Professional Documents
Culture Documents
2006 Aes Bavencoff PDF
2006 Aes Bavencoff PDF
INTRODUCTION
1240 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 42, NO. 4 OCTOBER 2006
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
[42]. Let us emphasize that, quite similarly to this deterministic one (Appendix B). Both have their own
paper, their aim is precisely range estimation [30]. advantages. The first one has the definite advantage
However, to perform satisfactorily they need a to take into account the statistical characteristics of
substantial observer maneuver [42]. We address here the problem. However, it is fundamentally a local
the problem of weakly observable scenarios. To analysis. The aim of the deterministic investigation
overcome this fundamental difficulty, the key is to is to provide state parameter bounding. It is a
include constraints on target motion which have been global approach which allows us to investigate the
incorporated via appropriate Monte Carlo Markov performance of the MCMC methods, especially the
chain (MCMC) methods. It is shown that speed effects of constraints. Geometry of nonobservable
constraints are very useful for reducing the uncertainty BOT TMA has been analyzed both from a nonlinear
about the target trajectory. More precisely, uncertainty perspective (envelope geometry) and a linear one
can be reduced by a factor proportional to the quotient (the geometry of the observable subspace). In
of the target speed interval length by the observer this setup, ghosting has been also analyzed (see
speed. Appendix B).
We restrict our analysis to a special class of target This paper is organized as follows. Problem
motion: namely the “conditionally deterministic one.” formulation is given in the Section II, followed by a
presentation of the MCMC method for constrained
This means simply that conditional on a fixed number
estimation in Sections III (Metropolis algorithm)
of parameters, the target trajectory is deterministic.
and IV (hit-and-run algorithm). In these sections, the
In this setup, it has been recognized that the problem
emphasis is on application to the BOT TMA context.
is fundamentally one of nonlinear regression, for
Section V deals with the HPD estimation. Results are
which batch algorithms are a convenient framework;
presented in Section VI. Finally, proofs of the main
see [28] for a seminal contribution. A vast and rich
results are available in Appendices A, B, and C.
literature has been devoted to this subject [28, 6, 31].
However, inclusion of constraints in the estimation
process for BOT TMA is rarely considered [14, 19, Notations and Acronyms
29]. For instance, the use of Uzawa’s like method has
AT Transposition of the A matrix
been mentioned in [28] and [14]. However, practically,
Id Identity matrix
results seem to indicate a neat tendency to “push” the
μ̂1:N Set of observed bearings: μ̂1 , : : : , μ̂N
solution toward the constraint bounds [33]. The reason
is simple, the likelihood functional is flat relative to Lμ̂1:N (x) Likelihood of the parameter vector x,
the unobservable component(s). So, it is appropriate given observations μ̂1:N
to replace point estimation by interval estimation Fr(S) Frontier of S domain (1)
and to include state constraints in the estimation tgt Target
process. obs Observer
To a certain extent, this is also the aim of BOT Bearings only tracking
geometric methods like “semi-azimetrie” [37, 4]; TMA Target motion analysis
very simple in principle, but providing robust and MCMC Monte Carlo Markov chains
often adequate performance.3 However, these methods HPD Highest probability density.
are only able to accommodate simple kinematic
constraints for computing intervals and do not give II. PROBLEM FORMULATION
a useful statistical insight. They are not designed for
The problem of estimating the parameters of the
estimating highest probability density (HPD) intervals
target trajectory under the assumption of rectilinear
[9] or dealing with general and complex constraints,
and uniform (target) motion from a history of noisy
which is precisely the objective of this work. Another
passive bearing measurements is considered. The
fundamental reason to use MCMC methods is that
general scenario is depicted in Fig. 1. It is assumed
marginal maximum a posteriori estimation for that the observer motion is rectilinear and uniform on
Bayesian models can be achieved very simply in each leg, and that the problem is planar.
this setting [13, 10]. Thus, it is not surprising that Firstly, let us briefly recall the problem in
its importance has been recognized also in target Cartesian coordinates. Let X be the relative (target)
tracking, especially for stressing scenarios [11]. state vector, defined by
The effects of constraint inclusion on state
¢
estimation have been investigated, both from X = Xtgt ¡ Xobs = [rx , ry , vx , vy ]T : (2)
a statistical perspective (Appendix A) and a
The discrete time equation of the relative motion takes
3 For
the following form:
a mathematical analysis of these methods we refer to
Appendix B. Xk = FXk¡1 + Uk
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
where μk [(r, μ, v, ®)] is the true kth angle associated
with the constant velocity target trajectory defined by
the state vector (r, μ, v, ®). With the notations of Fig. 1,
we have5 (see (28)):
μ ¶
vtgt sin ®
sin μ0 + k
r
tan(μk ) = μ ¶ : (6)
vtgt cos ® ¡ vobs
cos μ0 + k
r
This parameterization of the target trajectory is chosen
(instead of the Cartesian one) simply because it is
more convenient for including constraints. Note,
furthermore, that the parameter μ represents the target
Fig. 1. BOT TMA scenario: μ: bearing; r: target range; ®: target bearing at an arbitrary reference time. So, the suffix 0
heading; vtgt : target speed. (μ0 ) is omitted. Since there does not exist an explicit
method for solving (4), iterative methods are generally
used for solving this nonlinear regression problem,
where μ ¶ e.g. gradient methods [28] or modified instrumental
Id2 Id2
F= : (3) variable [28, 6] (MIV) methods. So, a vast literature
0 Id2 has been devoted to this subject. Even if the likelihood
In the above formula the vector Uk accounts for the is not concave, convergence of these methods has
effects of the observer accelerations, while Id2 is the been analyzed [26].
two-dimensional identity matrix. Let us consider a However, estimability is characterized by the
slightly different parameterization (see Fig. 1). The Cramer-Rao Lower Bound (CRLB), usually calculated
initial target position is defined by its initial range within the framework of asymptotic unbiased
r and bearing μ; while it is moving with a constant estimators (typically MLEs). The CRLB provides a
velocity v and heading ®. The target trajectory can theoretical lower bound for the mean square error
be also defined by the state vector (r, μ, v, ®), at an of the state vector estimator. It means that under
arbitrary reference time. regularity hypotheses, we have a (matrix) inequality
¢
Let μ̂1:N =(μ̂1 , : : : , μ̂N ) be N noisy bearing cov(£̂) º I ¡1 (£)
measurements, at time periods f¿1 , ¿2 , : : : , ¿N g. The where the notation A º B means that the (A ¡ B)
bearing measurements μ̂k are the exact bearings matrix is semi-definite positive. £ is the vector of the
corrupted by a sequence of independent and trajectory (true) parameters, while £̂ is any unbiased
identically (normally) distributed (IID) noises.
