gain matrix. The model
switching process cons:idered is of the semi-Markov
with mode-dependent noisy measurements type. The process is specified by a family of transition
matrices where T~ is the system's sojourn
z ( k ) = h [ ( k , x ( k ) , m ( k ) l + w[k,m(k)l (2) time in model i, Le., it is a "sojourn-time-dependent
and a (possibly base state dependent) Markovian Markov" (STDM) chain, which belongs to the
transition of the system mode semi-Markov class. The current probabilities of
transition for the STDM process (chain) are defined
P { m j ( k + 1) I mi@)} = 4[k,4k),mi,mjl as
'd mi,mj EM, (3) pi;(.) = P{mj(k + 1) I mi(k), ~ ~ (=k7 ))
where x ( . ) is the base state, m(k) is the modal state v i,jEMf (8)
(system mode index) at time k , which denotes the
mode in effect during the sampling period ending ) sojourn time in state i at time k . It
where ~ ~ (iskthe
A is assumed that at k = 0 the sojourn time (in whatever
at k, P { . } is probability, wtJ(k)= { m ( k ) = j } is the state the system model is) is 7 = 1. Thus, the values 7
event that mode j is in effect at time k , M, is the are taken from 1 to the maximum, which at time k is
set of all modal states at all times, v[.] and w [ . ] are
the mode-dependent process and measurement noise 21f the two noise sequences ;ire correlated one can modify the
sequences with means and Fj and covariances estimator modules (see e.g., [14]).

MAZOR ET AL.: INTERACTING MULTIPLE MODEL METHODS IN TARGET TRACKING: A SURVEY 105

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 then k + 1. The process and measurement noises are appropriate stochastic models, developing the exact
Gaussian mutually uncorrelated with zero mean and filter solution from the Bayesian estimation theory
known covariances. for Markov processes, and finally introducing
The simplest hybrid system is one that can efficient numerical approximations of the exact
be described by a set of linear models as (6), (7) solution. By following these three steps, several
with the mode transition governed by a first-order approximate Bayesian methods were developed
homogeneous Markov chain for each of the above subproblems. The various
approximate Bayesian methods show a trade-off
qi{mj(k + 1) I mi(k)}= p i j 'd i , j E Mf (9) between computational load and performance,
where p i j is the Markov transition probability from depending on the particular application.
mode i to mode j . It is assumed, as before, that To cope with the first difficulty, the well-known
the process and measurement noises are Gaussian extended Kalman method of linearization along the
mutually uncorrelated with zero mean and known predicted path is adopted as a good compromise.
covariances. To cope with the second difficulty, the aircraft
When the Markovian switching coefficients can be behavior is modeled as a hybrid state Markov process.
isolated to a system bias, such linear hybrid system An aircraft trajectory can be subdivided into distinct
can be represented as [21] segments, corresponding to modes of flight. An
appropriate model for the switching between modes
x(k + 1) = ~ ( k ) x ( k+) rj(k)b(k) + r x ( k ) v x ( k ) (io) is a finite-state (semi-) Markov process. An aircraft
is then modeled as having a hybrid state; consisting
b(k + 1) = ~ , ( k ) b ( k+) rg(k)vb(k) (1 1) of a continuous Euclidean-valued (diffusive) state
z ( k ) = N ( k ) x ( k )+ Gjb(k)b(k)+ w ( k ) (12) component and a discrete-valued state component.
The discrete state describes the mode of flight. The
with bias transition governed by a finite state Markov diffusive state describes the horizontal position,
chain ground speed, course, and, for some modes, bank
angle, flight-path angle, and transversal acceleration.
The evolution in time of the discrete state (i.e.,
where b ( k ) is the system bias, p,, is the Markov switching between modes) and the diffusive state are
transition probability from bias model mp to bias determined by the physical relations between the state
model mjb; v"(k) and vb(k) are white Gaussian errors components, the jump-type changes of the control
for the system state and bias process respectively; commands of the pilot, and disturbances caused by
w(k) is the white Gaussian error in the measurement both the pilot and the wind. In this model, switching
and independent of v"(k) and v b ( x ) .It is assumed from one mode to another generally causes a
that covariances of v"(k), vb(k) and w ( k ) are known simultaneous random jump in bank angle, flight-path
(Q"(k), Qb(k) and R ( k ) respectively); the functions angle, and acceleration. The combination of discrete
F ( k ) , r j ( k ) , F " ( k )are known also. and diffusive states in a maneuvering trajectory model
The problem of hybrid state estimation is to causes the associated tracking problem to fall within
estimate the base state and the modal state based on the class of so-called hybrid-state estimation problems.
the (noisy base state) measurement sequence. From the various approximate methods to cope
Following [33] the outline principles of the track with the third difficulty, an MM probabilistic data
maintenance system are as follows. Once a track association (PDA) method [7, 15, 541 is judged to be
has been initiated, the track maintenance system an excellent compromise to update an aircraft track
updates that track as long as it is observed by at least from new observations. Moreover, the PDA approach
one sensor. Hence, a multisensor track-continuation allows to include an efficient solution to the fourth
problem is reduced to a "single-sensor" problem, difficulty (outliers).
where the updating is sequential across the sensors. Efficient cooperation between the KFs is
Track continuation has to cope with several realized by an interaction between the estimates (the
uncertainties, of which the following four cause the conditional mean and covariance given a particular
major difficulties: mode is in effect) for the different modes at the
1) nonlinear target dynamics during a turn, beginning of each filter cycle. The interaction is
2) sudden starts and stops of maneuvers (mode determined by the conditional probabilities of
switching), switching between modes. For problems like tracking,
3) the association of measurements with existing the IMM interaction is so effective, that IMM
tracks, algorithm performs almost like the exact Bayesian
4) Gaussian-mixture-type measurement noise. filter [30]. Moreover, the IMM requires a far lower
computational power than other high-performance
In developing a high-quality track continuator the algorithms for tracking maneuvering aircraft [8,
Bayesian approach is employed. It consists of building 30, 341.

106 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 1 JANUARY 1998

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 Ill. THE IMM ESTIMATOR AND ITS VARIETIES algorithms, which yield model probabilities, provide
the true target probability for each track under
The IMM estimator has been shown to be one of consideration. Thus, these algorithms can assess
the most cost-effective and simple schemes for the their own reliability (track loss); that is, they can be
estimation in hybrid systems [5, 10-12, 15, 55, 581 called “intelligent.” A number of track formation
and therefore it is suitable to be used in MTMST. The examples with the IMMPDAF on real data can be
IMM design parameters are as follows. found in [ 151.
1) The set of models for the various regimes and
their structure. A. Baseline IMM Algorithm
2) The process noise intensities for the various
models, in particular the nonmaneuvering model with The baseline IMM ,algorithm [8, 14, 28, 301
low-level process noise and the maneuvering model(s) assumes the simplest form of hybrid system (6), (7),
with certain higher noise levels determined by the (9); each mode-matched filter is a standard KF.
assumed maneuverability of the targets. Nutations: Z k denotes the measurement sequence
3) The jump structure (usually Markov) and the through time k ; ;(i 11) denotes the state estimate at
transition probabilities between the models from the time i conditioned on Z’, and P(i I 1 ) is the associated
selected set. These probabilities are chosen according covariance matrix; quantities pertinent to mode j
to the designer’s beliefs about the frequency of the are denoted with subscript j ; N ( y ; L , P ) denotes the
regime switches and can be subsequently adjusted (multivariate) Gaussian, density function of y with
based on Monte Carlo simulation results. When the mean j j and covariance P ; E,
stands for
transition probabilities of the Markov chain are zero i j ( k I k ) and Pj(k I k ) are the state estimate and its
the IMM algorithm reduces to the well-known static covariance in mode-matched filter j at time k ;
MM algorithm. i O j ( kI k ) , G j ( k I k ) are the mixed initial condition
The IMM algorithm has three desirable properties: for mode-matched filter j at time k ; ;(k 1 k ) , P(k I k )
it is recursive, modular, and has fixed computational are the combined state estimate and its covariance;
requirements per cycle. In each cycle it consists of p j ( k ) is the mode probability at time k ; pili@I k ) is the
three major steps: interaction (mixing), filtering, and mixing probability at tiime k (the weights with which
combination. At each time, the initial condition for the estimates from the previous cycle are given to
the filter matched to a certain mode (a module) is each filter at the beginining of the current cycle); h,(k)
obtained by mixing the state estimates of all filters is the likelihood function of mode-matched filter j .
at the previous time under the assumption that this The basic IMM algorithm (one cycle) is as
particular mode is in effect at the current time. This is follows.
followed by a regular filtering (prediction and update)
Interaction. V i, j 6: Mf,p i i j ( k- 1 I k - 1) =
step, performed in parallel for each mode. Then, a ( l / F j ) p i j p i ( k- 1) (mixing probability), where E j is a
combination (weighted sum) of the updated state normalization factor
estimates of all filters yields the state estimate. The
probability of a mode being in effect plays a key role
in the weighting of the mixing and the combination of
Fj = CPij&(k
- 1)
i
states and covariances.
The main feature of this algorithm is its ability to i o j ( k - 1 I k - 1) = Cqk-1 I k - l ) p i V ( k - 1 I k - 1)
estimate the state of a dynamic system with several i
behavior modes which can “switch” from one to
another. In particular, the IMM estimator can be a Poj(k- 1 I k - 1 ) = C{gk-1 I k - 1)
self-adjusting variable-bandwidth filter, which makes i

it natural for tracking maneuvering targets. +[qk- 1 I k - l)-lZoj(k- 1 Ik- 111

The combinations of the IMM approach with
PDAF and JPDAF methods for single and multiple x [ q k - 1 I k - l ) - i o j ( k - 1 I k - 1)IT)
targets in clutter are discussed extensively in [9, 11, x q j ( k - I I k - 1).
13, 33, 48, 49, 58, 631. These recursive algorithms
have fixed computational and memory requirements Filtering. V j E M ,
and can initiate tracks, maintain tracks in the presence
of maneuvers, and terminate tracks if warranted. 2j(k 1 k - 1) = Fj(k - l.)iOj(k- 1 1 k - 1)
The algorithms are useful for situations of low +rj(k -- i)vj(k - 1)
signal-to-noise ratios (SNRs), where the detection
threshold has to be set low to detect the targets, Pj(k I k - 1) = Fj(k - I1)POj(k- 1 I k - l ) F j ( k -
this leading to a high rate of false alarms for which
logic-based techniques are not adequate. These + ri(k - i ) ~ ~ -( ik) r j ( k - i)T

