Hindawi Publishing Corporation

Mathematical Problems in Engineering

Volume 2013, Article ID 970158, 8 pages
http://dx.doi.org/10.1155/2013/970158

Research Article
IMM Iterated Extended H∞ Particle Filter Algorithm

Yang Wan, Shouyong Wang, and Xing Qin

Air Force Early Warning Academy, Wuhan 430019, China

Correspondence should be addressed to Yang Wan; [email protected]

Received 2 March 2013; Accepted 28 March 2013

Academic Editor: Xiaojie Su

Copyright © 2013 Yang Wan et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

In order to solve the tracking problem of radar maneuvering target in nonlinear system model and non-Gaussian noise background,
this paper puts forward one interacting multiple model (IMM) iterated extended H∞ particle filter algorithm (IMM-IEHPF). The
algorithm makes use of multiple modes to model the target motion form to track any maneuvering target and each mode uses
iterated extended H∞ particle filter (IEHPF) to deal with the state estimation problem of nonlinear non-Gaussian system. IEHPF is
an improved particle filter algorithm, which utilizes iterated extended H∞ filter (IEHF) to obtain the mean value and covariance of
each particle and describes importance density function as a combination of Gaussian distribution. Then according to the function,
draw particles to approximate the state posteriori density of each mode. Due to the high filter accuracy of IEHF and the adaptation
of system noise with arbitrary distribution as well as strong robustness, the importance density function generated by this method
is more approximate to the true sate posteriori density. Finally, a numerical example is included to illustrate the effectiveness of the
proposed methods.

1. Introduction When the mode probability increases again, particles need

to be reconfirmed, which certainly will result in a signif-
In the field of manoeuvring target tracking, IMM [1, 2] icant increase of estimation errors in certain period. This
algorithm is a popular method. It uses multiple modes to phenomenon is known to cause numerical problems and is
model the target motion form to track any maneuvering also discussed in [14]. To overcome this problem, Boers and
target and modes transition governed by a first-order Markov Driessen [15] combine IMM with particle filter (IMM-PF)
chain. However, in every mode a Kalman filter is running that and there is an interaction between modes. Each mode uses
the performance is deteriorated by nonlinearities and non- a fixed number of particles and runs a regularised particle
Gaussian. Although EKF [3, 4] and UKF [5] can deal with filter. The state posteriori density of the system is a weighted
the nonlinear problem, they cannot avoid the limitation of sum of posteriori density of each mode. Subsequently, the
Gaussian hypotheses. approach has been extensively used in variety manoeuvring
Recently, particle filter (PF) [6] has been paid attention target tracking scenarios. For example, in [16, 17], IMM-PF
to due to its capacity of dealing with the state estimation was applied to the problem of tracking ground target, and
problem in any nonlinear non-Gaussian system and it has in [18, 19], it was applied to solve the maneuvering target
been widely applied [7–10]. It represents the required state tracking problem in non-Gaussian noise background. The
posteriori density by a stochastic particles (or samples) multirate IMM-PF developed in [20] placed emphasis on
with associated weights and to compute estimates based computation savings, which enabled the particles in each
on these particles and weights. Particle filter methods for model to be updated at a different rate. An IMM Gaussian
Markovian switching systems have also been proposed in [11– particle filter was proposed in [21], which made particle filter
13]. However, a major drawback of these methods is that able to concurrently process and saved operation time.
the number of particles in each mode is proportional to However, in IMM-PF, the particles in each mode are
the mode probability. If a mode probability is close to 0, drawn according to the prior density without any consider-
the particles in that mode will be lost in a large number. ation of current measurements, which have a large deviation
2 Mathematical Problems in Engineering

