2005 7th International Conference on Information Fusion (FUSION)
Sequential Detection of Target Maneuvers*
Jifeng Ru Vesselin P. Jilkov X. Rong Li A. Bashi
Department of Electrical Engineering
University of New Orleans, New Orleans, LA 70148, USA
{jru, vjilkov, xli } @uno.edu, (504) 280-{7416, 3950 (fax)}
Abstract - This paper addresses target maneuver onset de-
tection based on sequential statistical tests. Cumulative
sums (CUSUM) type and Shiryayev sequential probability ratio
(SSPRT) tests are developed by using a likelihood marginaliza-
tion technique to cope with the difficulty that the target maneu-
ver accelerations are unknown. The approach essentially utilizes
a priori information about the maneuver accelerations in typical
tracking engagements and thus allows to improve detection per-
formance as compared with traditional maneuver detectors. Sim-
ulation results are presented that demonstrate the capabilities of
the maneuver detectors developed.
Keywords: Target Tracking , Maneuver Detection, Hypothesis
Testing, SPRT, CUSUM
1 Introduction
Most maneuver detectors developed for maneuvering target
tracking are based on either the chi-square significance test
or the generalized likelihood ratio test [5]. Six such maneu-
ver onset detection algorithms have been studied recently
in [12] using different scenarios. Another class of statis-
tical tests that can be applied to maneuver detection is the
sequential tests for change point detection, as pointed out in
[5, 12]. Sequential detection procedures have been success-
fully applied to fault detection (see, e.g., [8, 15]) but not to
maneuver detection, to our knowledge, except for a quick-
est detector in [16]. Some solutions based on min-max ap-
proach have been proposed in [9, 10, 11] with applications
for navigation system integrity monitoring. It optimizes the
worst case situations with given decision error rates.
For maneuver detection, sequential testing procedures
are actually preferable because measurements are available
sequentially. Moreover, a sequential test does not need
to determine the sample size in advance unlike the non-
sequential tests. A sequential test involves a stopping rule
and a final decision to achieve a trade-off between sam-
ple size and decision accuracy. We consider detecting a
target's maneuver as a binary hypothesis testing problem.
Once a target starts maneuvering, it should be detected as
quickly as possible under certain constraints such as the
rates of false alarms and missed detections. It is well known
that for binary simple hypothesis testing, Wald's sequential
*Research supported in part by ARO grant W911NF-04-1-
0274, and NASA/LEQSF grant (2001-4)-01.
0-7803-9286-8/05/$20.00 ©2005 IEEE
Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on January 30, 2009 at 09:14 from IEEE Xplore. Restrictions apply.
probability ratio test (SPRT) is optimal in the sense that it
makes the quickest detection under both hypotheses (Ho,
H1) given any fixed false alarm and missed detection prob-
abilities. The SPRT makes the decision by comparing the
likelihood ratio with two thresholds, which are obtained
to guarantee the specified error probabilities. Recently
the Shiryayev sequential probability ratio test (SSPRT) has
been applied to the fault detection problem. The SSPRT
is optimal in the sense that it provides the quickest detec-
tion of a change in a sequence of conditionally indepen-
dent measurements under the given decision error rates.
The SPRT and SSPRT are different in problem formula-
tion. In particular, the SPRT assumes all data relates to
