You are on page 1of 15

1.

INTRODUCTION
The conventional problem of bearings-only target
motion analysis consists of estimating the trajectory of a
Leg-by-leg Bearings-Only target (or source) whose velocity is constant during the
period of measurement [13]. This requires an efficient
Target Motion Analysis maneuver of the observer to guarantee observability
[6][12] and to obtain an accurate estimate [11][15]. In
Without Observer Maneuver a recent paper [7], we proved that, conversely, if the
observer has a constant velocity and the source changes
its heading (so its trajectory is composed of two legs at
constant speed–see Fig. 5), then, subject to a condition
CLAUDE JAUFFRET on velocity vectors of the two mobiles, the source
is observable. For this problem, called bearings-only
DENIS PILLON
maneuvering target motion analysis (BOMTMA), we
ANNIE-CLAUDE PIGNOL
proposed the maximum likelihood estimate (MLE) and
compared its performance with the Cramér-Rao lower
bound (CRLB), revealing that this estimate is relatively
efficient. The major criticisms are
In a previous paper [7], the problem of bearings-only tracking
of targets whose trajectory is composed of two legs from a non- 1) The operator must wait until the source has
maneuvering observer was addressed and the maximum likelihood changed its heading to run the computation of the esti-
estimate (MLE) proposed. We named it bearings-only maneuvering mate.
target motion analysis (BOMTMA). Recently in [9], we proposed 2) The computation by a numerical routine needs a
another estimate based on leg-by-leg tracking and compare its
“good guess” (to reduce the risk of converging toward
a local minimum).
performance to the MLE. We give here the extended version of [9],
3) The computation takes time.
together with some comparison between the conventional bearings-
only target motion analysis (BOTMA) and the BOMTMA. In this paper, we propose a new approach to this
problem which consists of estimating what is observable
during the first leg of the source, then during the second
one, and finally of fusing these two estimated state vec-
tors to obtain an estimate of the source trajectory. Note
that in the classic BOTMA, the leg-by-leg approach has
been employed for the same reasons [2][14][16]. We
will assume that the maneuver time is known, hence we
will not address the problem of detecting the maneuver
(see [5] and [17] for this topic).
The paper is composed of three main sections:
² In Section 2, we present the problem of target motion
analysis (TMA) when neither the source nor the
observer maneuvers.
² A new bearings-only maneuvering target motion anal-
ysis by a non-maneuvering observer is proposed in
Section 3.
² In Section 4, some examples are provided to compare
respective performances of BOTMA and BOMTMA,
Manuscript received November 12, 2009; revised October 27, 2010; in terms of estimated range accuracy.
released for publication February 08, 2011.
Refereeing of this contribution was handled by Huimin Chen. 2. PROBLEM FORMULATION WHEN NEITHER THE
SOURCE NOR THE OBSERVER MANEUVERS
Authors’ addresses: C. Jauffret, Université Sud Toulon-Var, CNRS,
IM2NP (UMR 6242), Bâtiment X, BP 132, 83957 La Garde Cedex, We consider in this section the case where the source
France, Email: (jauffret@univ-tln.fr); D. Pillon, Thales Underwater and the observer are moving in the same plane with their
Systems, 525 Route des Dolines, 06903 Sophia-Antipolis, France,
Email: (denis.pillon@fr.thalesgroup.com); A. C. Pignol, Université
own constant velocity vectors (see Fig. 1).
Sud Toulon-Var, CNRS, IM2NP (UMR 6242), Bâtiment X, BP 132,
83957 La Garde Cedex, France, Email: (pignol@univ-tln.fr). 2.1. Measurement Equation and Trajectory Model
Consider a source and a passive observer (also called
1557-6418/11/$17.00 °
c 2011 JAIF own ship). From here on, the subscript S is used to

24 JOURNAL OF ADVANCES IN INFORMATION FUSION VOL. 6, NO. 1 JUNE 2011


Fig. 1. Example of observer and source trajectories.