with the particles generated from the true posteriori density. vector is x(𝑘) = [ 𝑥(𝑘) 𝑥(𝑘)̇ ̇ ]T , where 𝑥(𝑘), 𝑦(𝑘), and
𝑦(𝑘) 𝑦(𝑘)
Particularly in the case of the measurement model with ̇
𝑥(𝑘), ̇
𝑦(𝑘) denote the target position, velocity in Cartesian
high accuracy or target maneuvering, this issue will be coordinates. T represents transpose. z(𝑘) is the measurements
more serious. To import the current measurements into at time 𝑘. ℎ(⋅) is the nonlinear measurement function in mode
the particle sampling process, an EKPF and an UPF were 𝑟(𝑘). The process noise k(𝑘) and the measurement noise w(𝑘)
applied in each mode of the IMM algorithm presented in are mutually independent. The mode transition of the system
[20] and in [22], respectively. They used EKF or UKF to is modelled by a Markov chain with
construct an importance density function for approximating
the posteriori density for the state in each mode, during
Pr {𝑟 (𝑘) = 𝑗 | 𝑟 (𝑘 − 1) = 𝑖} = 𝜋𝑖𝑗 , 𝑖, 𝑗 ∈ {1, 2, . . . , 𝑈} , (2)
which the measurements were utilized. And then the particles
were drawn from the function. Nevertheless, due to the
limitation of Gaussian hypothesis, their filter performance where 𝜋𝑖𝑗 is the transfer probability and 𝑈 is the number of
has no obvious improvement in non-Gaussian condition. modes in the IMM.
On the other hand, over the past decades, considerable The state estimation problem of system described in
attention has been devoted to H∞ filtering, and many impor- (1)∼(2) solved by IMM can be divided into four parts,
tant results have been reported in [23–34], and references which are interaction, filtering, mode probability update, and
therein. One of its main advantages is the fact that it is combination. Among them, the combination is a weighted
insensitive to the exact knowledge of the statistics of the sum of filtering estimation and mode probability of each
noise signals. The idea of substituting an extended H∞ filter mode. Therefore, filtering is the key element. Take model
(EHF) for EKF or UKF in the PF framework (called PF- 𝑖 as an example and in view of Bayesian estimation, the
EHF) was first proposed in [24]. However, direct utilization filtering estimation of this mode is to obtain the posteriori
of EHF can face the same problem of model linearisation density 𝑝(x𝑖 (𝑘) | Z(𝑘)) of state x𝑖 (𝑘) in model 𝑖 given all
error as EKF. To eliminate this error and further enhance measurements Z(𝑘) = {z(1), z(2), . . . , z(𝑘)}. Based on this a
the filter accuracy, this paper uses an iterative technique posteriori density, an estimate of the state is, easily obtained
to improve the Jacobian linearisation of nonlinear system as
on basis of EHF, namely, linearize the dynamical model at
smoothed values and linearize measurement model at filtered
x̂𝑖 (𝑘) = ∫ x𝑖 (𝑘) 𝑝 (x𝑖 (𝑘) | Z (𝑘)) 𝑑x𝑖 (𝑘) . (3)
value, respectively (called IEHF). Meanwhile, the iterative
technique can repeat many times at one step to improve the
degree of linearisation approximation of nonlinear system. So, the final combination is given by
As the importance density function generated by IEHF, the
resulting filter called IEHPF. 𝑈
In this paper, we propose to combine IMM with IEHPF x̂ (𝑘) = ∑𝑢𝑖 (𝑘) x̂𝑖 (𝑘) , (4)
(IMM-IEHPF) for the tracking problem of maneuvering 𝑖=1
target in nonlinear non-Gaussian system. The filter accuracy
of IEHF is higher than that of EKF and UKF, and IEHF is where x̂𝑖 (𝑘) and 𝑢𝑖 (𝑘) represent the state estimation and mode
insensitive to the exact knowledge of the noise processes, probability of mode 𝑖 at time 𝑘, respectively.
therefore the importance density function generated by it is It is well known that Kalman filter can get the optimal
more approximate to the true state posteriori density. state estimation in the linear Gaussian system. However,
The structure of this paper is as follows. Section 2 presents in practice, most cases belong to nonlinear non-Gaussian
the problem formulation of the dynamical system model system, so the analytic expression of posteriori density
for maneuvering target. Section 3 describes the IMM-IEHPF 𝑝(x𝑖 (𝑘) | Z(𝑘)) is difficult to get. Even though we can get it, the
algorithm to be used for the modes. In Section 4, through integral operation involved in (3) is very complex. Therefore,
simulation example, several algorithms are compared and it is urgent to find out an effective filtering method to solve
analyzed. Finally, the conclusion is provided in Section 5. the problem of state estimation of nonlinear non-Gaussian
system.

2. System Model and Problem Description 3. Proposed Algorithm

Consider the following discrete nonlinear time-varying sys- IMM-IEHPF algorithm proposed in this paper, in fact, is
tems for maneuvering target tracking: the combination of IMM algorithm and IEHPF algorithm.
Compared with IMM-PF, IMM-IEHPF has considered the
x (𝑘 + 1) = 𝑓 (x (𝑘) , 𝑟 (𝑘)) + 𝑔 (x (𝑘) , 𝑟 (𝑘)) k (𝑘) , 𝑘 ∈ N, latest measurements in the particle sampling process by
utilizing the prediction and updating mechanism of IEHF.
z (𝑘) = ℎ (x (𝑘)) + w (𝑘) , 𝑘 ∈ N, Therefore, the distribution of particles is more approximate to
(1) the true state posteriori density and the filter accuracy can be
enhanced. We first briefly describe the H∞ filter [23] ahead
where 𝑓(⋅) and 𝑔(⋅) represent the system function and process of IEHF and then provide the mathematical model and the
noise function in mode 𝑟(𝑘), respectively. The target state concrete implementation steps of IMM-IEHPF algorithm.
Mathematical Problems in Engineering 3

3.1. H∞ Filter. Consider the following state space model of where P(0) = Π0 is a given positive definite matrices and
linear time-varying system: P(𝑘 + 1/𝑘) satisfies the Riccati recursion [24]

x (𝑘) = F (𝑘) x (𝑘 − 1) + G (𝑘) k (𝑘) , 𝑘+1 𝑘

P( ) = F (𝑘) P ( ) FT (𝑘) + G (𝑘) GT (𝑘)
𝑘 𝑘−1
z (𝑘) = H (𝑘) x (𝑘) + w (𝑘) , (5)
𝑘
s (𝑘) = L (𝑘) x (𝑘) 𝑘 = 0, 1, 2, . . . , − F (𝑘) P ( ) [HT (𝑘) IT𝑛 ] (10)
𝑘−1
where F(𝑘) ∈ C𝑛×𝑛 , G(𝑘) ∈ C𝑛×𝑚 , H(𝑘) ∈ C𝑎×𝑛 , and H (𝑘) 𝑘
L(𝑘) ∈ C𝑏×𝑛 are known and the definition of other parameters × R𝑒−1 (𝑘) [ ]P( ) FT (𝑘) ,
I𝑛 𝑘−1
is the same as the above. The statistical properties of v(𝑘) and
w(𝑘) are unknown and mutually independent. The system where
initial state x(0) is unknown. s(𝑘) is the signal we intend to I𝑎 0 𝑘
H (𝑘)
estimate. If we want to directly estimate x(𝑘) then we set R𝑒 (𝑘) = [ ]+[ ]P( ) [HT (𝑘) IT𝑛 ] ,
0 −𝛾2 I𝑛 I𝑛 𝑘−1
L(𝑘) = I𝑛 (I𝑛 represents the 𝑛 × 𝑛 identity matrix). Let x̂(𝑘) =
F(z(0), z(1), . . . , z(𝑘)) denote the estimate of x(𝑘) for a given (11)
set of measurements Z(𝑘). The estimation error e(𝑘) can be I𝑎 represents the 𝑎 × 𝑎 identity matrix.
described as The covariance updating is
e (𝑘) = s (𝑘) − ̂s (𝑘) = x (𝑘) − x̂ (𝑘) , (6) 𝑘 𝑘
P (𝑘) = P ( ) − P( ) [HT (𝑘) IT𝑛 ]
and the H∞ norm [23] can be defined as 𝑘−1 𝑘−1
(12)
H (𝑘) 𝑘
𝑘 × R𝑒−1 (𝑘) [ ]P( ).
𝐽= sup (( ∑ eT (𝑝) e (𝑝)) I𝑛 𝑘−1
x(0),w,v∈ℎ2 𝑝=0
H∞ filter gain is given by
−1
× ( (x (0) − x̂ (0)) Π−1 ̂ (0)) 𝑘 𝑘
0 (𝑥 (0) − 𝑥 K = P( ) HT (𝑘) (I𝑎 + H(𝑘)P ( ) HT (𝑘)) .
𝑘−1 𝑘−1
(7) (13)
𝑘
T
+ ∑ v (𝑝) v (𝑝) So, the state updating equation of H∞ filter is given by
𝑝=0