one of the two hypotheses so that it simply chooses one of
the hypotheses that there is a maneuver or that there is no
maneuver. The SSPRT does not make this assumption and
tests for an occurrence of a change. One note to empha-
size is that both SPRT and SSPRT hold the optimality only
when there is no unknown parameter, which is often not the
case for maneuver detection where maneuver magnitude is
usually unknown. In order to apply these sequential tests,
modifications have to be made.
This paper presents two target maneuver onset detectors
based on sequential tests. Cumulative Sums (CUSUM) type
and SSPRT tests are developed by using a likelihood mar-
ginalization technique to cope with the difficulty that target
maneuver accelerations are unknown. The proposed ap-
proach essentially utilizes a priori information about the
maneuver accelerations in typical tracking engagements
and thus allows to improve the detection performance, es-
pecially for normal accelerations, as compared with two
widely used maneuver detectors.
The rest of the paper is organized as follows. In Sec.
2, the problem of maneuvering target detection is formu-
lated as binary composite hypothesis testing. In Sec. 3,
sequential detection procedures based on a repeated-SPRT
and SSPRT are formulated and discussed. In Sec. 4, test
statistics are derived by means of likelihood marginaliza-
tion using two typical prior models of maneuvers. The per-
formance of the proposed maneuver detectors is evaluated
by simulations and compared with that of two widely used
detectors in Sec. 5. Conclusions are provided in Sec. 6.
345
 2 Problem Formulation Clearly, the standard SPRT presented above cannot be
applied directly to maneuver onset detection as formulated
The target-measurement model is given by by (3). We modify this test to address several issues next.
First, the SPRT assumes that measurements are indepen-
Xk+l = FkXk + Uk + Wk
dent. However, measurements are correlated in the target
Zk = HkXk +Vk, k = 1,2, ... tracking problem (1)-(2). In this case the above test still
works provided the sequence (1k) of marginal likelihood
where Xk is the target state, Uk is the maneuver control ratios is independent, where
input, and Zk is the measurement. Wk N K (0, Qk) and
Vk rv K (0, Rk) are independent process and measurement
noises, respectively, and the initial state xo K (0, Po) is f1 f (jHo,zK-1)
independent of Wk and Vk.
It is assumed that uk 0 when the target is not maneu- Fortunately, this is approximately the case since the mea-
vering at time k and Uk =L 0 when the target is maneuvering. surement residual sequence is approximately Gaussian dis-
If the target begins a maneuver at an unknown time n < k tributed and weakly coupled under some conditions [4].
then Thus measurement residuals, instead of measurements,
should be used to compute likelihood ratios .
(Uk) ={. .O. O. . . OOUn ±1Un+l ... ,Uk}
Second, the SPRT only decides which hypothesis is to
In general, it is not necessary for u to remain constant accept. Once H0 or H1 is accepted the test will terminate.
during the maneuver. The focus of the maneuver onset de- In fact, wethis
However, is not the goal of maneuver onset detection.
tection is to decide on a maneuver and estimate the onset
want to know "when H1 occurs". Therefore,
time n, which can be formulated as a binary hypothesis test- the test should continue to the next cycle with more mea-
ing problem: surements if Ho is deemed true, that is, we should keep
restarting (resetting) the SPRT as long as Ho is accepted.
Ho: U-n 0 for m 11 k ...
The standard CUSUM algorithm was first proposed by Page
for classical
I