represent source quantities and O to represent observer 2.2. The Set of Homothetic Trajectories
quantities.
The trajectory of a vehicle moving at a constant
At time t, the respective location vectors of the _ T is described by the following
velocity vector [x_ y]
source and of the observer are PS (t) = [xS (t) yS (t)]T classic equations:
and PO (t) = [xO (t) yO (t)]T relative to a Cartesian coor-
dinate system. Similarly, the vectors VS = [x_ S y_ S ]T and x(t) = x(t¤ ) + (t ¡ t¤ )x_
VO = [x_ O y_ O ]T denote the source and own ship velocity (2)
y(t) = y(t¤ ) + (t ¡ t¤ )y:
_
vectors, respectively. We define also the relative velocity
vector of the source w.r.t. the observer as VR = VS ¡ VO . Such a trajectory is hence defined by the vector X =
The corresponding speeds and headings (or courses) are [x(t¤ ) y(t¤ ) x_ y]
_ T.
denoted vS , vO , vR , cS , cO and cR . The equation of the noise-free bearings being μ(t) =
At time tk , the observer measures the azimuth of the atan[(xS (t) ¡ xO (t))=(yS (t) ¡ yO (t))], it is straightforward
line of sight in which it detects the source: to check that the set of trajectories producing the same
· ¸ noise-free-bearings from the observer is
x (t ) ¡ xO (tk )
¯k = atan S k + "k (1) ¤ = fX(¸) = ¸(XS ¡ XO ) + XO , for ¸ > 0g:
yS (tk ) ¡ yO (tk )
If μ(t) is not constant, there is no other trajectory
where "k is assumed to be a zero-mean Gaussian random
set that generates the same noise-free data when the
noise of variance ¾k2 .
source and observer are moving at constant velocity
The BOTMA aims to estimate
vectors [10]. From now on, we will assume that μ(t)
XS = [xS (t¤ ) yS (t¤ ) x_ S y_ S ]T is not constant.
Note that the vector
from the collected measurement set f¯1 , ¯2 , : : : , ¯N g pro-
X(¸) = [x1 (¸) x2 (¸) x3 (¸) x4 (¸)]T
vided observability is guaranteed (t¤ is an arbitrary ref-
erence time). The vector XO = [xO (t¤ ) yO (t¤ ) x_ O y_ O ]T defines a ¸-homothetic trajectory (in particular, X(1) =
is known. It is well known that if the observer does XS ).
not maneuver or if its maneuver is ambiguous (see [6] It follows that XS is not observable from the bearing
and [10]), then the vector XS is not observable. In the measured by the observer. In short, in this context,
coming paragraph, we explore the situations where the the BOTMA is impossible. We can however estimate
observer keeps its velocity vector during the scenario. a parameter (or a state vector) that characterizes ¤. We

LEG-BY-LEG BEARINGS-ONLY TARGET MOTION ANALYSIS WITHOUT OBSERVER MANEUVER 25


call the estimation of this parameter (or any equivalent vector of ¤ as state vector is more convenient. A way to
parameter) partial bearing-only target motion analysis “stay” in ¤ is to fix one coordinate of a 4-dimensional
(see Subsection 2.4). vector X and “adjust” the remained coordinates to Y:
The previous set ¤ can be characterized by any of its For example, if we fix the first coordinate of an par-
elements X(¸). At any X(¸) of ¤, there is a correspond- ticular element of ¤ to the value xfix , the corresponding
ing unique three dimensional vector Y = [y1 y2 y3 ]T de- ¸ will be ¸ = (xfix ¡ xO (t¤ ))=(xS (t¤ ) ¡ xO (t¤ )); however,
fined by we must choose xfix such that ¸ be positive. This will
· ¸
x1 (¸) ¡ xO (t¤ ) help us in Subsection 2.4.
y1 = atan
x2 (¸) ¡ yO (t¤ )
p 2.3. Properties of v(½) and c(½)
(x3 (¸) ¡ x_ O )2 + (x4 (¸) ¡ y_ O )2
y2 = p (3) First of all, we note that v(0) and c(0) are equal
[x1 (¸) ¡ xO (t¤ )]2 + [x2 (¸) ¡ yO (t¤ )]2
· ¸ to the observer’s speed and heading, respectively. This
x3 (¸) ¡ x_ O corresponds to the degenerate case where the observer
y3 = atan :
x4 (¸) ¡ y_ O and the source are located at the same position.
Indeed, the coordinates of Y are independent of ¸: 2.3.1. Study of vS (½)
v Let us compute its derivative w.r.t. ½.
y1 = μ(t¤ ), y2 = R¤ and y3 = cR
½a (t )
d 1
where ½a (t¤ ) is the actual range between the source and v(½) = [(½y2 sin y3 + x_ O )y2 sin y3
d½ v(½)
the observer at time t¤ (see [6, 7]). Note that because
+ (½y2 cos y3 + y_ O )y2 cos y3 ]
μ(t) is not constant, we have VS 6= VO , hence y2 > 0.
We can plot the set of homothetic trajectories ¤ y
= 2 [½y2 + x_ O sin y3 + y_ O cos y3 ]
using the graphs of the two functions v(½)
q · ¸
y 1
½ 7! v(½) = (½y2 sin y3 + x_ O )2 + (½y2 cos y3 + y_ O )2 = 2 ½y2 + VOT (VS ¡ VO ) :
v(½) vR
(4a)
· ¸ The sign of this derivative is hence the sign of ½y2 +
½y2 sin y3 + x_ O x_ O sin y3 + y_ O cos y3 . It is equal to 0 when ½ = ¡(1=y2 vR )
½ 7! c(½) = atan (4b)
½y2 cos y3 + y_ O ¢ VOT (VS ¡ VO ).
where [v(½) c(½)]T are the polar coordinates of the If VOT (VS ¡ VO ) ¸ 0 or equivalently VOT VS ¸ vO2 , the
velocity vector of any source of ¤ at a distance ½ (¸ 0) at function vS (½) is injective, i.e. the mapping ½ 7! v(½)
time t¤ (the corresponding ¸ is equal to ½=½a (t¤ )). Note satisfies the one-to-one condition.
that v(½a (t¤ )) = vS and c(½a (t¤ )) = cS . We draw two other conclusions:
We insist on the fact that 1) The one-to-one condition holds if and only if
1) any element of ¤ allows us to construct the vector _xS sin cO + y_ S cos cO ¸ vO .
Y (see Eq. (3)); 2) The set of source’s velocity vectors VS satisfying
2) conversely, the vector Y allows us to construct the one-to-one condition of the speed is
any element of ¤, thank to the following equation ½ · ¸ · ¸
μ ¶ sin cO cos cO
½ ½ VS = ® ¡¯ ,
X = (X ¡ XO ) + XO cos cO sin cO
½a (t¤ ) ½a (t¤ ) S ¾
2 3
½a (t¤ ) sin y1 for any ® ¸ vO (no condition for ¯) and VS 6= VO
½ 6 ¤
6 ½a (t ) cos y1 7
7
= ¤ 6 7 + XO : Note that ® and ¯ that help define the above set are
½a (t ) 4 ½a (t¤ )y2 sin y3 5
dummy variables.
½a (t¤ )y2 cos y3
This equivalence between Y and ¤ is the fundamental 2.3.2. Study of c(½)
property of the partial bearings-only TMA which will First of all, note that c(½) goes to cR when ½ ! 1.
be developed in Section 3. As a consequence, we can A basic computation yields
choose the vector Y as well any vector in ¤. The choice
d y (y_ sin y3 ¡ x_ O cos y3 )
of a state vector must be guided by simplicity. c(½) = 2 O
d½ v2 (½)
Because it is expressed in polar coordinates, Y is
subject to a constraint: y2 > 0, whereas any vector of which has the same sign as
¤ (expressed in Cartesian coordinates) is not. So, from
the point of view of the estimation, choosing a particular vR (y_ O sin y3 ¡ x_ O cos y3 ) = ¡ det[VO , (VS ¡ VO )],