−1
x̂ (𝑘) = F (𝑘) x̂ (𝑘 − 1) + K (z (𝑘) − H (𝑘) F (𝑘) x̂ (𝑘 − 1)) .
𝑘
T (14)
+ ∑ w (𝑝) w (𝑝)) ) ,
𝑝=0 H∞ filter can be extended to nonlinear systems by using
Jacobian linearisation technique. The resulting filter is called
where Π0 is a positive-definite matrix which reflects a priori
EHF. Compared to EKF and UKF, the main advantage of EHF
knowledge as to how close x(0) is to the initial guess x̂(0). The
is that there is no limitation about system noise distribution
H∞ norm can thus be interpreted as the maximum energy
and it treats the system noise as an energy-bounded signal.
gain from the input {Π−1/2
0 (x(0) − x̂(0)), {v(𝑝)}𝑘𝑝=0 , {w(𝑝)}𝑘𝑝=0 } Through adjusting parameter 𝛾, the state estimation error can
to the output {e(𝑝)}𝑘𝑝=0 . However, the closed-form solution be minimized.
of the H∞ optimal estimation problem is available only in
specific cases, so the realization of H∞ optimal filter is a 3.2. IEHF. Although EHF can adapt to the case of system
suboptimal filtering. noise with non-Gaussian distribution, it also will face the
Suboptimal H∞ filtering problem: given a scalar problem of model linearisation error. Therefore, to further
𝛾 > 0, find an H∞ suboptimal estimation strategy enhance the degree of linearisation approximation of non-
F(z(0), z(1), . . . , z(𝑘)) that achieves 𝐽 < 𝛾2 . This clearly linear system, this paper makes improvements on model
requires checking whether linearisation technique and proposes IEHF algorithm. Its
fundamental idea is as below.
𝛾 ≥ inf 𝐽. (8) Given the filtered estimation x̂(𝑘 − 1) at time 𝑘 − 1. Take
the general nonlinear system described in (1) as an example,
The solution of the optimal H∞ filtering problem can be linearize the dynamical model at x̂(𝑘 − 1), and retain first-
obtained by iterating on 𝛾 of the suboptimal H∞ problem. order small quantity (for simplicity, omit model variable 𝑟(𝑘))
For a given 𝛾 > 0, if the [F(𝑘), G(𝑘)] is full rank, then an
estimator that achieves 𝐽 < 𝛾2 exists if, and only if [24], 𝜕𝑓 (̂x (𝑘 − 1))
x (𝑘) ≈ 𝑓 (̂x (𝑘 − 1)) +
𝜕̂x (𝑘 − 1) (15)
𝑘+1
P−1 ( ) + HT (𝑘 + 1) H (𝑘 + 1) − 𝛾−2 I𝑛 > 0, (9)
𝑘 × (x (𝑘 − 1) − x̂ (𝑘 − 1)) + v (𝑘) .
4 Mathematical Problems in Engineering

𝑘
Similarly, linearize measurement model at state predictive P𝑐 ( ) = f 𝑐−1 P𝑑 (𝑘 − 1) f (𝑐−1)T + Q (𝑘 − 1) ,
value x̂(𝑘/𝑘 − 1) 𝑘−1
𝑘 𝜕ℎ (̂x (𝑘/𝑘 − 1)) 𝑘 𝑘
z (𝑘) ≈ ℎ (̂x ( )) + K𝑐 ( ) = P𝑐 ( ) h(𝑐−1)T
𝑘−1 𝜕̂x (𝑘/𝑘 − 1) 𝑘−1 𝑘−1
(16) −1
𝑘 𝑘
× (x (𝑘) − x̂ ( )) + w (𝑘) . × {h(𝑐−1)T P𝑐 ( ) h(𝑐−1)T + I𝑎 } ,
𝑘−1 𝑘−1
It is not difficult to see that x̂(𝑘 − 1) and x̂(𝑘/𝑘 − 1) have 𝑘 𝑘
P𝑐 (𝑘) = P𝑐 ( ) − P𝑐 ( )
nothing to do with the current measurement z(𝑘). However, 𝑘−1 𝑘−1
the smoothed value x̂(𝑘−1/𝑘) and filtered value x̂(𝑘) obtained
from measurement z(𝑘) are obviously superior to x̂(𝑘 − 1) h𝑐−1 𝑐 𝑘
× [h(𝑐−1)T IT𝑛 ] R𝑒−1 (𝑘) [ ]P ( ),
and x̂(𝑘/𝑘 − 1). If we linearize the dynamical model and I𝑛 𝑘−1
linearize measurement model at the smoothed value and
filtered value, respectively, the approximation degree of the 𝜕𝑓 (̂x𝑐−1 (𝑘 − 1/𝑘)) 𝜕ℎ (̂x𝑐−1 (𝑘))
f 𝑐−1 = , h𝑐−1 = .
model linearisation will be further improved. Therefore, this 𝜕̂x𝑐−1 (𝑘 − 1/𝑘) 𝜕̂x𝑐−1 (𝑘)
paper adopts the linearisation form as follows. (20)
Linearize the dynamical model at the smoothed value
x̂(𝑘 − 1/𝑘) The smoothing value required by 𝑐 + 1 iteration is

