You are on page 1of 9

Multiple-Model Detection of Target Maneuvers∗

Jifeng Ru X. Rong Li Vesselin P. Jilkov


Department of Electrical Engineering
University of New Orleans, New Orleans, LA 70148, USA
{jru, xli, vjilkov }@uno.edu, (504) 280-{7416, 3950 (fax)}

Abstract
This paper proposes a multiple-model (MM) hypothesis testing approach for detection of unknown target maneuvers that
may have several possible prior distributions. An MM maneuver detector based on sequential hypothesis testing is developed.
Simulation results that compare the performance of the proposed MM detector to that of traditional maneuver detectors are
presented. They demonstrate that the new sequential MM detector outperforms traditional multiple hypothesis testing based
detectors when the prior acceleration distributions are unknown.

1 Introduction
Maneuvering target tracking (MTT) is a complicated problem due to the fact that accelerations are generally unknown and
not available directly through measurements. The decision-based techniques to MTT have been studied extensively, where
the state estimation is based on a hard decision on the target motion model made by the maneuver detector. There exist quite
a few algorithms to detect unknown target maneuvers [1, 8, 14]. Essentially they can be classified into two principal types of
tests: chi-square significance tests, such as measurement residuals (MR) and input estimates (IE) based, and likelihood ratio
tests, including the generalized likelihood (GLR) test and the marginalized likelihood ratio (MLR) test. A survey and analysis
of those results were made in [8] and a simulation based comparison of such maneuver onset detection algorithms were
conducted in [14]. Those tests, however, are based on simple nonmaneuvering model and ignore possible prior information
about target motions, which may be available in some cases. Besides, most existing detectors such as IE and GLR based
are computationally demanding due to window based batch processing. These limitations motivate us to explore sequential
testing procedures, which are actually preferable for maneuver detection because measurements are available sequentially.
In [15], a sequential maneuver detection scheme was proposed, which utilizes the likelihood marginalization technique to
cope with difficulty that target accelerations are usually unknown, by assuming that a priori information of the maneuver
accelerations is available in typical tracking engagements.
In practice, the true prior distribution of target accelerations might be partially known, and is usually dependent on
some parameters related to the target class and maneuver behavior in typical engagements. For example, in [3] the normal
acceleration of evasive targets is modeled by an asymmetric probability distribution with three design parameters. Here we
consider detection of a target that may have multiple possible distributions after a maneuver, corresponding to different target
classes and/or types. The goal is to detect maneuvers on-line as quickly as possible. While traditionally such a problem
would be formulated as M-ary hypothesis testing, we consider it still as a binary composite hypothesis testing problem with,
however, a multiple model distribution. For example, suppose we are detecting an unknown target that has two possible
distributions F11 (z) and F12 (z) of different families. A traditional formulation in this case is

⎨ H0 : z ∼ F0 (z) (target absent)
F: H11 : z ∼ F11 (z) (target 1 present)

H12 : z ∼ F12 (z) (target 2 present)
and, we formulate it as a binary hypothesis test with two models in H1