0 for m :11 1 change detection problems. It guarantees the

H .| Um =
n
quickest
...

decision given the decision error rate for simple

,

1 *|um r 0 for m n, k ... I

hypotheses, i.e., if u is known. Page's test can be inter-

where u and n are unknown parameters, and n is referred preted as a repeated SPRT with the lower threshold logA
to as the change point. This is known as change point de- equal to 0 and the upper threshold equal to A determined
tection in the statistical literature. by error probabilities. The key idea is to restart the SPRT
Since both u and n are usually unknown in practice, hy- algorithm once Ho is accepted, which makes it naturally fit
pothesis H1 is clearly composite. This makes the maneuver our problem. The CUSUM algorithm [1] can be written in
detection difficult since in general there is no existing opti- a recursive manner:
mal non-Bayesian solution for composite hypothesis testing
problems. Lk =max {Lkl1 + log f( O}, Lo 0 oHjz_-1)
(6)
3 Sequential Detection of Maneuvers and the decision rule is
3.1 Repeated SPRT-based detector
1) Accept H1 (declare maneuver) if Lk > A. Then stop-
Before proposing sequential test procedures for target ma- ping time n = min{k : Lk > A} is the time maneuver
neuver detection, we briefly present the SPRT for binary is detected.
simple hypothesis testing. Let
2) Continue the test (k ~-* k + 1) if Lk < A.
E
Lk -logof(Zk
f (Zk Ho)
H) log
*c_-1 g(loz
f (ZnIHi,zK-l) Compared with the standard SPRT, it is shown that
two tests differ only in the value of the lower thresholds:
denote the log-likelihood ratio of two simple hypotheses SPRT uses A determined by two error probabilities while
H1 and Ho based on the measurements up to k, where CUSUM uses which was also formally proven for op-
0 its
timality.
ZA-( Zl *...*, Zk ), f IZ Hi ) and f (Z,;, Hi, zlt1)( =O Third, the SPRT is optimal only for simple hypothesis
are the joint and marginal likelihoods of the hypotheses, re-
spectively. Further, let A and B be two thresholds such that testing, which is obviously not the case for maneuver detec-
tion since maneuver input u is usually unknown. This
P("H11"IHo) < a,) P("Ho" JH1) < ,l (O < a,,3 < 1), makes the the marginal likelihood ratio in the test statistics un-
where "Hi" stands for "deciding on Hi". Then the SPRT
decision rule is
known. There are two common approaches td solving this
problem. In the first one, the generalized likelihood ratio
Accept H1 if Lk > log B (GLR) test, the unknown u (treated as nonrandom) is re-
Accept H0 if Lk < logA placed with the maximum likelihood estimate ui (see [5]).
Continue (k k + 1) otherwise
F-* The second technique is a Bayesian approach, where u is
346

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on January 30, 2009 at 09:14 from IEEE Xplore. Restrictions apply.
 treated as random and the likelihood f(Pk IH) is deter- probability of Hi is given by
mined by an appropriate marginalization-averaging over
all possible values of the unknown u [2]. Specifically,
Pk = 1 Olf(k1H1Zk-1)) I Pk 1 pk (9)
ff (zklHi, z k ) = E [f (4 H, (u) , zk )] ZlP-1 +PVVfiIHz
fk-1 T(1-Pk1) 00-1 =1
1
I
-

