www.ietdl.

org
Published in IET Radar, Sonar and Navigation
Received on 2nd December 2013
Revised on 16th May 2014
Accepted on 2nd June 2014
doi: 10.1049/iet-rsn.2014.0142

ISSN 1751-8784

Switched Kalman filter-interacting multiple model

algorithm based on optimal autoregressive model for
manoeuvring target tracking
Biao Jin, Bo Jiu, Tao Su, Hongwei Liu, Gaofeng Liu
National Laboratory of Radar Signal Processing, Xidian University, 710071 Shaanxi, People’s Republic of China
E-mail: [email protected]

Abstract: A manoeuvring target tracking algorithm based on the autoregressive (AR) model is proposed. First, the AR model is
incorporated into the Kalman filter (KF) for target tracking. The closed-form solution of the AR model coefficients is obtained by
minimising the mean-square tracking error, and subject to the polynomial constraint of target motion. Then, based on the AR
model, the proposed algorithm is constructed by combining the KF with the interacting multiple model (IMM) filter, coupled
with the proposed detection schemes for manoeuvre occurrence and termination, as well as for switching initialisation.
Simulations are performed to demonstrate the effectiveness of the AR model, and the proposed algorithm is compared with
the IMM filter and variable-dimension filter in the manoeuvring scenario.

1 Introduction one is the interacting multiple model (IMM) algorithm [8–

11, 13–17]. The IMM filter consists of a first-degree model
The kinematic state estimation of a non-cooperative target is a for the target non-manoeuvring motion and one or two
challenging problem in practice, since the prior information of second-degree models with different process noise levels
target motion is not available. Maybe the target is in the for the manoeuvring motions. It utilises multiple different
quiescent mode, or turns to a manoeuvring motion at models to estimate the state of single target and combines
unknown time. Many adaptive algorithms for tracking a the estimates of all the models with non-zero model
manoeuvring target have been considered in the literature. probability. This procedure can be efficient in tracking a
Therein two main classes of techniques are researched: manoeuvring target, but may result in some performance
serial model methods [1–6] and parallel model methods [7– degradation because of over-modelling in the
17]. The serial model methods are also called non-manoeuvring case. Generally, all the current tracking
decision-based methods [1], such as the variable-dimension algorithms utilise the discrete-time differential models, such
(VD) filter [2, 3] and the input estimation (IE) filter [4–6]. as CV and CA models [18]. These models have the only
The VD filter [2] is a typical decision-based method, which function of predicting the target motion, rather than of
changes the target state model according to whether the reducing the noise, which has effect on the tracking
manoeuvre is detected. Switching from the performance to some extent.
constant-velocity (CV) model to the constant-acceleration Aiming at the issues mentioned above, a tracking algorithm
(CA) model is triggered by a detection of manoeuvre onset, based on the autoregressive (AR) model is proposed, and the
and the CA model is switched back to CV model when the major contributions of this work are 2-fold:
manoeuvre is declared over. However, the VD filter has to
completely reconstruct the estimate of the state variables (1) Sub-model design: An AR model is incorporated into the
when changing the model. It means that the reinitialising Kalman filter (KF) for target tracking. The exact solution of
procedure will cause sharp discontinuities in the processing AR model coefficients can be derived from a convex
load [5], which may increase the tracking error. In the IE quadratic programming. The degrees of freedom of AR
approach [4–6], the tracking filter is in normal mode for the model can not only satisfy the polynomial constraint of
CV motion, and detects the time at which the target starts target motion, but also reduce the noise by minimising the
to manoeuvre. Then the unknown input (i.e. the target mean-square error (MMSE). The AR model can make the
acceleration) is estimated by the least-squares estimation filter converge more rapidly. Since it can utilise more past
with the data in a sliding window. It is based on the data effectively, rather than only use the latest data as the
assumption that the target acceleration is slowly varying. traditional differential model [18].
Therefore, if the target acceleration changes rapidly, this (2) Switched KF-IMM technique: The KF based on AR
approach does not give satisfactory tracking performance. model (KF-AR) is used to tackle the non-manoeuvring
As for the parallel model methods, the most representative motion, while the IMM filter based on AR model

IET Radar Sonar Navig., 2015, Vol. 9, Iss. 2, pp. 199–209 199
doi: 10.1049/iet-rsn.2014.0142 & The Institution of Engineering and Technology 2015
 17518792, 2015, 2, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rsn.2014.0142, Wiley Online Library on [23/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
(IMM-AR) deals with the manoeuvring motion. It is virtually and the process noise wCV k is a white Gaussian random
a simplified form of the variable-structure multiple-model process with the covariance matrix [20]
(VSMM) algorithm [16, 17]. The KF-AR and IMM-AR are
  CV T T 
switched by the proposed detection scheme. A fading T −t

memory average of innovations is utilised to detect the QCV
k = E wCV
k wk = T −t 1 qv dt
manoeuvre onset in KF mode and in the IMM mode 0 1
whether the effective probability of the quiescent model is T 3 /3 T 2 /2
the largest is added to the detection method of manoeuvre = qv (4)
termination. The length of sliding window for detecting the T 2 /2 T
manoeuvre termination is also different from that for
detecting the manoeuvre onset. These methods have given where the parameter qv is the process noise intensity and
facilities for the fast response to the manoeuvre controls the size of the deviations of the predictive velocity
occurrences, and for reducing the false alarm of manoeuvre from the actual one.
terminations.

