Mathematical and Computer Modelling 44 (2006) 332–341

www.elsevier.com/locate/mcm

Use of the interacting multiple model algorithm with multiple

Received 5 October 2005; received in revised form 10 January 2006; accepted 11 January 2006

Abstract

In a tracking system with multiple sensors, fusion usually plays a critical role in combining information. There exist many
fusion techniques and most of them fall into two categories — measurement fusion and track fusion, depending on what kind
of information is to be shared among sensors. These two techniques are typically applied to centralized or distributed systems,
respectively. In this paper, our main interest lies in the centralized systems which generally have a central processor where
measurement fusion is performed. A common question arising in measurement fusion is the following: is it necessary for the
system to process all the measurements, or does more data always mean better estimates? Blair et al. have shown that the use of
multiple sensor data can sometimes degrade performance when a single model filter is used, and that the use of a multiple model
filter, such as the interacting multiple model (IMM) algorithm, is preferred in order to best utilize the multiple sensor data. Since the
IMM algorithm is the state-of-the-art technique for tracking maneuvering targets, we investigate the IMM algorithm for multiple
sensor tracking in more detail. We have found that, even with the IMM, more data may not give us a better estimate and, in some
cases, it may give us a worse estimate.
c 2006 Elsevier Ltd. All rights reserved.

Keywords: Target tracking; Multi-sensor fusion; Interacting multiple model; Kalman filter; Performance prediction

1. Introduction

In recent years, surveillance systems have relied more and more on multiple sensors working in a coordinated
way to provide more accurate and reliable estimates of targets than isolated sensors. Advantages also include greater
area coverage, robustness, and cooperative sensing. In these systems, fusion usually plays a critical role in combining
information. There exist many fusion techniques [1] and most of them fall into two categories — measurement fusion
and track fusion, depending on what kind of information is to be shared among sensors.
Measurement fusion is often used in a centralized system where measurements are shared directly. Some recent
work [5,6] has found that the data communication requirements of measurement fusion systems could be similar to that
of track fusion systems, and that further bandwidth improvement could be achieved through the use of an intelligent
quantizer. A centralized system may be preferable in some applications, since the global optimal estimation can be

∗ Corresponding author. Tel.: +1 937 775 5054.

c 2006 Elsevier Ltd. All rights reserved.

obtained [1]. In this paper, we will explore a fundamental question in the centralized system: is it always necessary,
for the sake of estimation accuracy, to communicate all the measurements? The particular problem of interest here is
the tracking of maneuvering targets. Blair et al. [3,7] have shown that the use of multiple sensor data can sometimes
degrade performance when a single model filter (e.g., the Kalman filter) is used, and that the use of a multiple model
filter, such as the interacting multiple model (IMM) approach, is required in order to best utilize multiple sensor data.
Since the IMM is the state-of-the-art technique for tracking maneuvering targets, in this paper we investigate in
more detail how the performance of the IMM is affected when it is used with multiple sensors. We illustrate the
problem through an example involving a two-sensor system with a constant sampling rate and target motion in a
two dimensional coordinate. We evaluate the performance of the IMM through both Monte Carlo simulation and
non-simulation approaches.
This paper is organized as follows. The IMM algorithm and the way to measure its performance are discussed
briefly in Section 2. A tracking example is given in Section 3 to illustrate the problem of the IMM performance
associated with tracking maneuvering targets with multiple sensors. Concluding remarks are given in Section 4.