26 JOURNAL OF ADVANCES IN INFORMATION FUSION VOL. 6, NO. 1 JUNE 2011


Fig. 2. (a) Some elements of ¤. (b) The corresponding speed and heading graphs.

which is independent of ½. Hence the mapping ½ 7! c(½) corresponding to (®, ¯) = (4, 2); the initial positions are
is monotonic while the mapping ½ 7! v(½) can be not. PS (t0 ) = [3:8 1:4]T (km) and PO (t0 ) = [2 0]T (km). Note
Fig. 2 gives an example of an increasing speed that the condition x_ S sin cO + y_ S cos cO ¸ vO is satisfied.
function for VO = [3 0]T (m/s) and VS = [4 2]T (m/s) In Fig. 2(a), letters O and S denote the initial position

LEG-BY-LEG BEARINGS-ONLY TARGET MOTION ANALYSIS WITHOUT OBSERVER MANEUVER 27


Fig. 3. (a) Some elements of ¤. (b) The corresponding speed and the heading graphs.

of the observer and that of the actual source and several Fig. 3 illustrates the case of a non-monotonic speed
homothetic solutions. Fig. 2(b) depicts ½ 7! v(½) and function for VO = [3 0]T (m/s) and VS = [1 2]T (m/s)
½ 7! c(½); the small circles correspond to the actual corresponding to (®, ¯) = (1, 2). The initial positions are
speed and heading. PS (t0 ) = [6:2 6:7]T (km) and PO (t0 ) = [2 0]T (km). Here,