k-1 (10)
I f (Zk H1 (u) , zk -1) f (u) du (7)
Then, the test statistic of the SSPRT is
where f (u) is the a priori probability density function A Pk f(k Hl,zk1) Pk-,±7 + A (11)
(PDF) of u. Pk f(ZEk|Ho,Zk-1) 1-7T ' ° p
Clearly, this approach provides a good means of utilizing
the a priori information if it is available. A major diffi- and the SSPRT becomes
culty for its application is how to evaluate the integral in (7)
which is rarely possible analytically. In Sec. 4 we will use 1) Accept H1 (declare maneuver) if Pk > PT
typical prior distributions of target accelerations and pro-
pose a fairly general way to approximately evaluate this ex- 2) Continue the test (k F - k + 1) if Pk < PT.
pectation integral based on a Gaussian sum approximation
of the prior PDFs. Note that no reset mechanism is necessary for the SSPRT
due to its nature of the problem formulation that determines
when a disruption H1 is true. Its recursive nature and the
3.2 SSPRT-based detector problem formulation make it well suited to maneuver de-
tection.
The SSPRT focuses on the detection in a series of condi-
tionally independent measurements by noting the change in
the probability density function of the measurement. Good 4 Maneuver Detection Algorithms
results have recently been reported for fault detection us- Development
ing SSPRT [8, 151, however we are not aware of any work
that addresses the adaptation and effectiveness of SSPRT 4.1 Maneuver model
for maneuver detection. A 2D target maneuver acceleration vector can be decom-
The SSPRT minimizes an expected cost at each time step. posed along two directions: tangential (along the veloc-
This cost includes the measurement cost and the cost due ity vector) and normal (orthogonal to the velocity vector).
to a terminal decision error by false alarm or miss detec- Such a target maneuver motion can be described as
tion. The decision function can be written in terms of a
likelihood ratio or posterior probability, which are equiva- Xk+l = FXk + F(Xk)ak + Wk (12)
lent once a prior is known. Since a Bayesian framework
is used, it is necessary to define prior information. This where the state in the Cartesian coordinates is x =
includes the a priori probabilities of Hi and the transition [x,I , y, y]', and the acceleration is a = [at, a,]' with tan-
probabilities of Ho to Hi from k - 1 to k. When the ma- gential and normal components at and a,, respectively.
neuver onset detection is formulated as binary hypothesis The system matrices are
testing, the transition probabilities degenerate into a single
number, assumed to be time-invariant for simplicity. F = diag {F2, F2}, F(x) -GT(x), G = diag {G2, G2}
The decision rule of the SSPRT is obtained by defining
the posterior probability ratio Pk. Let Pk = P{n < klzk} F2 = -s(x) 1
denote the posterior probability that a maneuver starts at theG2 T(
i ) i(x) c(x) j
an unknown time n at or before time k given the available where T(x) is the rotation matrix with
measurements zk, and
c(x) =cos4 *+ s(x) sin X =
PkPkA- 11_Pk ,PT-=1 PT(8
PT X2 ± S'2
(8)
-Pk 1-PT
that maps ak to the Cartesian coordinates, and q =
where PT is a preset threshold. Then the SSPRT detector arctan(*/S') is the target heading angle. A target maneuver
either declares a maneuver if Pk > PT or takes another is described in this system model through the control input
measurement if 0 < Pk < PT. The choice of PT is re- Uk = (xkk)ak. Different models of the acceleration ak can
lated to the desirable decision error rate. More details and be used [6] depending on the maneuver capabilities of the
examples are discussed in [8, 15]. targets of interest in the tracking application. To be more
Clearly, calculating the posterior probability is the key to specific we develop the maneuver detectors for manned ma-
SSPRT. Fortunately, Pk can be calculated recursively. Let neuvering aircraft in mind.
pO denote the prior probability of hypothesis Hi being true The normal acceleration is induced by the lift forces and
(i = 0,1), 7r the transition probability from Ho to H1, and is usually the dominant one during maneuver. Its direction
A-P(n < k+ ltzk). The recursive form of the posterior is determined by the target aspect angle and its magnitude
347

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on January 30, 2009 at 09:14 from IEEE Xplore. Restrictions apply.
 For the particular implementation of the maneuver de-
tectors, in the sequel it is assumed also that direct
position measurements of the target are available, i.e.,
H diag{[ 1 0 ] 41 0 ]}in(2).

4.2 Test statistics

As discussed in Sec. 3, the key to computing test sta-
tistics of the detectors is to obtain likelihood functions
f(Zk IH, Zk-l) under each hypothesis (i 0O, 1).
Denote by Zk and Sk the measurement residual and
its covariance, respectively, which are provided at time
k by a nonmaneuver Kalman filter (i.e., u, = 0, Ki
Fig. 1: Asymmetric PDFs of normal acceleration. 0, 1, ..., k-1)forthesystem(1)-(2).
can be modeled as a colored random process with an asym- The marginal likelihood of Ho is then (under the linear
metrical distribution. It was proposed in [3] that the nor- Gaussian assumptions)
mal acceleration an (t) be modeled as an asymmetric func-
tion an(t) aa + /3eYb(t) where a, 3, -y are design para- f(4k Hozk k-1)
Z 1) AN(2k;,OSk) WS-~
1
e-2k
1
Sk Zki
meters, depending on the particular target type and b(t) is a 27r"Sk
zero-mean first-order Gauss-Markov process. The marginal (15)
probability density function, essential for the further devel- To obtain the likelihood f (ikHl, zk-l) of H1, we im-
opment of the maneuver detector, can be easily derived (see plement the marginalization approach (7) discussed in Sec.
Appendix A) as 3 with the acceleration PDF models (13)-(14). Specifically,
we write f(ktHl,zk 1) = f(ikIH1lXk-ljk-1). Then,