2. The IMM and its performance prediction

To track a maneuvering target, we need to address the problem of discrete structure/parameters model uncertainty,
in addition to the continuous uncertainties arising from motion itself as well as from observations. This is the so-called
hybrid problem. The approaches dealing with this kind of problem include the IMM and the generalized pseudo-
Bayesian (GPB) method [2]. Both schemes have a structure that consists of a bank of filters for the state and a filter
for the parameters. A GPB algorithm of order l (GBPl) needs r l filters in the bank, where r is the number of models
and l is the number of sampling periods over which model jumps are considered. The IMM has mixtures of the
previous model-conditioned estimates as prior knowledge for the current model-conditioned estimate. Although it is
conceptually similar to the GPB2, it requires only r filters, which amounts to the computational load of the GPB1,
and the performance is nearly as good as the GPB2. Due to this property of cost-effectiveness, the IMM becomes the
state-of-the-art approach for the estimation of hybrid systems. Actually, it has been applied successfully to a number
of problems, especially in maneuvering target tracking and automatic track formation. The reader may refer to [2] for
a detailed description of the algorithm.
When we design a centralized tracking system, we need a global filter that is updated with observations from
multiple sensors. This also poses some subtle issues, including that the measurements are out of sequence, and does
more data mean better performance? In this paper, our focus is on the latter, and we are interested in seeing how the
IMM performs with multiple sensors.
The traditional way to evaluate the performance of the IMM is the Monte Carlo simulation method. In [4], a
non-simulation alternative is proposed. The authors introduce a system model sequence as an essential description of
the scenario and take the expectation of estimation error conditioned on the entire true model history of the system.
This scenario-dependent performance measure is of a hybrid nature, in the sense that it is a continuous-valued matrix
function of a discrete-valued random sequence — the system mode sequence. The performance measure is calculated
in an off-line recursion. This conditional expectation method is ideal for the performance prediction of the IMM,
since the randomness of the error covariance due to the uncertainties in the continuous subspace is averaged out,
whereas the essential information concerning the scenario is retained. Since our goal is to compare a variety of IMM
implementations, we obtain the performance of the IMMs through both this non-simulation method and the Monte-
Carlo simulations. In the sequel, we will study the IMM through two illustrative examples and, to keep our research
focused, we assume a two-sensor system with a constant sampling rate.

3. Illustrative examples

In this section, we consider two examples to investigate how the IMM works with multiple sensors. In the first one,
the target performs a “manipulated maneuver” — the maneuver is created in simulation by significantly increasing the
process noise covariance. This amounts to a mathematically ideal scenario, in the sense that the maneuver model that
the IMM uses can perfectly match the true maneuver model of the target. The second example is a more realistic one:
the target performs a coordinated turn. The IMM needs to select a suitable mode and parameters to model the target
motion.
334 Y. Ruan, L. Hong / Mathematical and Computer Modelling 44 (2006) 332–341

3.1. Maneuver: Process noise level inflation

The target motion is described by a nearly constant velocity model and the surveillance system includes two
tracking radars and a fusion center. Specifically, the model equations can be written as:
x(k + 1) = F x(k) + v(k)
zi (k) = Hi x(k) + wi (k), i = 1, 2 (1)
for k = 1, 2, . . . , N . Here, as usual, x(k) represents the trajectory of the target at time tk , and zi (k) its corresponding
observation from sensor i; the matrices F and Hi are known, and are assumed to represent an observable and
controllable system. The random sequences {v(k), wi (k)} are assumed to be white, zero-mean, Gaussian, and mutually
independent, with E{v(k)vT (k)} = Q(k), E{wi (k)(wi (k))T } = Ri (k).
Suppose that the time interval between two consecutive sampling points of a sensor is ∆t ; the plant equation in (1) is
T T
defined in Cartesian coordinates with x(k) ≡ x(k) ẋ(k) y(k) ẏ(k) and observations zi (k) ≡ x i (k) y i (k)
 

are positions measured by the ith sensor in Cartesian coordinates. Thus, we can further specify the model
as:
 ∆3 ∆2t 
t
qx qx 0 0
 3 2 
 2 
   ∆t
1 ∆t 0 0

 qx ∆t qx 0 0 
0 1 0 0   2
 
F= , Q = (2)

3 2
 
0 0 1 ∆t 
 ∆t ∆t 
 0 0 qy qy 
 
0 0 0 1  3 2 
 
 ∆2t 
0 0 q y ∆t q y
2
where qx and q y are the process noise covariance factors.
The measurement model is given by