28 JOURNAL OF ADVANCES IN INFORMATION FUSION VOL. 6, NO. 1 JUNE 2011


the condition x_ S sin cO + y_ S cos cO ¸ vO is violated. together with the 95% confidence bands (deduced from
In this case, due to the non-monotony of the map- the CRLB). Detail of the computation of these bands is
ping ½ 7! v(½), a speed v would yield two corresponding given in the Appendix.
estimated ranges and courses as shown in Fig. 3(b).
3. LEG-BY-LEG BOMTMA
2.4. Estimation of the Set of Homothetic Trajectories
(partial bearings-only TMA) 3.1. Problem Formulation
The available measured bearings (¯1 , ¯2 , : : : , ¯N ) are Suppose now that the trajectory of the source is com-
taken at times (t1 , t2 , : : : , tN ). Without loss of general- posed of two legs at constant speed (cf. Fig. 5): the first
ity, we choose t¤ = tN . Assuming the covariance ma- leg starts at t1 and finishes at time tM (assumed to be
trix of the random vector ("1 "2 ¢ ¢ ¢ "N )T to be diagonal, known). Similarly, the second leg starts at tM and fin-
the log-likelihood function to be maximized is propor- ishes at tK . This model of trajectory is simple, but it has
tional to the least squares criterion. The vector Y or been widely adopted in the past, especially in submarine
any vector in ¤ can be chosen as state vector. Because environment (see [1] pp. 175—176 and [4]). The time
of the simplicity of use of Cartesian coordinates, we of the maneuver is assumed to be known; in reality, it
propose to estimate a particular element of ¤ by fix- has to be estimated, for example by a sequential test;
ing its first coordinate x1 to ½¯ sin ¯N + xO (tN ) for con- this point, which is out of the scope of this paper, has
venience, ½¯ being arbitrarily chosen. We call it X. So been addressed in [8]. Such a trajectory is hence param-
we only have to compute the last three coordinates of eterized by the vector Z = [xS (tK ) yS (tK ) vS cS,1 cS,2 ]T
X = [½¯ sin ¯N + xO (tN ) y x_ y] _ T for which the criterion (coordinates of position at time tK , speed, courses of the
PN
C(X) = k=1 (1=¾k2 )[¯k ¡ μk (X)]2 is minimal. first and of second leg). Provided that VOT (VS,1 ¡ VS,2 ) 6= 0
The Gauss-Newton method [3] is used for the min- (observability condition-see its proof in [7]), the entire
imization, initialized at source trajectory is observable. For this problem, we
2 3 proposed the maximum likelihood estimate in [7].
½¯ sin ¯N + xO (tN )
6 7 We propose here another estimate denoted Z̃ the
6 ½¯ cos ¯N + yO (tN ) 7
6 7 principle of which is as follows: First, we compute, for
6 ½¯ sin ¯N + xO (tN ) ¡ ½¯ sin ¯1 ¡ xO (t1 ) 7 leg #1, the estimate
Xinit = 6 7:
6 tN ¡ t1 7
6 7
4 ½¯ cos ¯ + x (t ) ¡ ½¯ cos ¯ ¡ x (t ) 5 X̂1 = [½¯ sin ¯M + xO (tM ) ŷ1 xˆ_ 1 yˆ_ 1 ]T
N O N 1 O 1
tN ¡ t1 and for leg #2, the estimate
As pointed out previously, the maximum likelihood
estimate X̂ allows us to construct the set of estimated X̂2 = [½¯ sin ¯M + xO (tM ) ŷ2 xˆ_ 2 yˆ_ 2 ]T
homothetic trajectories presented as the graphs of the with the common reference time tM and after having
pair of functions fixed their respective first coordinates to a common
q
value, say ½¯ sin ¯M + xO (tM ). These estimates are com-
½ 7! v̂(½) = (½ŷ2 sin ŷ3 + x_ O )2 + (½ŷ2 cos ŷ3 + y_ O )2
puted following the partial BO-TMA principle as pre-
(5a) sented in Section 2.4. Second, we compute the two ho-
· ¸ mothetic estimates, ¹(X̂1 ¡ XO ) + XO for the first leg and
½ŷ2 sin ŷ3 + x_ O
½ 7! ĉ(½) = atan (5b) ¹(X̂2 ¡ XO ) + XO for the second, such that the estimated
½ŷ2 cos ŷ3 + y_ O
velocities on each leg are equal, i.e.
where Ŷ = [ŷ1 ŷ2 ŷ3 ]T is the vector corresponding to X̂
(with (3)). k¹(V̂1 ¡ VO ) + VO k = k¹(V̂2 ¡ VO ) + VO k (6)
The behavior of the estimator X̂ has been evaluated
for the following scenario: given a coordinate system, with V̂1 = [xˆ_ 1 yˆ_ 1 ]T and V̂2 = [xˆ_ 2 yˆ_ 2 ]T in order to satisfy
the initial location of the source is [1000 2300]T (m) the constraint (6) (which is the strong assumptions of
with a velocity vector of [1 1:5]T (m/s). The ob- the BOMTMA). The solution of (6), denoted ¹̃, and
server starts from [0 0]T (m) with a velocity vector equal to
[1 0]T (m/s). The number of measurements is N = 450, 2(V̂1 ¡ V̂2 )T VO
the time tk is equal to k £ ¢t, with ¢t = 4 s. The time ¹̃ = ¡
(kV̂1 k2 ¡ kV̂2 k2 ¡ 2(V̂1 ¡ V̂2 )T VO )
of reference is chosen to be tN . The standard deviation
of the measurements is 1± . The final range is 5,099 m. allows us to compute the corresponding homothetic
For the initialization of the Gauss-Newton method, we estimates on each leg
have chosen ½¯ = 20 km.
Fig. 4 depicts an example of a 500-run Monte-Carlo X̃1 = ¹̃(X̂1 ¡ XO ) + XO = [x̃1 ỹ1 x˜_ 1 y˜_ 1 ]T
simulation: the 500 graphs are plotted in grey, while
the graphs of the functions v(½) and c(½) are in black, X̃2 = ¹̃(X̂2 ¡ XO ) + XO = [x̃2 ỹ2 x˜_ 2 y˜_ 2 ]