MAZOR ET AL.: INTERACTING MULTIPLE MODEL METHODS IN TARGET TRACKING: A SURVEY 107

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 i j ( k I k ) = i j ( k k - 1) + Wj(k)rj(k) transition probability is a known function of the
sojourn time given by (8) and has a sojourn time
q(k I k) =q(k k - 1) - y ( k ) S j ( k ) W , ( k ) T r,(k) in state i which is, however, not known. The
conditional probability mass function (pmf) of the
q k ) = z ( k ) - i j ( k I k - 1) (residual) sojourn time in state m(k) = i based on the available
information Z k at time k , is
A
(measurement prediction) ) T I m(k) = i , Z k }
g,k(T>= P { T , ( ~ =

I k - l)Hj(k)T + R j ( k )
S,;(k) = Hj(k)Ti(k ) T I m(k) = i,Zk-'}
= P { T , ( ~=
(residual covariance)
= P { m ( k - 1) = i,. . .,m(k - r + 1)
y(k)= q ( k I k - l)EZi(k)TSj(k)-' (filter gain)
= i,m(k - T ) # i I m(k) = i , Z k - ' } (14)
A,;@) = N(rj(k);O,Sj(k)) (likelihood function)
where the perfect knowledge of the state m(k) allows
one to go down to one index less in the conditioning,
i.e., Z k - ' . The conditional pmf of the sojourn time r
in state i at time k (14) is given by the expressions
1 in [37, 391. Following (8) the conditional probability
= -Aj(k)E,;
c of transition from i to j at time k - 1, given the
observations Zk-' is, in terms of (14),
(mode probability, c is a normalizing factor).
Combination. V j E Mf
j i j ( k - 1) =A P { m ( k ) = j I m(k - 1) = i,Zk}
k

T= 1

Note that the argument of pi,, defined in (8), is the

sojourn time, while the argument of jij, defined
above is the current time. The filter has access only
to the observations from which the conditional pmf
of the sojourn time can be obtained: this, in turn, is
A second-order dependence of the base state on to be used in calculation of the conditional transition
the modal state was considered in [22, 29, 331 and probabilities (15).
designated as "generalized IMM', With the notations There are numerous examples of systems which
of (l), this is written as have discrete models that randomly vary with time
and experience switchings, between different models,
after a random sojourn time. In some situations the
switching probabilities depend on the sojourn time.
Such situations are commonly experienced in MTMST
[7, 15, 18, 35, 961. The model switching process is
characterized by a family of transition probability
With the complexity of the (slightly modified) IMM matrices p . . ( ~ ~i.e.,
) , it is STDM chain, which belongs
estimator being the same, this formulation allows a to the semEMarkov class. The example in [37, 391
different state equation at the instant of the jump, Le., shows that STDM-based IMM estimator gives a stable
when m(k + 1) # m(k), without the need for an extra estimation filter for both two- and three-model cases.
model. Furthermore, it can operate without predefined The three-model filter was shown to give a small
acceleration levels. It was indicated in [94] that an probability of error in system structure detection and
IMM with second-order dependency yielded lower thus the true system model becomes rapidly known,
errors than a conventional IMM. Smoothing versions which is critical in failure detection schemes.
of the IMM have been studied in [31, 561. Summary of the STDM-based IMM estimator (one
cycle) is as follows.

B. STDM-Based IMM Estimator Interaction. V i,j E Mf

1,
In this approach [29, 31-33, 37-39, 781 the hybrid /!Liij(k- 1 I k - 1) = =pi;@ - l)Pi(k - 1)
c
system is modeled by the equations (6)-(8); each
mode-matched filter is a standard KF. Each model (mixing probability)

108 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 1 JANUARY 1998

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 The conditional probability of transition from i to j at an underlying Markov chain, the mixer uses the model
time k - 1, given the observations Zk-’ is probabilities and the mlDdel switching probabilities
k
to compute a mixed bias estimate for each bias
filter. The bias-free filter error covariance is also
modified to reflect the spread of the means of the
r=1
bias filters. At the begiinning of filtering cycle, each
bias filter uses a mixed bias estimate and the bias-free
i measurement residual to compute a new estimate
of the bias and a likelihood for the bias model. The
i o j ( k - 1 I k - 1) and Poj(k - 1 I k - 1) are defined in likelihoods, prior model probabilities, and the model
Section IIIA. switching probabilities are then used to compute new
Filtering. From Section IIIA model probabilities. The overall bias estimate and
1 1 compensated state estimate are then computed with
p j ( k ) = - A j ( k ) C j i j ( k - 1)pi(k - 1) = -Aj(k)Zj. the output of the bias-free filter, the outputs of the
c C
i bias filters, and their model probabilities.
Combination. See Section IIIA. The IMBM algorithm was applied to tracking
maneuvering targets, where the target acceleration
More work on semi-Markov switching systems can was treated as a system bias with Markovian
be found in [321. switching coefficients. This algorithm was called the
Interacting Multiple Acceleration Model (IMAM)
C. IMM Algorithms for Systems with Markovian algorithm. In the IMAM algorithm the bias-free filter
Switching Coefficients Isolated to a System Bias corresponds to the constant velocity filter, while
the bias filters correspond to a zero acceleration
The idea of using a two-stage filter to implement model and two constant acceleration models. The
an augmented state filter was introduced in [53], constant acceleration models will have different
where the bias vector is a part of the augmented process noises covariances, while one constant
system state. The central filter is decoupled into two acceleration filter will utilize the kinematic constraint
parallel filters. The first filter, the “bias-free” filter, is for constant speed targets [2, 31. Thus, this IMAM
based on the assumption that the bias is nonexistent. algorithm includes a constant velocity model and
The second filter, the bias filter, produces an estimate three bias models. The first bias model, ml ( k ) , which
of the bias vector. The output of the first filter is then corresponds to a zero mean and zero covariance bias,
corrected with the output of the second. If the bias is does not require a separate filter. The second bias
deterministic and constant but unknown, the two-stage model, m,(k), corresponds to a constant state with
filter is equivalent to the augmented state filter [53]. the kinematic constraint, while the third bias model,
As proposed in [3], the two-stage Kalman estimator m,(k), corresponds to a constant state. This IMAM
is equivalent to the augmented state filter under algorithm which utilizes a zero bias model, m l ( k ) ,and
an algebraic constraint on the correlation between two constant acceleratilon models is presented in [21].
the state process noise and the bias process noise. The tracking performance of the IMAM algorithm is
An important problem is the estimation of a linear compared with the tracking performance of a similar
system with Markovian switching coefficients. When baseline IMM algorithm. Simulation results show
the Markovian switching governs the system bias that IMAM provides tracking that is very similar
and the IMM algorithm is used for state estimation, to that provided by the baseline IMM algorithm,
the “bias-free’’ portion of the system state must be while requiring about 43% of the computations of
duplicated in each model and its corresponding filter the IMM algorithm, when a constant velocity model
[2, 31. Such a linear system can be represented as in and two constant acceleration models are used. The
equations (lo)-( 13). computational requirements of the IMAM algorithm
A novel approach to state estimation for systems may allow MM algorithms to be implemented for
with Markovian switching system bias was presented target tracking in existiing systems, where the IMM
in [21]. This Interacting Multiple Bias Model (IMBM) algorithm cannot be justified from a computational
algorithm integrates the baseline IMM algorithm standpoint.
and the recent developments of Alouani and Blair in A special case of the IMAM algorithm is presented
two-stage estimation. The IMBM algorithm consists of in [98]. This algorithm was called the Interacting
a filter for the bias-free portion of the state model, Acceleration Compensation (IAC) algorithm. It
a filter for each bias model, a model probability was suggested that this IAC algorithm can further
evaluator for each bias model filter, an estimate reduce the computational cost of the IMM algorithm
mixer at the input of the bias filters, and an estimate while maintaining similar performance. The IAC
combiner at the output of the bias filters. With the algorithm incorporates the concept of the IMM
assumption that the model switching is governed by algorithm for two motiion models into the framework

MAZOR ET AL.: INTERACTING MULTIPLE MODEL METHODS IN TARGET TRACKING: A SURVEY 109

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 of the two-stage estimator. In the IAC algorithm, matched to these modes provides no gain but also the
the two-stage estimator is viewed as having two unnecessary “competition” among modes degrades
acceleration models, The first acceleration model, the performance. This degradation arises from the
ml ( k ) , corresponds to the zero acceleration of the fact that the same (large) mode set is used for all
constant velocity model and does not require a filter time periods. To find a way out of this dilemma,
(as in IMAM) and the second acceleration model, [65] introduced the concept of variable structure, i.e.,
mz(k),is a constant state model. The IMM algorithm adaptation of the model set. The work [65] proposed
is used with the acceleration models to compute the several variable-structure algorithms, in contrast to the
acceleration estimate that compensates the constant existing efforts of developing better implementable
velocity filter estimate. The tracking performance versions of the so-called optimal (fixed-structure)
of the IAC algorithm is compared with an IMM estimator. One solution is to use a time-varying
algorithm containing a constant velocity model and mode set. That is, Mf in (4)-(5) is replaced by a
a constant acceleration model. The IAC and IMM time-varying set M ( k ) , to be determined from all
algorithms were simulated with a radar tracking the information available-in particular, the
system measuring the target range, bearing, and sequence of the measurements. This is further
elevation at periodic intervals of 1.O s. Simulation detailed in [69].
results showed that the tracking performance of the The “full mode-set’’ estimators described above
IAC algorithm approaches that of a comparable IMM have been considered to yield the performance bound
algorithm, while requiring approximately 50% of of all IMM estimators. Unfortunately, these “optimal”
the computations. The tracking performance of the estimators may give unsatisfactory results in some
IMM algorithm has been found to be better than the cases because they are the best only within the class
IAC algorithm when the targets maneuver through of fixed-structure algorithms. By using a variable
constant-speed turns and the data rates are low (Le., structure, it is possible to develop algorithms that can
four or five measurements per maneuver). As data outperform these estimators.
rates are increased, the tracking performance of both The variable-structure (or adaptive mode-set)
algorithms becomes similar. The performance of the estimator can be viewed as a two-level hierarchical
IAC algorithm may be improved by using a turning algorithm: MM at the lower level and multiple
rate model to estimate the target acceleration for mode-set at the higher level. As was shown in [65,
constant speed maneuvers by treating the acceleration 691, the fixed-structure estimator is only a special
as a dynamical bias [3]. The turning rate model case of the variable-structure estimator using a trivial
would improve filter response during a maneuver partition consisting of the following single sequence
by providing improved prediction over the constant of mode sets
velocity model. The use of turning rate models or the
use of a kinematic constraint in the IMM algorithm ...’Mf}
M k = {Mf’Mf, (16)
for such targets has proved very effective. Multiple where M k denotes a sequence of the mode sets
acceleration models have been incorporated into through time k. That is, the variable-structure
the bias or acceleration estimation in the IMBM estimator reduces to one of the fixed-structure
algorithm [21]. The use of multiple acceleration estimators if only sequences of fixed mode sets
models (IMAM algorithm) will aid in responding are allowed. Note that (16) is not the only possible
to target maneuvers. Since the IAC algorithm is partition. Suppose M, = {1,2,3,4}. The following two
most applicable for suddenly accelerating targets, sequences of mode sets constitute one of the many
the use of multiple acceleration models would possible partitions at time k = 2:
improve the estimation of targets performing slight
maneuvers. M? = {{1,2,3,4},{1,2}}
M: = {{1,2,3,4},{3,4}}.
D. IMM Estimators with Mode-Set Adaptation
Thus, the choice of the partition can be made after
Often, for practical problems an algorithm using the arrival of the measurements. The full-fledged
a fixed set of a small number of models cannot yield variable-structure estimator is too complicated to
accurate results [65, 691. Apart from the increase in be feasible. Nevertheless, it suggests the use of
computation, use of more models does not guarantee time-varying mode-set sequences since such sequences
better performance-actually, it may yield even poorer can lead to better results than a fixed sequence. The
results. The major reason for the poor performance number of mode histories in the adaptive mode-set
of the fixed “full” mode-set algorithms in some algorithm at k is n,“=,
N ( t ) ( N ( t ) is the number of
cases is that when a large number of modes are modes at time t ) , which is much smaller than Nf”
used at any time, many of them are so different since N ( t ) < Nf.Thus, the adaptive mode-set approach
from the system mode in effect at a particular can provide a tremendous saving in computation
time, that not only the computation for the filters over a fixed mode-set approach, apart from superior