The rest of this paper is organised as follows. In Section 2, 2.2 Proposed AR model for target tracking
the traditional model and proposed AR model are introduced. The key to target tracking lies in the effective extraction of
Based on the optimal AR model, a switched KF-IMM useful information about the target state from observations.
algorithm for target tracking is proposed in Section 3. A An effective model will certainly facilitate this information
simulation study is presented in Section 4 to show the extraction to a great extent. From Section 2.1, the
effectiveness of the proposed model and algorithm, discrete-time form of differential model can definitely
followed by concluding remarks in Section 5. The depict the target kinematic motion. However, it does not
mathematical derivations of some equations are detailed in have the function of reducing the noise and cannot adjust to
Appendices 1–3. the process and measurement noise intensities adaptively,
which results in the performance reduction to some extent.
2 System model description For this problem, the AR model is utilised to predict the
target state, and on the premise of satisfying the constraint
Considering one-dimensional (1D) tracking scenario, the of target motion the AR model coefficients are obtained in
target model is a discrete-time dynamical motion model of the sense of MMSE.
the form In the AR model, the target state vector is different from
that of the traditional differential model. It includes the
xk+1 = F k+1|k xk + wk (1) ranges from time k to time k − M + 1, which can be written as

T
where xk is the target state at time kT, and T is the sampling k = rk
xAR rk−1 . . . rk−M +1 (5)
interval. For clarity, in the following, the notation T is
omitted, that is, time k means time kT. Fk + 1|k is the state The state transition matrix in the AR model, F AR
k+1|k , is defined
transition matrix, and the process noise wk is a white by
Gaussian random process with covariance matrix Qk.
With the range-only measurements, the measurement ⎡ ⎤
h1 h2 ... hM
vector of target state at time k is given by ⎢1
⎢ 0 ... 0 ⎥⎥
F AR =⎢ . .. .. .. ⎥ (6)
zk = Hxk + vk (2)
k+1|k
⎣ .. . . . ⎦

 0 ... 1 0 M ×M
where the measurement matrix H = 1 0 · · · 0 1×K
and K is the dimension of the target state vector. The The AR model coefficients hm (m = 1, 2, …, M ) to be
measurement noise vk is a white Gaussian random process, optimised should satisfy the polynomial constraint of target
independent of wk, with covariance matrix Rk. In the motion [21, 22]
tracking applications, target dynamics are usually modelled
in the Cartesian coordinates, while target measurements are Auk = b (7)
directly available in the polar or spherical coordinate. When
the sensor measurements are converted to the ones in the
T
with uk = h1 h2 · · · hM , b = [ 1 0 ··· 0 ]T and
Cartesian coordinate by some measurement-conversion A is a Vandermonde matrix
techniques [19], the target dynamic model can be utilised in
the engineering application. ⎡ ⎤
1 1 ··· 1
⎢1 2 ··· M ⎥
⎢ ⎥
2.1 Traditional discrete-time differential model ⎢ 22 ··· M2 ⎥
A = ⎢1 ⎥ (8)
⎢ .. .. .. .. ⎥
For simplicity, the CV model

is takenT as an example. In (1), ⎣. . . . ⎦
k = rk
the target state is xCV ṙk , where rk is the target 1 2N ··· MN (N +1)×M
range, and the dot notation denotes the differentiation with
respect to time. The state transition matrix is
where M is the number of AR model coefficients, and N is the
  highest degree of polynomial used to approximate the range.
1 T
k+1|k =
F CV (3) The derivation of (7) is given in Appendix 1. When M = N +
0 1 1, that is, (7) is a consistent equation [23], the target motion

200 IET Radar Sonar Navig., 2015, Vol. 9, Iss. 2, pp. 199–209
& The Institution of Engineering and Technology 2015 doi: 10.1049/iet-rsn.2014.0142
 17518792, 2015, 2, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rsn.2014.0142, Wiley Online Library on [23/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
can also be depicted by the solution of (7) equivalent to
 
uk = A−1 b (9) minimise
uk
P k|k−1 (1,1)
(13)
as the traditional discrete-time differential model. For subject to Auk = b
example, the AR model of N = 1, M = 2 can exactly
describe the CV motion, which is equivalent to the CV where (Pk|k−1)(1,1) can be obtained from (10b)
model (as shown in the simulation results). Note that when    
N and M are given, the vector uk is defined. It cannot adjust P k|k−1 (1,1)
= uTk P k−1|k−1 uk + Qk−1 (1,1) (14)
to the process and measurement noise intensities either.
In the following, the derivation of the optimal AR model is Since uk is also independent of (Qk−1)(1,1), the cost function
given in the framework of KF by the criterion of MMSE. (13) is equivalent to
Owing to the dependence of the AR model on uk, the
Kalman recursive equations are also explicitly dependent on
minimise uTk P k−1|k−1 uk
uk. Specifically uk
(15)
    subject to Auk = b
xk|k−1 uk = F k|k−1 uk · xk−1|k−1 (10a)
      The optimisation problem (15) is a convex quadratic
P k|k−1 uk = F k|k−1 uk · P k−1|k−1 · F Tk|k−1 uk + Qk−1 programming problem [24]. It can be solved by Lagrange
(10b) multiplier technique [24]. The closed-form solutions of (15)
in the cases of N = 1, M = 4 and N = 2, M = 4 are derived in
    Appendix 2.
S k uk = HP k|k−1 uk H T + Rk (10c)
      2.3 Consistency of process noise covariance
K k uk = P k|k−1 uk · H T · S −1
k uk (10d)
In the

AR model, the  process noise is
     
  wk−1 . . . wk−M +1 ,
T
xk|k uk = xk|k−1 uk + K k uk · zk − Hxk|k−1 uk (10e) k = wk
wAR which is an
independent identically distributed (i.i.d.) zero-mean white
 
     Gaussian sequence with covariance matrix QAR k . Since all
P k|k uk = I − K k uk · H · P k|k−1 uk (10f ) the components of the target state in the AR model are
the ranges, the process noise in AR model represents the
where xk|k − 1 and Pk|k − 1 are the state vector’s a priori estimate stochastic change of the target range. Assume that the
and covariance matrix, respectively. Sk is the innovation changes of the ranges at different sampling intervals are
covariance, and Kk is the filter gain matrix. xk|k and Pk|k are mutually independent, the covariance matrix of process
the state vector’s a posterior estimate and covariance matrix, noise in the AR model is
respectively. I is an identity matrix.   AR T
The state vector in the AR model consists of M elements as QAR
k = E wAR
k wk = qr T · I (16)
shown in (5), however, the element we concern about most at
time k is the first one. The corresponding mean-square
tracking error of rk is the element at the first row and first where the parameter qr is the process noise intensity in terms
column of covariance matrix Pk|k. So the objective function of range, and I is an M-dimensional identity matrix.
in the sense of MMSE can be specially expressed as Hence, as to maintain consistency with the process noise
covariance of AR model, another form of the process noise
  covariance in the CV model is given by
minimise P k|k
uk (1,1)  
(11) 1 1/T
Auk = b QCV = qr T · (17)
subject to k
1/T 2/T 2