LEG-BY-LEG BEARINGS-ONLY TARGET MOTION ANALYSIS WITHOUT OBSERVER MANEUVER 29


Fig. 4. Results of 500 Monte-Carlo runs (on left v̂(½), and on right ĉ(½)).

30 JOURNAL OF ADVANCES IN INFORMATION FUSION VOL. 6, NO. 1 JUNE 2011


Fig. 5. Example of observer and source trajectories composed by two legs.

and the corresponding leg-by-leg BOMTMA Z̃ = it suddenly changes its course and its new heading is
[x̃S (tK ) ỹS (tK ) ṽS c̃S,1 c̃S,2 ]T , with 240± . The total duration of the scenario is 30 min cor-
responding to 450 measurements (the sampling time is
[x̃S (tK ) ỹS (tK )]T = [x̃2 ỹ2 ]T ¢t = 4 s). The standard deviation of the measurement
r noise is 1± .
2 2 The average values of the coordinates of the two es-
Sṽ = x˜_ + y˜_
2 2 timates, their respective biases and their empirical stan-
à ! dard deviations are given in Table I. They are compared
x˜_ 1 to the true vector and the minimum standard deviations
c̃S,1 = atan
y˜_
1
deduced from the CRLB.
à ! We observe an increase in the bias and the standard
x˜_ 2 deviation, but the quality of the leg-by-leg BOMTMA
c̃S,2 = atan : estimator is only weakly degraded. Moreover, the com-
y˜_
2 putation time of the leg-by-leg BOMTMA estimator is
q 2.5 times less than the BOMTMA computation time. A
2 2
qEMARK by construction, x̃1 = x̃2
R and x˜_ 1 + y˜_ 1 = compromise can probably be found: the leg-by-leg es-
2 2
x˜_ 2 + y˜_ 2 , but there is no reason that ỹ1 = ỹ2 . So, an- timate can be used as an initial point for the BOMTMA
numerical routine. It will reduce the risk of stalling at
other solution must be [x̃S (tK ) ỹS (tK )]T = [x̃2 ỹ1 ]T .
a local minimum. Fig. 7 shows the 500 estimates to-
3.2. Problem Formulation gether with the 90% confidence ellipsoid deduced from
the CRLB.
A 500-run Monte Carlo simulation allows the be-
havior of this new estimator to be appreciated. To com- 4. COMPARISON OF THE RESPECTIVE
pare the MLE BOMTMA estimate Ẑ and the leg-by-leg PERFORMANCES OF THE BOMTMA AND THE
BOMTMA estimate Z̃, we use the scenario presented in CONVENTIONAL BOTMA
[7] which is illustrated in Fig. 6: let us recall that the One can ask a relevant question concerning tactical
observer starts from the origin with a speed of 5 m/s aspects: do situations exist in which the performance
and a heading of 90± . Meanwhile, the source, with a of the BOMTMA is superior to the performance of
speed of 4 m/s, starts its trajectory at [0 km, 10 km]T the BOTMA in terms of estimated range accuracy,
with an initial heading of 90± . At time tM = 20 min, assuming that the antennas are the same?

LEG-BY-LEG BEARINGS-ONLY TARGET MOTION ANALYSIS WITHOUT OBSERVER MANEUVER 31


Fig. 6. 500 MLE BOMTMA estimates of the final position.