110 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 1 JANUARY 1998

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 performance. Meanwhile, the same management algorithm with the IMM algorithm. The performance
techniques of mode histories as in the fixed mode-set of the AIMM algorithm in tracking is better than
algorithms can be retained. Practical suboptimal the IMM algorithm, which uses submodels with
algorithms with variable structure are presented in different noise levels, when accuracy and computation
[69]. A graph-theoretic formulation of MM estimation time are used in the assessment. A major advantage
which leads to a systematic treatment of model-set of the AIMM algorithm is that it does not need
adaptation and opens up new avenues for the study predefined fixed deterministic accelerations for the
and design of the IMM estimation algorithms was also submodels. The IMM algorithm may provide better
given. accuracy when one of the submodels is well matched
An ad-hoc variable-structure method named with the motion of the target. However, the AIMM
Selected Filter IMM (SFIMM) has been proposed algorithm can provide better coverage of variable
in [7 11. The SFIMM algorithm is a modified IMM target accelerations using fewer submodels.
algorithm which uses a subset of filters with the
specific subset chosen using decision rules. Particular E. IMM Algorithm witlh Correlated Measurement
attention was focused on the design parameters, Noise
namely, the subset type and the decision rule in
this study. The coverage of thirteen filters was In tracking airborne or missile targets using radar
investigated. Two different subset types were data, the measurement noise is significantly correlated
examined. In subset type 1, suggested in [65], each when the measurement frequency is high. In the paper
subset “connects” to another one. This was found to [55], a simple decorrelation process is proposed to
have a higher possibility of shift and so when shifts enhance the IMM algorithm to track a maneuvering
occur a larger portion of models needed initialization target with correlated measurement noise. It assumed
and larger tracking errors ensued. Subset 2, also the simplest form of hybrid system (6), (7), (9) and
suggested in [65], used a symmetric structure, with that the noise can be modeled as a first-order Markov
some small overlapping, and yielded better tracking process given by
performance. Two different decision rules were
used to decide whether a shift was required or not,
Wj(k + 1) = Xw,(k) + W p ) (17)
first the maximum likelihood (ML) criterion which where X is the correlation coefficient (0 5 X 5 1)
checked the most likely among all the models in between successive measurements and W: is a white
the same subset to make a decision. In the IMM s e q ~ e n c eTo
. ~ decorrelate the measurement noise
algorithm the likelihood function for each model a new measurement z,i(k), denoted as “artificial
has some contributions from all the models in measurement” in [90], is generated. Let X be the
the same subset, so it is a mixed likelihood. The estimated value of noise correlation, then the
second criterion was the maximum a posteriori measurement equation can be rederived as follows
(MAP) criterion which included the prior probability.
Again, as the probability of each model has some z,(k) = z ( k ) - Xz(k - 1) (18)
contributions from all the models in the same subset, = (HjX(k)+ Wj(k))- X(HjX(k - 1) + Wj(k- 1))
the a posteriori probability also has mixed terms.
Since more information is used in this decision =[ ~ ~ ~-(Dkq )~ j ) - l ( ~ ( k-)r,vj(k - 1)1
rule, the tracking performance was improved. The
simulation results showed that the performances for + [Wj(k)- X W j ( k - l)]
all four designs of the SFIMM algorithm considered =[H~ - X H ~ ( ~ ; ) - ~ I X+(XHj(F;)-lrjvj(k
~) - 1)
were acceptable, although design 4 which used
subset type 2 and the MAP decision rule had the best + [(A - X)Wj(k 1) + w3k)l. - (19)
performance. The tracking accuracy was only slightly Let
worse than that of an IMM algorithm with thirteen
subfilters and took approximately one quarter of the Hj = H, -IH,(FJ)-’ (20)
computation time.
Another approach, suggest by [82], is to use
the so-called adaptive interacting multiple model i j ( k ) = F j V j ( k - 11) + (A -X)wj(k - 1) + W p ) (22)
(AIMM) algorithm. Each model is assigned a fixed
deterministic acceleration to cope with the different Then n

target accelerations. A rough estimate of the target +ij(k).

z,(k) := HjX(k) (23)
acceleration is taken from a biased filter which If X M X the new measurement noise Gj(k) would
uses combined information from the models. This be white, but it is correlated with the process noise
estimated acceleration is then taken into account in the
models. Different maneuvering target scenarios were
generated to compare the performance of the AIMM 3The notations in [55] were erroneous.

MAZOR ET AL.: INTERACTING MULTIPLE MODEL METHODS IN TARGET TRACKING: A. SURVEY 111

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 vj(k - 1). By reformulating the dynamic equation (sequential fusion). The paper [411 presents a JPDAF
(6) properly, the process noise can be made to be in DSN structure, where each local node processes
uncorrelated with new measurement noise [ 141. the local sensor measurements and sends the local
In some practical systems, this procedure can be estimates periodically to the global processor (fusing
omitted with little degradation in performance since node, parallel fusion). After processing the results, the
the term f j v J ( k- 1) is usually small. Thus, the global processor sends back the global estimates to
baseline IMM algorithm can be applied to the case each local node.
with correlated measurement noise by the following In many multisensor systems the number and type
substitutions: of sensors supporting a particular target track can
A vary with time due to the mobility, type, and resource
Hj --f HJ w,(k)-+ kj(k) limitations of the individual sensors. This variability
z(k) --f z,(k) for j = 1,2,. ..,Nf. in the configurations of the sensor system poses
a significant problem when tracking maneuvering
The results of computer simulations [55] indicated targets because of the uncertainty in the target motion
that the decorrelation process may improve system model. In the paper [27], the IMM estimator is
performance significantly, especially in the velocity applied to the problem of tracking maneuvering
and acceleration estimates. These large improvements targets with multiple, possibly intermittent, sensors.
in velocity and acceleration are particularly useful in For the IMM estimator, the following two models
situation when the target is suddenly accelerated by were chosen: 1) the nearly constant velocity model
the pilot or missile guidance program. with piecewise constant acceleration errors, and
2) constant acceleration model with piecewise constant
F. Multisensor IMM Tracking Algorithms acceleration increment errors. The IMM algorithm
provides significantly better tracking performance than
In some military and civilian applications, the a nearly constant velocity KF (this filter is based on
data are collected by a network of sensors distributed the assumption that the target is moving with constant
over a large geographic region. In such distributed velocity plus zero-mean, white Gaussian acceleration
sensor networks (DSNs), because of considerations errors that are piecewise constant) even when the
such as reliability, survivability, and communication latter runs at a higher rate.
bandwidth, centralized processing is either undesirable In the case where "track-to-track" fusion [ 151
or infeasible. Instead, the sensors supply data to a set is carried out, it is very important that the state
of local processorshodes which are connected by estimators have realistic (consistent) covariances [ 141.
a communication network. The nodes process the While the KF cannot be consistent since it uses only
local sensor data and exchange processing results one model, the IMM has been shown to be nearly
with other nodes. The key problem in multitarget consistent. Lack of consistency of the KF can be a
tracking is data association, i.e., the association of major problem in multisensor fusion.
the measurements from single or multiple sensors to Sequential Multisensor IMMPDAF Algorithm: The
the set of targets. Many different approaches have sequential multisensor (MS) IMMPDAF algorithm
been proposed for data association in a centralized is presented in [15, 581. This algorithm assumes
framework [41], in which all the sensor measurements the simplest form of hybrid system (6), (7), (9);
are transmitted to a centralized site where they each mode-matched filter is a standard PDAF [15].
are processed. One such approach is JPDAF (for The sensors (two-one infrared and one radar)
multiple targets in clutter) and PDAF (for a single are in the same location, assumed, for simplicity,
target in clutter); see [15]. In these algorithms the synchronized and with the same sampling rate. Track
data association problem is handled by computing initialization is assumed to have been made. At each
the posterior probabilities of the measurement scan, a validation gate, centered around the predicted
associations with targets in clutter. These probabilities measurement of the target, is set up for each sensor
are used to update a set of Gaussian distributions for the next scan.
representing (suboptimal) Bayesian estimates of The summary of the sequential MS IMMPDAF
the target locations. Several distributed tracking algorithm (one cycle) is as follows.
algorithms for DSN have been discussed in [41].
The paper [44] presents a distributed tracking scheme 1) Mixing of the estimates from the previous time
based on the JPDAF where both data association and (see Section IIIA).
distributed estimation are considered. In this scheme, 2) Predicted states and predicted measurements.
each node first performs the tracking functions with 3) Measurement validation from sensor 1. The
the JPDAF using the local measurements and sends validation region has to be taken the same for each
the processed results to other nodes. The receiving of the models as the largest of them.
node then fuses the information from other nodes 4) Estimation with sensor 1 measurements in each
with its local information to arrive at a better estimate filter. At the end of the estimation, one computes