where (·)(i, j ) represents the (i, j)th element of the matrix in the Note that it is a little different from the form in (4). Since the
bracket. (Pk|k)(1,1) can be obtained from (10c), (10d) and (10f) velocity in original model is assumed to be a Wiener process
as follows – the integral of white noise [20], while (17) is derived based
on the premise that the process noise is depicted as a white
  2 noise process. However, it has no effect on the use of the
    P k|k−1 (1,1) CV model, and it is only that the conditions of parameter
P k|k (1,1) = P k|k−1 (1,1) −     setting are different. It affords a benchmark for the fair
P k|k−1 (1,1) + Rk (1,1)
    comparison between the AR model and CV model in the
Rk (1,1) · P k|k−1 (1,1) simulation. The derivations of (16) and (17) are given in
=    Appendix 3 in detail.
P k|k−1 (1,1) + Rk (1,1)
 
Rk (1,1) 3 Switched KF-IMM algorithm based on AR
=     (12) model
1 + Rk (1,1) / P k|k−1 (1,1)
During the past several decades, multiple-model estimation
Since uk is independent of (Rk)(1,1), the cost function (11) is algorithms for target tracking have received extensive

IET Radar Sonar Navig., 2015, Vol. 9, Iss. 2, pp. 199–209 201
doi: 10.1049/iet-rsn.2014.0142 & The Institution of Engineering and Technology 2015
 17518792, 2015, 2, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rsn.2014.0142, Wiley Online Library on [23/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
Table 1 Algorithm for the KF based on AR model (KF-AR) 1; di = zi − Hxi|i−1 is the innovation and Si is the
corresponding covariance matrix which is given in (10c).
Initialisation
Under the Gaussian assumption, d Ti S −1 i d i  xnd holds where
2
Given M, N with M ≥ N
+ 2. T
Initial estimate: x 0|0 = z0 z−1 · · · z−M+1 , where
nd denotes the dimension of measurement vector (nd = 1

T herein). As a weighted sum of i.i.d. Gaussian variables, the
z0 z−1 · · · z−M+1 is the range measurements before the
test statistics εk is not chi-square distributed, but by moment
filter start time k = 1;
Initial covariance: P0|0 = R·I, where R is the measurement noise
matching it can be approximately treated as a scaled version
variance; of a chi-square variable, that is, [1]
Process noise covariance: Q0 = qrT·I.
For k = 1, 2, … 1
Step 1. Calculation of transition matrix 1k  x2 (19)
Solve the convex quadratic programming problem (15), and 1 + l s·nz
obtain the transition matrix by (6).
Step 2. Predict the target state using AR model with nz = nd(1 + l)/(1 − l). Then the detection of a manoeuvre
Projected estimate: xk|k−1(uk) = Fk|k−1(uk)·xk−1|k−1; is declared if
Projected
 covariance:   
P k|k−1 u k = F k|k−1 u k · P k−1|k−1 · F Tk|k−1 u k + Q k−1 .
1
Step 3. Update the target state using the measurements 1k . x2 (a) (20)
Compute the gain matrix: Kk(uk) = Pk|k−1(uk)·H T·[HPk|k−1(uk)H T + 1 + l s·nz
Rk]−1;
Updated estimate: xk|k(uk) = xk|k−1(uk) + Kk(uk)·[zk − Hxk|k−1(uk)]; where 1 − α is the level of confidence, which should be set
Updated covariance: Pk|k(uk) = [I − Kk(uk)·H]·Pk|k−1(uk).
Output the target state estimate xk|k, and let k = k + 1, return to
quite high (e.g. 95% or 99%).
Step 1. A reasonable window length should be provided in
advance. If the window length s is too large, it will lead to
the delay of manoeuvre detection. Otherwise, if the window
length is too small, the estimated manoeuvre onset time
attention in the theoretic research and engineering application, may not be reliable. For a fast response to the manoeuvre
which have experienced two generations generally: onset, a relatively small length in terms of the fading factor
fix-structure multiple-model (FSMM) [8–11, 13–15] and l1 is selected as [2]
VSMM [16, 17]. The model sets of FSMM are assumed
unchanged over time, while VSMM algorithms adjust their  
1
model sets in real time with the online mode information. s1 = ceil (21)
1 − l1
Model set adaptation for VSMM can be decomposed into
two complementary and closely related tasks: activation of
where ceil (·) stands for the nearest integer greater than or
a set of new models and termination of a subset of the
equal to the value in the bracket.
current model set [16, 17]. However, the computing
complexity of the VSMM algorithms is very huge. In this
section, we propose a tracking algorithm consisting of the 3.2 Manoeuvre tracking and manoeuvre
KF-AR and the IMM-AR, which is virtually a simplified termination detection
form of the VSMM algorithm. The algorithm performs
three main tasks: non-manoeuvre tracking, manoeuvre In Section 3.1, assume that a manoeuvre is declared at time k,
detection and manoeuvre tracking. The KF-AR mode deals the manoeuvre onset time is estimated as k̂ = k − s − 1. A
with the non-manoeuvre motion, whereas the IMM-AR switch from the KF-AR (N = 1) to the IMM-AR is activated
mode handles the manoeuvre motion. at time k̂ and the target estimates and covariance matrices
of the IMM-AR are initialised as follows
3.1 Non-manoeuvre tracking and manoeuvre
x(k|k
i)
= xk̂|k̂ , P (k|k
i)
= P k̂|k̂ (for i = 1, 2, . . . , L) (22)
onset detection