 H0 : z ∼ F0 (z) (target absent)
F :
H1 : z ∼ F1 (z) (target present)
∗ Research supported in part by ARO grant W911NF-04-1-0274, and NASA/LEQSF grant 2001-4-01.
where F1 (z) is related to F11 (z) and F12 (z), that is, H1 is a hypothesis with multiple models. Since F1 (z) could consist
of different distribution families, H1 is not a traditional composite hypothesis which is usually thought of as a function of a
known form but only with one or more unknown parameters. For such a multiple model detection formulation, existing tech-
niques such as generalized likelihood and Bayesian marginalization would not work well. Clearly, formulation F  is superior
to formulation F since the problem is primarily target detection (group decision) rather than classification or identification. It
is quite possible that H0 beats H11 and H12 individually, but loses to H1 as a team formed by H11 and H12 .
There are several formulations in the literature regarding to the sequential change detection. A non-Bayesian approach first
proposed by Page, the cumulative sums (CUSUM) test minimizes the detection delay given the decision error probabilities.
It can be interpreted as a repeated sequential probability ratio test (SPRT), that is, restart the SPRT algorithm whenever H0 is
accepted. Many modifications and extensions of CUSUM-type tests have been developed for online change-point detection
in stochastic control and signal processing [5, 4, 12]. Some solutions based on the min-max approach have been proposed in
[11, 12, 13] with applications to navigation system integrity monitoring. Generally, they optimizes the worst case situations
with given decision error rates. Shiryaev sequential probability ratio test (SSPRT) is a Bayesian approach which minimizes
the risk function at each time step. It is proven to provide the quickest detection of a change in a sequence of conditionally
independent measurements under the given decision error rates. Due to the Bayesian formulation, it needs to define prior
information including a priori probabilities of Hi (i = 0, 1) being true and the transition probabilities of H0 to H1 from k − 1
to k. For each hypothesis under consideration, the SSPRT yields the probability of that hypothesis being the correct one. The
SSPRT makes decision by noting the change in the probability density function of the measurements. Also, the recursive,
probabilistic framework of the SSPRT provides the estimation of all those hypothesis probabilities based on data, while the
CUSUM-based techniques do not provide such information. Some results have recently been reported for fault detection
using SSPRT [10, 16], however we are not aware of any work that addresses the adaptation and effectiveness of SSPRT for
maneuver detection.
The rest of this paper is organized as follows. In Section 2, the problem of maneuvering target detection is formulated
as one of binary composite hypotheses testing with multiple models. In Section 3, a sequential detection procedure based on
SSPRT is formulated and the test statistics are derived. Section 4 presents simulation results and performance comparison of
the proposed MM detector with some traditional maneuver detectors. Conclusions are provided in Section 5.

2 Problem Formulation
Assume the target-measurement model is given by

xk+1 = Fk xk + Gk uk + wk (1)
zk = Hk xk + vk , k = 1, 2, . . . (2)

where xk is the target state, uk is the maneuver control input, and zk is the measurement. wk ∼ N (0, Qk ) and vk ∼
N (0, Rk ) are independent process and measurement noises, respectively, and the initial state x0 ∼ N (0, P0 ) is independent
of wk and vk .
It is assumed that uk = 0 when the target is not maneuvering at time k and uk = 0 when the target is maneuvering.
Although the acceleration u is usually unknown in target tracking applications, prior distributions f (u) of some target ma-
neuver motions may be available, which are dependent on the target class and/or type. Then maneuver onset detection can be
formulated as a binary hypothesis testing problem:

H0 :  um = 0 for m = 1, . . . , k (no maneuver)


um = 0 for m = 1, . . . , n − 1 (3)
H̄1 : (maneuver)
um ∼ f (u) for m = n, . . . , k

where f (u) is the a priori probability function of u, and unknown onset time n is referred to as the change point. This is
known as change point detection in the statistical literature. The focus of the maneuver detection is to decide on a maneuver
and estimate n. Since both u and n are usually unknown in practice, hypothesis H̄1 is clearly composite.
Assume H1 : u ∼ f (u) is the hypothesis with M distribution models f1i (u) (i = 1, ..M )

H11 : u ∼ f11 (u)


.. (4)
.
H1M : u ∼ f1M (u)
where each f1i (u) is defined by different distributions with possible unknown parameters. According to [7], the proposed
detector in the following based on multiple models provides an optimal solution when the true distribution of accelerations
is from H1i , i.e., one of H1 . When the hypotheses are significantly mismatched with the true distribution of accelerations,
we propose to construct a distribution model of H1 based on data in order to make hypothesis testing. For example, the truth
could be a mixture distribution rather than being one of f1i (u)


M 
M
f (u) = pi f1i (u), pi = 1 (5)
i=1 i=1

where pi is the probabilistic weight (model probability) of the ith mixture component. In this case, if we can estimate pi
on-line based on measurements, naturally we think H1 will be closer to the true motion distribution than each H1i , which
will consequently improve detection performance. To this point, our formulation is more general and better than traditional
M-ary hypothesis testing for maneuver detection especially for model mismatch cases.
The weight estimation problem in (5) can be formulated as a well-known classification problem referred to as prior
probability estimation (PPE) of finite mixtures in a general statistical setting. There exists a large number of publications and
results for solving this problem [2, 17] . Incorporating those techniques into our multiple model detection is currently under
investigation. Clearly the estimates of the weights serve to adapt the distribution of the hypothesis as well as provide valuable
information for target maneuver identification/estimation.