𝑘−1 𝜕𝑓 (̂x (𝑘 − 1/𝑘)) 𝑘−1

x (𝑘) ≈ 𝑓 (̂x ( )) + x̂𝑐 ( ) = x̂𝑑 (𝑘 − 1) + P𝑑 (𝑘 − 1) f (𝑐−1)T
𝑘 𝜕̂x (𝑘 − 1/𝑘) 𝑘
(17) −1
𝑘−1 𝑘 𝑘
× (x (𝑘 − 1) − x̂ ( )) + v (𝑘) . × (P𝑐 ( )) (̂x𝑐 (𝑘) − x̂𝑐 ( )) .
𝑘 𝑘−1 𝑘−1
(21)
Linearize the measurement model at the filtered value
x̂(𝑘)
3.3. IMM-IEHPF. On the basis of IMM-PF [15] algorithm,
𝜕ℎ (̂x (𝑘)) this section provides the implementation process of IMM-
z (𝑘) ≈ ℎ (̂x (𝑘)) + (x (𝑘) − x̂ (𝑘)) + w (𝑘) . (18) IEHPF algorithm.
𝜕̂x (𝑘)
In order to further improve the filtering accuracy, the Step 1 (interaction stage). Calculate the mixing probability
linearisation technique can be repeated many times at one
step. All these contribute to IEHF algorithm. However, for 1
many nonlinear filtering problems, the performance is not 𝑢𝑖𝑗 (𝑘 − 1) = 𝜋 𝑢 (𝑘 − 1) ,
𝑐𝑗 𝑖𝑗 𝑖
significantly improved after repeating iterations. Usually it
ends after two or three times of iterations. According to the (22)
𝑈
above thoughts, the implementation steps of IEHF algorithm 𝑐𝑗 = ∑𝜋𝑖𝑗 𝑢𝑖 (𝑘 − 1) ,
are as follows. 𝑖=1
Assume that the iterative filtering value x̂𝑑 (𝑘 − 1) and where 𝑢𝑖𝑗 (𝑘 − 1) represents the mixed probability of the mode
covariance P𝑑 (𝑘 − 1) at time 𝑘 − 1 are known, the iterative at time 𝑘 − 1. The interactive input state x̂𝑜𝑗 (𝑘 − 1) and the
filtering value x̂𝑐 (𝑘) in 𝑐 (1 ≤ 𝑐 ≤ 𝑑) iteration process at time ̂ 𝑜𝑗 (𝑘 − 1) are given by
corresponding covariance P
𝑘 can be solved according to the following expression:
𝑈
𝑐 𝑐 𝑘 x̂𝑜𝑗 (𝑘 − 1) = ∑x̂𝑖 (𝑘 − 1) 𝑢𝑖𝑗 (𝑘 − 1) , 𝑖, 𝑗 = 1, 2, . . . , 𝑈,
x̂ (𝑘) = x̂ ( )
𝑘−1 𝑖=1

𝑈
+ K𝑐 {z (𝑘) − ℎ (̂x𝑐−1 (𝑘)) (19) ̂ 𝑜𝑗 (𝑘 − 1) = ∑ 𝑢𝑖𝑗 (𝑘 − 1)
P
𝑖=1
𝑘
−h𝑐−1 (̂x𝑐 ( ) − x̂𝑐−1 (𝑘))} , ̂ 𝑖 (𝑘 − 1) + (̂x𝑖 (𝑘 − 1) − x̂𝑜𝑗 (𝑘 − 1))
× {P
𝑘−1
T
where × (̂x𝑖 (𝑘 − 1) − x̂𝑜𝑗 (𝑘 − 1)) } ,
𝑘 𝑘−1 (23)
x̂𝑐 ( ) = 𝑓 (̂x𝑐−1 ( ))
𝑘−1 𝑘 ̂ 𝑖 (𝑘 − 1) are the state estimate and
where x̂𝑖 (𝑘 − 1) and P
𝑘−1 covariance estimation in mode-matched filter 𝑖 at time 𝑘 − 1,
+ f 𝑐−1 (̂x𝑑 (𝑘 − 1) − x̂𝑐−1 ( )) , respectively.
𝑘
Mathematical Problems in Engineering 5

Step 2 (filtering stage). (1) According to the Gaussian distri- Mode likelihood function
bution constructed by x̂𝑜𝑗 (𝑘 − 1) and P̂ 𝑜𝑗 (𝑘 − 1), each mode
draws particles randomly, 𝑁
1 𝑠
𝐿 𝑖 (𝑘) = ∑ 𝑁 (𝛿𝑙𝑖 (𝑘) ; 0, S𝑖 (𝑘)) . (32)
̂ 𝑜𝑗 (𝑘 − 1)) ,
x𝑖𝑙 (𝑘 − 1) ∼ 𝑁 (̂x𝑜𝑗 (𝑘 − 1) , P 𝑁𝑠 𝑙=1
(24)
𝑙 = 1, 2, . . . , 𝑁𝑠 , 𝑖 = 1, 2, . . . , 𝑈, Mode probability
where 𝑁𝑠 is the number of particles.
(2) Importance sampling. Use IEHF algorithm 1 𝑈

in Section 3.2 to update every particle x𝑖𝑙 (𝑘 − 1) and 𝑢𝑖 (𝑘) = 𝐿 (𝑘) 𝑐𝑖 , 𝐶 = ∑ 𝐿 𝑖 (𝑘) 𝑐𝑖 . (33)
𝐶 𝑖 𝑖=1
obtain a combination of a bank of Gaussian distribution
̂ 𝑙,𝑑 (𝑘)), 𝑙 = 1, 2, . . . , 𝑁𝑠 , 𝑖 = 1, 2, . . . , 𝑈. Then
𝑁(̂x𝑖𝑙,𝑑 (𝑘), P 𝑖
draw particles according to this distribution Step 4. Combination