estimator of £ and cov(£̂) is the covariance matrix
Thus, the mean4 of μ̂k is μk , while its variance is ¾2 . of £̂. The CRLB I ¡1 (£) is the inverse of the Fisher
Classically, the batch BOT TMA problem is solved information matrix (FIM) I(£). Under the asymptotic
by a maximum likelihood estimator (MLE) which efficiency assumption, the accuracy of the MLE
consists in finding the state vector (r, μ, v, ®)¤ which estimator is evaluated by the CRLB. It may be shown
maximizes the likelihood given the observed bearings that the MLE is asymptotically efficient in the BOT
μ̂1:N , i.e., TMA context, though observations are fundamentally6
¢ not IID [34]. Let us recall below various equivalent
(r, μ, v, ®)¤ = arg max Lμ̂ [(r, μ, v, ®)]: (4)
(r,μ,v,®)
1:N
parameterizations for the TMA problem:
8
The likelihood of a state vector given the observed > x = (rx , ry , vx , vy )T Cartesian coordinates
>
>
bearings is defined by the following expression: < μ ¶T
x = μ, μ, _ r_ , 1 modified polar coordinates
Lμ̂ [(r, μ, v, ®)] >
> r r
1:N >
:
x = (r, μ, v, ®)T constraint coordinates:
= f(r,μ,v,®) (μ̂1:N )
μ ¶N " # (7)
1 X
N
1 In Fig. 2, we present the r (diagonal) coefficient of
= p ¢ exp ¡ 2 (μ̂k ¡ μk [(r, μ, v, ®)])2
¾ 2¼ 2¾ _ matrix (i.e., the variance of r) as
the I ¡1 (r, μ, r_ =r, μ)
k=1
(5)
5 Notice that it is assumed here that the interval between
1242 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 42, NO. 4 OCTOBER 2006
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Fig. 2. Value of range CRLB (in y axis) versus observer maneuver angle ¹.
a function of the angle ¹ of the observer maneuver7 is recovered and there is an interaction between the
(moving with constant speed). _ and r [1].
observable part (μ, r_ =r, μ)
Obviously, when the amplitude of the observer What is the underlying difficulty for estimating
maneuver is below a certain value (here, about the r component? It is well known that the FIM is
10 deg), then the r CRLB coefficient grows straightforwardly related to the likelihood curvature
dramatically, thus showing that the TMA problem [22]. To illustrate this dependency we present in Fig. 3
becomes practically inestimable (see Fig. 2). If the log-likelihood versus the range r, the three others
instead of heading change, we can consider a change _ being fixed at their true values,
parameters (μ, r_ =r, μ)
in speed, the conclusion is similar: a very strong the amplitude of the observer maneuver being very
acceleration is necessary for estimating the range small. We can notice that the likelihood functional
r. Actually, without observer maneuver, the three is not sharp, not symmetric and that its maximum
_ are observable and uncoupled
parameters (μ, r_ =r, μ) is achieved for a false range value (56,500 m versus
with the fourth (unobservable) parameter. However, 74,080 m).
when there is an observer maneuver, full observability Now, it is quite common for the amplitude of
the observer’s maneuvers to be very limited. This
may be due to operational factors, themselves related
7 i.e., the difference of the observer headings, associated with two to the role of the observer itself (e.g. maritime
consecutive legs. surveillance) or to physical constraints. Indeed, it is
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Fig. 4. Parabolic envelope of bearings and target tracks.
1244 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 42, NO. 4 OCTOBER 2006
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
The ratio 2[(vtgt,max ¡ vtgt,min )=vobs ] cos(®) acts as a
“reduction” factor for estimating the distance r, under
speed constraints; thus showing that the benefit of
including hard constraints for range estimation is
generally quite noticeable.
A more intuitive representation of this
phenomenon can be found in Fig. 5. For a given
(target) speed constraint (vtgt,min · vtgt · vtgt,max ),
admissible target trajectories are inside the two shaded
areas. One area corresponds to trajectories forming
an acute angle (close to zero) with the true target
trajectory, while the the other (ghost) corresponds
to obtuse angles (close to ¼). So, they are named
“ghost” target trajectories. Obviously, this reduction
factor is really interesting when the observer speed
largely exceeds the target one. Therefore, the key
idea consists in incorporating prior information, Fig. 6. General architecture for sampling, marginalization, and
as constraints on target speed. In the first case, estimation of confidence intervals.
constraints are defined by intervals and used in the
Metropolis MCMC algorithm. Then (Section IV), is defined by the integral of the posterior density
constraints are defined in terms of the target trajectory ¼[(r, μ, v, ®) j μ̂1:N ], according to
itself, for which the hit-an-run MCMC is especially Z
relevant. This is summarized below: ¼(r j μ̂1:N ) = ¼[(r, μ, v, ®) j μ̂1:N ]d(μ, v, ®):
8
< Constraints: target speed, viable domain,
Since this may be a very demanding calculation, we
: Target motion: rectilinear or piecewise linear shall use a different approach for approximating this
(leg-by-leg).
posterior density based on simulation. The strength of
(9)
this method is also to produce confidence intervals for
It is assumed that target speed is constant on each leg. constrained estimation of r based on HPD intervals.
Roughly, an HPD interval is a confidence interval of
minimum length, for a given probability of content
III. CONSTRAINTS ON TARGET SPEED GIVEN AS
(see Section V). For a given probability content,
INTERVAL
the HPD method provides a confidence interval
for the range r by using a sample fr(i) gi=1:n from
Prior information can be integrated in a Bayesian
¢ the marginal posterior density ¼(r j μ̂1:N ). This is
context. Given observations (μ̂1:N =fμ̂1 , : : : , μ̂N g, the
obtained by taking the first component of a sample
posterior distribution of the state vector is given as the f(r, μ, v, ®)(i) gi=1:n , itself generated from the posterior
product of the prior with the likelihood function via distribution ¼[(r, μ, v, ®) j μ̂1:N ] via MCMC sampling, as
the Bayes’s theorem: summarized in Fig. 6.
Lμ̂ (r, μ, v, ®) £ ¼(r, μ, v, ®)
¼[(r, μ, v, ®) j μ̂1:N ] = 1:N
A. Prior Modelling
¼(μ̂1:N )
(10) As will be shown, taking into account prior
where information on the target speed (v) is instrumental
¼[(r, μ, v, ®) j μ̂1:N ] is the posterior density of the in this solution. For instance, lower and upper
bounds (vmin and vmax ) are considered. Other priors
state vector,
are relative to the parameters r, μ, and ®. Thus, it
¼(r, μ, v, ®) is the prior density of the state vector,
is natural to assume lower and upper bounds for
Lμ̂ (r, μ, v, ®) is the likelihood of the state vector
1:N range r (namely rmin and rmax ).9 Similarly, we can
(r, μ, v, ®) given the observations,
suppose that the first bearing measured μ̂1 is not
¼(μ̂1:N ) is the density of the observations.
too far from the right bearing μ1 , i.e., in the interval
For determining an interval for range r, given a prior,
[μ̂1 ¡ m¾, μ̂1 + m¾] where ¾ is the standard deviation of
it is necessary to access to the marginal posterior
distribution ¼(r j μ̂1:N ) of the posterior density state 9 From an operational point of view, detecting measurements coming
vector ¼[(r, μ, v, ®) j μ̂1:N ]. The main difficulty lies in from a very remote target is not likely and has no operational
the fact that the marginal posterior density ¼(r j μ̂1:N ) interest.
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
the measurements. Typically, a value m = 5 is largely ALGORITHM The Basic MCMC Metropolis Algorithm:
sufficient. The target heading ® is let free in its own Given x(i) ,
interval [¡¼, +¼]. 1) Generate yi » q(y j x(i) ).