σmi
   2 
1 0 0 0 0
i
H = , R = i
2 , i = 1, 2. (3)
0 0 1 0 0 σmi
In the tracking system under consideration, there are two sensors that scan the surveillance region and transmit
reports to the central processor with a fixed rate 1/T . The scan time for Sensor 2 is shifted by t = T /2 from that of
Sensor 1. The report noise arising from the two sensors may be different, and their standard deviations are denoted by
σm1 and σm2 for Sensor 1 and Sensor 2, respectively.
The parameters of the scenario are defined as:
• Sampling time interval ∆t = T = 10 s for each sensor.
• For the most time the process noise covariance factor is qx = q y = (0.3)2 ; between time 200 s and 300 s (or
from the 20th scan to 30th scan for each sensor), the process noise variance is increased by a factor of 252 , i.e.,
qx = q y = (7.5)2 during this interval. This results in a significant change in direction and in speed, which
corresponds to the maneuvering effect on the target during this period of time.
• Measurement noise standard deviations are σm1 = 50 and σm2 = η σm1 , where η varies.
 T
• Initial target state x0 = 0 0 0 120 .
• Probability of detection p D = 1.0.
• No false alarms.
In the following, we take the mean squared error (MSE) of the estimate as the measure of tracking accuracy.
On each parameter setting, we first conduct Monte Carlo simulations, and then the hybrid conditional average
method. The analytic performance prediction provides a confirmation of the Monte Carlo results. In addition to
the MSE, we also show the filter calculated covariance, in that it reflects the confidence that the estimator has
regarding the estimate accuracy, and a consistent filter should have a covariance genuinely reflecting the estimation
error.
 Y. Ruan, L. Hong / Mathematical and Computer Modelling 44 (2006) 332–341 335

Fig. 1. The illustration of the two IMM schemes.

Let us consider two schemes: in Scheme 1, the fusion center processes the measurements from both Sensor 1 and
Sensor 2 in the order of their arrival time; and in Scheme 2, the fusion center processes measurements from Sensor
1 only. The timings of measurements for the two schemes are illustrated in Fig. 1. We can see that the measurement
scan rates are 1/t for Scheme 1 and 1/(2 t) for Scheme 2, where t = T /2, or the actual sampling interval is ∆t1 = t
for Scheme 1 and ∆t2 = 2 t for Scheme 2.
In the IMM, the evolution of the hybrid system among models is modeled as a first order Markov chain. The
transition probability matrix of this chain is designed on the basis of the expected sojourn time in each model of the
hybrid system [1], i.e.,
∆t
τi = (4)
1 − pii
where τi is the expected sojourn time of model i, pii is the probability of transition from model i to itself, and ∆t is
the sampling interval. Thus
∆t
pii = 1 − . (5)
τi
The above equation is valid for small values of ∆t /τi . The transition probabilities pi j for i 6= j are selected using the
identity
X
pi j = 1 − pii (6)
j6=i

and “apportioning” the probability mass 1 − pii to the various possible jumps according to the designer’s intuition.
Considering the given scenario, we choose, respectively, the following transition probability matrices for Scheme 1
and Scheme 2:
 
0.98 0.02
p1 = ,
0.05 0.95
 
0.95 0.05
2
p = . (7)
0.1 0.9
Intuitively, Scheme 1 should yield the improved estimation accuracy or at least undegraded performance compared
with Scheme 2, since in Scheme 1 more information is fed into the processor.
Consider the following situations. We fix σm1 and vary σm2 such that σm2 = η σm1 , η = 1, 3, 5, . . . , 20. This
amounts to a change in the measurement accuracy for Sensor 2. We then inspect the performance of Schemes 1 and
2. The results are shown in Figs. 2–6, where results from both the Monte Carlo simulations and the performance
prediction method are shown. The solid curves are for the performance prediction method and the dashed curves are
from the Monte Carlo simulations (200 runs).
Figs. 2 and 3 show, respectively, the MSE of the estimate and the filter calculated covariance when σm2 = σm1 .
These results show that the filters in both schemes are consistent and Scheme 1 significantly outperforms Scheme 2,
336 Y. Ruan, L. Hong / Mathematical and Computer Modelling 44 (2006) 332–341

Fig. 2. The position MSEs of the IMMs with Scheme 1 and Scheme 2. (Pd = 100%, q1 = 0.32 , q2 = 7.52 , σm1 = 50, σm2 = 50, T = 10.)