̂ 𝑙,𝑑 (𝑘)) ,
x̃𝑖𝑙 (𝑘) ∼ 𝑁 (̂x𝑖𝑙,𝑑 (𝑘) , P 𝑖 = 1, 2, . . . , 𝑁𝑠 . (25) 𝑈
𝑖
x̂ (𝑘) = ∑𝑢𝑖 (𝑘) x̂𝑖 (𝑘) ,
Calculate the particle weights [6] 𝑖=1

𝑈
𝑝 (z (𝑘) | x̃𝑖𝑙 (𝑘)) 𝑝 (̃x𝑖𝑙 | x𝑖𝑙 (𝑘 − 1)) ̂ (𝑘) = ∑ 𝑢𝑖 (𝑘) (P
̂ 𝑖 (𝑘) + (̂x𝑖 (𝑘) − x̂ (𝑘)) (̂x𝑖 (𝑘) − x̂ (𝑘))T ) .
𝑤𝑖𝑙 (𝑘) = , (26) P
̂ 𝑙,𝑑 (𝑘))
𝑁 (̂x𝑖𝑙,𝑑 (𝑘) , P 𝑖=1
𝑖
(34)
where 𝑝(z(𝑘) | x̃𝑖𝑙 (𝑘))
is likelihoods function, | − 𝑝(̃x𝑖𝑙 x𝑖𝑙 (𝑘
𝑙,𝑑 ̂ 𝑙,𝑑
1)) is the priori density, and 𝑁(̂x𝑖 (𝑘), P𝑖 (𝑘)) represents 4. Simulation Results and Analysis
importance density function generated by IEHF.
Normalizing In order to validate the filtering performance of IMM-IEHPF
algorithm, this paper compares standard IMM, IMM-PF, and
𝑤𝑖𝑙 (𝑘) IMM-IEHF algorithm. The experiment scene is designed
̃𝑖𝑙 (𝑘) =
𝑤 𝑁
. (27) as follows. Radar scanning interval is 𝑇 = 1 s and repeat
∑𝑙=1𝑠 𝑤𝑖𝑙 (𝑘)
for 100 times. The target moves in straight line in the
(3) Resampling. According to the normalized weight, former 34 s, then it turns anticlockwise with turning speed
𝑙=1,2,...,𝑁 of 𝜔 = 0.1 rad/s until 67 s and moves in straight line in
resample to get the new particles {x𝑙𝑖 (𝑘), (1/𝑁𝑠 )}𝑖=1,2,...,𝑈𝑠 .
the last 33 s. Adopt the system model in (1)∼(2) and the
(4) State estimation of each mode
target state x(𝑘) = [ 𝑥(𝑘) 𝑥(𝑘)̇ ̇ ]T , which includes the
𝑦(𝑘) 𝑦(𝑘)
𝑁 position (m) and velocity component (m/s) in 𝑥-axis and
1 𝑠 𝑙
x̂𝑖 (𝑘) = ∑ x (𝑘) . (28) 𝑦-axis. Mode 𝑟(𝑘) ∈ {1, 2, 3}, 𝑟(𝑘) = 1 corresponds to the
𝑁𝑠 𝑙=1 𝑖 uniform rectilinear model. The 𝑟(𝑘) = 2 and 𝑟(𝑘) = 3
correspond to the anticlockwise turning mode and clockwise
Covariance estimation turning mode, respectively. With the known initial state
𝑁 x(0) = [ 200 m 15 m/s 150 m 6 m/s ]T , the target state was evoluted
̂ 𝑖 (𝑘) = 1 ∑ (̂x𝑖 (𝑘) − x𝑙 (𝑘)) (̂x𝑖 (𝑘) − x𝑙 (𝑘))T .
𝑠

P 𝑖 𝑖 (29) according to different system functions as follows.

𝑁𝑠 𝑙=1 Mode 1 (CV mode, 𝑟(𝑘) = 1):
Step 3 (update the mode probability). Particle residual of
each mode 1 𝑇 0 0
[0 1 0 0]
𝑓 (x (𝑘) , 1) = [
[0
],
𝛿𝑙𝑖 (𝑘) = z (𝑘) − ℎ (̂x𝑖𝑙 (𝑘)) , (30) 0 1 𝑇]
[0 0 0 1]
where ℎ(̂x𝑖𝑙 (𝑘)) is measurement predicted output.
𝑇2
Covariance of residual [2 0]
[ ]
[ ] (35)
𝑁 [𝑇 0]
1 𝑠 [ ]
S𝑖 (𝑘) =
T
∑ (z (𝑘) − ℎ (̂x𝑖𝑙 (𝑘))) (z𝑖 (𝑘) − ℎ(̂x𝑖𝑙 (𝑘)) , 𝑔 (x (𝑘) , 1) = [ ],
𝑁𝑠 𝑙=1 𝑖 [ 2]
𝑇 ]
[
[0 ]
(31) [ 2]
𝑁 [0 𝑇]
where z𝑖 (𝑘) = (1/𝑁𝑠 ) ∑𝑙=1𝑠 ℎ(̂x𝑖𝑙 (𝑘)) is mean of predicted
output over the sample set. Q1 (𝑘) = diag ([52 22 52 22 ]) .
6 Mathematical Problems in Engineering

Mode 2 (CT mode, 𝑟(𝑘) = 2): 550

𝜔𝑇 (cos 𝜔𝑇 − 1)

𝑦-axis, position (m)