In the absence of prior about the kinematic 2) Take
parameter densities and for simplifying the sampling ½
(i+1)
yi with probability (½(x(i) , yi ))
process, we avoid too specific prior, so that uniform x =
x(i) with probability (1 ¡ ½(x(i) , yi ))
distributions are the more convenient:10
(11)
¼(r) » U[rmin , rmax ], ¼(μ) » U[μ̂1 ¡ m¾, μ̂1 + m¾] where ½(x, y) = minf(f(y)=f(x)), 1g:
¼(v) » U[vmin , vmax ], ¼(®) » U[¡¼, +¼]:
C. Sampling from the Posterior Density
so, that (independence of the prior):
The problem we are dealing with now, is to obtain
¼(r, μ, v, ®) = ¼(r) £ ¼(μ) £ ¼(v) £ ¼(®), an MCMC sample X (i) = (r, μ, v, ®)(i) drawn from the
joint posterior distribution f(r, μ, v, ®) = ¼[(r, μ, v, ®) j
= C1[rmin ,rmax ] (r) £ 1[μmin ,μmax ] (μ) £ 1[vmin ,vmax ] (v) μ̂1:N ]. Firstly, a transition kernel is defined by
where C = 1=(4m¼¾(rmax ¡ rmin )(vmax ¡ vmin )), and q((r0 , μ 0 , v 0 , ®0 ) j (r, μ, v, ®)(i) )
while 1 is the indicator function. Note that the great
= q(r0 j r(i) ) £ q(μ 0 j μ (i) ) £ q(v 0 j v (i) ) £ q(®0 j ®(i) )
interest (r, μ, v, ®) parameterization is precisely to
render explicit operational constraints. where
· μ ¶ μ ¶¸
0 (i) (i) rmax ¡ rmin (i) rmax ¡ rmin
B. Generation of Multidimensional Random Variables. q(r j r ) » U r ¡ ,r +
·1 ·1
The MCMC Metropolis-Hastings Algorithm · ¸ (12)
0 (i)10¾ (i) 10¾ (i)
q(μ j μ ) » U μ ¡ ,μ +
The aim of this section is to provide a specific ·2 ·2
setting for the application of the MCMC methods · μ ¶ μ ¶¸
[18] for constrained estimation [41]. In this part, the vmax ¡ vmin vmax ¡ vmin
q(v 0 j v (i) ) » U v (i) ¡ , v(i) +
only constraint deals with target speed and priors are ·3 ·3
defined as intervals on parameters r, μ, v, and ®. · ¸
0 2¼ (i) 2¼
(i) (i)
In this setup, an MCMC method is used for q(® j ® ) » U ® ¡ ,® + :
·4 ·4
drawing a sample fX1 , X2 , : : : , Xn g approximately
distributed from f without directly simulating from Scalar parameters (see (12)) ·1 , ·2 , ·3 , and ·4 (e.g.
f (see [35]). More precisely, its aim is to produce equal to 20) are introduced for algorithmic reasons,
an ergodic Markov chain (X (t) ) with stationary simply for avoiding that new values (r0 , μ0 , v0 , ®0 ) be
distribution f, i.e., (X (t) ) converges in distribution to drawn out of the posterior distribution support and be
a random variable drawn from f. For convergence rejected. Indeed, the above transition kernel satisfies
analysis, we refer to [17]. In our context, the the symmetry condition since we have
stationary distribution is the posterior distribution of ¡ a · y ¡ x · a , ¡a · x ¡ y · a,
the target trajectory ¼[(r, μ, v, ®) j μ̂1:N ].
For an arbitrary starting value x(0) , a chain (X (t) ) so that (13)
is generated using a transition kernel with stationary 1 1
distribution f. Thus, for a large enough T0 , X (T0 ) the q(x j y) = 1 (x) = 1 (y) = q(y j x):
2a [y¡a,y+a] 2a [x¡a,x+a]
method produces a dependent sample X T0 , X T0 +1 , : : :
which is generated from f. Now, the next step is the computation of the (test)
A symmetric conditional density q(y j x) (transition functional ½((r, μ, v, ®)(i) , (r0 , μ0 , v0 , ®0 )) which reduces to
kernel) is then chosen (“symmetric” means that the computation of the ratio of the posterior densities:
q(y j x) = q(x j y)), which defines the density to move ¼[(r0 , μ0 , v0 , ®0 ) j μ̂1:N ]
from a state x to a state y. The transition kernel must
¼[(r, μ, v, ®)(i) j μ̂1:N ]
allow every point in the support of the density f to be
reached. The Metropolis-Hastings algorithm associated Lμ̂1:N (r0 , μ0 , v0 , ®0 ) £ ¼(r0 , μ0 , v0 , ®0 )
with the objective target density f and the conditional =
Lμ̂ (r, μ, v, ®)(i) £ ¼(r, μ, v, ®)(i)
density q produces a Markov chain (X (i) ) through the 1:N
1246 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 42, NO. 4 OCTOBER 2006
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
itself requiring the calculation of the measurement
likelihood. Assuming that the observations are
normally distributed, i.e., μ̂k = μk + ²k with ²k »
N(0, ¾2 ), the elementary likelihood of the state vector
(r, μ, v, ®) for the single bearing μk is given by
" #
1 1 (μ̂k ¡ μk [(r, μ, v, ®)])2
Lμ̂k (r, μ, v, ®) = p exp ¡
¾ 2¼ 2 ¾2
(15)
where μk [(r, μ, v, ®)] is the exact bearing at the
time period ¿k under the hypothesis that the target
trajectory is defined by the state vector (r, μ, v, ®).
Fig. 7. Complex constraints on target trajectory.
Then, under the independence assumption, the ratio
(R) of the posterior densities reduces to
IV. INCLUSION OF GENERAL CONSTRAINTS. THE
0 0 0 0
¼[(r , μ , v , ® ) j μ̂1:N ] MCMC HIT-AND-RUN ALGORITHM
where ½ = min(R, 1) with R the ratio of the posterior described by a set of equations, in a classical nonlinear optimization
densities expressed by (16). problem.
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Fig. 8. Left: Hit-and-run algorithm. Right: rectilinear target trajectory under constraints.
1248 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 42, NO. 4 OCTOBER 2006
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Fig. 9. Rectilinear target trajectory under constraints.
4) Draw a signed distance ¸i from the density: identically defined. The following proposition, though
elementary, is especially relevant for applying the
1 hit-and-run algorithm (see Appendix C, for a proof).
gi (¸ j di , Ti ) = 1 (¸)
m(Si (di , Ti )) Si (di ,Ti )
PROPOSITION 3 Assume that the constraint domain C
where m is the (Lebesgue) measure of the segment is convex, then the set of feasible trajectories S is also
Si (di , Ti ). convex.