112 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 1 JANUARY 1998

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 a new “predicted” measurement for sensor 2, as in update with the measurements from the various
step 2. sensors (assumed synchronized) [84]. The reason
5) Measurement validation from sensor 2. This the sequential update is superior is that it reduces
step is similar to step 3. the uncertainty faster and allows for better data
6) Estimation with sensor 2 measurements in each association.
filter. This step is similar to step 4.
7) Updating of the model probabilities. (See G. Square Root Algori.thms for IMM Filters
Filtering in Section IIIA).
8) Combination of the model-conditioned Theoretically, the K F gives the unbiased, minimum
estimates. (See Combination in Section IIIA). variance estimate of the state vector of a linear
dynamic system, disturbed by additive white noise
Parallel Multisensor IMMPDAF Algorithm: The when measurements of the state vector are linear, but
parallel MS IMMPDAF algorithm is presented in corrupted by Gaussian white noise. Such performance
[42, 431. This algorithm assumes the simplest form of is hardly ever realized in actual practice since the
hybrid system (6), (7), (9) too; each mode-matched information required tal construct the KF is only
filter is a PDAF. The two-node scenario similar to approximately known. The noise parameters and
that given in [41] is considered, where each node models may be based upon only relatively few data
processes the local measurements from its own sensor points, computer round-off errors may be significant,
and sends the local estimates to the fusion processor and the system model may not be adequate. The
periodically. The fusion processor then sends back conventional KF algorithm is not numerically
the processed results after each communication robust due to round-ofl errors and ill-conditioning
time. Lossless communication is assumed and the problems. Numerical sludies in the past [16, 621 have
information communicated is the sufficient statistics. shown that the Kalman algorithm is so sensitive to
For each local node, the centralized IMMPDAF computer round-off errors that numerical accuracy
algorithm, where all measurements are sent to and can degrade to the point where results cease to
processed with one processor, is described similar to be meaningful. Typical problems include the loss
baseline IIvlM algorithm with a standard PDAF. The of positive-definiteness of the covariance matrix
goal is to compute the conditional state distribution resulting from numerical errors such as finite
given the local accumulated measurements. With the computer word length and cancelation errors due
local conditional pdf of node i, one can derive the to the subtraction term in the Kalman covariance
fusion algorithm for the baseline IMM filter. The update. This type of error may be manifested as a
information needed to be communicated from local covariance matrix computed erroneously as having
nodes to the fusion node consists of a) the model negative eigenvalues. 111 a significant class of filtering
probabilities; b) the association event probabilities; problems, propagation of the error covariance matrix,
and c) the corresponding probability density functions by means of the KF equations, results in a matrix
(pdfs) (mean and covariance for Gaussian case). which is not positive semidefinite (a theoretical
Three different scenarios were tested. First, each impossibility). This may occur when 1) the covariance
sensor tracks the target independently using the matrix is rapidly reduced by processing very accurate
IMM algorithm from Section IIIA. Second, the measurements, or 2) a linear combination of state
measurements from sensors were first concatenated vector components is known with great precision,
[loo] into one, then the standard IMM algorithm, as while other combinations are essentially unobservable.
in the first case, was applied. Finally the distributed The source of trouble in both cases is the numerical
(parallel MS IMMPDAF) case was simulated. computation of ill-conditioned quantities in finite
Simulation results show that the performance of the word length. These errors can cause filter divergence
distributed case depends explicitly on the amount of or computations that are so erroneous as to be
information being communicated between nodes. The meaningless. In order to overcome these difficulties,
results of centralized processing scheme represent square root formulations of the Kalman filtering have
the upper bound of the performance. When two been proposed since they improve the numerical
nodes communicate every scan, the results of the conditioning (robustness) of the solution and achieve
distributed case are the same as the centralized case twice the effective precision given by conventional
(but has the advantages of increased reliability), Kalman algorithms [17, 59, 1031. One class of
which confirms their theoretical equivalence. When algorithms that yields satisfactory results uses a
two nodes communicate every two scans, the results square root factorizaticin of the covariance matrix.
of the distributed case are better than that of the Square root factorizations generally have less dynamic
decentralized case (no communication), which shows range associated with their calculations, and are less
the advantage of interchanging information. susceptible to round-off error. Another advantage of
For a centralized MS PDAF the sequential update the square root representation is that the covariance
across sensors is preferable over the simultaneous matrix does not need to be calculated directly. Instead,

MAZOR ET AL.: INTERACTING MULTIPLE MODEL METHODS IN TARGET TRACKING: A SURVEY 113

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 it is represented in terms of its square root matrices. present an alternate way of implementing the
This factored form cannot represent a matrix with covariance prediction step, when the process noise
negative eigenvalues. There are some disadvantages to covariance is time invariant. This results in substantial
factored forms of the KF. One significant disadvantage computational savings when compared with the
is that they normally require more computations, and MWG-S algorithm. The covariance prediction step
therefore more computer time, than the conventional involves rank-one corrections to the L-D factors. This
KF. The Potter covariance square root, the Carlson also can be accomplished by successively using the
covariance square root, and Bierman-Thornton’s U-D L-D rank-one correction algorithm. The proposed
filter are three of the more common factorization algorithm is approximately two times cheaper than
methods for discrete-time Kalman filtering. Of these MWG-S when the process noise covariance is time
factored forms, the U-D filter offers a high degree invariant. However, if the process noise covariance is
of numerical precision along with a reasonable time varying, then MWG-S algorithm is preferable
increase in the required number of mathematical since the rank-one corrections approach needs the
operations. Because of these properties, the U-D decomposition at every filter iteration. It can lead
filter is a standard for comparison. The U-D filter is to numerical instabilities, because the semidefinite
a factorization of the covariance matrix in the form Cholesky might not be numerically robust [88]. By
U D U T , where U is an upper triangular unit matrix (it using the results of [87, 881, the need to compute
has ones along the main diagonal) and D is a diagonal the filter gain matrix y(k)and the residual
matrix. The efficient L-D (or U-D) factorization covariance S j ( k ) is obviated. Given the L-D factors
algorithms for IMM and IMMPDAF are presented of q ( k I k - l), the L-D factors of q ( k I k ) are
in [87, 881. The system dynamics are modeled by the determined. The output of this stage are the model
equations (6)-(7); it is assumed that the model jump conditioned estimates (see Section IIIA (Filtering)),
process is a Markov process with known transition L-D factored covariances for time k [87, 881,
probabilities pij (see (9)). The new version of the likelihood functions
square root PDAF is approximately two times faster
than Kenefic’s algorithm [ 5 9 ] .
1) Square Root IMM Filter: The baseline IMM
algorithm was presented in Section IIIA. Here, we
summarize the square root covariance updating for the where m’ is the number of measurements originated
IMM filter (one cycle). from the object in track at time k , { p I j } and {az,}
are the sequential innovation and sequential
Interaction. Starting with L-D factors of the innovation covariance of the ith scalar component of
model conditioned covariances of time k - 1 measurement vector for the filter matched to the jth
model (these are computed during the covariance,
P , ( k - 1 / k - 1) = L , ( k - 1 I k - l ) D , ( k - 1 Ik-1) update recursion of the filter matched to the jth
xL,(k-l Ik-1)T j = I , ...,Nf model). The likelihoods are used to update the model
probabilities (see Section IIIA (Filtering)).
we compute the L-D factors of the mixed initial Combination. The overall state estimate is
covariances for each of the filters. Note that the factor computed as usual. The covariance mixture equation
pi,,(k - 1 I k - 1) can be absorbed into D,(k - 1 I k - 1). is very similar in structure to the interaction mixture
To compute the factorization of Pq(k - 1 I k - 1) equation of Interaction mixing step. Therefore, the
a rank-one correction is applied to factors of each strategy followed here is the same. Let
P,(k - 1 I k - 1) and the resulting sum of Nf L-D
factored terms are combined to form a single L-D q ( k I k ) = ~ ~I k () ~k ~I k( )k ~ ~I klT
( k i = 1,2,...,Nf .
factored term. The rank-one correction can be done Note that the factor p j ( k ) can be absorbed into
using the generalized L-D rank-one correction D,(k I k ) . To compute the factorization of P(k I k ) , it
algorithm given in [ 8 8 ] .The merging of Nf L-D is needed to apply a rank-one correction to factors
factored terms can be accomplished by recursively of each of P,(k I k ) and the resulting sum of Nf L-D
applying the algorithm, which is also given in [88]. factored terms must be combined to form a single
The mixed initial condition for each of the model L-D factored term. The rank-one correction can be
conditioned filters is computed as usual (see Section done using the generalized L-D rank-one correction
IIIA) . algorithm also. The merging of Nf factored terms can
Filtering. The state estimates and the factored be accomplished by recursively applying the algorithm
covariances computed in the interaction step form which is also given in [88].
inputs to filter matched to each of the Nf models.
The covariance prediction step P,(k 1 k - 1) can be 2) Square Root IMMPDAF: The square root
performed using a modified weighted Gram-Schmidt IMMPDAF is as described above for the square root
(MWG-S) algorithm [16]. In [87, 881 the authors IMM except for the following modifications.

I14 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 1 JANUARY 1998

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 Model Conditioned Filtering. Since the standard one of a finite set of accelerations. The perturbation
KF are replaced by PDAFs, the covariance prediction upon this constant acceleration is modeled as white
and update recursions for the square root PDAF, given zero-mean Gaussian noise. The authors consider
in [87, 881 are used for the Nf filters. The output of two different tracks and twelve models, each being
the stage are the model conditioned estimates and L-D assigned a specific acceleration value ranging from 0
factored covariances for time k. to 4g (= 40 m/s2). In the first track the acceleration
Computation of Likelihoods. The likelihoods can level of the target almost matches the acceleration
be expressed as level of one of the system models. In the second
track, there is a significant difference between the
acceleration of the target and the acceleration level
of the closest model. It is shown that the performance
i= 1 depends strongly on the parameter
q=’ IA Z (AT*>
~
3

ar

where in which IAZI the maxiimal difference between the

acceleration levels of any two models, AT is the time
interval between measurements and 0; is the standard
deviation of the measurement error. In general, when
this parameter is relatively large, both algorithms
perform well. When q is relatively small, the Viterbi
algorithm is better. However, during the beginning
where V, is the volume of the common validation of a track and after the start of a maneuver the IMM
region at time k , mi is the number of validated algorithm provides betiter estimates. This is especially
measurements at time k , PD is the probability important when data association is critical.
of detection, Pc is the probability that the target Recently the IMM lhas been tested on real
measurement falls in the validation region, { p i j l } and multisensor ATC data iin [102]. The IMM estimator
{ a i j }are the sequential innovation and its variance was implemented in cclnjunction with a 2-dimensional
corresponding to the ith scalar component of Zth assignment algorithm that associated the latest
validated measurement vector for the filter matched to scan measurements to the established tracks using
jth model (these are computed during the covariance a ML criterion. The performance of the IMM
update recursion of PDAF matched to jth model). was significantly superior to that of a KF used as
The rest of the algorithm is the same as that of the comparison in all the following aspects:
square root IMM filter. 1) horizontal prediction error,
2) vertical prediction error,
3) number of measurements correctly assigned to
IV. COMPARISON OF THE MAIN MM ALGORITHMS
tracks
Two maneuvering target tracking techniques on an extensive database consisting of over 50 targets
were compared in [SI. The first technique, called from 2 radars. Also, the IMWassignment algorithm
input estimation (see, e.g., the text [14]), models tracked with no difficulty the aircraft from the test
the maneuver as constant unknown input, estimates flight reported in [70].
its magnitude and onset time and then corrects
accordingly the state estimate. The second approach
models the maneuver as a switching of the target V. PARALLEL IMPLEMENTATION OF IMM
state model, where the various state models can be ALGORITHMS
of different dimension and driven by process noises of
different intensities, and estimates the state and mode The IMM algorithm basically consists of a
via the IMM algorithm. While the input estimation group of filters which run in parallel, and a global
algorithm requires around 20 parallel filters, it was computation process which collects the results
shown that the IMM performs equally or better with 2 of the filters and produces the output estimation.
or 3 filters. However, in real applications, processing time may
The paper [SI investigates and compares the become impractical. Multiprocessors seem to offer
performance of the IMM and Viterbi algorithms, the throughput needed for such applications. In the
applied to the problem of radar tracking of paper [6] the authors investigate the performance
maneuvering target. The problem considers a of IMM estimator, when implemented on a general
kinematic model in which the acceleration during a purpose shared-memoly and a shared-bus MIMD
maneuver is almost constant and its mean value is (Multiple Instruction Multiple Data) multiprocessor.