TABLE I
Comparison of Performances of the Two Estimators

ZK Units ZK,True Average of ẐK Average of Z̃K Bias of Ẑ Bias of Z̃ ¾CRLB ¾ˆ of Ẑ ¾˜ of Z̃

xS (tK ) km 2.921 2.897 2.872 0.024 0.049 0.153 0.175 0.213


yS (tK ) km 8.800 8.828 8.875 0.028 0.075 0.283 0.308 0.364
vS m/s 4 4.10 4.10 0.10 0.10 0.03 0.13 0.17
cS,1 degree 90 90.5 89.4 0.5 0.6 12 12 13
cS,2 degree 240 240 241 0 1 7.5 7 9

For the conventional BOTMA, the observer must “surrounding” trajectory (see Fig. 8(a)) and two “es-
correctly maneuver and it has to estimate a 4-dimen- caping” trajectories (see Figs. 9(a) and 10(a)). The total
sional state vector [13], whereas for the BOMTMA, the duration is 30 min. The maneuvering platform changes
observer does not maneuver but it has to estimate a 5- its course at 15 min. At this time, its location is aligned
dimensional vector. Because the number of unknown is with the course of the observer.
less in BOTMA than in BOMTMA, one can think that The bearings, collected at a sample time of 4 seconds
the BOMTMA returns a less accurate estimated range by each platform, are corrupted by an additive Gaussian
than the BOTMA. Surprisingly, for some scenarios, noise with the same standard deviation equal to 1 de-
this statement is wrong. We give three examples: the gree. We have computed the relative accuracy of the
first one contradicts the intuition; for the second, the estimated range (at the final time tK ) by the mean of the
performances are similar, and the last is an example of CRLB: In ¾½(tK ) =½(tK ), given in percentage, ¾½(tK ) is the
the superiority of the BOTMA to the BOMTMA. CRLB of the estimated final range.
We consider two mobiles: one maneuvers (denoted Typical scenarios are plotted in Figs. 8, 9 and 10,
hereafter by the letter “M”) and the second (denoted (“M” and “N” give the initial positions of the two plat-
“N”) does not. Each of them performs a TMA against forms) together with the corresponding ¾½(tK ) =½(tK ) vs.
the other. ½(tK ). In these figures, the lines joining circles are re-
For each example, the speeds of M and N are lated to the conventional BOTMA and the lines joining
4 m/s and 5 m/s, respectively. The heading of the non- the “+” are for the BOMTMA. We have changed the
maneuvering platform is equal to 90 degrees. The kinds final range by modifying the initial position of the non-
of maneuvering source trajectory have been chosen: a maneuvering platform along the x-axis.

32 JOURNAL OF ADVANCES IN INFORMATION FUSION VOL. 6, NO. 1 JUNE 2011


Fig. 7. 500 leg-by-leg BOMTMA estimates of the final position.

EXAMPLE 1 In this scenario, the first and the second APPENDIX. COMPUTATION OF THE CONFIDENCE
headings of the maneuvering source are respectively BANDS
¡135± and 135± . To conduct properly the computation, we need to
EXAMPLE 2 The first and the second headings of the define the following vector X̃ = [y x_ y] _ T whose com-
maneuvering source are respectively 135± and ¡135± . ponents are the last there components of X (defined
in Subsection 2.4). We rename the components of X̃ as
EXAMPLE 3 The first and the second headings of the follows: X̃ = [x1 x2 x3 ]T for convenience and we define
maneuvering source are respectively ¡135± and 135± .
the function
· ¸
5. CONCLUSIONS ¯ ¯t¤ + (tk ¡ t¤ )(x2 ¡ x_ O )
½sin
μk (X̃) = atan
We have presented the problem of bearing-only tar- x1 ¡ yO (t¤ ) + (tk ¡ t¤ )(x3 ¡ y_ O )
get motion analysis when neither the source nor the which would be the noise-free bearing of a source at
observer maneuvers. This yields the so called partial time tk , whose trajectory would be defined by X.
BOTMA, since the observability is missing. Then, we Under classic Gaussian assumption about the addi-
have proposed another estimate for a two-leg source tra- tive noise, the Fisher information matrix is
jectory by bearings-only TMA from a non-maneuvering
observer, which is an alternative solution of the maxi- XK
1
mum likelihood estimate proposed in [7]. The compu- F(X̃) = r μ (X̃)rTX̃ μk (X̃)
2 X̃ k
¾k
tation time is reduced by a factor of approximately 2.5. k=1

The price is a small degradation in the statistical per-


and the CRLB of X̃ is B(X̃) = F(X̃)¡1 .
formance. We have also shown, by three examples, that
We go into detail about the expression of rX̃ μk (X̃)
the superiority of the BOTMA to the BOMTMA is not
by defining the following quantities
always guaranteed.
Further work will be carried out to extend this ¢t (tk ) = tk ¡ t¤
estimation principle (leg-by-leg estimates, then fusion
of them) to the case of several legs. Robustness to ¢x (tk ) = ½¯ sin ¯t¤ + ¢t (tk )(x2 ¡ x_ O ),
the assumption of constant speed and of the immediate
¢y (tk ) = x1 ¡ yO (t¤ ) + ¢t (tk )(x3 ¡ y_ O ),
change of heading will be the topic of another paper [8].
q
But the challenge remains to construct a powerful
r(tk ) = ¢2x (tk ) + ¢2y (tk ):
test to detect the maneuver of the source.