500
[1 sin
𝜔
0
𝜔 ]
[ ] 450
[ ]
[0 cos 𝜔𝑇 0 − sin 𝜔𝑇 ]
[ ] 400
𝑓 (x (𝑘) , 2) = [ ],
[ (1 − cos 𝜔𝑇) sin 𝜔𝑇 ]
[0 1 ] 350
[ ]
[ 𝜔 𝜔 ]
300
[ 0 sin 𝜔𝑇 0 cos 𝜔𝑇 ]
550 600 650 700 750 800
2
𝑇 𝑥-axis, position (m)
[ 0] (36)
[ 2 ] True IMM-PF
[ ]
[ 𝑇 0] IMM IMM-IEHPF
[ ] IMM-IEHF
[ ]
𝑔 (x (𝑘) , 2) = [ 2],
[ 𝑇 ]
[ 0 ] Figure 1: Position estimate.
[ 2]
[ ]
[ 0 𝑇]
[ ]
For observing the state estimation diversity of the four
Q2 (𝑘) = diag ([102 52 102 52 ]) , methods, we have magnified Figure 1.
As seen from Figure 1, these four methods do well in the
where 𝜔 is the turning coefficient.
target state estimation in the two straight-line movements.
Mode 3 is the same as Mode 2; however, in mode 2, 𝜔 > 0
However, in the anticlockwise turning period, the deviation
corresponds to the anticlockwise manoeuvre, and in mode 3,
from the true values obtained from IMM is larger than those
𝜔 < 0 corresponds to the clockwise manoeuvre. The initial
of the other three methods. Among them, IMM-IEHPF has
mode probabilities are set as 𝑢1 (0) = 𝑢2 (0) = 𝑢3 (0) = 1/3,
the best state estimation in the maneuvering stage. We can see
and the mode transition matrix [𝜋𝑖𝑗 ] is given by [15]
that the trajectory of IMM-IEHPF is in accordance with the
0.96 0.02 0.02 true trajectory.
[𝜋𝑖𝑗 ] = [0.02 0.96 0.02] . (37) To compare the filtering accuracy of the previous four
methods, Figure 2 shows the position root mean square error
[0.02 0.02 0.96] of the four methods. From Figure 2, we observe that the IMM
We measure the range and bearing, and the measurement filter diverges after 𝑡 = 34 s until 𝑡 = 67 s. The reasons for
function ℎ(⋅) is mode- and time-independent and is given by this is that in the IMM filter the speed is wrongly estimated.
This is due to the nonlinearity in the measurements in
𝑦(𝑘) T the maneuvering mode. Compared with IMM, the position
ℎ (x (𝑘)) = [√𝑥2 (𝑘) + 𝑦2 (𝑘) arctan ] . (38)
RMSE of IMM-IEHF, and IMM-PF also increase to certain
𝑥(𝑘)
extent, but both are less than that of IMM. IEHF improves
Measurement noise standard deviations are 𝜎𝑟 = 30 m the degree of approximation of the nonlinear system through
for the range measurement, 𝜎𝜃 = 2 mrad for the bearing, and iteration and PF can deal with nonlinear filtering problem
the number of particles is 𝑁𝑠 = 1000 [15]. In IMM-IEHPF by describing posteriori density through a series of particles,
algorithm, parameter 𝛾 = 1.6. Usually, in order to ensure the which achieve better filtering accuracy than EKF. However,
robustness of the algorithm, parameter 𝛾 should be as small IEHF is just a linear approximation for nonlinear system
as possible when meeting the existence conditions of (9). At and improvement of filtering accuracy is limited. PF adopts
present, it can only be got through trial and error. Iteration the prior density as the importance density function with-
times 𝑑 = 3. We carried out Monte Carlo simulation for 𝑀 = out considering the latest measurement information. When
100 times and the position root mean square error (RMSE) at the target maneuvering and the measurement information
time 𝑘 is defined as change greatly, the particles generated from the function may
RMSE (𝑘) deviate from the true posteriori density, resulting in a bad
filtering performance.
1 𝑀 2
The IMM-IEHPF algorithm, however, can deal with
=( ∑ [(𝑥̂mot (𝑘) − 𝑥(𝑘)) target maneuvering and we see that the filter also performs
𝑀 mot=1 (39) well after 𝑡 = 34 s until 𝑡 = 67 s. IEHPF takes advantages
1/2 of IEHF and PF together in itself. In the filtering process,
mot 2 IEHPF not only considers the latest measurements in the
+(𝑦̂ (𝑘) − 𝑦 (𝑘)) ] ) .
particle sampling process but also reduces the influence
of error caused by linearizing measurement model. When
Figure 1 compares the state estimation generated from a these algorithms are directly applied to the IMM framework,
single run of IMM, IMM-IEHF, IMM-PF and IMM-IEHPF. posteriori density of each mode is described much more well
Mathematical Problems in Engineering 7

40 Table 1: Mean and variance of RMSE under different measurement

35 bearing deviations.
Position RMSE (m)