MAZOR ET AL.: INTERACTING MULTIPLE MODEL METHODS IN TARGET TRACKING: A SURVEY 115

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 The implementation of this algorithm on the parallel TABLE I
Efficiency of Parallelization of IMM on 4-CPU SPARCserver
processor is based on the concept of one processor
1000 Multiprocessor
acting as a manager which performs the global
computations, and all other processors acting as # Models in IMM Speed-up Efficiency
workers. Both the manager and the workers are
1 0.67 0.17
implemented as local tasks. The assignments of the 2 0.73 0.18
manager during each time slice are: getting new 3 0.83 0.21
measurements, generating the output based on results 5 1.17 0.29
computed by the other processors, synchronizing 9 1.98 0.49
11 2.24 0.56
between processors, and updating some variables
13 2.43 0.60
which are part of the EKF and are relevant to all
processors.
Each time slice consists of a large number of VI. NONS1MULATION PERFORMANCE PREDICTION
computations, most of which are an EKF step TECHNIQUES FOR IMM
performed independently for each possible current
Performance prediction of IMM algorithms
model. The input of each such filter is an estimate of
only via costly and time-consuming Monte-Carlo
the system state vector along with its covariance for
simulations does not provide sufficient insight into
the previous time step and a new measurement. The
the problems. Unfortunately, none of the earlier
output is an updated estimate. This computational load
nonsimulation techniques [52, 64, 81, 911 can be
can be divided among the processors almost without
applied to hybrid state estimation schemes based on
additional overhead by assigning each processor a
MM, such as the IMM algorithm. Error bounding
subgroup of the models.
techniques (see e.g. [47, 50, 60, 971) can provide
The computational complexity (both the time
at most only bounds. In the paper [66] an effective
and space (memory) complexity) as well as the
hybrid approach to the nonsimulation performance
speed-up and efficiency are examined in detail. The prediction of IMM systems was developed. In view
time complexity is evaluated using four categories: of the fact that the performance of a MM scheme
elementary arithmetic operations (+, -, ., /), complex is scenario dependent, this approach is based on a
arithmetic operations (exp( ), arctan( )), accesses to stochastic performance measure (Le., conditional
local memory, and access to shared memory. expectation of the error covariance) of hybrid nature
The space complexity is evaluated using two in the sense that it is a continuous-valued matrix
categories: shared memory and local memory. The function of a discrete-valued sequence-the system
basic unit in calculations is the space needed to store a mode sequence, which constitutes the scenario of
floating variable which is 4 bytes. the problem of interest. This hybrid approach is
In the study reported by [4] the objective was to obtained by using a conditional expectation operation,
implement the IMM algorithm in a parallel structure through which the randomness of the error covariance
(on a network of four transputers). The experimental due to the uncertainties in the continuous subspace
work has been done on transputer development system of the hybrid space is averaged out whereas the
(TDS) using the Occam programming language. Three information of the system mode jumping is retained.
different architectures for parallel implementation To calculate efficiently the performance measure,
of the IMM algorithm were examined, each with an off-line recursion has been obtained based on
different task partitions for the manager and the conditional successive one-step-ahead-predictions. The
workers. A speedup of approximately 2.7 was capability of the approach in predicting quantitatively
achieved for the 13 filter IMM implementation. This the average performance of the algorithm was
was further improved to 3.2 by using a time-slipping illustrated via two important examples: a generic
calculation method which was shown experimentally maneuvering target tracking problem in ATC and
to have a negligible effect on the average tracking a nonstationary noise identification problem. The
accuracy. predicted performance given by this approach agrees
More recent work on parallelization of an remarkably well with the simulated one, which
IMM-based large scale MTMST algorithm has been verifies the good accuracy of the approach and
presented in [86]. It was shown in this work that therefore it can be used as an efficient tool for design,
parallelization of an IMM becomes practical only if sensitivity analysis, and so on.
the number of models is large (see Table I). The philosophy of this conditional expectation
Highly efficient parallelization (around 100%) based approach is quite general and can be applied
was shown in [86] to be achievable on a large scale to many other algorithms for hybrid systems. For
MT MST problem with IMM trackers by doing the example, the performance of the STDM-based
parallelization across the IMMs, rather than within the IMM estimator [37, 391 can be evaluated using this
IMMS. approach with suitable modifications. The approach

1I 6 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 1 JANUARY 1998

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 can be readily extended to the variable structure IMM acceleration model, conresponds to the situations
estimator [65], square root IMM algorithm [87, 881, that the target startdterminates a maneuver (large
correlated measurement noise IMM estimator [55], acceleration increments). The tracking algorithm is
IMBM and IMAM algorithms [21]. an IMM estimator with three models corresponding
to the above three modes of the dynamic system. The
VII. USING AN IMM ESTIMATOR FOR A PHASED evolution of the hybrid system among the modes is
ARRAY RADAR modeled as a first order Markov chain. The covariance
of the noise, contributing to the evolution of the state
The use of the IMM algorithm for computing in each mode, is determined based on the maximum
measurement times for a phased array radar is acceleration corresponding to that mode. The tracking
illustrated by [22, 241 for the benchmark problem algorithm also selects the revisit (dwell) time for
[23], where the average sample was about 4 s for an the target. The interval to the next dwell is selected
aircraft in a holding pattern and about two seconds for from a predetermined set of values as the largest
a manned aircraft performing 7g maneuvers. A revisit interval, such that the prediction of the measurement
technique that responds to target maneuvers and errors are smaller than ia certain selected threshold
avoids the decision-directed approach to both tracking related to the radar’s beamwidth. Beam pointing
and sample rate adjustment can be accomplished using control is done based on the prediction of the target
the IMM algorithm. Since the output error covariance coordinates at the selected revisit time and pointing
of the IMM algorithm reflects better the uncertainty in the beam to this direction. The range gate of the radar
target motion during a maneuver than that of single is centered at the predicted range of the target at the
model filters, the predicted error covariance of the revisit time. The pointing accuracy (how much the
IMM algorithm can be used to compute the time for target is off the beam center) is accounted for in the
the next measurement, so that a given level of the next measurement since it affects the SNR and the
performance i s maintained during a maneuver. The measurement accuracies depend on the SNR.
output of the IMM algorithm was used to compute The probability transition matrix of the Markov
the time of the next measurement, or radar dwell, chain is designed based on the expected sojourn time
to keep track loss below a given level. By including in each mode of the hylbrid system [14]. In [45] a
maneuvering and nonmaneuvering models in the IMM predetermined set of sampling intervals is used, with
algorithm, the sample times automatically decreased the selection done based on the prediction of the radar
during maneuvers to maintain the desired level of angle innovation standard deviations compared with
performance. Thus, using the output of the IMM the radar beamwidth. At each time step, the estimator
algorithm gives measurement or dwell times that predicts the innovation covariance for the largest
automatically reflect target maneuvers, target range, possible sampling interval. The uncertainty, caused
missed detections, and weak and strong signal returns. by the fluctuations of the target cross-section, is also
Also, the probability of the nonmaneuvering mode considered in the calculations of the variances. Then,
within the IMM algorithm was used to adjust the the standard deviations of both angle innovations are
energy of the waveform. By increasing the energy compared with a fraction of the antenna beamwidth.
during maneuvers, the accuracy of the state estimates If any of the azimuthal or elevation angle innovation
was improved significantly. standard deviations excleeds this threshold, the
Another design and the performance of an IMM candidate sampling interval is rejected and the test
estimator for the benchmark problem [23] of highly is repeated for the next smaller sampling interval.
maneuvering targets are presented in [45]. The design Sampling for initialization is done with the maximum
parameters of the IMM algorithm as well as an allowable rate (10 Hz). This rate lasts for a total
adaptive sampling policy are described. It is shown of four steps to assure a fast decrease of the initial
that the IMM filter with adaptive sampling policy has errors. After the sampling interval is determined, the
a better performance in terms of the average track position of the target is predicted and used to point
loss, averaged dwells per run and noise reduction the beam.
in nonmaneuvering periods of time than the best The simulation results of the work reveal that
KF. The design was done based on the first four the IMM estimator in conjunction with the adaptive
benchmark trajectories. The evaluations were done sampling policy yields satisfactory results in terms of
for these four plus two additional trajectories. In [45] track loss percentage and average dwells per run. This
the authors considered three modes: the first mode is comes from the maneuver adaptability of the IMM
a nearly constant velocity motion (second order per algorithm. A single model KF with the same adaptive
coordinate with a low level white noise acceleration). sampling policy does nlot yield satisfactory results.
In the second mode, the target dynamic model Even in comparison with the KF with 1 s constant
has a relatively higher level of noise. This mode sampling, the IMM with the average sampling interval
corresponds to an on-going maneuver. Finally, the of about 2 s has less track loss. Another advantage of
third mode of operation, which is a Wiener process the IMM algorithm is tlhe noise reduction in the state