LEG-BY-LEG BEARINGS-ONLY TARGET MOTION ANALYSIS WITHOUT OBSERVER MANEUVER 33


Fig. 8. (a) Typical surrounding maneuvering platform trajectory for Example 1. (b) The relative accuracies of the estimated ranges for the
first example (“o” for the conventional BOTMA, “+” for the BOMTMA).

34 JOURNAL OF ADVANCES IN INFORMATION FUSION VOL. 6, NO. 1 JUNE 2011


Fig. 9. (a) Typical escaping maneuvering platform for the second example. (b) The relative accuracies of the estimated ranges for the
second example (“o” for the conventional BOTMA, “+” for the BOMTMA).

LEG-BY-LEG BEARINGS-ONLY TARGET MOTION ANALYSIS WITHOUT OBSERVER MANEUVER 35


Fig. 10. (a) Typical escaping maneuvering platform trajectory for the third example. (b) The relative accuracies of the estimated ranges for
the third example (“o” for the conventional BOTMA, “+” for the BOMTMA).

36 JOURNAL OF ADVANCES IN INFORMATION FUSION VOL. 6, NO. 1 JUNE 2011


We deduce that [5] J. Holst and A. Holtsberg
2 3 On target manoeuvres in bearings-only tracking.
¡¢x (tk ) In Proceedings of the 25th Asilomar Conference on Circuits
1 6 7
rX̃ μk (X̃) = 2 4 ¢t (tk )¢y (tk ) 5 : Systems, (Nov. 1991).
r (tk ) [6] C. Jauffret and D. Pillon
¡¢t (tk )¢x (tk ) Observability in passive target motion analysis.
IEEE Transactions on Aerospace and Electronic Systems,
The computation of the confidence bands is based
AES-32, 4 (Oct. 1996), 1290—1300.
on the Jacobians of the two following mappings [7] C. Jauffret, D. Pillon, and A. C. Pignol
2 3 2 3 Bearings-only TMA without observer maneuver.
x1 y1 · ¸
6 7 6 7 v(½) In Proceedings of the 11th International Conference on Infor-
X̃ = 4 x2 5 7! Y = 4 y2 5 7! V(½) = , mation Fusion, Cologne, Germany, (July 2008), 501—508.
c(½) [8] C. Jauffret, D. Pillon, and A. C. Pignol
x3 y3 Bearings-only maneuvering target motion analysis from a
with non maneuvering platform.
2 · ¸ 3 IEEE Transactions on Aerospace and Electronic Systems,
½¯ sin ¯t¤
atan AES-46, 4 (Oct. 2010), 1934—1949.
6 x2 ¡ yO (t¤ )7 [9] C. Jauffret, D. Pillon, and A. C. Pignol
6 p 7
6 (x3 ¡ x_ O ) + (x4 ¡ y_ O ) 7
2 2 Leg-by-leg bearings-only TMA without observer maneu-
6 7 ver.
Y=6p 7 and
6 (½¯ sin ¯t¤ )2 + [x2 ¡ yO (t¤ )]2 7 In Proceedings of the 12th International Conference on In-
6 · ¸ 7
4 _ 5 formation Fusion, Seattle, (July 2009), 2036—2043.
x3 ¡ xO [10] J. P. Le Cadre and C. Jauffret
atan
x4 ¡ y_ O Discrete-time observability and estimability analysis for
2 p 3 bearings-only target motion analysis.
v(½) = (½y2 sin y3 + x_ O )2 + (½y2 cos y3 + y_ O )2 IEEE Transactions on Aerospace and Electronic Systems,
V(½) = 4
· ¸ 5: AES-33, 1 (Jan. 1997), 178—201.
½y2 sin y3 + x_ O
c(½) = atan [11] J. P. Le Cadre and S. Laurent-Michel
½y2 cos y3 + y_ O Optimizing the receiver maneuvers for bearings-only track-
The CRLB of V(½) is then ing.
Automatica, 35 (1999), 591—606.
B(V(½)) = J2 J1 B(X̃)J1T J2T [12] S. C. Nardone and V. J. Aidala
Observability criteria for bearings-only target motion anal-
where J1 is the Jacobian of the mapping X̃ 7! Y and J2 ysis.
is the Jacobian of the mapping Y 7! V(½). We evaluate IEEE Transactions on Aerospace and Electronic Systems,
AES-17, 2 (Mar. 1981), 162—166.
the confidence bands of each component of v(½) and of
[13] S. C. Nardone, A. G. Lindgren, and K. F. Gong
c(½) from the diagonal terms of B(V(½)). Fundamental properties and performance of conventional
bearings-only target motion analysis.
ACKNOWLEDGMENT IEEE Transactions on Automatic Control, AC-29, 9 (Sept.
1984), 775—787.
We are grateful to Caroline Turner for having up- [14] S. C. Nardone and M. L. Graham
graded the writing of this paper. A closed-form solution to bearings-only target motion anal-
ysis.
REFERENCES IEEE Journal of Oceanic Engineering, OE-22, 1 (Jan. 1997),
168—178.
[1] Y. Bar-Shalom and X. R. Li [15] J. M. Passerieux and D. Van Cappel
Estimation and Tracking: Principles, Techniques, and Soft- Optimal observer maneuver for bearings-only tracking.
ware. IEEE Transactions on Aerospace and Electronic Systems,
Artech House (1993). AES-34, 3 (July 1998), 777—788.
[2] P. Blanc-Benon and J. M. Passerieux [16] D. T. Pham
Fusion of reduced-rank TMA estimates. Some quick and efficient methods for bearings-only target
IEEE Transactions on Aerospace and Electronic Systems, motion analysis.
AES-36, 3 (July 2000), 750—759. IEEE Transactions on Signal Processing, SP-41, 9 (Sept.
[3] J. E. Dennis 1993), 2737—2751.
A user’s guide to nonlinear optimization algorithms. [17] J. Ru, V. P. Jilkov, X. R. Li, and A. Bashi
In Proceedings of IEEE, 72, 12 (Dec. 1984). Detection of target maneuver onset.
[4] K. C. Ho and Y. T. Chan IEEE Transactions on Aerospace and Electronic Systems,
An asymptotically unbiased estimator for bearings-only and AES-45, 2 (Apr. 2009), 536—554.
Doppler-bearing target motion analysis.
IEEE Transactions on Signal Processing, 54, 3 (Mar. 2006),
809—821.