f (an)
f(n)- e ~~1 e- 1(In an -a )2 (13) according to (7) and (12) we havel
2-1(an- a)le
For three typical choices of a,/3, -y, this highly asym-
metrical density is shown in Fig. 1, including one with
(ao, 3, -y) = (8, -4,0.5) that is considered typical of mod-
em piloted aircraft in evasive maneuvers. This model is
more accurate than the usual symmetrical models [6] (e.g.,
the Singer model) at the cost of designing these parameters,
which requires knowledge of the target type, obtained ei-
ther a priori or a posteriori (e.g., with the help of an image-
sensor-aided target classification).
The tangential acceleration is determined by the thrust-
minus-drag force. Various random process models for
its magnitude are discussed and analyzed in detail in [6].
Most important for our maneuver detection approach is
the choice of marginal PDF of the process. Purely a pri-
ori symmetrical models (such as the ternary-uniform den-
sity in the Singer model) or a posteriori, adaptive (such as
the conditional Rayleigh density in the mean-adaptive "cur-
rent" model) are possible candidates. In this paper we use
Gaussian marginal PDF model
(at-at )2
f(at) = V(at; at, o2) e 2ct
with parameters at, rt2 either specified a priori (e.g., with
some nominal values for the targets of interest) or deter-
mined a posteriori through filter-based estimates at =
tk
a~tkklk I~~t2 7t~kiivetiatd
2-
* Both cases were investigated. It
should be noted that the Gaussian sum approximation tech-
nique used to cope with the non-Gaussian f(a,) in the
next section can be also applied to f(at) if a non-Gaussian
model is adopted.
Under the assumption of independence between tangen-
tial and nonnal maneuvers, the PDF of the total acceleration
vector is f (a) = f (at, a,,) = f (at) f (a,,).
f (iIH,I x^) = E[f (IH,Ix( a)]
M(i; H:F(&)a, S) f (a)da
=A
|JV(i; Hr'(x')[at, a,,]', S)f (at)f(an,)datdan,
(16)
where f(at) and f (an) are given by (13) and (14), respec-
tively.
However, due to the complex prior density f(a,), the
integral in (16) has no analytical solution available, which
makes the implementation of the proposed detectors highly
complicated. To overcome this difficulty we propose to em-
ploy the Gaussian sum approximation techniques (see, e.g.,
[14]). That is, we use
N
f1(an)
1E Ai(an; n n (17)
i=1
where N is the number of Gaussian components, Ai are
their weights (0 < Ai < 1, EN Ai- 1) and a(i), a (i)2
(14) are their means and variances, respectively. The determina-
tion of these parameters is a part of the design and does
not require on-line approximation. For our implementa-
tion of the maneuver detectors we obtained a fairly accu-
rate approximation of the asymmetric f (an) of (13) by a
sum of only two components, which is illustrated in Fig. 2.
We used standard Matlab functions for nonlinear multidi-
mensional optimization to obtain a locally best fit. Other
expectation maximization (EM) based mixture estimation
techniques can also be used. Higher accuracy can be ob-
tained by using more components of the sum. In general,
'We drop the time index to simplify the notation.
348

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on January 30, 2009 at 09:14 from IEEE Xplore. Restrictions apply.
 The initial states used for each scenario were generated
as
xO 60km, 5km
Yo0 2Af(Yo,
O 2) )
X(OXr Yo 90km, U0~V 5km
V0 300m/s, 5m/s
vo -A(¢'o,avo ) 45 deg, ldeg
with
a = 5m, w= 5m, av = 1m/s,8a1(Tw= 0.01 deg
The scenarios included deterministic and random ones.
The target track length starts from k = 30 and continued
for another 70sec. The sampling time T ls.
Fig. 2: Gaussian sum approximation of the asymmetric (DN) Deterministic normal acceleration scenario. The
PDF target makes a normal maneuver during the period k =
under some regularity conditions, a non-Gaussian density [80, 100] with a magnitude of 20 m/s2.
function can be approximated to any desired accuracy by a (DT) Deterministic tangential acceleration scenario. The
weighted sum of Gaussians provided the number of com- target makes a tangential maneuver during the period k =
ponents is sufficiently large [14]. This makes the approach [80, 100] with a magnitude of 20 m/s2.
generally applicable when a marginalization is hard to ob-
tain in an exact analytical form. (RN) Random normal acceleration scenario. In order to
Now that the approximation (17) is available evaluating test detectors on multiple maneuvering scenarios simul-
the likelihood (16) reduces to taneously, a random maneuver scenario was developed,
where the magnitude of the normal acceleration was fixed
N during k [80, 100] but random over runs: an f(an) as
=