The proposed algorithm utilises the KF-AR of N = 1 for the where L is the model number in IMM-AR, xk̂|k̂ and P k̂|k̂ are
CV motion tracking. The AR model of N = 1 matches with the target estimate and covariance of the KF-AR (N = 1) at
the CV motion, and it can exhibit high efficiency during the time k̂, respectively. The target states from time k̂ to k
target’s CV movements. The algorithm of KF-AR is shown which had been estimated by the KF-AR (N = 1) should be
in Table 1. re-processed by the IMM-AR. The process flow of the
When the KF-AR (N = 1) is working, a detection scheme is IMM-AR estimator is given in Table 2. So as to afford
used to determine the manoeuvre occurrence and the facilities for the combination of the target state, the number
manoeuvre onset time. In line with the VD filter, a of the elements in target state should be identical in every
detection method for the AR model is presented. The AR model. The degree of polynomials and the process
manoeuvre detector is chi-square significance test based. noise intensities should be different to cover the various
Since the occurrence of a manoeuvre can lead to large motions.
innovations for the KF-AR (N = 1), a fading memory The proposed algorithm utilises the IMM-AR for
average of innovations over a sliding window [k–s + 1, k] of manoeuvre tracking. Meanwhile the detection of manoeuvre
length s is utilised as the test statistic termination is undertaken. Since the manoeuvre termination
can result in large innovations for the IMM-AR estimator, a

k fading memory average of innovations over a sliding
1k = r(s) · ls−i d Ti S −1
i di (18) window is also employed as the test statistic. It is similar to
i=k−s+1 the detection of manoeuvre onset, given in (18). Herein, the
innovation is generated by subtracting the predicted state of
where ρ(s) = (1 − l)/(1 − ls + 1), l is the fading factor, 0 < l < the IMM-AR from the measurement. A longer window

202 IET Radar Sonar Navig., 2015, Vol. 9, Iss. 2, pp. 199–209
& The Institution of Engineering and Technology 2015 doi: 10.1049/iet-rsn.2014.0142
 17518792, 2015, 2, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rsn.2014.0142, Wiley Online Library on [23/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
Table 2 Algorithm for the IMM estimator based on AR model
(IMM-AR)

Initialisation (for i = 1, 2, …, L):

Given M, N (i), with M ≥ N (i) + 2;
Initial estimate x (0|0
i)
, and initial covariance P (0|0
i)
;
The process noise covariance: Q (i) (i)
0 = qr T · I ;

Initial model probability: m(i)

0 = 1/L;

Model transition probability: P = {pij} (for i, j = 1, 2, …, L);

For k = 1, 2, …
Step 1. Calculation of transition matrix (for i = 1, 2, …, L):
Solve the convex quadratic programming problem (15), and
obtain the transition matrix F (k|k−1
i)
by (6) in every AR model.
Step 2. Model-conditioned re-initialisation (for i, j = 1, 2, …, L):
( j)
j|i
Mixing probability: mk−1 = p ji mk−1 /m(k|k−1
i)
, where

L
(j)
mk|k−1 =
(i)
p ji mk−1 ;
j=1
L
( j)
Mixing estimate: x (k−1|k−1
i)
= j|i
x k−1|k−1 mk−1 ;
j=1    
 j|i
mk−1 P (k−1|k−1
j) (i,j ) (i,j ) T
(i ) L
Mixing covariance: P k−1|k−1 = + Dk−1 Dk−1 ,
j=1
(i,j ) ( j)
where Dk−1 = x (k−1|k−1
i)
− x k−1|k−1 . Fig. 1 Flowchart of proposed tracking algorithm
Step 3. Model-conditioned filtering (for i = 1, 2, …, L):
Predicted state: x (k|k−1
i)
= F (k|k−1
i)
x (k−1|k−1
i)
;
 T
P manoeuvre termination is declared, the KF-AR is enabled
(i)
Predicted covariance: P (i)
=F
k|k−1
(i)
k|k−1 k−1|k−1 F (i)
k|k−1 +Q (i)
k−1 ;
 T  −1 with the overall estimate and covariance of the IMM-AR
Filter gain: K (i) (i)
k = P k|k−1 H k
(i)
S k(i) , where estimator.
 T
(i) (i) (i) (i)
S k = H k P k|k−1 H k +R k ; (i) As the above development, the flowchart of the proposed
tracking algorithm is shown in Fig. 1. The proposed
Updated state: x (i) (i) (i) (i) (i) (i) (i)
k|k = x k|k−1 + K k d k , where d k = z i − H k x k|k−1 algorithm is comprised of the KF-AR and the IMM-AR. A

Updated covariance: P (i)
k|k = I
(i)
− K (i) (i) (i)
k H k P k|k−1
switch between them is trigged by the manoeuvre detection
Step 4. Model probability update (for i = 1, 2, …, L)
based on the chi-square test. For the fast response to the
  manoeuvre occurrence and the reliable identification of
(i)
Model likelihood: L(i) (i)
k = N d k ; 0, S k manoeuvre termination, the detection mechanisms are
(i)  (j)
L
Model probability: m(i)
k = mk|k−1 Lk /
(i)
mk|k−1 L(j)
k proposed for both the cases, respectively.
j=1
Step 5. Estimate fusion:

L
4 Simulation results
k|k mk
x (i) (i)
Overall estimate: x k|k =
i=1
Overall covariance: In this section, the proposed model and algorithm are verified
   T 

L
P k|k = m(i) (i) (i)
k P k|k + x k|k − x k|k
(i)
x k|k − x k|k by the simulations. In Case 1, the AR model is compared with
i=1
Output the target state estimate xk|k and return to Step 1 to
the traditional differential model (CV model) in 1D CV scene.
calculate the transition matrix F (i) (i)
k+1|k with P k|k for i = 1, 2, …, L,
In Case 2, the proposed algorithm competes with the
respectively traditional VD filter and IMM filter in a 2D manoeuvring
scenario.