3 Sequential Detection Algorithm Development


We first briefly discuss the general SSPRT scheme and then propose an SSPRT-based maneuver detection algorithm with
multiple models.

3.1 SSPRT-based Detector


The optimality of SSPRT is to minimize an expected cost at each time step. This cost includes the measurement cost and
the cost due to a terminal decision error by false alarm or miss detection. The decision rule of the SSPRT is defined by the
posterior probability ratio Pk . Let pk  P {n ≤ k|z k } denote the posterior probability that a change occurs at an unknown
time n at or before time k given the available measurements z k  {z1 , ..., zk }, and
 pk  pT
Pk = , PT = (6)
1 − pk 1 − pT
where a preset threshold PT is determined by the desirable decision error rate. The SSPRT becomes

1) Accept H1 (declare change) if Pk ≥ PT . The stopping time is n̂ = min{k : Pk ≥ PT }.

2) Continue the test (k → k + 1) if Pk < PT .

Calculating Pk is the key to SSPRT. Let pi0 denote the prior probability of hypothesis Hi being true (i = 0, 1), π the

 H1 , and φk = P (n ≤ k + 1|z ) is the predicted probability that a change occurs at
i k
 from Hκ−1
transition probability 0 to
time n ≤ k + 1, f zκ |Hi , z is the marginal likelihood of the hypotheses at time k. A recursive form of the posterior
probability of Hi is given by
 
1 φ1k−1 f zk |H1 , z k−1
p k = 1 , p0k = 1 − p1k (7)
φi
i=0 k−1 f (z k |H i , z k−1 )
 
φ1k−1 = p1k−1 + π 1 − p1k−1 , φ0k−1 = 1 − φ1k−1 (8)

Then, the test statistic of the SSPRT is

 p1k f (zk |H1 , z k−1 ) Pk−1 + π  p


1
Pk = 0 = , P0 = 00 (9)
pk f (zk |H0 , z k−1 ) 1−π p0

These results were extended to a general case of testing multiple hypotheses in [10].
3.2 SSPRT-based Detector with Multiple Models
 
As seen from (9), the key to compute the test statistics
 is to obtain posterior probability p1k or likelihood function f zk |H1 , z k−1
of hypothesis H1 since p0k and f zk |H0 , z k−1 are known given distribution model H0 . In our formulation, the whole mea-
surement space is partitioned by distribution models H0 and H1 , where the space of H1 is again partitioned by M models
H11 , . . . , H1M . Assume the prior probability of H1i is known as well as the transition probability from H0 to H1i . Then
the multiple SSPRT provides probability estimate for each hypothesis H1i which naturally fits our problem. The likelihood
and probability for hypothesis H1 can be obtained, following [7], as given below.
1i

Define φ̂k  P H1i |H1 , z k−1 as predicted probability of hypothesis H1i over H1 , likelihood f (zk |H1 , z k−1 ) 
p(zk |H1 , z k−1 ), and probability p1k−1  p(H1 |z k−1 ). We have
1i M 1j
φ̂k = φ1i k−1 / φk−1 (10)
j=1

then probability of H1 is the sum of the probabilities of all models


M
M M
p1k = P H1 |H1i , z k−1 P {H1i |z k−1 } = P {H1i |z k−1 } = p1i
k (11)
i=1 i=1 i=1

and the likelihood of H1 can be obtained as


M
M 1i
f (zk |H1 , z k−1 ) = f (zk |H1i , z k−1 )P H1i |H1 , z k−1 = f (zk |H1i , z k−1 )φ̂k (12)
i=1 i=1

where predicted probability φ1j 1i


k and probability pk of hypothesis H1i at time k are directly obtained by SSPRT recursions.
1 1i
Equation (11) shows that pk ≥ max pk , so does the corresponding probability ratio over H0 . In the sense of maximum
i
posterior probability (MAP) test, or equivalently minimum decision-error test, this supports that our formulation is better
than the traditional M-ary testing for maneuver detection, which usually makes decision based on a hypothesis that has the
maximum test statistics.