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Therefore, at step 3 of the above algorithm, Si (di , Ti ) 1) Choose an initial arbitrary trajectory satisfying
is simply a real interval [¸min , ¸max ] satisfying the the constraints. Define T0 as
following conditions: Ti + ¸min ¢ di 2 Fr(S) or Ti + 0 1(0)
¸max ¢ di 2 Fr(S). The whole algorithm is summarized Pinit
in Fig. 9. B C
T0 = @ Pmid A 2S
The above method deals with a target located
Pend
in a known area and moving in a straight line with
constant speed. We now see how this method can and set i = 0.
handle scenarios of increasing complexity. 2) Generate six (6) uniform variables in [¡1, 1]:
fD1 , D2 , D3 , D4 , D5 , D6 g that represent the evolutions
C. Hit-and-Run Algorithm and the Leg-by-Leg Target of directions of the points Pinit , Pmid , and Pend and
Trajectory Model define the normalized di vector (kdi k = 1) by dTi =
(d1(i) , d2(i) , d3(i) , d4(i) , d5(i) , d6(i) ).
def
We are now dealing with the estimation of 3) Determine the ¸ set: Si (di , Ti ) =f¸ 2 R j Ti + ¸ ¢
leg-by-leg target trajectory parameters. The target di 2 Sg. Then the constraints on ¸ stand as follows:
trajectory is supposed made of two legs. The
constraint expresses that the first leg and the second Ti + ¸di 2 S
leg of the target trajectory stand in given areas. A new °Ã (i) (i) ! Ã (i) !°
° r d3 ¡ d1(i) °
state vector has to be considered. It seems natural ° x,mid ¡ rx,init °
° (i) +¸ °
to consider the state vector made of the three points ° r (i) (i)
d4 ¡ d2(i) °
y,mid ¡ ry,init
that characterize the leg-by-leg trajectory. In this °Ã (i) ! Ã (i) !°
° r (i)
d5 ¡ d3(i) °
setting, the target trajectory T can be defined by ° x,end ¡ rx,mid °
+ ° (i) +¸ °·L
three “characteristic” points (Pinit , Pmid , and Pend ), ° r (i) (i)
d6 ¡ d4(i) °
y,end ¡ ry,mid
so that T
μ ¶ where di = (d1(i) , d2(i) , d3(i) , d4(i) , d5(i) , d6(i) ) : (26)
def rx,init
T = (Pinit , Pmid , Pend )T , Pinit = 2 R2
ry,init 4) Generate a signed distance ¸i from the density:
μ ¶ μ ¶ (24)
rx,mid r 1
Pmid = 2 R2 , Pend = x,end
2 R2 : gi (¸ j di , Ti ) = 1 (¸)
m(Si (di , Ti )) Si (di ,Ti )
ry,mid ry,end
Denote C1 the area where the first target trajectory where m is the (Lebesgue) measure of the segment
leg stands: [Pinit , Pmid ] and C2 the area for the second Si (di , Ti ).
leg: [Pmid , Pend ]. Moreover, it is assumed that Pmid 5) Define the © vector by © = Ti + ¸i ¢ di and
define Ti+1 by
stay in C1 \ C2 . Moreover, let us denote C = C1 [
C2 the area where all the target trajectories stand Ti+1 =
(without constraint on the maximal speed) and let 8 μ ¶
f(©) f(©) Lμ̂ (©)
S 2 R2 £ R2 £ R2 be the C subset representing all the <© with probability min 1, , = 1:N
f(Ti ) f(Ti ) Lμ̂ (Ti ) :
target trajectories satisfying both the constraint on the : 1:N
1250 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 42, NO. 4 OCTOBER 2006
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Fig. 10. Representation of HPD interval.
V. HIGHEST PROBABILITY DENSITY INTERVALS Tanner’s algorithm requires evaluating the marginal
posterior densities analytically or numerically,
Practically, it is of fundamental importance to have which renders its implementation rather complicated
an estimation of confidence intervals for high level and computationally intensive. To overcome such
information processing: e.g., track-to-track association, difficulties, Chen, Shao, and Ibrahim [9] developed
sensor management, situation assessment, etc. We see a simple Monte Carlo method for estimating HPD
now that MCMC-based methods give us estimates of intervals. This approach requires only an MCMC
these confidence intervals, as a by-product. sample generated from the marginal posterior
distribution of the parameter of interest and propose
A. HPD Intervals Method the following procedure [8] for calculating an HPD
interval for ¸:
The aim of this paragraph is to present the ALGORITHM Chen-Shao HPD Estimation Algorithm
calculation of the HPD intervals method and their 1) Obtain an MCMC sample f¸i gi=1:n from
connection with MCMC methods. Consider a ¼(¸ j D).
Bayesian posterior density having the form: 2) Sort f¸i gi=1:n to obtain the ordered values:
LD (¸, ') £ ¼(¸, ')
¼(¸, ' j D) = ¸(1) · ¸(2) · ¢ ¢ ¢ · ¸(n) :
c(D)
3) Compute the 100(1 ¡ ®)% credible intervals:
where D denotes data, ¸ is one-dimensional, ' may
be a multidimensional vector of parameters other than Rj (n) = (¸(j) , ¸(j+[(1¡®)n]) )
¸. The functional LD (¸, ') is a likelihood functional,
¼(¸, ') is a prior and c(D) is a normalization constant. for j = 1 : n ¡ [(1 ¡ ®)n] (where [¢] refers to the entire
An HPD interval has two main properties: part).
4) The 100(1 ¡ ®)% HPD interval is the one,
1) the density for every point inside the interval is denoted by Rj ¤ (n), with the smallest interval width
greater than that for every point outside the interval, among all credible intervals.
2) for a given probability content, say 1 ¡ ®, the
interval is of the shortest length. The main interest of this procedure is that it does
not require a closed form of the marginal density.
Especially, it is when the marginal distribution is The following proposition, given by Gelfand and
not symmetric (see Fig. 3), that an HPD interval is Smith [13] (1990), is particularly useful for our TMA
mostly desirable. Fig. 10 illustrates these properties. problem.
It presents a density function, an interval having
PROPOSITION 5 If f(¸i , 'i )gi=1:n is an MCMC sample
a probability content of 0.6413, and the HPD
interval for this given probability. Tanner [40] from the joint posterior distribution ¼(¸, ' j μ̂1:N )
provides a Monte Carlo algorithm to calculate the then f¸i gi=1:n is an MCMC sample from the marginal
content and boundary of the HPD region. However, posterior distribution ¼(¸ j μ̂1:N ).
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Let us consider now the application of this property to
our context.
1252 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 42, NO. 4 OCTOBER 2006
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Fig. 12. Left: MCMC samples versus time. Right: HPD interval width versus probability of content.
Fig. 13. Left: observations (sequence of estimated bearings). Right: 500 iterations of MCMC hit-and-run.
So, for abscissa 0, the HPD interval is 100% of the 2) Constraints on Target Speed and Target
MCMC values, since this value gives the minimum Trajectory: The second result concerns the
and maximum range of the MCMC sampling. For implementation of the hit-and-run algorithm using
every value of probability of content, the lower bound constraints about target trajectories (trajectory domain
(respectively higher bound) of the HPD interval is and target speed) (see Section IVB). The scenario
represented by the dash-dot line (respectively the is almost the same than previously, apart from the
dashed line). The true value of range falls outside the fact that the target range which is now 18,520 m. In
HPD interval only when the probability of content is this example, the trajectory constraints are simple:
as low as 0.3. Even for this value, the HPD interval all the target trajectories are supposed to lie inside a
is close to the true value, while the interval width is rectangle of length 37,000 m and width 10,000 m,
as small as 2.5 km. Moreover, for greater values of as depicted by the rectangle in Fig. 13, while it
the probability of content, we can see that the lower is assumed that target speed is bounded above by
bound is very close to the true range value and that vmax = 30 knots. The noisy bearings measured by
the interval width remains very reasonable, which is a the observer are also represented in Fig. 13 (left).
very satisfactory result. The benefit of including speed In Fig. 13 (right), we present 30,000 iterations
constraints is clear. of the MCMC hit-and-run, associated with this
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Fig. 14. Left: MCMC samples versus time. Right: HPD interval width versus probability of content.