30
Algorithm Bearing standard Mean Variance
25 deviations 𝜎𝜃 (mrad)
20
IMM 2 9.7293 98.2215
15
IMM-IEHF 2 4.0394 5.6625
10
IMM-PF 2 4.1397 2.4750
5
IMM-IEHPF 2 2.7593 0.3633
0
0 10 20 30 40 50 60 70 80 90 100 IMM 5 12.794 159.49
Time (s) IMM-IEHF 5 6.5694 7.5583
IMM IMM-PF IMM-PF 5 5.2413 2.9841
IMM-IEHF IMM-IEHPF IMM-IEHPF 5 5.3559 1.1192
Figure 2: Position RMSE. IMM 10 18.421 333.87
IMM-IEHF 10 8.6552 10.3592
IMM-PF 10 8.3949 8.2179
by IEHPF than EKF, IEHF, and PF. Therefore, the filtering IMM-IEHPF 10 7.7631 3.4886
accuracy of IMM-IEHPF is the highest.
To quantitatively analyse the filtering performances of the
four methods, Table 1 lists the mean value and variance of Acknowledgments
position RMSE of each method in different observed bearing
The authors thank the referees for their careful reading,
standard deviation, which stands for different disturbances
invaluable comments, and suggestions on improving the
of receiver noise and atmospheric. From Table 1, we can see
quality of this paper. This work is supported in part by the
that the mean value and variance of position RMSE of all
National Natural Science Foundation of China under Grant
four methods grow with the increase of the observed bearing
nos. 61179014 and 60872156.
standard deviation 𝜎𝜃 , nevertheless, which of IMM-IEHPF
are less than those of the others (which is in accordance with
the analytical results in Figure 2), especially when the bear- References
ing standard deviation is considerable high, our proposed
[1] H. A. P. Blom and Y. Bar-Shalom, “Interacting multiple model
method still demonstrates higher tracking precision (i.e., 𝜎𝜃 =
algorithm for systems with Markovian switching coefficients,”
10 mrad, taking the mean of RMSE as the metric, it reduces IEEE Transactions on Automatic Control, vol. 33, no. 8, pp. 780–
by 57.86% compared with IMM, and 7.53% compared with 783, 1988.
IMM-PF). [2] E. Mazor, A. Averbuch, Y. Bar-Shalom, and J. Dayan, “Interact-
ing multiple model methods in target tracking: a survey,” IEEE
5. Conclusions Transactions on Aerospace and Electronic Systems, vol. 34, no. 1,
pp. 103–123, 1998.
This paper has investigated the problem of maneuvering [3] Z. Wei and S. X. Yang, “Neural network based extended kalman
target tracking in the nonlinearities of the dynamic state- filter for localization of mobile robots,” in Proceedings of the 8th
space and the non-Gaussian measurement noise. We have World Congress on Intelligent Control and Automation, pp. 937–
presented a new IEHPF method in the IMM framework. The 942, Taipei, Taiwan, 2011.
performance of the proposed method is evaluated via sim- [4] M. H. Bahari, A. Karsaz, and N. Pariz, “High maneuvering
ulations and compared with IMM, IMM-IEHF, and IMM- target tracking using a novel hybrid Kalman filter-fuzzy logic
architecture,” International Journal of Innovative Computing,
PF, which illustrates that the tracking performance of our
Information and Control, vol. 7, no. 2, pp. 501–510, 2011.
proposed method is superior to the others. We have also pro-
[5] W. Zou, Q. X. Du, K. Yuan, and J. F. Li, “An attitude tracking
vided insight into the influence of various bearing standard
method considering acceleration compensation and based on
deviations. Three bearing standard deviations are given for UK,” in Proceedings of the 8th World Congress on Intelligent
comparing the performance of our proposed method with the Control and Automation, pp. 52–57, Taipei, Taiwan, 2011.
other methods. We have found that IMM-IEHPF is weakly [6] A. Doucet, S. Godsill, and C. Andrieu, “On sequential Monte
sensitive to noise levels. The simulation results have well Carlo sampling methods for Bayesian filtering,” Statistics and
verified the validity of the proposed method. Computing, vol. 10, no. 3, pp. 197–208, 2000.
This paper only considers a maneuvering target appearing [7] J. Sherrah, B. Ristic, and N. J. Redding, “Particle filter to track
in the scanned region of radar at whole time step. However, multiple people for visual surveillance,” IET Computer Vision,
in actual radar target tracking, target may disappear at any vol. 5, no. 4, pp. 192–200, 2011.
time step. Therefore in future work, we will study how to [8] A. D. Bimbo and F. Dini, “Particle filter-based visual tracking
construct a suitable model to describe target appearance or with a first order dynamic model and uncertainty adaptation,”
disappearance in maneuvering target tracking under a low Computer Vision and Image Understanding, vol. 115, no. 6, pp.
SNR environment. 771–786, 2011.
8 Mathematical Problems in Engineering