Fig. 3. The position covariances of the IMMs with Scheme 1 and Scheme 2. (Pd = 100%, q1 = 0.32 , q2 = 7.52 , σm1 = 50, σm2 = 50, T = 10.)

as expected. In Scheme 1, the updating rate of the IMM is twice as much as that of Scheme 2, while the measurement
accuracies of two sensors are the same. In this case, more data does mean a better estimate.
As another example, we consider σm2 = 9 σm1 , and Figs. 4 and 5 show, respectively, the MSE of the estimate
and the filter calculated covariance. We can see that Scheme 1 is slightly worse than Scheme 2. The situation in this
experiment is the same as in the previous one, except that the measurements from Sensor 2 have significantly lower
accuracy. In this case, more data yields a worse estimate.
To provide a complete picture, we vary the value of η and present the difference in MSE between the two schemes
in Fig. 6. It indicates that Scheme 1 outperforms Scheme 2 only when Sensor 2, the source of the additional data, has
a relatively low level of measurement noise — in this scenario, σm2 < 5 σm1 ; when σm2 > 5 σm1 , Scheme 1 is worse
than Scheme 2. This is a rather surprising result, since it is against common sense that more data, or information,
means better estimate.
 Y. Ruan, L. Hong / Mathematical and Computer Modelling 44 (2006) 332–341 337

Fig. 4. The position MSEs of the IMMs with Scheme 1 and Scheme 2. (Pd = 100%, q1 = 0.32 , q2 = 7.52 , σm1 = 50, σm2 = 450, T = 10.)

Fig. 5. The position covariances of the IMMs with Scheme 1 and Scheme 2. (Pd = 100%, q1 = 0.32 , q2 = 7.52 , σm1 = 50, σm2 = 450, T = 10.)

3.2. Maneuver: Coordinated turn

The parameters of the models are identical to the previous example except that: between scans 20 and 30, the target
performs a coordinated turn maneuver with the turn rate ωT = 18◦ , which results in a change of 180◦ in the course
of the target during the period of maneuvering. This is a more realistic maneuver, and we still use the inflated process
noise covariance, as in Model 2 of the IMM algorithm.
Since the motion of the target is nonlinear and its performance is measurement dependent, the previous prediction
method may not apply. Thus in this example we only show the results obtained from the Monte Carlo simulations
(200 runs). The results are given in Figs. 7 and 8 for σm2 = σm1 , and in Figs. 9 and 10 for σm2 = 9 σm1 .
These results are quite similar to those in the previous example. It reinforces the conclusion that more data may be
worse in this realistic scenario.
338 Y. Ruan, L. Hong / Mathematical and Computer Modelling 44 (2006) 332–341

Fig. 6. The difference of position MSEs of the IMMs with Scheme 1 vs. Scheme 2. (Pd = 100%, q1 = 0.32 , q2 = 7.52 , σm1 = 50, σm2 = 50,
T = 10.)

Fig. 7. The position MSEs of the IMMs with Scheme 1 and Scheme 2. (The target performs the coordinated turn maneuver.) (Pd = 100%,
q1 = 0.32 , q2 = 7.52 , σm1 = 50, σm2 = 50, T = 10.)

3.3. Explanation

In our view, the essential cause of the difference is obvious: the suboptimality of the IMM. As a contrast, it can be
shown that Scheme 1 is always no worse than Scheme 2 if they both use the Kalman filter, the optimal algorithm, to
track a single mode target. What we are really interested here is: why, under certain conditions, is Scheme 1 better
than Scheme 2? Scrutinizing the two schemes, we find that the main difference lies in the observation rate, or updating
rate: Scheme 1 updates the state every t seconds; Scheme 2 every T = 2 t seconds.
As we already knew, the IMM is conceptually similar to the GPB2 and is therefore a suboptimal algorithm, since
its memory of model history is two scans, i.e., only the models in the last two scans are taken into account and model
 Y. Ruan, L. Hong / Mathematical and Computer Modelling 44 (2006) 332–341 339

Fig. 8. The position covariances of the IMMs with Scheme 1 and Scheme 2. (The target performs the coordinated turn maneuver.) (Pd = 100%,
q1 = 0.32 , q2 = 7.52 , σm1 = 50, σm2 = 50, T = 10.)