Fig. 15. Left: MCMC samples versus time. Right: HPD interval width versus probability of content.
sequence of observations and these constraints. by considering the “hourglass” drawn by the feasible
Considering Fig. 13, we notice that the hit-and-run trajectories in Fig. 13 (right).
algorithm is rapidly “converging” toward a set of Then it is meaningful to consider an additional
feasible target trajectories, roughly representing an condition on the feasible trajectories directions so as
“hourglass” (see Fig. 13 right). Thus, the hit-and-run to eliminate ghosting. This is the object of Fig. 15
algorithm provides an exploration of the uncertainty (left) for which trajectory sampling is restricted to the
domain about target trajectory (see Proposition 2 and true direction (i.e., left to right). MCMC estimation
Appendix B). The choice of the initial trajectory (the is clearly improved by deghosting. More generally,
segment down left) is unimportant provided that it ghosting is also resulting from unobservability
satisfies the conditions 18. (see Appendix B) and deghosting is a fundamental
In order to evaluate the accuracy of HPD intervals concern for all these scenarios. Actually, it may be
calculated by this way, 30,000 samples (trajectories) shown that even a very limited observer maneuver
are drawn and represented by the histogram in Fig. 14 may be sufficient for ensuring good performance
(left). This histogram clearly exhibits two peaks: one deghosting thanks to the calculation of appropriate
is close to the true range value (dashed line), while likelihood ratio tests and reliable confidence intervals
the other is corresponding to ghost trajectories (see via HPD.
Fig. 5 for a geometric interpretation). In consequence,
HPD intervals are also “pulled” by these ghost B. Maneuvering Target
trajectories, as seen in Fig. 14 (right) which presents
the evolution of the HPD intervals obtained via the The last part of this section deals with the
hit-and-run algorithm. This event is also quite evident leg-by-leg target trajectory case (see Section IVC).
1254 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 42, NO. 4 OCTOBER 2006
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Fig. 16. Left: true target trajectory under set constraints. Right: noisy observed bearings.
Fig. 17. Left: HPD intervals for first range r1. Right: HPD intervals for third range r3.
For the first leg, the target is within a small rectangle The target trajectory is defined by 6 parameters,13
of length 3,000 m and width 1000 m (representing namely, ranges fr1 , r2 , r3 g and bearings fμ1 , μ2 , μ3 g.
a strait for example) and the second leg is assumed An arbitrary initial trajectory being chosen, 5000
to be within the larger rectangle of length 15,000 m iterations of the hit-and-run sampler provide a set of
and width 10,000 m as depicted in Fig. 16 (left). first range r1i , a set of middle range r2i and a set of
The observer is at the origin at time 0 and goes final range r3i related to the three points Pinit , Pmid and
along the x-axis with a constant speed of 180 Pend defining a leg-by-leg target trajectory. In the same
knots (92.6 m/s), while the target speed is 20 knots way, three sets of bearings are generated. It remains to
(10.3 m/s). The observer turns with a (very) weak apply the HPD intervals method to each of them so as
angle of 2 deg after 25 observed measurements. to provide six sets (r, μ) of confidence intervals about
In all, 50 bearings are taken, with 5 s between these parameters. Figs. 17 and 18 (left and right)
each measurement. The measurement error is illustrate the behavior of the proposed method.
modelled by a centered Gaussian noise, with 1 deg Obviously, the hit-and-run algorithm perform quite
standard deviation as presented in Fig. 16 (right). satisfactorily, even for this difficult scenario. Even if
The target is approximately 36,000 m far from the first range HPD is far smaller than the third one, the
observer. The total target (respectively observer)
trajectory length is thus 2,524 m (respectively 13 It
is simply assumed that the target maneuver is constrained to
22,678 m). occur in a given domain (here the intersection of two domains).
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Fig. 18. Left: HPD intervals for first bearing μ1 . Right: HPD intervals for third bearing μ3 .
third range estimation remains quite acceptable. For appealing. Second is the avoidance of gradient-based
the angle estimation, HPD intervals are of the same methods, whose definite drawback is to perform
order. It is remarkable that such performance has been poorly as soon as the problem is poorly conditioned.
achieved with a very weak observer maneuver.
ACKNOWLEDGMENTS AND DEDICATION
VII. CONCLUSION
The authors are greatly indebted to the reviewers
This paper deals with range estimation estimation and the editor for an insightful review process.
under poor observability conditions via operational This paper is dedicated to the memory of Philippe
constraints reinforcement. The algorithms we have Nicolas (Thales), our friend and colleague, who was a
developed are based on simulating constrained constant source of inspiration and enthusiasm.
MCMC converging toward the posterior distribution
of the target trajectory. Effects of constraints have APPENDIX A. BOT TMA PERFORMANCE ANALYSIS
been investigated, both from a statistical perspective (NONMANEUVERING OBSERVER)
(Cramér-Rao bounds) and a deterministic one
(intervals). Geometry of nonobservable BOT TMA has The aim of this appendix is to quantify the
been analyzed both from a nonlinear viewpoint (the effect of the BOT TMA scenario parameters, for a
envelope geometry) and a linear one (the geometry of nonmaneuvering observer. We refer to Fig. 1 for the
the observable subspace). Performance is illustrated meaning of the parameters (t: time index). Elementary
by simulation results, showing the considerable calculations yield
improvements offered by these methods. ¡! ¡!
Due to an exhaustive exploration of the posterior det(M0 Mt , Mt Tt )
density, these algorithms are both reliable and tan(μt ) = ¡! ¡!
feasible. Using them, it is possible to exploit complex hM0 Mt , Mt Tt i
kinematic and geographic constraints on target μ ¶
vtgt sin ®
behavior and thus to provide a convenient estimate sin μ0 + t
for hardly estimable parameter(s). Moreover, reliable r0
= μ ¶ : (28)
confidence intervals are obtained as a by-product and vtgt cos ® ¡ vobs
cos μ0 + t
are a definite advantage of simulation-based methods. r0
This is of fundamental importance from an operational
We can notice that (vtgt sin ®) is the radial (relative)
viewpoint; more specifically for problems like threat
target velocity, while (vtgt cos ® ¡ vobs ) is the
assessment, sensor management, or tactical aids. From
longitudinal one. From (28), we see that the
an estimation perspective, these algorithms provide a
target bearing history is described by a set of
basic ground for a wide variety of problems like track
three parameters (e.g. fμ0 , (vtgt sin ®)=r, 1=rg or
initialization, track-to-track association, or deghosting. ¢
Other important features of this approach are fμ0 , ®, r0 = 1=rg). This is certainly not surprising
also particularly noticeable. First, the robustness to since in the absence of observer maneuvering the
the choice of the initial starting point is particularly dimension of the observable space is generally 3.