MAZOR ET AL.: INTERACTING MULTIPLE MODEL METHODS IN TARGET TRACKING: A !SURVEY 117

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 estimate, especially, in the velocity components of the algorithm with two filters. One filter is matched to the
state. dynamic system with Gaussian measurement noise
The second, more realistic, benchmark problem and the other is matched to the same dynamic system
[26] included false alarms and ECM (electronic but with a high variance Laplacian distributed noise.
countermeasures), specifically RGPO (range With target motion described in Cartesian coordinates
gate pull-off) and SOJ (stand-off jammer). The and measurements in the radar coordinates, the
solutions for this problem presented in [61, 981 former filter is an EKE Even though one can obtain
used the IMMPDAF while the one of [20] used the a nonlinear filter that accounts explicitly for the
MHT and an IMM. The latter also proposed the Laplacian distributed measurement noise, it is
combination of the MHT with the IMM, but this was expensive due to the need to perform numerical
not implemented. integration in real time. The results of this work show
that a W with moment-matched measurement noise
VIII. IMM ALGORITHMS FOR TARGET variance, performs close to the optimum. Therefore,
TRACKING WITH GLINT NOISE AND NOISE the filter matched to the second model is also an EKE
IDENTIF I CAT10N The authors showed that this method performs better
than the score function method and it is also robust to
The optimality of the KF is based on the the variations of the glint noise parameters.
assumption of having zero-mean white Gaussian In [68] an approach to noise identification is
noises. If this assumption is violated, the KF is no proposed which uses hybrid system models with a
longer the optimal filter. A non-Gaussian noise arising suitable estimation scheme. This approach is valid
in a radar system is known as a glint noise. The glint for non-stationary noise with rapidly or slowly
noise distribution is long tailed and seriously affects varying statistics as well as stationary noise. The IMM
the tracking performance. Filtering in non-Gaussian algorithm was used in the approach. The proposed
environments has been studied by many researchers. approach is powerful in the sense that it can be used
Robust preprocessing methods for KF have been to identify simultaneously various noise parameters
proposed in [57, 751. In the method of [74] a in linear and nonlinear systems. It is also flexible in
nonlinear score function was introduced which yields terms of model design.
a corrective term in the state estimate equation, while
retaining the computationally appealing recursive IX. CONCLUSION
structure of the KF. In this method it is assumed that
either the process noise or the measurement noise is While no tracking algorithms can work in an
non-Gaussian. However, since it requires a numerical arbitrarily dense multitarget-multisensor environment,
convolution operation to evaluate the score function , the recently developed IMM techniques described in
this limits the practicality of this method. this paper extend the range in which one can achieve
New algorithms were developed by Wu [loll. reliable tracking performance, Le., small track loss
These algorithms incorporate the nonlinear Masreliez percentage. Compromises between computational
filter [75] into the IMM method. They substitute requirements and performance will be necessary in
the KF in the IMM algorithm with the nonlinear many cases, especially where the data rate is high
Masreliez filter and modify the estimate of model and the available computing power is limited, as in
probabilities. Simulations show that these new airborne systems. The assumption made in the various
algorithms not only can filter the glint noise algorithms discussed here will probably be modified
efficiently, but also respond quickly to maneuvering. according to the needs of the specific applications.
Although Wu has used the score function method with A number of problems that are yet open can be
combination of IMM approach, he has avoided the seen at this time. One such problem is the treatment
convolution operation and simplified the evaluation of multisensor group tracking of maneuvering targets
of the score €unction by applying a normal expansion in clutter [18]. At this point it is not known whether
for the distribution of the measurement prediction. filtering techniques can be coinbined with any of
However, the approximate spherical model [79] the methods cited above. The application of new
used to decouple the state components is not a good techniques to phased-array radars, with special
approximation [46]. consideration to the necessary scanning patterns,
The application of the IMM approach to the should also be continued [21, 23, 36, 45, 951. The
target tracking problem, when the measurements are performance of the IMM estimator can be further
perturbed by glint noise, is considered also in [46]. improved by fine-tuning of the design parameters.
A good approximation for a glint noise with heavy The IMM procedure has been established on
tails is a mixture of a Gaussian noise with a moderate a solid theoretical basis and proved to be quite
variance and a Laplacian distributed noise, with high appropriate for the maneuvering target tracking
variance and low occurrence probability. Therefore, problem [94]. The investigations documented in
for this problem the authors considered an IMM the literature indicate that the IMM algorithm is the

118 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 1 JANUARY 1998

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 superior technique for multisensor-multitarget tracking Bar-Shalom, Y. (Ed.) (1992)
and data association. Multitarget-Multisensor Tracking: Applications and
Advances, Vol. 11.
The recent activity in the area of IMM methods
Norwood, MA: Airtech House, 1992.
indicates that this emerging body of knowledge Bar-Shalom, Y., Chang, K. C., and Blom, H. A. P. (1992)
experiences a rapid transition to becoming relevant Tracking splitting targets in clutter by using an interacting
to engineering practice by solving real problems. multiple model joint probabilistic data association filter.
In Y. Bar-Shalom (Ed.), Multitarget-Multisenor Tracking:
Applications and Advances, Vol. 11.
ACKNOWLEDGMENTS Norwood, MA: AI-techHouse, 1992, 93-110.
Bar-Shalom, Y., and Li, X. (1993)
E. Mazor wishes to thank S. Rogers, S. Levin and Estimation and tracking: Principles, techniques, and
B. Golubev from Elta Corp., Israel, for introducing the software.
problem and continuing support through all the stages Norwood, MA: Artech House, 1993.
of the work. Bar-Shalom, Y., and Li, X. R. (1995)
Multitarget-Multissnsor Tracking: Principles and
Techniques.
REFERENCES
Storrs, CT: YBS F'ublishing, 1995.
Ackerson, G. A., and Fu, K. S. (1970) Bierman, G. J., and Thornton, C. L. (1977)
On state estimation in switching environments. Numerical comparison of Kalman filter algorithms: Orbit
IEEE Transactions on Automatic Control (Feb. 1970), determination case: study.
10-17. Automatica, 13 (Jan. 1977), 23-27.
Alouani, A. T., and Blair, W. D. (1991) Bierman, G. J. (1977)
Use of a kinematic constraint in tracking constant speed, Factorization Methods for Discrete Sequential Estimation.
maneuvering targets. New York: Academic Press, 1977.
In Proceedings of the 30th Conference on Decision and Blackman, S. S. (19186)
Control, Brighton, England, Dec. 1991, 2055-1058. Multiple Target Trucking with Radar Applications.
Alouani, A. T., Xia, P., Rice, T. R., and Blair, W. D. (1991) Norwood, MA: Artech House, 1986.
A two-stage Kalman estimator for state estimation in the Blackman, S. S. (1986)
presence of random bias and for tracking manuevering Multitarget tracking with an agile beam radar.
targets. In Y. Bar-Shalom (Ed.), Multitarget-Multisenor Tracking:
In Proceedings of the 30th IEEE Conference on Decision Applications and Advances, Vol. 11.
and Control, Brighton, England, Dec. 1991, 1059-1062. Norwood, MA: Artech House, 1986, 237-269.
Atherton, D. P., and Lin, H.-J. (1994)
Blackman, S. S., Busch, M. T., and Popoli, R. F. (1995)
Parallel implementation of IMM tracking algorithm using
IMMMHT tracking and data association for benchmark
transputers.
tracking problem.
IEE Proceedings on Radar Sonar Navigation, 141, 6 (Dec.
In Proceedings of ithe American Control Conference,
1994), 325-332. Seattle, WA, June 1995, 2606-2610.
Averbuch, A,, Itzikowitz, S., and Kapon, T. (1991)
Blair, W. D., and Watson, G. A. (1992)
Radar target tracking-Viterbi versus IMM.
Interacting multiple bias model algorithm with application
IEEE Transactions on Aerospace and Electronic Systems,
to tracking maneuvering targets.
27, 3 (May 1991), 550-563.
In Proceedings of i'he 3Ist Conference on Decision and
Averbuch, A., Itzikowitz, S., and Kapon, T. (1991)
Control, Tucson, AZ, Dec. 1992, 3790-3795.
Parallel implementation of multiple model tracking
algorithms. Blair, W. D. (1994)
IEEE Transactions on Parallel and Distributed Systems, 2, Toward the integration of tracking and signal processing
2 (Apr. 1991), 242-251. for phased array radar, signal and data processing for
Bar-Shalom, Y., and Fortmann, T. E. (1988) small targets.
Tracking and data association. Orlando, FL, Apr. 1994.
New York: Academic Press, 1988. Blair, W. D., Watson, G. A,, and Hoffman, S. A. (1994)
Bar-Shalom, Y., Chang, K. C., and Blom, H. A. P. (1989) Benchmark problem for beam pointing control of phased
Tracking a maneuvering target using input estimation array radar against. maneuvering targets.
versus the interacting multiple model algorithm. In Proceedings of i'he I994 American Control Conference,
IEEE Transactions on Aerospace and Electronic Systems, Baltimore, MD, June 1994.
25, 2 (Mar. 1989), 296-300. Blair, W. D., and Watson, G. A. (1994)
Bar-Shalom, Y. (1989) IMM algorithm for solution to benchmark problem for
Recursive tracking algorithms: From the Kalman filter to tracking maneuvering targets.
intelligent trackers for cluttered environment. In Proceedings of ,4cquisition, Tracking, and Pointing
In Proceedings of the I989 IEEE International Conference Conference, SPIE Orlando '94, Apr. 1994.
on Control and Applications, Jerusalem, 1989. Blair, W. D., and Kazakos, D. (1994)
Bar-Shalom, Y. (Ed.) (1990) Estimation and detection for systems with second-order
Multitarget-multisensor tracking: Advanced applications. Markovian switching coefficients.
Norwood, MA: Artech House, 1990. In Proceedings of 1994 American Control Conference,
Bar-Shalom, Y., Chang, K. C., and Blom, H. A. P. (1990) Baltimore, MD, June 1994.
Automatic track formation in clutter with a recursive Blair, W. D., Watsom, G. A,, Gentry, G. L., and Hoffman,
algorithm. S. A. (1995)
In Y. Bar-Shalom (Ed.), Multitarget-Multisensor Tracking: Benchmark problem for beam pointing control of phased
Advanced Applications. array radar againsit maneuvering targets in the presence of
Norwood, MA: Artech House, 1990,2542. ECM and FA.