3.3 Multiple Model Detection of Target Maneuvers


Note that the SSPRT assumes that the measurements are independent, which is usually not the case for the target tracking
problem (1)–(2). Fortunately, the measurement residual sequence is weakly coupled and approximately  Gaussian distributed
under some conditions [6]. Thus measurement residuals should be used to compute likelihood z̃κ |Hi , z κ−1 instead of
measurements. Denote z̃k and Sk the measurement residual and its covariance, provided at time k by a nonmaneuvering
Kalman filter for the system (1)–(2).
According to (9), the key to obtain test statistics is to determine the hypothesis probability pik or the marginal likelihood
of each hypothesis. Under the linear Gaussian assumptions, the marginal likelihood of H0 is then determined by
  1 1  −1
f z̃k |H0 , z k−1 = N (z̃k ; 0, Sk ) = e− 2 z̃k Sk z̃k (13)
|2πSk |
 
and the marginal likelihood f z̃k |H1i , z k−1 for H1i are
  1 1  −1
f z̃k |H1i (u), z k−1 = e− 2 (z̃k −HGu) Sk (z̃k −HGu) , u ∼ f1i (u) (14)
|2πSk |
Then 
     
f z̃k |H1i , z k−1
= E f z̃k |H1i (u), z k−1
= f z̃k |H1i (u) , z k−1 f1i (u) du (15)
 
If the prior f1i (u) is Gaussian or uniform, f z̃k |H1i , z k−1 can be obtained in analytical forms given in [15]. Otherwise
it is rarely possible to analytically calculate the integral in (15). To overcome this difficulty, a fairly general way was proposed
to approximately evaluate this expectation integral off-line based on a Gaussian mixture of the prior PDFs [15], that is

N
f1i (u) ≈ λj N (u; ūj , Λj ) (16)
j=1
N
where N is the number of mixture components, λj are their weights (0 < λj < 1, j=1 λj = 1) and ūj , Λj are their means
and variances, respectively.
N
Proposition If f1i (u) is a Gaussian mixture f1i (u) = j=1 λj N (u; ūj , Λj ) then the marginal likelihood
f (z̃k |H1i , z k−1 ) is given by
  N
f z̃k |H1i , z k−1 = λj Ij (z̃k ) (17)
j=1

where Ij (z̃k ), j = 1, . . . , N are given explicitly



Ij (z̃k ) = N (z̃k ; HGu, Sk )N (u; ūj , Λj )du = N (z̃k ; HGūj , Sk + HGΛj (HG) ) (18)

A proof is given in [15]. Thus the test statistic of the SSPRT detector with multiple models can be obtained.

4 Simulation
Two simulation examples were designed for performance evaluation. The first one is simple—it just illustrates the idea.
It considers known Gaussian distributions fi (z) under different hypotheses Hi with a white noise measurement sequence.
The second example considers application of the proposed MM detection scheme for a realistic target maneuver detection
scenario. The threshold used by the detector in both scenarios was determined via simulations so as to have a false alarm rate
of Pf a = 1%.

4.1 Performance Indices


In order to evaluate the performance of the algorithms, the following measures were used: false alarm, missed detection,
average onset detection delay (n̂ − n), and relative computation time. All results are averages over 100 Monte Carlo runs.

4.2 Example 1
We consider three different hypotheses

H0 : z ∼ N (0, 1)
H1i : z ∼ N (i, 1) 1≤i≤2

A white Gaussian measurement sequence was simulated with a bias of 1 or 2 plugged in starting at n = 10 corresponding
to a change in the hypothesis from H0 to H11 or H12 . The measurement sequence was generated as

z =s+w (19)

where w ∼ N (0, 1) and s = 0, 1 or 2. The a priori probabilities of H1i were assumed to be 0.01, i = 1, 2, 3.
In Scenario 1, the bias was generated from the distribution H11 . The posterior hypothesis probabilities are shown in
Fig. 1 (a). Clearly for t = 15, 16, 17 approximately, H0 beats H11 and H12 individually but loses to H1 . This indicates
that the multiple-model based detector provides a quicker decision compared with traditional the M-ary hypothesis testing if
maximum posterior probability test is used. Due to the fact that H11 is closer to H0 than H12 is, the likelihood of H11 is
larger than that of H12 before the bias is inserted, which is also true for the probabilities. The bias in Scenario 2 was generated
from the distribution H12 . The probabilities are shown in Fig. 1 (b). Compared with Scenario 1, the larger bias results in a
quicker detection. Detection results are listed in Table 1.