1256 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 42, NO. 4 OCTOBER 2006
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Thus, estimability analysis is achieved by considering vtgt or r) and consider the following decomposition
only the estimation of 3 parameters, the fourth xT = (x1 , x1 ). For instance, consider the x1 = fμ0 , ®, r0 g
being fixed. For this appendix, we consider the parametrization (x1 = vtgt fixed), then the diagonal
¢
target trajectory parameterization fμ0 , ®, r0 = 1=rg, the terms of FIM¡1 have relatively simple expressions, i.e.,
parameter vtgt being fixed (r ´ r0 ).
Practically, an estimability measure is FIM¡1 (1, 1) = 1
obtained in the following way [27]. Let gt be FIM¡1 (2, 2) =
the gradient vector of the measurement, i.e., gt =
2
f(@=@μ0 )μt , (@=@®)μt , (@=@r0 )μt gT and define the (3 £ 3) [1 + 16r0 (vtgt cos(® ¡ μ0 ) ¡ vobs cos(μ0 ))2
M matrix as M = (gt , g_ t , g̈t ). Then, we consider as a 4
+4r0 (vobs
2 2
+ vtgt
2
¡ 2vobs vtgt cos ®) ]
measure of the uncertainty (also termed estimability), 2
(32)
4r0 4 vtgt
2 (v ¡ v
tgt obs cos ®)
the determinant of the FIM, itself approximated by
(three consecutive measurements: f0, 1, 2g, Gaussian FIM¡1 (3, 3) =
hypothesis): 2
[cos2 (® ¡ μ0 ) + 4r0 (¡vtgt cos(2(® ¡ μ0 )) + vobs
det(FIM) ¼ [det(gt , g_ t , g̈t )]2t=0 cos(® ¡ 2μ0 )) + 4r0 (vobs
2 2 4 2
+ vtgt ¡ 2vobs vtgt cos ®) ]
2
:
yielding: 4r0 2 (vtgt ¡ vobs cos ®)2 (vtgt sin(® ¡ μ0 ) + vobs sin μ0 )2
2 04
4vtgt r Furthermore, admitting that the assumption vobs r0 ¿ 1
det(FIM0,2 ) = (v ¡ vobs cos ®)2
12 tgt is reasonable, an approximation of FIM¡1 (3, 3)
¾w6 (cos μ0 )
(29) (associated with the r0 estimation) becomes as simple
£ (vtgt sin(® ¡ μ0 ) + vobs sin μ0 )2 as:
r2 cos2 (® ¡ μ0 )
2 02
4vtgt r FIM¡1 (3, 3) ¼ 4 : (33)
= (vtgt ¡ vobs cos ®)2 μ_2 4vobs (cos ® sin μ0 )2
¾w6 (cos μ0 )12
Now, a classical and important result is that the
à param.: fμ0 , ®, r0 g: following chain of (matrix) inequalities holds true [2]:
From (29), we see that the uncertainty volume cov(x̂1 ¡ x1 ) º [FIMx1 ]¡1 º (FIMx ¡1 )1̄,1̄ º [fFIMx g1̄,1̄ ]¡1
(i.e., [det(FIM)]¡1 ) varies with r2 , μ_¡2 .14 Moreover,
(34)
considering the basic assumption of this manuscript where FIMx1 is the FIM matrix relative to the
(i.e., vobs À vtgt ) and assuming vtgt known, yields subvector x1 (parameter x1 being fixed), and
μ ¶4 (FIM¡1
x )1̄,1̄ is the 3 £ 3 left-upper block of the
r (cos μ0 )12
det(FIM¡1 ) / 2
, Ã param.: fμ0 , ®, r0 g: ¡1
FIM matrix. Then, assuming a random prior about
vobs vtgt (sin μ0 cos ®)2
parameter x1 (here vtgt ), whose support is denoted C1 ,
(30) we have:
This formula needs some comments. First, the Z
importance of the value of cos μ0 is noteworthy E[cov(x̂1 ¡ x1 )] º [FIM¡1
x1 ]x1 =u px1 (u)du: (35)
C1
(exponent: 12), as well as that of the ratio (r=vobs )
(exponent: 4). The above analysis can be extended The diagonal terms of the matrix [FIM¡1x1 ]x1 =u are
in two principal ways. First, for a larger number of T given by (32). Considering for instance (33), we
consecutive measurements, the Binet-Cauchy formula obtain
[27] allows us to treat any measurement number, and 2
Z
cos (® ¡ μ0 )
the preceding approximations becomes var(r̂0 ) ¸
4r0 2 C1
6 2
det[FIM(1,:::,T) ] / [T ] det(FIM0,2 ): (31) pvtgt (u)
£ du:
This approximation is valid for a reasonably long (vtgt ¡ vobs cos ®) (vtgt sin(® ¡ μ0 ) + vobs sin μ0 )2
2
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Fig. 19. Left: geometric definitions. Right: parabolic envelope of bearing lines and its axis.
are those of Fig. 19. The vector N(t) is orthogonal From (41), we see that the envelope of the Lt lines is
to the bearing line Lt . Thus, we can write (Q: ¼=2 a parabola. Indeed, g(X, t) is a quadratic form in X
rotation matrix): whose associated matrix is (QV)(QV)T . This matrix is
nonnegative definite while its rank is 1 which proves
N(t) = Q[X0 + tV] where
that this envelope is a parabola (see [15], chapter
μ ¶ X). Its image is one-dimensional and generated by
0 ¡1
Q= the vector QV, while its kernel is spanned by the
1 0 (37) vector V (direction of relative motion (DRM), see
0¢ Fig. 19). Thus, we have just shown the first part of
X = X0b ¡ X0a
Proposition 1.
¢ ¢ ¡! ¢ ¡! The second part of Proposition 1 is much more
V = Vb ¡ Va (Vb = vtgt , Va = vobs ):
intricate. An analytic proof is feasible but seems
Denoting X the generic point of Lt , an (implicit) (at best) very tedious. Instead, a purely geometric
equation of the bearing line Lt is then given by proof can be found in classical (but ancient) treatises
on geometry. Reference [24] does not require any
¢ prerequisite, while a large part is devoted to conic
g(X, t) = (X ¡ Xa (t))T Q(X0 + tV) = 0: (38)
properties. Now, let us refer to the following figure:
Then, the envelope of the bearing lines fLt gt is let ¢(AP) and ¢(AN) be two fixed tangent lines to
defined by the following system of equations: the parabola, and ¢(BC) a generic line, tangent in
M. B and C are the intersections of this tangent
@ line with ¢(AP) and ¢(AN). Then, with the
g(X, t) = 0, g(X, t) = 0: (39)
@t notations of Fig. 20, the following equality holds
From the second equation ((@=@t)g(X, t) = 0), we true:
obtain the following condition which is necessarily MB BP CA
= = : (42)
satisfied by a point X belonging to the envelope of the MC BA CN
Lt lines: Actually, this remarkable property is valid whatever
1 T the point M and is a consequence of Poncelet
t= T
[(X ¡ X0a ) QV ¡ VTa QX0 ]: (40) theorems [24]. So, to end the proof it remains to
2(Va QVb )
prove that the target trajectory is also a tangent to
Injecting (40) in g(X, t), we deduce an equation of the the envelope parabola. This is easily achieved by
envelope (see (39)), i.e., considering the following equalities:
XT (QVVT QT )X + ®1 XT QX0 + ¯1 XT QV + °1 = 0
(X ¡ Xb (t))T Q(X0 + tV)
where
= g(X, t) + (Xa ¡ Xb (t))T Q(X0 + tV)
®1 = 4(VTa QVb ) (41)
¯1 = ¡2(X0,T T 0
a QV + Va QX ) = g(X, t) + (X0 + tV)T Q(X0 + tV) = g(X, t):
°1 = (X0,T T 0 2 T 0,T
a QV + Va QX ) + 4(Va QVb )(Xa QXb ): (43)
1258 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 42, NO. 4 OCTOBER 2006
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Fig. 20. Algebraic properties of the tangent lines of a parabola.