Case 1: A target moves with the CV v = 340 m/s and the

length [10] initial range r0 = 10 km. Assume that only the target range
can be measured, the measurement noise variance R = 100
    m2 and the sampling interval T = 1 s. In the AR models for
ln 1 − l2 − ln nd
s2 = ceil   −1 (23) estimating the range, the degree of polynomial N = 1, and
ln l2 the number of AR model coefficients M = 2, 3, 4,
respectively. The corresponding low-degree AR model is
is suggested to reduce the false alarm of manoeuvre utilised to estimate the velocity. The AR model obtained by
terminations. Thus 1k  (1/(1 + l2 ))x2s2 ·nz , if s = s2. In (9) is denoted as ‘AR model 0’ in Figs. 2 and 3, which
conjunction with the test statistic, the termination of a only satisfies the constraint of target motion. The process
manoeuvre is declared when the effective probability of the noise covariance matrices of the AR model and CV model
AR model (N = 1) is the largest for M sampling intervals, are given in (16) and (17), respectively. The Monte Carlo
that is simulation with 1000 runs is carried out for a period of
100 s. The two cases of parameter qr matching (qr = 0) and
   not matching (qr = 1.0) with the actual target motion are
arg max m(i)
j = 1, for j = k − M + 1, . . . , k (24) considered in Figs. 2 and 3, respectively.
1≤i≤L Since only the range can be estimated by a single AR
model, we regard the difference of the estimated range as
where the sign ‘1’ denotes the AR model of N = 1. When the the measurement value of velocity, that is, ṙk = r̂k − r̂k−1 .

IET Radar Sonar Navig., 2015, Vol. 9, Iss. 2, pp. 199–209 203
doi: 10.1049/iet-rsn.2014.0142 & The Institution of Engineering and Technology 2015
 17518792, 2015, 2, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rsn.2014.0142, Wiley Online Library on [23/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org

Fig. 2 RMSE of estimated range and velocity for CV and AR models (qr = 0)
a Range
b Velocity

Then with the measurements, the velocity can be estimated motion, the optimal AR model outperforms the CV model
by another low-degree AR model. The root-mean-square more evidently.
error (RMSE) comparisons of estimated range and Since the measurement of the AR model for estimating the
velocity for the models in two cases are shown in Figs. 2 velocity is different from that of the CV model, the results
and 3. shown in Figs. 2b and 3b are a little different from the
From Figs. 2a and 3a, it can be seen that the AR model 0 Fig. 2a and 3a. However, it can be seen that the optimal
has the same performance as the CV model, since it is AR model performs better than the CV model.
equivalent to the CV model as mentioned in Section 2.2.
The AR model 0 and CV model can equally depict the CV Case 2: We consider a 2D scene with the fast and slow turn
motion. The optimal AR model (M = 3, 4) performs better manoeuvres in (x, y) plane as in [10]. The target moves
than the CV model in the aspect of estimated accuracy. As with CV unless otherwise stated, and its state evolves with
the AR model (M = 3, 4) cannot only satisfy the constraint the initial state x0 = [21689 m −8.3 m/s 0 m/s2 10840 m
of target motion as the CV model and AR model (M = 2), −340 m/s 0 m/s2]T. The position of the target is sampled at
but also minimise the mean-square tracking error with the every T = 1 s. The measurement noise variance of the x- and
extra degrees of freedom. y-dimensional position is R = 100 m2. During the 25th to
In addition, the longer the length of the AR model 33rd sampling periods, the target makes a fast turn with
coefficient is, the better the tracking performance of the x-axis acceleration 0.3 m/s2 and y-axis acceleration 0.3 m/s2.
KF-AR has. As the longer the length of the AR model During the 52nd to 57th sampling periods, the target makes
coefficients is, the more information the filter can utilise. a slow turn with x-axis acceleration 0.08 m/s2 and y-axis
However, the target may manoeuvre at unknown time in acceleration 0.02 m/s2.
practice, it is suggested that the appropriate length is N + We choose the traditional IMM filter and VD filter to
1 ≤ M ≤ 5. The comparison of Figs. 2 and 3 reveals that compare with the proposed algorithm. In the IMM filter, a
when the parameter setting does not match with the actual CV model and two CA models with different process noise

Fig. 3 RMSE of estimated range and velocity for CV and AR models (qr = 1.0)
a Range
b Velocity

204 IET Radar Sonar Navig., 2015, Vol. 9, Iss. 2, pp. 199–209
& The Institution of Engineering and Technology 2015 doi: 10.1049/iet-rsn.2014.0142
 17518792, 2015, 2, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rsn.2014.0142, Wiley Online Library on [23/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org

Fig. 4 RMSE of estimated position for VD, IMM and proposed algorithms
a x-axis
b y-axis

covariance matrices are utilised. The process noise covariance Table 3 Data compress ratio against sampling interval when
matrices are, respectively (see (25)), with q1 = 0.01, q2 = 0.1 R = 100 m2
and q3 = 1.0. The derivation of the process noise covariance Algorithm Sampling interval, s
is given in Appendix 3. The transition probability matrix of
the three models is 0.1 0.5 1 2 5

⎡ ⎤ VD 0.5152 0.7405 0.9615 1.0537 1.8524

0.70 0.15 0.15 IMM 0.4763 0.6045 0.8318 0.9721 1.2135
P tracker = ⎣ 0.20 0.60 0.20 ⎦ (26) proposed 0.4262 0.5758 0.7651 0.9563 1.2005
0.25 0.25 0.50
Fig. 4 gives the RMSE comparison of the estimated (x, y)
In the VD filter, the confidence level of the chi-square position for the three algorithms. It can be seen that the
variable with a degree of freedom at 95% is 3.841. When proposed algorithm outperforms the traditional VD filter in
the test statistic does not exceed the threshold, the CV the non-manoeuvring motion and performs better than the
model with QCV k in (25) is utilised and otherwise the CA traditional IMM filter in the manoeuvring motion. This is
model with QCA2k . In the proposed algorithm, the parameters because that the optimal AR model is employed in the
for manoeuvre detection are l1 = l2 = 0.9, α1 = α2 = 0.05. proposed algorithm while the traditional differential models
The model base of the IMM-AR is similar to the traditional are utilised in the traditional algorithms. The proposed
IMM filter. There are an AR model of N = 1 and two AR algorithm alleviates sharp discontinuities effectively in the
models of N = 2. The coefficient number of the three AR processing load, which attributes to the AR models having
models M = 4. For consistency with the process noise the same target state variables and the switching
intensity in (25), the process noise covariance matrices of initialisations. It does not require to reconstruct the state
AR models are variables, giving facilities for switching between the KF-AR
and IMM-AR. Through the proposed detection schemes of
manoeuvre occurrence and termination, the proposed
QAR1
k = q1 T · I, QAR2
k = q2 T · I and algorithm has a fast response to the manoeuvre onset and
(27)
QAR3 = q3 T · I reduces the false alarm of manoeuvre terminations, which
k
has improved the tracking performance significantly.
Tables 3 and 4 give the data compress ratio of the three
respectively. The transition probability matrix of the three algorithms against the sampling interval (inverse of the data
models is the same as (26). The Monte Carlo simulation rate) and the measurement noise variance, respectively. The
with 1000 runs is carried out for a period of 100 s. data compress ratio is the RMSE averaged over time and