Table 1: Detection Results

n̂ − n False Alarm Miss Detection


Scenario 1 5.10 1.0% 2%
Scenario 2 1.63 1.1% 0%
Average Probabilities of Hypotheses Average Probabilities of Hypotheses
1 1

H0
0.9 H11 0.9
H12
H1
0.8 0.8

0.7 0.7

0.6 0.6
Probability

Probability
0.5 0.5
H0
H11
0.4 0.4 H12
H1

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Time Time

(a) Change from H0 to H11 (b) Change from H0 to H12

Figure 1: Bias change in a Gaussian measurement sequence at n = 10

4.3 Example 2
The proposed detection scheme was applied to detect target maneuvers. The performance was evaluated on typical target
maneuver scenarios and compared with that of two traditional detectors.

4.3.1 Target maneuver model


The model of 2D target maneuver motion is given by

xk+1 = Fk xk + Γk (xk )ak + wk (20)

where the state in Cartesian coordinates is x = [x, ẋ, y, ẏ] , and the acceleration is a = [at , an ] with decomposed tangential
and normal components at and an , respectively. A target maneuver is described in this model through the input Γk (xk )ak .
The system matrices are given in [15].
In [3, 9], the normal acceleration an (t) was modeled as an asymmetric function an (t) = α + βeγb(t) and its marginal
probability density function was derived in [15], as
1 − 1 (ln anβ−α )2
f (an ) = √ e 2γ 2 (21)
2π |γ(an − α)|

where α, β, γ are design parameters, depending on the particular target type and b(t) is a zero-mean first-order Gauss-Markov
process. Fig. 2 includes two typical choices of α, β, γ.
A Gaussian marginal PDF model was used to model the tangential acceleration in this paper
2
1 (a −ā )
− t 2t
f (at ) = N (at ; āt , σ 2t )  √ e 2σ t
(22)
σ t 2π

where parameters āt , σ 2t are specified a priori (e.g., with some nominal values for the targets of interest).
Under the assumption of independence between tangential and normal maneuvers, the PDF of the total acceleration
vector is f (a) = f (at , an ) = f (a

t )f (an ). It is also assumed that direct position measurements of the target are available,
i.e., H =diag 1 0 , 1 0 in (2).
The maneuver model (20) is nonlinear   nonlinear dependency of Γk (xk ) on xk [15]. Consequently, in order to
due to the
compute the marginal likelihoods f z̃k |H1i , z k−1 given by (15) and (14), the implementation of the MMSSPRT detector
(1)
in this example used Gk = Γk (xk )|x =x̂(1) where x̂k|k−l denotes the l time-step state prediction via the maneuver model
k k|k−l
(20) started at time k − l with the estimate x̂k−l|k−l provided by the nonmanever Kalman filter and ak = āk = E[ak ]. Note
that the nonmanever Kalman filter (i.e., with ak = 0) is the only filter which is run in the MM maneuver detection algorithm.
The number of predictions l is a design parameter. In the simulations it was set to l = 3. A further enhancement of this
0.5

0.45 α, β, γ = (4,−2,0.5)

0.4

0.35

0.3
α, β, γ = (8,−4,0.5)

Pr(a)
0.25

0.2

0.15

0.1

0.05

0
−2 −1 0 1 2 3 4 5 6 7 8
a[g]

Figure 2: Asymmetric PDFs of normal accelerations

detection algorithm that avoids the permanent prediction at each time-step k is to enable the above prediction procedure only
if a threshold, lower than PT , is exceeded. This scheme amounts to a two-stage detection logic and it can significantly reduce
the computational load of the algorithm.