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
2
so, that vy,max 2
= vmax =(1 + (Bμ¡1 t)22 ) and the extremal The scalar (Bμ¡1 t)2 only depends of the scenario, i.e., it
vectors of the (convex) set of feasible trajectories (see is common to any choice of the vy parameter). Thus,
Fig. 21) are given by we have
2 2
vy,max 1 vy,min
2
rx,max = (Bμ¡1 t)21 vy,max
2
= = : (53)
2
vx,max (Bμ¡1 t)22 2
vx,min
2
rx,min = (Bμ¡1 t)21 vy,min
2
(Bμ¡1 t)21 + (Bμ¡1 t)23 Gathering (53) and (54), we deduce finally:
k~rmax k2 ¡ k~rmin k2 = (vmax
2 2
¡ vmin ) Ã !
1 + (Bμ¡1 t)22 2
vy,max (v2 ¡ v2 ) vy2
μ 2¶ = 1 + max2 min ,
2 2 r 2
vy,min vy,min v2
= (vmax ¡ vmin ) 2 : (49)
v à !
(vtgt,max ¡ vtgt,min ) cos(®) vy2
This result is surprisingly simple. The effect of speed ¼ 1+2 2
2 2 vobs vy,min
constraints (vmax ¡ vmin ) as well as scenario parameters
2 2
on the ratio (r =v ) is noteworthy. Notice also that the and (55)
ratio r2 =v2 is independent of the the particular choice à x ! à x
! Ã !
vtgt,max vtgt,min x
of the “free” parameter vy (0) (see (46)). Hence, it is vobs
¼ +¯
well defined. y
vtgt,max y
vtgt,min
y
vobs
From an operational viewpoint, this result is
particularly enlightening. Actually, it is worth to where the scalar ¯ is an appropriate factor, such that
simply rewrite it as speed constraints remain satisfied.
2 2
(vmax ¡ vmin ) 2
k~rmax k2 ¡ k~rmin k2 = r (50)
v 2 APPENDIX C. CONVEXITY OF THE SET OF FEASIBLE
2 2
TRAJECTORIES
and to examine the ratio (vmax ¡ vmin )=v2 . Invoking the
validity of the assumption vtgt ¿ vobs , we have The aim of this appendix is to prove the convexity
of feasible trajectories and associated propositions
2
(vmax 2
¡ vmin ) (vtgt,max ¡ vtgt,min ) cos(®)
¼2 : (51) (Propositions 3 and 4). First, let us define trajectories
v2 vobs
μ 1 ¶ μ 2 ¶
Pinit Pinit
Thus, this ratio acts as a “reduction” factor for T1 = and T2 = :
1 2
estimating the distance r, under speed constraints; thus Pend Pend
showing that the benefit of including hard constraints
Then, from the definition of S, we have
for range estimation is generally quite noticeable.
It remains to show that the set of feasible 8 1
> Pinit 2 C
trajectories is a convex set. Actually, this results >
<
immediately from the norm triangular inequality. 1
T1 2 S () Pend 2C
Proving that (48) corresponds to extremal vectors is >
>
:
1 1
2
a direct consequence of the fact that vy,max 2
= vmax = kPinit ¡ Pend k·L
¡1 2 2
(1 + (Bμ t)3 ) is an extremal value of the vy velocities, (56)
8 2
which is itself resulting from (47). > Pinit 2 C
>
<
Let us consider now the feasible set of target 2
T2 2 S () Pend 2C
trajectories (see Fig. 5). From the previous results, >
>
:
the relative velocity vector (~v) for a candidate target 2
kPinit 2
¡ Pend k · L:
trajectory is given by
μ ¡1 ¶ Let u 2 [0, 1] and let T3 = (1
(Bμ t)2 ¡ ¡3
u) ¢ T1 ¢+ u ¢ T2 ,
3
~v = ~vtgt ¡~vobs = vy : (52) associated with the trajectory Pinit Pend . We show
1 that T3 2 S, where
à 1 2 !
2 (1 ¡ u) ¢ Pinit + u ¢ Pinit
16 Elementary calculations yield: (Bμ¡1 t)1 = rx2 (0)=vy2 , (Bμ¡1 t)22 =
T3 = :
vx2 =vy2 , (Bμ¡1 t)23 = ry2 (0)=vy2 . 1
(1 ¡ u) ¢ Pend 2
+ u ¢ Pend
1260 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 42, NO. 4 OCTOBER 2006
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
1 2 1 2 [13] Gelfand, A. E., and Smith, A. F. M.
Indeed, as C is convex and Pinit , Pinit , Pend and Pend 2C
Sampling based approaches to calculating marginal
we have
densities.
3 1 2 Journal of the American Statistical Association, 85 (1990),
Pinit = (1 ¡ u) ¢ Pinit + u ¢ Pinit 2C
398—409.
(57)
3 1 2 [14] Gong, K. F., Lindgren, A. G., and Nardone, S. C.
Pend = (1 ¡ u) ¢ Pend + u ¢ Pend : 2 C: Passive localization and motion analysis with a state
3 3 parameter constraint.
Then it remains to show that kPinit ¡ Pend k · L, but In Proceedings of the 15th Asilomar Conference on
3 3 1 1 2 2
Circuits, Systems and Computers, Nov. 9—11, 1981,
kPinit ¡ Pend k = k(1 ¡ u) ¢ (Pinit ¡ Pend ) + u ¢ (Pinit ¡ Pend k 366—374.
1 1 2 2
[15] Greub, W.
· (1 ¡ u) ¢ kPinit ¡ Pend k + u ¢ kPinit ¡ Pend k · L: Linear Algebra (4th ed.).
New York: Springer-Verlag, 1975.
(58)
[16] Grossman, N.
REFERENCES The Sheer Joy of Celestial Mechanics.
Boston: Birkhauser, 1996.
[1] Aidala, V. J., and Hammel, S. E. [17] Häggström, O.
Utilization of modified polar coordinates for Finite Markov Chains and Algorithmic Application.
bearings-only tracking. London: Cambridge University Press, Cambridge, London
IEEE Transactions on Autmatic Control, AC-28, 3 (Mar. Mathematical Society, Student Text 52, 2002.
1983), 283—294. [18] Hastings, W. K.