⎡ ⎤
  1 1/T 1/T 2
1 1/T ⎢ ⎥
k = q1 T ·
QCV , QCA1
k = q2 T · ⎣ 1/T 2/T 2 3/T 3 ⎦,
1/T 2/T 2
1/T 2 3/T 3 6/T 4
⎡ 2
⎤
1 1/T 1/T
⎢ ⎥
QCA2
k = q3 T · ⎣ 1/T 2/T 2
3/T 3 ⎦ (25)
1/T 2 3/T 3 6/T 4

IET Radar Sonar Navig., 2015, Vol. 9, Iss. 2, pp. 199–209 205
doi: 10.1049/iet-rsn.2014.0142 & The Institution of Engineering and Technology 2015
 17518792, 2015, 2, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rsn.2014.0142, Wiley Online Library on [23/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
Table 4 Data compress ratio against measurement noise 4 Bogler, P.L.: ‘Tracking a maneuvering target using input estimation’,
variance when T = 1 s IEEE Trans. Aerosp. Electron. Syst., 1987, 23, (3), pp. 298–310
5 Park, Y.H., Seo, J.H., Lee, J.G.: ‘Tracking using the variable-dimension
Algorithm Measurement noise variance, m2 filter with input estimation’, IEEE Trans. Aerosp. Electron. Syst., 1995,
31, (1), pp. 399–408
52 102 202 302 502 6 Khaloozadeh, H., Karsaz, A.: ‘Modified input estimation technique for
tracking manoeuvring targets’, IET Radar Sonar Navig., 2009, 3, (1),
VD 0.9867 0.9645 0.9451 0.9377 0.9137 pp. 30–41
IMM 0.8512 0.8326 0.8026 0.8023 0.7703 7 Li, X.R., Jilkov, V.P.: ‘Survey of maneuvering target tracking. Part V:
proposed 0.7704 0.7647 0.7314 0.6822 0.6615 multiple-model methods’, IEEE Trans. Aerosp. Electron. Syst., 2005,
41, (4), pp. 1255–1321
8 Mazor, E., Averbuch, A., Bar-Shalom, Y., Dayan, J.: ‘Interacting
multiple model methods in target tracking: a survey’, IEEE Trans.
Aerosp. Electron. Syst., 1998, 34, (1), pp. 103–123
the Monte Carlo runs, which is defined as 9 Fu, X., Jia, Y., Du, J., Yu, F.: ‘New interacting multiple model
 algorithms for the tracking of the manoeuvring target’, IET Control
 Theory Appl., 2010, 4, (10), pp. 2184–2194
 1  Nd Nm
 2
h= rk − r̂ki /R (28) 10 Ho, T.J.: ‘A switched IMM-extended Viterbi estimator-based algorithm
for manuevering target tracking’, Automatica, 2011, 47, pp. 92–98
Nd Nm k=1 i=1
11 Foo, P.H., Ng, G.W.: ‘Combining the interacting multiple model method
with particle filters for manoeuvring target tracking’, IET Radar Sonar
where Nm is the number of Monte Carlo runs, Nd is the Navig., 2011, 5, (3), pp. 234–255
12 Jing, L., Vadakkepat, P.: ‘Interacting MCMC particle filter for tracking
number of the data for a Monte Carlo run and R is the maneuvering target’, Digital Signal Process., 2010, 20, pp. 561–574
variance of the measurement noise. The two tables show 13 Boers, Y., Driessen, J.N.: ‘Interacting multiple model particle filter’, IEE
that the integral performance of the proposed algorithm is Proc. Radar Sonar Navig., 2003, 150, (5), pp. 344–349
better than the traditional VD and IMM algorithms on the 14 Laneuville, D., Bar-Shalom, Y.: ‘Maneuvering target tracking: a
same conditions of data rate and measurement noise variance. Gaussian mixture based IMM estimator’. Proc. 2012 IEEE/AIAA
Aerospace Conf., Big Sky, MT, March 2012, pp. 1–12
15 Lan, J., Li, X.R., Jilkov, V.P., Mu, C.: ‘Second-order Markov chain
5 Conclusions based multiple-model algorithm for maneuvering target tracking’,
IEEE Trans. Aerosp. Electron. Syst., 2013, 49, (1), pp. 3–19
In this paper, a switched KF-IMM algorithm based on AR 16 Li, X.R.: ‘Multiple-model estimation with variable structure. Part II:
model-set adaptation’, IEEE Trans. Autom. Control., 2000, 45, (11),
model is proposed for target tracking. First, an AR model is pp. 2047–2060
incorporated into the KF for target tracking. The degrees of 17 Lan, J., Li, X.R., Mu, C.: ‘Best model augmentation for
freedom of the AR model can satisfy the target motion on variable-structure multiple-model estimation’, IEEE Trans. Aerosp.
the one hand, and reduce the noise in the sense of MMSE Electron. Syst., 2011, 47, (3), pp. 2008–2025
18 Li, X.R., Jilkov, V.P.: ‘Survey of maneuvering target tracking. Part I:
on the other hand. Then, a switching mechanism is dynamic models’, IEEE Trans. Aerosp. Electron. Syst., 2003, 39, (4),
employed to adjust to the non-manoeuvring and pp. 1333–1364
manoeuvring motions, combined with the detection scheme 19 Zhao, Z., Li, X.R., Jilkov, V.P.: ‘Best linear unbiased filtering with
of manoeuvre onset and termination. Simulations nonlinear measurements for target tracking’, IEEE Trans. Aerosp.
demonstrate that the AR model has enhanced the tracking Electron. Syst., 2004, 40, (4), pp. 1324–1336
20 Bar-Shalom, Y., Li, X.R., Kirubarajan, T.: ‘Estimation with applications
accuracy and the proposed algorithm has a better to tracking and navigation: theory algorithms and software’ (John Wiley
performance than the VD and IMM filters in the & Sons, 2001)
manoeuvring scene. However, the proposed AR model has 21 Väliviita, S., Ovaska, S.J., Vainio, O.: ‘Polynomial predictive filtering in
some disadvantages as follows: (i) the range and velocity control instrumentation: a review’, IEEE Trans. Ind. Electron., 1999, 46,
(5), pp. 876–888
cannot be obtained synchronously in a single AR model as 22 Xu, L., Li, X.R., Duan, Z., Lan, J.: ‘Modelling and state estimation for
in the traditional differential model and (ii) the coefficients dynamic systems with linear equality constraints’, IEEE Trans. Signal
of the AR model need to be computed in real time, which Process., 2013, 61, (11), pp. 2927–2939
has added to the computing amount a bit. 23 Meyer, C.D.: ‘Matrix analysis and applied linear algebra’ (SIAM, 2000)
In the future work, a manoeuvring target tracking algorithm 24 Singiresu, S.R.: ‘Engineering optimization: theory and practice’ (John
Wiley & Sons, 2009, 4th edn.)
based on the neural network or fuzzy logic are suggested for
further studies.
8 Appendices
6 Acknowledgments 8.1 Appendix 1: derivation of (7)
This work is partially supported by the National Natural It is well known that any continuous target trajectory can be
Science Foundation of China (nos. 61271291 and approximated by a polynomial of a certain degree to an
61201285), Programme for New Century Excellent Talents arbitrary accuracy. As such, it is possible to model target
in University (NCET-09-0630), and the Fundamental motion by Nth-degree polynomial in the Cartesian
Research Funds for the Central Universities. coordinates [18]