4.3.2 Ground-truth model


The standard 2D curvilinear-motion model

ẋ(t) = v(t) cos φ(t) + wx (t) (23)


ẏ(t) = v(t) sin φ(t) + wy (t) (24)
v̇(t) = at (t) + wv (t) (25)
φ̇(t) = an (t)/v(t) + wφ (t) (26)

was used to generate the ground truth trajectories, where (x, y), v, φ denote the target position, speed and heading. This model
is fairly general since it accounts for along- and cross-track accelerations. The initial states x0 = [60km, 90km, 300m/s, 30deg]
was used in scenarios. The target track starts from k = 40 and continued for another 40 sec. The sampling time T = 0.3s.
In order to test detectors on multiple maneuver scenarios simultaneously, a random maneuver scenario was designed,
where the magnitude of the normal acceleration during k = [65, 80] for each realization was random: an ∼ f (an ) with
α = 8, β = −4, γ = 0.5.

4.3.3 Simulation results


Two normal acceleration models, according to (21), were used to represent two types of target maneuver: f11 (an ) with
α = 4, β = −2, γ = 0.5, and f12 (an ) with α = 8, β = −4, γ = 0.5. Their Gaussian sum approximations (16) obtained by
the local fit technique described in [15] are, respectively

f11 (an ) ≈ 0.44N (an ; 0.65g, (1.3g)2 ) + 0.56N (an ; 2.2g, (0.6g)2 )
f12 (an ) ≈ 0.57N (an ; 3g, (1.9g)2 ) + 0.43N (an ; 5g, (g)2 )

The distributions of the tangential accelerations are f11 (at ) = f12 (at ) = N (at ; 0, (0.5g)2 ). For the SSPRT detector, the
prior probability for H1 being true was 0.01 and transition probability π from H0 to H1i was 0.005. Two point differencing
was used to initialize the Kalman filter in state estimation. The corresponding noise covariances were Q = (1.8)2 I and
R = (11)2 I.
In the sequel, MMSSPRT denotes the proposed detector. Its detection performance was evaluated in comparison with two
traditional maneuver detectors. MR denotes a detector based on measurement residuals, and IE represents one based on the
input estimation technique. Details for the MR and IE detectors can be found in [8, 14]. The size of data window used in IE
and MR detectors was 5.
Average Probabilities of Hypotheses
1

H0
0.9 H11
H12
H1
0.8

0.7

0.6

Probaility
0.5

0.4

0.3

0.2

0.1

0
40 45 50 55 60 65 70 75 80
Time

Figure 3: Change from nonmaneuver to maneuver at k = 65

The hypothesis probabilities for Scenario 4 is shown in Fig. 3. It is seen that H0 beats H11 and H12 individually but loses
to H1 for k = 71, 72, 73. Note that the posterior probability of H12 keeps increasing due to the accelerations acting on the
target. Detection performance comparison is given in Table 2.

Table 2: Detection Performance Comparison

n̂ − n False Alarms Miss Detection


MMSPRT 4.88 1% 0%
MR 5.50 1% 0%
IE 5.11 1% 0%

The ROC curves were generated by computing Pd at the time of interest with different Pf a using 100 Monte Carlo runs.
The ROC curves for each maneuver onset detector at k = 68 and k = 69 are given in Fig. 4 (a) and (b), respectively. It is
seen that overall MMSSPRT outperforms the two other detectors. The simulations also showed that the IE has a competitive
performance for really very small Pf a as time goes due to the improved accuracy of the input estimate as more maneuver data
become available. Compared with the two other detectors, MMSSPRT has a potential to improve its detection performance
by further fine-tuning of some design parameters, e.g., the a priori probabilities and the interval of computing the predicted
estimate under H1 .
ROC Curves of Detectors at k=68 ROC Curves of Detectors at k=69
1 1

0.9 0.9

0.8 0.8

0.7 0.7 MMSSPRT


MMSSPRT MR
MR IE
IE
0.6 0.6
Pd

Pd

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Pfa Pfa

Figure 4: ROC curves of all detectors


5 Conclusions
In this paper, a multiple-model hypothesis testing approach has been proposed for detection of unknown target maneuvers.
The multiple-model maneuver detector developed is based on sequential hypothesis testing. Performance evaluation and
comparison of the proposed MM maneuver detector with other detectors have been conducted by simulation. The simulation
results have demonstrated that the new detector outperforms traditional maneuver detectors. Future study includes exploration
of techniques for efficient mixture weight estimation, which will be very useful for cases where the prior information is
limited, and when the hypotheses are mismatched with the truth.