LEG-BY-LEG BEARINGS-ONLY TARGET MOTION ANALYSIS WITHOUT OBSERVER MANEUVER 37


Claude Jauffret born in France on March 29, 1957, received the diplôme d’Etudes
Appronfondies in applied mathematics from Saint Charles University, Marseille,
France, in 1981, the Diplôme d’Ingénieur from Ecole Nationale Supérieure
d’Informatique et de Mathématiques Appliqués de Grenoble, Grenoble, France, in
1983, the title of Docteur de l’Université in 1993 and the Habilitation à Diriger des
Recherches from the Université de Toulon et du Var, France.
From November 1983 to November 1988, he worked on passive sonar systems,
more precisely on target motion analysis at GERDSM, France. After a sabbatical
year at the University of Connecticut (from November 1988 to December 1989) dur-
ing which he worked of tracking problems in cluttered environment, he developed
researches in tracking, data fusion, and extraction in CERDSM. Since September
1996, he has been at the Université de Toulon et du Var where he teaches statistical
signal processing. His current researches are about observability, estimation in non
linear systems as they appear in tracking problems.
He is an associate editor for IEEE Transactions on Aerospace and Electronic
Systems.

Denis Pillon comes from a small French mountain called “Le Jura.” He received the
Diplôme d’Ingénieur from Ecole Supérieure de Physique et de Chimie Industrielle
(ESPCI-Paris) in 1977 and the Diplôme d’Etudes Approfondies in probability
theory from the Jussieu University, Paris in 1977.
From 1978 to 1979 he worked in the area of operational research at the Centre
InterArmées de Recherche Opérationnelle (CIRO) agency of the ministère de la
Défense. Since 1980 he has been employed by Thomson Compagny (now Thales
Compagny), France. From 1980 to 1981 he worked on sonar array processing. From
1982 to 1986, he was consultant at the GERDSM for which he worked on sea trial
analysis and TMA. From 1987 to 1989, he worked on tracking and track clustering
and in 1990 he was appointed group leader of signal and data processing laboratory
at Thomson Sintra company, Sophia Antipolis, France. From 1992 to 1999, he
was consultant at the DCN sonar laboratory. He worked on new sonar arrays and
various submarine combat systems. Since 2000, he has been sonar expert at Thales
Underwater Systems (Sophia-Antipolis). His current researches are about passive
sonar performances estimation, submarine and torpedo array processing, and long
term prospective studies for submarine warfare.

Annie-Claude Pignol born in France on November 10, 1965, received the Diplôme
d’Etudes Appronfondies in optics and image processing from the Université de
Toulon et du Var, Toulon, France, in 1988 and the title of Docteur de l’Université
in 1991 from the Université de Toulon et du Var, France.
Since September 1994, she has been at the Université de Toulon et du Var
where she teaches electronic systems. Her researches were focused signal processing
applied to biomedical systems before turning them to Target Motion Analysis.

38 JOURNAL OF ADVANCES IN INFORMATION FUSION VOL. 6, NO. 1 JUNE 2011