f zHi x) =I Ai-Ti (i; (18) given by (13) with (a, 3, y) = (4, 2, 0.5).
5.2 Simulation results
where
There were some parameters to be designed. The thresh-
old for each detector was determined by simulations with
IiP(; X) = |X(i; Hr (')[at, a,,,]' S)Ar(at; at, at) given false alarm rate Pf a = 1%. Two point differencing
was used to initialize the Kalman filter in state estimation.
NV(an; d(i), a(i)2 )datdan (19) The corresponding noise covariances were Q = (0.5)21
As derived in the Appendix B, Ii(i; x) is explicitly deter- and R = (50)2j.
mined to be For the SSPRT detector, the prior probability for H1 be-
ing true was 0.01 and transition probability 7r from Ho
(z; x)-A(z; HF(&)a(i), S + HF(&)A(') (HIr())') (20) to H1 was 0.005. In addition, prior densities for test-
&I
ing normal accelerations were used by the CUSUM and
where a()- [at, a()]' and A(i) diag{f,2, r(i)2}. SSPRT detectors: f (at) = (at; 0, (0.5g)2 ) and f (a.) 1

0.44A/(an; 0.65g, (1.3g)2) + 0.56.A(an; 2.2g,(0.6g)2)

5 Simulation which were determined by local fit. Such design is obvi-
ously better to detect normal maneuvers since prior f(at)
Four scenarios were designed to compare different aspects will not affect normal accelerations much. It is possible
of the detection algorithms using ground truth trajectories that we design a model with prior densities particularly
generated by a curvilinear-motion model. All results are good for tangential accelerations. In this paper, we de-
averages over 100 Monte Carlo runs. sign f (at) = (at; 0, (3.5g)2) and f(a,) with zero mean
and very small variance to test tangential accelerations. We
5.1 Ground-truth model could have different design if we had better prior informa-
The standard 2D curvilinear-motion model tion for the tangential accelerations. For targets with both
maneuvers, naturally we consider combination with two or
x(t) = v(t) cos (t) + w. (t) (21) more models, which is under our study [13].
The CUSUM and SSPRT represent the two proposed de-
y(t) = v (t) sin 0(t) + wy (t) (22) tectors. The MR represents a detector based on measure-
iv(t) = at (t) + w, (t) (23) ment residuals. The IE represents one based on input esti-
+(t) = a.(t)/v(t) + wqs(t) (24) mation technique. Details for MR and IE can be found in
[5, 12]. The size of data window used in IE and MR de-
was used to generate the ground truth trajectories, where tectors was 5. The performance of four detectors was com-
(x, y), v, 0 denote the target position, speed and heading. pared in terms of average onset detection delay ( -n),
This model is fairly general since it accounts for along- and receiver operating characteristics (ROC) curves and com-
cross-track accelerations. putational load.
349

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on January 30, 2009 at 09:14 from IEEE Xplore. Restrictions apply.
 5.2.1 Average onset detection delay
The average onset detection delay under four scenarios is
shown in Table I. It is clear that SSPRT detectors have
smaller detection delay in both scenarios with normal ac-
celerations as well as tangential acceleration. For the cases
of tangential acceleration, the performance of the CUSUM
and SSPRT detectors should be improved If we have the 04 _
proper prior knowledge for tangential accelerations, .