7 References ⎡ ⎤
⎡ ⎤ ⎡ ⎤ 1 ⎡ ⎤
x(t) a0 a1 . . . aN ⎢ t ⎥ w (t)
1 Li, X.R., Jilkov, V.P.: ‘A survey of maneuvering target tracking. Part IV: ⎣ y(t) ⎦ = ⎣ b0 ⎢ ⎥ ⎣ x ⎦
b1 . . . b N ⎦⎢ .. ⎥ + wy (t) (29)
decision-based methods’. Proc. SPIE Conf. on Signal and Data ⎣ . ⎦
Processing of Small Targets, Orlando, USA, April 2002, pp. 4728–4760 z(t) c0 c1 . . . cN N
wz (t)
2 Bar-Shalom, Y., Birmiwal, K.: ‘Variable dimension filter for t
maneuvering target tracking’, IEEE Trans. Aerosp. Electron. Syst.,
1982, 18, (5), pp. 621–629
3 Koteswara Rao, S.: ‘Modified gain extended Kalman filter with
with a certain choice of the coefficients an, bn, cn (n = 0, 1,…,
application to bearings-only passive manoeuvring target tracking’, IEE N), where (x, y, z) are the position coordinates and (wx, wy, wz)
Proc. Radar Sonar Navig., 2005, 152, (4), pp. 239–244 are the corresponding noise terms. Such an Nth-degree

206 IET Radar Sonar Navig., 2015, Vol. 9, Iss. 2, pp. 199–209
& The Institution of Engineering and Technology 2015 doi: 10.1049/iet-rsn.2014.0142
 17518792, 2015, 2, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rsn.2014.0142, Wiley Online Library on [23/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
polynomial amounts to assuming the Nth time derivative of When n = 2, (35) is
the position is (nearly) constant. The CV and CA models
[18] are special cases (for N = 1, 2, respectively) of this 
M

general Nth-degree model with white noise. (k + 1)2 = hm (k + 1 − m)2 (39)

m=1
For simplicity, considering the 1D situation without the
noise term, after sampling uniformly the range at time tk
can be depicted as Using (36) and (38) in (39), we obtain the second-degree
 constraint
N  n
rk = an tk (30)

M
n=0
h m m2 = 0 (40)
where tk = kT. The AR model is used to predict the range at m=1
time k + 1 through the M latest samples rk, rk−1,…, rk+1−M
[21] In similarity, we can infer the nth-degree constraint from (38)

M and (40)
rk+1 = hm rk+1−m (31)
m=1

M
hm mN = 0 (41)
with the AR model coefficients hm (m = 1, 2,…, M ) to be m=1
optimised in the sense of MMSE. According to (30), rk+1
and rk+1−m in (31) can be written as follows
Now we gather all the nth-degree constraints of the AR model

N coefficients in an equation
rk+1 = an (k + 1)n T n , ⎧
n=0 ⎪
⎪ M
⎪
⎨ hm = 1, n=0

N m=1
rk+1−m = an (k + 1 − m)n T n (32) (42)
⎪
⎪ 
M
n=0 ⎪
⎩ hm m = 0, n = 1, 2, . . . , N
n
m=1
Substituting (32) into (31), we obtain

N 
M 
N
The
AR model coefficients
T are denoted by
an (k + 1)n T n = hm an (k + 1 − m)n T n (33) uk = h1 h2 · · · hM , b = [ 1 0 · · · 0 ]T and
n=0 m=1 n=0
⎡ ⎤
Rearranging the two summations on the right-hand side, (33) 1 1 ··· 1
⎢1 2 ··· M ⎥
is ⎢ ⎥
⎢ 2 ⎥
A = ⎢1 2 ··· M ⎥
2
(43)

N 
N 
M ⎢ .. .. . . .. ⎥
⎣. . . . ⎦
an (k + 1)n T n = an hm (k + 1 − m)n T n (34)
n=0 n=0 m=1 1 2N ··· MN (N +1)×M

Extracting the like terms (the terms having the same degree), so that (42) can be rewritten as
we have
Auk = b (44)

M
(k + 1) = n
hm (k + 1 − m) , n
n = 0, 1, . . . , N (35)
m=1
8.2 Appendix 2: closed-form solution of
When n = 0, the zero-degree constraint of the AR model optimisation problem (15)
coefficients is obtained from (35)
When N = 1, M = 4, using the method of Lagrange multipliers