Fig. 9. The position MSEs of the IMMs with Scheme 1 and Scheme 2. (The target performs the coordinated turn maneuver.) (Pd = 100%,
q1 = 0.32 , q2 = 7.52 , σm1 = 50, σm2 = 450, T = 10.)

histories before that are disregarded. As shown in Fig. 11, the two schemes exhibit different time spans over which
the IMM “remembers” the mode history.
Suppose that we intend to estimate the state at time t0 : in Scheme 1, all the models at time t0 and t0 − T are
exploited by the IMM; in Scheme 2, however, the models being considered are at time t0 and t0 − 2T . Therefore,
Scheme 2 has a larger depth in the model history than Scheme 1. In Scheme 1, when the additional measurements
(those from Sensor 2) are relatively less accurate, less information about the true model can be extracted from these
measurements and the IMM tends to approach the GPB1. As an extreme example, suppose that the standard deviation
of the measurement noise σm2 is infinite; the IMM in Scheme 1 is the GPB1. This could explain its inferiority to
Scheme 2, which is literally a GPB2 regardless of the data quality of Sensor 2.
340 Y. Ruan, L. Hong / Mathematical and Computer Modelling 44 (2006) 332–341

Fig. 10. The position covariances of the IMMs with Scheme 1 and Scheme 2. (The target performs the coordinated turn maneuver.) (Pd = 100%,
q1 = 0.32 , q2 = 7.52 , σm1 = 50, σm2 = 450, T = 10.)

Fig. 11. Illustration of the two IMM performance.

4. Concluding remarks

The IMM estimator is a suboptimal hybrid filter that was shown to achieve an excellent compromise between
performance and complexity. Its complexity is nearly linear (with a small quadratic component) in the size of the
problem (number of models) while its performance is almost the same as that of the GPB2. To further improve
tracking performance, multiple sensors can be used to provide more information about the target. However, when the
IMM is used with multiple sensors, a subtle issue involved in the efficient use of the measurements is that more may
not be better.
In this paper, we have examined two examples of the two sensor systems that perform a tracking task with a constant
measurement rate and where the fusion center processes the measurements obtained from each sensor. Through
both Monte Carlo simulation and the analytic method, we have shown that the IMM can gain some improvement
in accuracy by processing additional data only if those data are accurate enough. If the noise that corrupts the data
 Y. Ruan, L. Hong / Mathematical and Computer Modelling 44 (2006) 332–341 341

exceeds a certain level, the IMM performance could degrade by processing this data. We have also attempted to
explain the mechanism that lies behind this anti-intuitive phenomenon.
Thus, to implement the IMM in a multiple sensor system, one has to decide which data can be used to update the
estimate. A general rule is to choose those with better or at least similar accuracy, and discard those with much worse
accuracy. The idea can also be extended to any IMM design; to abandon measurements with poor quality may be a
better way for estimation than to process the measurements.
Our study is based on systems with linear observations. Issues such as nonlinearity and/or data association could
certainly complicate the problems, and additional data from other sensors will likely be more favorable, since the
information they carry includes not only target positions but also those eagerly needed in report-target association and
nonlinearity estimation.

Acknowledgment

This work is supported in part by the Air Force Research Laboratory, Sensor Directorate, ATR Division, Wright
Patterson Air Force Base.

References

[1] Y. Bar-Shalom, X.R. Li, Multitarget-Multisensor Tracking: Principles and Techniques, YBS Publishing, 1995.
[2] Y. Bar Shalom, X.R. Li, Estimation and Tracking: Principles, Techniques and Software, Artech House, Inc., 1993.
[3] W.D. Blair, Y. Bar Shalom, Tracking maneuvering targets with multiple sensors: does more data always mean better estimates? IEEE
Transactions on Aerospace and Electronic Systems 32 (1) (1996) 450–456.
[4] X.R. Li, Y. Bar Shalom, Performance prediction of the interacting multiple model, IEEE Transactions on Aerospace and Electronic Systems
29 (3) (1993) 755–771.
[5] F. Palmieri, S. Marano, P. Willett, Measurement fusion for target tracking under bandwidth constraint, in: Proceedings of the 2001 IEEE
Aerospace Conference, Big Sky MT, March 2001.
[6] Y. Ruan, P. Willett, A. Marrs, Fusion of quantized measurements via particle filtering, in: Proceedings of the 2003 Aerospace Conference, Big
Sky MT, March 2003.
[7] G.A. Watson, W.D. Blair, Tracking maneuvering targets with multiple sensors using the interacting multiple model algorithm, in: Proceedings
of Signal and Data Processing for Small Targets 1993, SPIE Orlando’92, Orlando, FL, April 1992.