Table 1: Average of maneuver onset detection delay

nt- n DN RN DT
CUSUM 5.33 5.57 5.75
SSPRT 4.68 4.69 5.10 Fig. 4: ROC curves of all detectors for scenario DN
MR 5.19 5.12 5.15
IE 5.04 5.03 5.11

Furthermore, the SSPRT detector also provides the pos-

terior probability of a maneuver. The posterior hypothesis
probability for scenario DN is shown in Fig. 3. We note
that posterior probability of H1 quickly increases after the
change time. Similar results were observed for scenario DT.
Ave-ge p bililites o Hypoteo DN

-1-1-1 ----------------
------ -, _' -'
-'.1.
F ---.-o
.'
I

Fig. 5: ROC curves of all detectors for scenario DT

Fig. 3: Posterior probability: normal accelerations change

at k = 80
5.2.2 ROC curves
The ROC curves for deterministic normal acceleration sce- Pt.

narios were generated by computing the Pd at the time of Fig. 6: ROC curves of all detectors for scenario DT
interest with different Pf a using 100 Monte Carlo runs. The
ROC curves for each maneuver onset detector at k = 83 Table 2: Computational complexity of different algorithms
and k = 84 for scenario DN are given in Fig. 4 and Fig. | CPU(s) 1 MR CUSUM SSPRT IE
5, respectively. It is apparent from the curves that SSPRT 2.97 3.27 7.61
has the best performance for tested scenarios. The CUSUM
detector outperforms IE and MR detectors for scenario DN MR detector which is the simplest. It shows that the pro-
with Pfa > 0.2. Note that IE performed better as time posed detection algorithms have a much less computational
load than the IE-based algorithms.
goes due to the improved accuracy of the input estimate
as more maneuver data become available. For the scenario
DT, shown in Fig. 6 for ROC curve at k = 83, IE and 6 Conclusions
SSPRT have compatible performance when Pf, is really In this paper, two sequential detectors for maneuver onset
small. Clearly, the ROC curves verify what we observed have been developed and evaluated over various scenarios
for maneuver onset detection delay. designed for different aspects of maneuvers. The proposed
detection algorithms are based on sequential likelihood ra-
5.2.3 Computational complexity tio test procedures, CUSUM and SSPRT, and use properly
The computational complexity of different algorithms were marginalized likelihoods to cope with the unknown magni-
compared by the ratio of the CPU processing time per iter- tude of maneuver accelerations - a key difficulty in maneu-
ation in Table II. All other algorithms were compared with ver onset detection. Both CUSUM and SSPRT detection
350