[9] S. Aihara, A. Bagchi, and E. Imreizeeq, “Identification of Transactions on Signal Processing, vol. 54, no. 11, pp. 4471–4475,
electricity spot models by using convolution particle filter,” 2006.
International Journal of Innovative Computing, Information and [26] L. Wu and D. W. C. Ho, “Fuzzy filter design for Itô stochastics
Control, vol. 7, no. 1, pp. 61–72, 2011. systems with application to sensor fault detection,” IEEE Trans-
[10] M. H. Refan, B. Sareh, and B. Mahdi, “A particle filtering- actions on Fuzzy Systems, vol. 17, no. 1, pp. 233–242, 2009.
based framework for on-line fault diagnosis in hybrid systems,” [27] L. Wu, P. Shi, H. Gao, and C. Wang, “𝐻∞ filtering for 2D
International Journal of Innovative Computing, Information and Markovian jump systems,” Automatica, vol. 44, no. 7, pp. 1849–
Control, vol. 6, no. 8, pp. 3669–3680, 2010. 1858, 2008.
[11] S. McGinnity and G. W. Irwin, “Multiple model bootstrap [28] L. Wu, P. Shi, H. Gao, and C. Wang, “Delay-dependent robust
filter for maneuvering target tracking,” IEEE Transactions on 𝐻∞ and 𝐿 2 -𝐿 ∞ filtering for LPV systems with both discrete
Aerospace and Electronic Systems, vol. 36, no. 3, pp. 1006–1012, and distributed delays,” IEE Proceedings: Control Theory &
2000. Applications, vol. 153, no. 4, pp. 483–492, 2006.
[12] Y. Zhai, M. B. Yeary, S. Cheng, and N. Kehtarnavaz, “An object- [29] L. Wu, Z. Wang, H. Gao, and C. Wang, “𝐻∞ and 𝐿 2 -𝐿 ∞ fil-
tracking algorithm based on multiple-model particle filtering tering for two-dimensional linear parameter-varying systems,”
with state partitioning,” IEEE Transactions on Instrumentation International Journal of Robust and Nonlinear Control, vol. 17,
and Measurement, vol. 58, no. 5, pp. 1797–1809, 2009. no. 12, pp. 1129–1154, 2007.
[13] T. Sathyan and M. Hedley, “Efficient particle filtering for [30] R. Yang, P. Shi, and G.-P. Liu, “Filtering for discrete-time
tracking maneuvering objects,” in Proceedings of the IEEE/ION networked nonlinear systems with mixed random delays and
Position, Location and Navigation Symposium, pp. 332–339, packet dropouts,” IEEE Transactions on Automatic Control, vol.
Indian Wells, Calif, USA, May 2010. 56, no. 11, pp. 2655–2660, 2011.
[14] N. Gordon, S. Makell, and T. Kirubarajan, “Efficient particle [31] R. Yang, H. Gao, and P. Shi, “Delay-dependent 𝐿 2 -𝐿 ∞ filter
filters for joint tracking and classification,” in Signal and Data design for stochastic time-delay systems,” IET Control Theory
Processing of Small Targets, vol. 4728 of Proceedings of SPIE, pp. & Applications, vol. 5, no. 1, pp. 1–8, 2011.
439–449, 2002. [32] F. B. Li and X. Zhang, “Delay-range-dependent robust 𝐻∞
[15] Y. Boers and J. N. Driessen, “Interacting multiple model particle filtering for singular LPV systems with time variant delay,”
filter,” IEE Proceedings: Radar, Sonar and Navigation, vol. 150, International Journal of Innovative Computing, Information and
no. 5, pp. 344–349, 2003. Control, vol. 9, no. 1, pp. 339–353, 2013.
[16] R. H. Guo, Z. Qin, X. N. Li, and J. L. Chen, “Interacting [33] X. J. Su, P. Shi, L. G. Wu, and Y. D. Song, “A novel approach
multiple model particle-type filtering approaches to ground to filter design for T-S fuzzy discrete-time systems with time-
target tracking,” Journal of Computers, vol. 3, no. 7, pp. 23–30, varying delay,” IEEE Transactions on Fuzzy Systems, vol. 20, no.
2008. 6, pp. 1114–1129, 2012.
[17] S. Maresca, M. Greco, F. Gini, and L. Verrazzani, “Radar [34] M. S. E. Abadi, V. Mehrdad, and M. Norouzi, “A family of
tracking of a maneuvering ground vehicle using an airborne set-membership affine projection adaptive filter algorithms,”
sensor,” in Proceedings of the International Radar Conference International Journal of Innovative Computing, Information and
Surveillance for a Safer World, pp. 1–6, Bordeaux, France, Control, vol. 8, no. 2, pp. 1313–1326, 2012.
December 2009.
[18] H. W. Li and J. Wang, “Particle filter for manoeuvring target
tracking via passive radar measurements with glint noise,” IET
Radar, Sonar & Navigation, vol. 6, no. 3, pp. 180–189, 2012.
[19] M. R. Morelande and S. Challa, “Manoeuvring target tracking
in clutter using particle filters,” IEEE Transactions on Aerospace
and Electronic Systems, vol. 41, no. 1, pp. 252–270, 2005.
[20] L. Hong, N. Cui, M. Bakich, and J. R. Layne, “Multirate interact-
ing multiple model particle filter for terrain-based ground target
tracking,” IEE Proceedings: Control Theory and Applications, vol.
153, no. 6, pp. 721–731, 2006.
[21] Z. G. Liu and J. K. Wang, “Interacting multiple model gaussian
particle filter,” in Proceedings of the 8th World Congress on
Intelligent Control and Automation, pp. 270–273, Taipei, Taiwan,
2011.
[22] X. L. Deng, J. Y. Xie, and H. W. Ni, “Interacting multiple model
algorithm with the unscented particle filter (UPF),” Chinese
Journal of Aeronautics, vol. 18, no. 4, pp. 366–371, 2005.
[23] B. Hassibi, A. H. Sayed, and T. Kailath, “Linear estimation in
Kreı̆n spaces. II. Applications,” IEEE Transactions on Automatic
Control, vol. 41, no. 1, pp. 34–49, 1996.
[24] K. Nishiyama, “Fast and effective generation of the proposal
distribution for particle filters,” Signal Processing, vol. 85, no. 12,
pp. 2412–2417, 2005.
[25] J. Seo, M. J. Yu, C. G. Park, and J. G. Lee, “An extended robust
𝐻∞ filter for nonlinear constrainted uncertain systems,” IEEE
Advances in Advances in Journal of Journal of
Operations Research
Hindawi Publishing Corporation
Decision Sciences
Hindawi Publishing Corporation
Applied Mathematics
Hindawi Publishing Corporation
Algebra
Hindawi Publishing Corporation
Probability and Statistics
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

The Scientific International Journal of

World Journal
Hindawi Publishing Corporation
Differential Equations
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Submit your manuscripts at

http://www.hindawi.com

International Journal of Advances in

Combinatorics
Hindawi Publishing Corporation
Mathematical Physics
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

Journal of Journal of Mathematical Problems Abstract and Discrete Dynamics in

Complex Analysis
Hindawi Publishing Corporation
Mathematics
Hindawi Publishing Corporation
in Engineering
Hindawi Publishing Corporation
Applied Analysis
Hindawi Publishing Corporation
Nature and Society
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014

International
Journal of Journal of
Mathematics and
Mathematical
Discrete Mathematics
Sciences

Journal of International Journal of Journal of

Hindawi Publishing Corporation Hindawi Publishing Corporation Volume 2014

Function Spaces
Hindawi Publishing Corporation
Stochastic Analysis
Hindawi Publishing Corporation
Optimization
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014 http://www.hindawi.com http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014 http://www.hindawi.com Volume 2014