MAZOR ET AL.: INTERACTING MULTIPLE MODEL METHODS IN TARGET TRACKING: A SURVEY 119

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 In Proceedings of the American Control Conference, Chang, K. C., Chong, C. Y., and Bar-Shalom, Y. (1986)
Seattle, WA, June 1995,2601-2605. Joint probabilistic data association in distributed sensor
Blair, W. D., and Bar-Shalom, Y.(1996) networks.
Tracking maneuvering targets with multiple sensors: Does IEEE Transactions on Automatic Control, 31, 10 (Oct.
more data always mean better estimates? 1986), 889-897.
IEEE Transactions on Aerospace and Electronic Systems, Chang, K. C., and Bar-Shalom, Y. (1987)
32, 1 (Jan. 1996), 450-455. Distributed multiple model estimation.
Blom, H. A. P. (1984) In Proceedings of the American Control Conference, 1987,
An efficient filter for abruptly changing systems. 797-802.
In Proceedings of the 23rd IEEE Conference on Decision Chang, K. C., and Bar-Shalom, Y. (1989)
and Control, Las Vegas, NV, Dec. 1984, 656-658. Distributed adaptive estimation with probabilistic data
Blom, H. A. P. (1985) association.
An efficient decision-making-free filter for processes with Automatica, 25, 3 (1989), 359-369.
abrupt changes. Chong, C. H., Mori, S., and Chang, K. C. (1985)
In Proceedings of the IFAC Symposium on Identijcation Information fusion in distributed sensor networks.
and System Parameter Estimation, 1985, 63 1-636. Proceedings of the American Control Conference (1985),
Blom, H. A. P., and Bar-Shalom, Y. (1988) 830-835.
The interacting multiple model algorithm for systems Daeipour, E., Bar-Shalom, Y., and Li, X. (1994)
with Markovian switching coefficients. Adaptive beam pointing control of a phased array radar
IEEE Transactions on Automatic Control, 33, 8 (Aug. using an IMM estimator.
1988), 780-783. In Proceedings of the 1994 American Control Conference,
Baltimore, MD, June 1994.
Blom, H. A. P., and Bar-Shalom, Y.' (1990)
Time-reversion of a hybrid state stochastic difference Daeipour, E., and Bar-Shalom, Y. (1995)
system with a jump-linear smoothing application. An interacting multiple model approach for target
IEEE Transactions on Information Theory, 36, 4 (July tracking with glint noise.
1990), 836-847. IEEE Transactions on Aerospace and Electronic Systems,
31, 2 (Apr. 1995).
Blom, H. A. P. (1991)
Hybrid state estimation for systems with semi-Markov Daum, F. E. (1990)
switching coefficients. Bounds on performance for multiple target tracking.
In Proceedings of the European Control Conference, IEEE Transactions on Automatic Control, 35,4 (Apr.
Grenoble, July 1991. 1990), 443446.
Dufour, E, and Mariton, M. (1991)
Blom, H. A. P., Hogendoorn, R. A., and Van Doom, B. A.
Tracking a 3D maneuvering target with passive sensors.
(1992)
IEEE Transactions on Aerospace and Electronic Systems,
Design of a multisensor tracking system for advanced air
27, 4 (July 1991), 725-738.
traffic control.
In Multitarget-Multisensor Tracking: Applications and Dufour, F., and Mariton, M. (1992)
Advances, Vol. 11. Passive sensor data fusion and maneuvering target
Norwood, MA: Artech House, 1992, 31-63. tracking.
In Y. Bar-Shalom (Ed.), Multitarget-Multisensor Tracking:
Bogler, P. L. (1987)
Applications and Advances, Vol. 11.
Tracking a maneuvering target using input estimation.
Norwood, MA: Artech House, 1992,65-92.
IEEE Transactions on Aerospace and Electronic Systems,
AES-23, 3 (May 1987), 298-310. Ezzine, J., and Haddad, A. H. (1988)
Error bounds in the averaging of hybrid systems.
Bogler, P. L. (1990)
In Proceedings of the 27th IEEE Conference on Decision
Radar Principles with Applications to Tracking Systems.
and Control, Austin, TX, Dec. 1988, 1787-1791.
New York: Wiley, 1990.
Farina, A., and Pardini, S. (1980)
Brookner, E. (Ed.) (1991) Survey of radar data-processing techniques in
Practical Phased Array Antenna Systems. air-traffic-control and surveillance systems.
Boston: Artech House, 1991. IEE Proceedings, 127, Pt. F, 3 (June 1980), 190-204.
Campo, L., Mookerjee, P., and Bar-Shalom, Y. (1988) Fortmann, T. E., Bar-Shalom, Y., Scheffe, M., and Gelfand,
Failure detection via recursive estimation for a class of S. (1985)
semi-Markov switching systems. Detection thresholds for tracking in clutter-A connection
In Proceedings of the 27th IEEE Conference on Decision between estimation and signal processing.
und Control, Austin, TX, Dec. 1988, 1966-1971. IEEE Transactions on Automatic Control, AC-30, 5 (Mar.
Campo, L., and Bar-Shalom, Y. (1990) 1985), 221-229.
Control of discrete-time hybrid stochastic systems. Friedland, B. (1969)
In Proceedings of the 1990 American Control Conference, Treatment of bias in recursive filtering.
Boston, June 1990, 7-12. IEEE Transactions on Automatic Control, AC-14, 4 (Aug.
Campo, L., Mookerjee, P., and Bar-Shalom, Y. (1991) 1969), 359-367.
State estimation for systems with sojourn-time-dependent Gauvrit, M. (1984)
Markov model switching. Bayesian adaptive filter for tracking with measurements
IEEE Transactions on Automatic Control, 36, 2 (Feb. of uncertain origin.
1991), 238-243. Automatica, 20 (1984), 217-224.
Chang, C. B., and Athans, M. (1978) Guu, J. A., and Wei, C. H. (1991)
State estimation for discrete system with switching Maneuvering target tracking using IMM method at high
parameters. measurement frequency.
IEEE Transactions on Aerospace and Electronic Systems, IEEE Transactions on Aerospace and Electronic Systems,
AES-14 (May 1978), 418-425. 27, 3 (May 1991), 514-519.

120 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 1 JANUARY 1998

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 Helmick, R. E., Blair, W. D., and Hoffman, S. A. (1995) In Proceedings of the 1991 Joint Service Data Fusion
Interacting multiple model approach to fixed-interval Symposium, Oct. 1991.
smoothing. Lin, J., and Atherton, D. P. (1993)
IEEE Transactions on Information Theory (Nov. 1995). An investigation of the SFIMM algorithm for tracking
Hewer, G. A., Martin, R. D., and Seh, J. (1987) maneuvering targets.
Robust preprocessing for Kalman filtering of glint noise. In Proceedings of the 32nd Conference on Decision and
IEEE Transactions on Aerospace and Electronic Systems, Control, San Antonio, TX, Dec. 1993, 930-935.
AES-23, 1 (Jan. 1987), 120-128. Lu, M., Qiao, X., and Chen, G. (1992)
Houles, A., and Bar-Shalom, Y. (1989) A parallel square-root algorithm for modified extended
Multisensor tracking of a maneuvering target in clutter. Kalman filter.
IEEE Transactions on Aerospace and Electronic Systems, IEEE Transactions on Aerospace and Electronic Systems,
25, 2 (Mar. 1989), 176-189. 28, 1 (Jan. 1992), 153-163.
Kenefic, R. (1990) Mariton, M., and Sworder, D. (1992)
Active sonar application of a U-D square root PDAF. Maneuvering target tracking: Imaging and non-imaging
IEEE Transactions on Aerospace and Electronic Systems, sensors.
26, 5 (Sept. 1990), 850-857. In C. T. Leondes (Ed.), Control and Dynamic Systems:
Ken; T. H. (1989) Advances in Theoiy and Applications.
Status of CR-like lower bounds for nonlinear filtering. New York: Academic Press, 1992, 483-516.
IEEE Transactions on Aerospace and Electronic Systems, Masreliez, C. J. (1975)
25, 5 (Sept. 1989), 590-601. Approximate non-Gaussian filtering with linear state and
observation relations.
Kirubarajan, T., and Bar-Shalom, Y. (1995)
IEEE Transaction:i on Automatic Control (Feb. 1975),
Adaptive beam pointing control of a phased array radar in
107-1 10.
the presence of ECM and FA using IMMPDAF.
In Proceedings of the American Control Conference, Masreliez, C. J., and Martin, R. D. (1977)
Seattle, WA, June 1995,2616-2620. Robust Bayesian estimation for the linear model and
robustifying the Kalman filter.
Leondes, C. T., and Pearson, J. 0. (1973)
IEEE Transactions on Automatic Control (1977), 361-371.
Kalman filtering of systems with parameter
Maybeck, P. S., and Schore, M. R. (1990)
uncertainties-A survey.
Robustness of a moving-bank multiple model adaptive
International Journal of Control, 17, 4 (1973), 785-801.
algorithm for control of a flexible spacestructure.
Lerro, D., and Bar-Shalom, Y. (1993)
In Proceedings of the National Aerospace and Electronics
Interacting multiple model tracking with target amplitude
Conference, NAECON 90, Dayton, 1990, 368-374.
feature.
Millnert, M. (1980)
IEEE Transactions on Aerospace and Electronic Systems,
An approach to recursive identification of abruptly
29,2 (Apr. 1993), 494-508.
changing systems
Li, X. R., and Bar-Shalom, Y. (1991) In Proceedings of the 19th IEEE Conference on Decision
Stability evaluation and track life of PDAF for tracking in and Control (1980), 1026.
clutter. Mookerjee, P., Camlpo, L., and Bar-Shalom, Y. (1987)
IEEE Transactions on Automatic Control, 36, 5 (May Estimation in systems with a semi-Markov switching
1991), 588-602. model.
Li, X. R., and Bar-Shalom, Y. (1992) In Proceedings of the 26th Conference on Decision and
Mode-set adaptation in multiple-model estimators for Control, Los Angeles, Dec. 1987, 332-334.
hybrid systems. Moose, R. L. (1975)
In Proceedings of the 1992 American Control Conference, An adaptive state estimation solution to the maneuvering
Chicago, IL, June 1992, 1794-1799. target problem.
Li, X. R., and Bar-Shalom, Y. (1992) IEEE Transactions on Automatic Control, AC-20 (1975),
Performance prediction of the interacting multiple model 359.
algorithm. Moose, R. L., Vanlandingham, H. F., and McCabe, D. H.
In Proceedings of the I992 American Control Conference, ( 1979)
Chicago, IL, June 1992, 2109-2113; also in IEEE Modeling and estimation for tracking maneuvering
Transactions on Aerospace and Electronic Systems, 29, 3 targets.
(July 1993), 755-771. IEEE Transaction3 on Aerospace and Electronic Systems
Li, X. R., and Bar-Shalom, Y. (1993) AES-15,3 (May 1979), 448-455
Design of an interacting multiple model algorithm for air Mori, S., Chang, K C., and Chong, C. Y. (1992)
traffic control tracking. Performance analysis of optimal data association-With
IEEE Transactions on Control Systems Technology, 1, 3 applications to multiple target tracking.
(Sept. 1993), 186-194. In Y. Bar-Shalom (Ed.), Multitarget-Multisensor Tracking:
Li, X. R., and Bar-Shalom, Y. (1994) Applications and Advances, Vol. 11.
A recursive multiple model approach to noise Norwood, MA: Artech House, 1992, 183-235.
identification. Munir, A., and Atherton, D. P. (1994)
IEEE Transactions on Aerospace and Electronic Systems, Maneuvering target tracking using an adaptive interacting
30, 3 (July 1994), 671-684. multiple model algorithm.
Li, X. R., and Bar-Shalom, Y. (1996) In Proceedings of the 1994 American Control Conference,
Multiple-model esitamtion with variable structure. Baltimore, MD, June 1994, 1324-1328.
IEEE Transactions on Automatic Control, 41, 4 (Apr. Nagy, P., and Bier, S. (1990)
1996), 1-16. Intelligent internelled sensor management systems for
Liggins, M. E., et al. (1991) tactical aircraft.
Multispectral fusion for low observable surveillance: A In Proceedings of the National Aerospace and Electronics
system perspective. Conference, NAECON 90, Dayton, 1990, 321-327.