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on January 30, 2009 at 09:14 from IEEE Xplore. Restrictions apply.
 algorithms developed are explicit, recursive and general. It can be directly verified that
Utilizing prior distributions of the accelerations is an essen- C3- C2Cl1C2 ( - m)'(S + Am)-1(i - m)
tial task involved in the application of the proposed scheme.
Performance evaluation and comparison have demonstrated SCiAm = S + Am
that the proposed detectors are more effective than two pop- Thus
ular detectors when appropriate prior distributions of accel-
erations are known.
I27i SCAmI
Appendix A: Proof of (13) K(mI S + Am)
According to the following theorem [7]: the PDF fy(y) of References
Y = g(X) can be determined from the PDF fx (x) of X [1] M. Basseville and I. Nikiforov. Detection of Abrupt
by Changes: Theory and Application. Prentice Hall, Engle-
wood Cliffs, NJ, 1993.
fy (y) = fEf(xi) [2] F. Gustafsson. Adaptive Filtering and Change Detection.
Wiley, 2001.
where g'(x) = g(x) and xi's are the real roots of the [3] J. D. Kendrick, P. S. Maybeck, and J. G. Reid. Estimation
equation y = g(x) in terms of y: y = g(xi). of Aircraft Target Motion Using Orientation Measurements.
Let g(b) = a + oe`b. Then g'(b) = /3yeYb. The root of IEEE Trans. Aerospace and Electronic Systems, 17(2):254-
a. = a±+ eyb is bi = a ln a,. Thus we have (13) 260, Mar. 1981.
[4] X. R. Li. Multiple-Model Estimation with Variable
JVF(ly1n ad-3 c; 0,1) Structure-Part II: Model-Set Adaptation. IEEE Trans. Au-
fg'(bl)l 1an- a)jI tomatic Control, AC-45(11):2047-2060, Nov. 2000.
1 e .-
(In a -c )2 [5] X. R. Li and V. P. Jilkov. A Survey of Maneuvering Target
Tracking-Part IV: Decision-Based Methods. In Proc. 2002
v'7f 1y(an - a)I SPIE Conf on Signal and Data Processing of Small Targets,
vol. 4728, Orlando, Florida, USA, April 2002.
Appendix B: Proof of (20) [6] X. R. Li and V. P. Jilkov. A Survey of Maneuvering Target
According to (19) Tracking-Part I: Dynamic Models. IEEE Trans. Aerospace
and Electronic Systems, AES-39(4), Oct. 2003.
1 e I(i-Hra)'S-l'(-Hra)a)d [7] X. Rong Li. Probability, Random Signals, and Statistics.
-(iP, &) I V12w7S[e2f()d [8]
CRC Press, Boca Raton, FL, 1999.
D. P. Malladi and J. L. Speyer. A Generalized Shiryaev
I, -2
I
-Hra)'S-1(i-Hra) Sequential Probability Ratio Test for Change Detection and
Isolation. IEEE Trans. Automatic Control, AC-44(8):1522-
1 e (a-a)'A- 1
¢a)da 1534, 1999.
VIj2rA~je2 [9] I. Nikiforov. A Simple Recursive Algorithm for Diagnosis
of Abrupt Changes in Random Signals. IEEE Trans. Infor-
Let m = Hra, Am = HFA(Hr)', and m = Hra. Then mation Theory, 2000.
dm IIHIl da, and [10] I. Nikiforov. Optimal Sequential Change Detection and Iso-
lation. In Proc. 15th IFAC Word Congress, Barcelona, Spain,
July 2002.
[11] I. Nikiforov. A Lower Bound for the Detection/Isolation De-
e
-2[Za)'S- ( a) + (7n- n)'Am -' (ra- i-n
)lnn lay in a Class of Sequential Tests. IEEE Trans. Information
+/ 2r2SAm Theory, 2003.
[12] J.-F. Ru, A. Bashi, and X. R. Li. Performance Comparison
= 1 [~e-'fDdm
I(2W)2SAmI Dd of Target Maneuver Onset Detection Algorithms. In Proc.
2004 SPIE Conf. on Signal and Data Processing of Small
Define Targets, vol. 5428, 2004.
[13] J.-F. Ru, X. R. Li, and V. P. Jilkov. Multiple-Model Detection
C1AS+A-1,
- +mX2-Zs-1
A
- C2 S --A1 li+A-1mn of Target Maneuvers. In Proc. 2005 SPIE Conf on Signal
and Data Processing of Small Targets, vol. 5913, 2005.
C3A i'S-1z + m'A-1m, D Tm'C1m - 2m'C2 + C3 [14] H. W. Sorenson and D. L. Alspach. Recursive Bayesian
Estimation Using Gaussian Sums. Automatica, 7:465-479,
and let t C1/2( - C-1C02) dt -C1/2 di. Then 1971.
[15] J. L. Speyer and J. E. White. Shiryaev Sequential Probability
Ratio Test for Redundancy Management. AIAA Journal of
x) = e +12WSCm
2(C3-C'Cl 02 eI- t'tdt Guidance, Control, and Dynamics, 7(5):588-595, 1984.
[16] T. C. Wang and P. K. Varshney. A Tracking Algorithm for
e- - (C3 - C2 Cl-C2) Maneuvering Targets. IEEE Trans. Aerospace and Elec-
V/ 12-RSC, Am tronic Systems, AES-29(3):910-924, July 1993.
351

Authorized licensed use limited to: UNIVERSITY OF NOTTINGHAM. Downloaded on January 30, 2009 at 09:14 from IEEE Xplore. Restrictions apply.