[2] Bobrosky, B. Z., Mayer-Wolf, E., and Zakai, M. Monte Carlo sampling methods using Markov chains and
Some classes of global Cramér-Rao bounds. their applications.
The Annals of Statistics, 15, 4 (1987), 1421—1438. Biometrika, 57 (1970), 97—109.
[3] Boneh, A., and Golan, A. [19] Holtsberg, A.
Constraints redundancy and feasible region boundedness A statistical analysis of bearings-only tracking.
by random feasible point generator (RFPG). Ph.D. dissertation, Dept. of Math. Statistics, Lund
Presented at the Third European Congress on Operations Institute of Technology, 1992.
Research, EURO III, The Annals of Statistics, [20] Jauffret, C., and Pillon, D.
Amsterdam, 1979. Observability in passive target motion analysis.
[4] Broman, V., and Shensa, M. J. IEEE Transactions on Aerospace and Electronic Systems,
Bearings tracking of targets with unobservable states. 32, 4 (Oct. 1996), 1290—1300.
In Proceedings of the American Control Conference (ACC), [21] Jauffret, C.
June 1986, 1850—1854. Observabilité et inversibilité de la matrice de Fisher dans
[5] Belisle, C. J. P., Romeijn, H. E., and Smith, R. L. les problèmes de régression non-linéaire.
Hit-and-run algorithms for generating multivariate Actes du XVII-eme colloque GRETSI, Vannes, France,
distributions. Sept. 1999, 685—688.
Mathematics of Operations Research, 18 (1993), 255—266. [22] Kass, R. E., and Vos, P. W.
[6] Chan, Y. T., and Rudnicki, S. W. Geometrical Foundations of Asymptotic Inference.
Bearings-only and Doppler-bearing tracking using New York: Wiley, 1997.
instrumental variables. [23] Kaufman, D. E., and Smith, R. L.
IEEE Transactions on Aerospace and Electronic Systems, Optimal direction choice for Hit-and-Run sampling.
28, 4 (Oct. 1992), 1076—1083. Operations Research, 46 (1998), 84—95.
[7] Chen, M-H., and Schmeiser, B. W. [24] Lebossé, C. and Hémery, C.
General hit-and-run Monte Carlo sampling for evaluating Géométrie, classe de mathématiques. Programme de
multidimensional integrals. 1945.
Operations Research Letters, 19 (1996), 161—169. Fernand Nathan 1955, Jacques Gabay (reprint, 1997).
[8] Chen, M-H., and Shao, Q-M. [25] Le Cadre, J. P., and Jauffret, C.
Monte Carlo estimation of Bayesian credible and HPD Discrete-time observability and estimability for
intervals. bearings-only target motion analysis.
Journal of Computational and Graphical Statistics, 8 IEEE Transactions on Aerospace and Electronic Systems,
(1999), 69—92. 33, 1 (Jan. 1997), 178—201.
[9] Chen, M-H., Shao, Q-M., and Ibrahim, J. G. [26] Le Cadre, J. P., and Jauffret, C.
Monte Carlo Methods in Bayesian Computation. On the convergence of iterative methods for bearings-only
New York: Springer-Verlag, corrected second printing tracking.
2002. IEEE Transactions on Aerospace and Electronic Systems,
[10] Doucet, A., Godsill, S. J., and Robert, C. P. 35, 3 (July 1999), 801—817.
Marginal maximum a posteriori estimation using Markov [27] Le Cadre, J. P.
chain Monte Carlo. Properties of estimability criteria for target motion
Statistics and Computing, 12 (2002), 77—84. analysis.
[11] Everett, N., Tham, S. T., and Salmond, D. IEE Proceedings on Radar, Sonar Navigation, 145, 2 (Apr.
A novel MCMC tracker for stressing scenarios. 1998), 92—99.
In O. Drummond (Ed.), Proceedings of the SPIE, Signal [28] Nardone, S. C., Lindgren, A. G., and Gong, K. F.
and Data Processing of Small Targets, vol. 5428, 2004. Fundamental properties and performance of conventional
[12] Gauss, K. F. bearings-only target motion analysis.
Theoria Motus Corporium Coelestium. IEEE Transactions on Automatic Control, AC-29, 9 (Sept.
Boston: Little, Brown & Co., 1857. 1984), 775—787.
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
[29] Mallick, M., Maskell, S., Kirubrajan, T., and Gordon, N. [36] Runnals, A., and Sirlantzis, K.
Littoral target tracking using particle filter. Low-observable maritime tracking using Monte Carlo
In Proceedings of the Fifth International Conference on Markov chain methods.
Information Fusion, Annapolis, MD, July 8—11, 2002, In IEE Colloquium on Algorithms for Target Tracking,
935—942. 1995.
[30] McCabe, B. J. [37] Siffredi, E.
Accuracy and tactical implications of bearings-only Transformations sur les systèmes dynamiques. Modélisation
ranging algorithms. et estimation de la cinématique d’un mobile dans le plan à
Operations Research, 33 (1985), 94—103. partir de mesures angulaires bruitées et échantillonnées.
[31] Nardone, S. C., and Graham, M. L. Thèse de doctorat de 3-ème cycle, Marseille, Janvier
A closed-form solution to bearings-only target motion 1982.
analysis. [38] Smith, R. L.
IEEE Journal of Oceanic Engineering, 22, 2 (Jan. 1997), Monte Carlo procedures for generating random feasible
168—178. solutions to mathematical programming.
[32] Nardone, S. C., and Aidala, V. J. In Proceedings of ORSA/TIMS Conference, Washington,
Observability criteria for bearings-only target motion D.C., 1980.
analysis. [39] Smith, R. L.
IEEE Transactions on Aerospace and Electronic Systems, Efficient Monte Carlo procedures for generating points
AES-17, 2 (Mar. 1981), 162—166. uniformly distributed over bounded regions.
[33] Passerieux, J. M. Operations Research, 32 (1984), 1296—1308.
Estimation par régions de confiance en trajectographie [40] Tanner, M. A.
passive. Tools for Statistical Inference (3rd ed.).
Actes du XV-ème colloque GRETSI, Juan-les-Pins, Sept. New York: Springer-Verlag, 1996.
1995, 153—1156. [41] Tierney, L.
[34] Poor, H. V. Markov chains for exploring posterior distributions (with
An Introduction to Signal Detection and Estimation (2nd discussion).
ed.). Journal of the American Statistical Association, 82 (1987),
New York: Springer-Verlag, 1998. 528—550.
[35] Robert, C. P., and Casella, G. [42] Wagner, D. H., Mylander, W. C., and Sanders, T. J.
Monte Carlo Statistical Methods. Naval Operations Analysis, (3rd ed.).
New York: Springer, Texts in Statistics, 2000. Annapolis, MD: Naval Institute Press, 1999.
1262 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 42, NO. 4 OCTOBER 2006
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.
Frédéric Bavencoff received an M.S. in statistics from the Lille 1 University,
France, in 2001, and the Ph.D. from the University of Rennes, France, in 2005.
He is with Thales Airborne Systems, France, where he is involved in
performance analysis of multitarget tracking algorithms, Bayesian methods, and
the theory of evidence.
Authorized licensed use limited to: UR Rennes. Downloaded on July 10, 2009 at 11:15 from IEEE Xplore. Restrictions apply.