References
[1] Y. Bar-Shalom and X. R. Li. Estimation and Tracking: Principles, Techniques, and Software. Artech House, Boston, MA, 1993.
(Reprinted by YBS Publishing, 1998).
[2] T. S. Ferguson. Bayesian Density Estimation by Mixtures of Normal Distributions. In M. H. Rizvi, J. Rustagi, and D. Siegmund,
editors, Recent Advances in Statistics, pages 287–302. Academic Press, 1983.
[3] J. D. Kendrick, P. S. Maybeck, and J. G. Reid. Estimation of Aircraft Target Motion Using Orientation Measurements. IEEE Trans.
Aerospace and Electronic Systems, 17:254–260, Mar. 1981.
[4] T. L. Lai. Sequential Multiple Hypothesis Testing and Efficient Fault Detection-Isolation in Stochastic Systems. IEEE Trans. on
Information Theory, 46(2):595–608, Mar. 2000.
[5] T. L. Lai and J. Z. Shan. Efficient Recursive Algorithms for Detection of Abrupt Changes in Signals and Controls Systems. IEEE
Trans. on Automatic Control, 44(5):952–966, May 1999.
[6] X. R. Li. Model-Set Adaptation in Variable-Structure MM Estimation by Hypothesis Testing. In Proc. 1998 American Control Conf.,
pages 2439–2443, Philadelphia, Pennsylvania, USA, June 1998.
[7] X. R. Li. Multiple-Model Estimation with Variable Structure—Part II: Model-Set Adaptation. IEEE Trans. Automatic Control,
AC-45(11):2047–2060, Nov. 2000.
[8] X. R. Li and V. P. Jilkov. A Survey of Maneuvering Target Tracking—Part IV: Decision-Based Methods. In Proc. 2002 SPIE Conf.
on Signal and Data Processing of Small Targets, vol. 4728, Orlando, Florida, USA, April 2002.
[9] X. R. Li and V. P. Jilkov. Survey of Maneuvering Target Tracking—Part I: Dynamic Models. IEEE Trans. Aerospace and Electronic
Systems, AES-39(4):1333–1364, Oct. 2003.
[10] D. P. Malladi and J. L. Speyer. A Generalized Shiryaev Sequential Probability Ratio Test for Change Detection and Isolation. IEEE
Trans. Automatic Control, AC-44(8):1522–1534, 1999.
[11] I. Nikiforov. A Simple Recursive Algorithm for Diagnosis of Abrupt Changes in Random Signals. IEEE Trans. Information Theory,
46(7):2740–2746, 2000.
[12] I. Nikiforov. Optimal Sequential Change Detection and Isolation. In Proc. 15th IFAC Word Congress, Barcelona, Spain, July 2002.
[13] I. Nikiforov. A Lower Bound for the Detection/Isolation Delay in a Class of Sequential Tests. IEEE Trans. Information Theory,
49(11):3037–3047, 2003.
[14] J.-F. Ru, A. Bashi, and X. R. Li. Performance Comparison of Target Maneuver Onset Detection Algorithms. In Proc. 2004 SPIE
Conf. on Signal and Data Processing of Small Targets, vol. 5428, Orlando, FL, USA, Apr. 2004.
[15] J.-F. Ru, V. P. Jilkov, X. R. Li, and A. Bashi. Sequential Detection of Target Maneuvers. In Proc. 2005 International Conf. on
Information Fusion, Philadelphia, PA, USA, Jul. 2005.
[16] J. L. Speyer and J. E. White. Shiryaev Sequential Probability Ratio Test for Redundancy Management. AIAA Journal of Guidance,
Control, and Dynamics, 7(5):588–595, 1984.
[17] D. M. Titterington, A. F. M. Smith, and U. E. Makov. Statistical Analysis of Finite Mixture Distributions. John Wiley & Sons, 1985.

You might also like