MAZOR ET AL.: INTERACTING MULTIPLE MODEL METHODS IN TARGET TRACKING: A SURVEY 121

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 Pao, L., and Frei, C. W. (1995) Vacher, P., Barret, I., and Gauvrit, M. (1992)
A comparison of parallel and sequential implementations Design of a tracking algorithm for an advanced ATC
of a multisensor multitarget tracking algorithm. system.
In Proceedings of the American Control Conference, In Y. Bar-Shalom (Ed.), Multitarget-Multisensor Tracking:
Seattle, WA, June 1995, 2601-2605. Applications and Advances, Vol. 11.
Pattipati, K. R., and Sandell, N. R. (1983) Norwood, MA: Artech House, 1992, 1-29.
A unified view of state estimation in switching 1951 VanKeuk, G., and Blackman, S. S. (1993)
environments. On phased-array radar tracking and parameter control.
In Proceedings of the 1983 American Control Conference, IEEE Transactions on Aerospace and Electronic Systems,
1983,458-465. 29, 1 (Jan. 1993), 186-194.
Popp, R., Pattipati, K. R., and Bar-Shalom, Y. (1995) Waltz, E., and Llinas, J. (1990)
The parallelization of a large-scale IMM-based multitarget Multisensor Data Fusion.
tracking algorithm. Norwood, MA: Artech House, 1990.
In Proceedings of the SPIE Conference on Signal and Data Wahsbum, R. B., Allen, T. G., and Teneketzis, D. (1985)
Processing of Small Targets, July 1995. Performance analysis for hybrid state estimation
1871 Raghavan, V., Pattipati, K. R., and Bar-Shalom, Y. (1992) problems.
Efficient square-root algorithms for PDA, IMM and In Proceedings of the 1985 American Control Conference,
IMMPDA filters. Boston, MA, June 1985, 1047-1053.
In Proceedings of the 31st Conference on Decision and Watson, G. A,, and Blair, W. D. (1995)
Control, Tucson, AZ, Dec. 1992, 3680-3685. Solution to second benchmark problem for tracking
Raghavan, V., Pattipati, K. R., and Bar-Shalom, Y. (1993) maneuvering targets in the presence of FA and ECM.
Efficient L-D factorization algorithms for PDA, IMM, In Proceedings of the SPIE Conference on Signal and Data
and IMMPDA filters. Processing of Small Targets, July 1995.
IEEE Transactions on Aerospace and Electronic Systems, Watson, G. A,, and Blair, W. D. (1995)
29,4 (Oct. 1993), 1297-1310. Interacting acceleration compensation algorithm for
[891 Raisbeck, G. (1983) tracking maneuvering targets.
Viewing the issue. IEEE Transactions on Aerospace and Electronic Systems,
IEEE Journal of Oceanic Engineering (Special Issue on 31, 3 (July 1995).
Positioning, Localization and Tracking), OE-8, 3 (July Willner, D., Chang, C. B., and Dunn, K. P. (1976)
1983), 110-1 12. Kalman filter algorithms for a multi-sensor system.
Rogers, S. R. (1987) In Proceedings of the ISth IEEE Conference on Decision
Alpha-beta filter with correlated measurement noise. and Control (Dec. 1976), 570-574.
IEEE Transactions on Aerospace and Electronic Systems, Wu, W. R., and Cheng, P. P. (1994)
AES-23, 4 (July 1, 1987), 592-594. A nonlinear IMM algorithm for maneuvering target
Rogers, S. R. (1991) tracking.
Diffusion analysis of track loss in clutter. IEEE Transactions on Aerospace and Electronic Systems,
IEEE Transactions on Aerospace and Electronic Systems, 2, 30, 3 (July 1994), 875-886.
2 (Mar. 1991), 380-387. Yeddanapudi, M., Bar-Shalom, Y., and Pattipati, K. (1995)
Sung, T. Y., and Hu, Y. H. (1987) IMM estimation for multitarget-multisensor air traffic
Parallel VLSI implementation of the Kalman filter. surveillance.
IEEE Transactions on Aerospace and Electronic Systems, Proceedings ofIEEE, 85, 1 (Jan. 1997), 80-94.
AES-23, 2 (Ma. 1987), 215-224. Zeitz, E H., and Maybeck, P. S. (1993)
Tugnait, J. K. (1982) An alternate algorithm for discrete-time filtering.
Detection and estimation for abruptly changing systems. IEEE Transactions on Aerospace and Electronic Systems,
Automutica, 18, 5 (1982), 607-615. 29,4 (Oct. 1993), 1123-1 136.

Efim Mazor was born in Kazan, Russia. He received his M.S. and Ph.D. degrees in
computer science from Kazan Institute of Aeronautics in 1962 and 1973, respectively.
From 1962 to 1969 he was employed by the Computer Devices Plant and by the
Electromechanical Plant, in Kazan, where he was engaged in radar systems work. From
1969 to 1971 he was a postgraduate student and a scientist at the R&D Institute of
Radio-Electronics, Moscow. He served as an Associate Professor and Group Leader at
the Kazan Institute of Aeronautics. In 1991 Dr. Mazor immigrated to Israel and joined
the Technion-Israel Institute of Technology in Haifa, where he pursued his interests in
the development of efficient algorithms for multisensor target tracking, aircraft group
recognition, image processing and machine vision. Since 1996 he has been working for
Kulicke & Soffa Co., Israel, in their Vision Engineering Group. His research interests
include estimation theory, target tracking, data association and micro-processor based
systems.

122 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 34, NO. 1 JANUARY 1998

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.
 Joshua Dayan received his BSc. and MSc. degrees in chemical engineering from
Technion-Israel Institute of Technology, in 1961 and 1’?64,respectively, and his Ph.D.
from Illinois Institute of Technology, Chicago, in 1967.
After completing his studies he spent several years with Amoco, near Chicago and
with The Industrial Automation Institute in Israel, working in the area of computer
control of industrial processes. Since 1971 he has been on the Faculty of Mechanical
Engineering at Technion in Haifa. His teaching and research are in control applications
and robotics. He is extensively involved in industrial consulting work both in Israel
and the United States. His interests and research areas are in power generation systems,
process control and robotics.
Prof. Dayan had supervised the research work of over 30 graduate students and
published over 80 papers in journals and in conference proceedings,

Amir Averbuch was born in Tel Aviv, Israel. He got his BSc. and MSc. dcgrees in
mathematics from the Hebrew University in Jerusalem, in 1971 and 1975, respectively.
He received his Ph.D. degree in computer science from Columbia University, New York,
in 1983.
During 1976-1986 he was a research staff member in IBM T. J. Watson Research
Center, Dept. of Computer Science. Since October 1987, he has been an Assistant
Professor in the Department of Computer Science, School of Mathematical Sciences,
Tel Aviv University. His research interests include parallel processing and wavelet for
signavimage processing and numerical computation.
Yaakov Bar-Shalom (S’63-M’66-SM’SO-F’84) war; born on May 11, 1941. He
received the B.S. and M.S. degrees from the Technion, Israel Institute of Technology, in
1963 and 1967, and the Ph.D. degree from Princeton University, Princeton, NJ, in 1970,
all in electrical engineering.
From 1970 to 1976 he was with Systems Control, Inc., Palo Alto, CA. Currently he is
Professor of Electrical and Systems Engineering and Director of the ESP Lab (Estimation
and Signal Processing) at the University of Connecticut, His research interests are in
estimation theory and stochastic adaptive control and he has published over 200 papers
in these areas.
In view of the causality principle between the name of a person (in this case,
“(he) will track”, in the modern version of the original language of the Bible) and the
profession of this person, his interests have focused on tracking. He coauthored the
monograph Tracking and Data Association (Academic Press, 1988), the graduate text
Estimation and Tracking: Principles, Techniques and Software (Artech House, 1993), the
text Multitarget-Multisensor Tracking: Principles and Techniques (YBS Publishing, 1995),
and edited the books Multitarget-Multisensor Tracking: Applications and Advances (Artech
House, Vol. I 1990; Vol. 11 1992). He has been elected Fellow of IEEE for “contributions
to the theory of stochastic systems and of multitarget tracking.” He has been consulting
to numerous companies, and originated the series of M~iltitarget-Multisensor Tracking short
courses offered via UCLA Extension, at Government Laboratories, private companies,
and overseas. He has also developed the commercially available interactive software
packages MULTLDATTMfor automatic track formation and tracking of maneuvering or
splitting targets in clutter, PASSDATTMfor data association from multiple passive sensors,
BEARDATTMfor target localization from bearing and fi-equency measurements in clutter,
IMDATTMfor image segmentation and target centroid tracking and FUSEDATTMfor
fusion of possibly heterogeneous multisensor data for tracking. During 1976 and 1977
he served as Associate Editor of the IEEE Transactions on Automatic Control and from
1978 to 1981 as Associate Editor of Automatica. He was Program Chairman of the 1982
American Control Conference, General Chairman of the 1985 ACC, and Co-Chairman of
the 1989 IEEE International Conference on Control and Applications. During 1983-1987
he served as Chairman of the Conference Activities Board of the IEEE Control Systems
Society and during 1987-1989 was a member of the Board of Governors of the IEEE
CSS. In 1987 he received the IEEE CSS Distinguished Member Award. Since 1995 has
been a Distinguished Lecturer of the IEEE AESS. He is co-recipient of the M. Barry
Carlton Award for the best paper in the IEEE Transactions on Aerospace and Electronic
Systems in 1996.

MAZOR ET AL.: INTERACTING MULTIPLE MODEL METHODS IN TARGET TRACKING: A SURVEY 123

Authorized licensed use limited to: Universitaetsbibliothek der RWTH Aachen. Downloaded on April 06,2022 at 09:23:37 UTC from IEEE Xplore. Restrictions apply.