M [24], the Lagrange function can be written as
hm = 1 (36)
m=1   4 
4  
L h1 , h2 , h3 , h4 , l0 , l1 = hi P k−1|k−1 (i,j) hj
i=1 j=1
When n = 1, (35) is # $ # $

4 
4
+ l0 hi − 1 + l1 hi i (45)

M
i=1 i=1
k+1= hm (k + 1 − m) (37)
m=1

Setting the partial derivatives of L(h1, h2, h3, h4, l0, l1) with
Using (36) in (37), we obtain the first-degree constraint of the respect to all the arguments equal to zero, we obtain
AR model coefficients
∂L 4  
=2 hi P k−1|k−1 (i,j) +l0 + jl1 = 0, for j = 1, 2, 3, 4

M ∂hj
hm m = 0 (38) i=1

m=1 (46)

IET Radar Sonar Navig., 2015, Vol. 9, Iss. 2, pp. 199–209 207
doi: 10.1049/iet-rsn.2014.0142 & The Institution of Engineering and Technology 2015
 17518792, 2015, 2, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rsn.2014.0142, Wiley Online Library on [23/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
Note that there is an assumption as mentioned in Section
∂L  4
2.3, which is the stochastic changes of the range at every
= h −1=0 (47) sampling interval are mutually independent, that is
∂l0 i=1 i

∂L  4  
= hi=0 (48) cov rk , ri = 0, for i = k (55)
∂l1 i=1 i

The coefficients h1 and h2 are solved from (47) and (48) as Thus the covariance matrix of the process noise in the AR
model is
h1 = h3 + 2h4 + 2, h2 = −2h3 − 3h4 − 1 (49)
  AR T
k = E wk
QAR = qr T · I
AR
wk (56)
Substituting (49) to the partial derivatives ∂L/∂hj ( j = 1, 2, 3,
4), we have (see (50) and (51))

T
T
where
g1 = 1 −2  1 0 , g2 = 2 −3 0 1 and where I is an M-dimensional identity matrix.
T We can also obtain
g3 = 2 −1 0 0 .
Also using the same method, we obtain the solution of
optimisation problem (15) when N = 2, M = 4    r −r 
cov rk , ṙk = cov rk , k k−1
T
gT1 P k−1|k−1 g2 + gT2 P k−1|k−1 g1   1  
h4 = − , 1
= cov rk , rk − cov rk , rk−1
2 · gT1 P k−1|k−1 g1 T T
h1 = −h4 + 3, h2 = 3h4 − 3, h3 = −3h4 + 1 (52) 1  
= cov rk , rk (57)

T
T T
where g1 = −1 3 −3 1 , and g2 = 3 −3 1 0 .
and (see (58))
8.3 Appendix 3: derivation of process noise
covariance Thus another form of the process noise covariance in the CV
model can be obtained
According to (4), the error covariance of predicted velocity is

  T    
cov ṙk , ṙk = qv dt = qv T (53) cov rk , rk cov rk , ṙk
0 QCV
k =    
cov rk , ṙk cov ṙk , ṙk
 
where cov(a, b) denotes the covariance of statistic a and b. 1 1/T  
Similarly, the error covariance of predicted range can be = 2 · cov rk , rk
defined as 1/T 2/T
 
1 1/T
  T = · qr T (59)
cov rk , rk = qr dt = qr T (54) 1/T 2/T 2
0

where qr is the process noise intensity in terms of range. In the same way, another form of the process noise covariance

 T     
2g2 P k−1|k−1 g2 gT1 P k−1|k−1 g3 + gT3 P k−1|k−1 g1 − gT1 P k−1|k−1 g2 + gT2 P k−1|k−1 g1 gT2 P k−1|k−1 g3 + gT3 P k−1|k−1 g2
h3 =  T 2 (50)
g1 P k−1|k−1 g2 + gT2 P k−1|k−1 g1 − 4gT1 P k−1|k−1 g1 · gT2 P k−1|k−1 g2

 T     
2g1 P k−1|k−1 g1 gT2 P k−1|k−1 g3 + gT3 P k−1|k−1 g2 − gT1 P k−1|k−1 g2 + gT2 P k−1|k−1 g1 gT1 P k−1|k−1 g3 + gT3 P k−1|k−1 g1
h4 =  T 2 (51)
g1 P k−1|k−1 g2 + gT2 P k−1|k−1 g1 − 4gT1 P k−1|k−1 g1 · gT2 P k−1|k−1 g2

  r − r 
k−1 rk − rk−1
cov ṙk , ṙk = cov k ,
T T
1
     
= 2 cov rk , rk − 2cov rk , rk−1 + cov rk−1 , rk−1
T
2  
= 2 cov rk , rk (58)
T

208 IET Radar Sonar Navig., 2015, Vol. 9, Iss. 2, pp. 199–209
& The Institution of Engineering and Technology 2015 doi: 10.1049/iet-rsn.2014.0142
 17518792, 2015, 2, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-rsn.2014.0142, Wiley Online Library on [23/11/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
www.ietdl.org
in the CA model can be obtained as follows Note that (59) and (60) are derived based on the assumption
⎡      ⎤ that the process noise is depicted as a white noise process,
cov rk , rk cov rk , ṙk cov rk , r̈k which are different from those obtained in Wiener process
⎢      ⎥ [20]. However, they have no effect on the use of models,
k = ⎣ cov rk , ṙ k
QCA cov ṙk , ṙk cov ṙk , r̈k ⎦
      and it is only that the conditions of parameter setting are
cov rk , r̈k cov r̈k , ṙk cov r̈k , r̈k different.
⎡ ⎤
1 1/T 1/T 2
⎢ ⎥  
= ⎣ 1/T 2/T 2 3/T 3 ⎦ · cov rk , rk
1/T 2 3/T 3 6/T 4
⎡ ⎤
1 1/T 1/T 2
⎢ ⎥
= ⎣ 1/T 2/T 2 3/T 3 ⎦ · qr T (60)
1/T 2 3/T 3 6/T 4

IET Radar Sonar Navig., 2015, Vol. 9, Iss. 2, pp. 199–209 209
doi: 10.1049/iet-rsn.2014.0142 & The Institution of Engineering and Technology 2015