Multi Frequencies And Bearing Target Motion Analysis :
Properties and Sonar Applications
Claude Jauffret Pierre Blanc-Benon
Université Sud Toulon-Var Thales Underwater Systems
CNRS, IM2NP (UMR 6242) Route Sainte-Anne-du-Portzic
Bâtiment X, BP 132, F-83957 29601 Brest, France
La Garde Cedex, France pierre.blancbenon@fr.thalesgroup.com
jauffret@univ-tln.fr
Abstract - Multi frequency lines, concurrently processed
with bearing measurements, are widely used for passive
sonar target motion analysis (TMA). This paper analyzes
the impact of utilizing additional narrow band (NB) lines
on the target range accuracy obtained after TMA. More
specifically, the TMA range and velocity accuracy
improvement that is obtained from adding multiple
frequency lines is demonstrate mathematically. This is a
new result, since to date this performance improvement
has only been shown numerically via Monte Carlo
simulations. Data recorded at sea is utilized to validate
the results and demonstrate the improvement obtained by
using additional frequency tracks in the TMA process.
This is a good example of the benefit of the association
and fusion of multiple frequency bearing tracks.
Keywords: TMA, Cramer Rao lower bound, passive
sonar, narrow band, tracking, track-to-track association.
1 Introduction.
The most standard target motion analysis (TMA) consists
in estimating target's position and (constant) velocity from
bearings-only measurements corrupted by noise [1]. When
frequency line is also available, own ship’s maneuver is
not necessary provided that the emitted and unknown
frequency is a constant and the bearing rate is not equal to
zero [5][16]. This well known method will be called in the
sequel FBTMA1 for frequency and bearing target motion
analysis.
In some cases, the source can emit several frequencies; in
such a situation, taking simultaneously into account several
frequency lines must improve the accuracy of the final
result as we conjectured in a previous paper [2]. The
fundamental question is then the following: is the
improvement of the accuracy (in terms of covariance
matrix) always guaranteed whereas the state dimension
augments with the number of frequency lines?
In [2], a partial answer was given by Monte Carlo
simulations, but not rigorously demonstrated
mathematically.
1
In the literature [2, 17], its name is also Doppler and
bearing TMA.
509
Denis Pillon
Thales Underwater Systems
525 Route des Dolines
06903 Sophia-Antipolis, France
denis.pillon@fr.thalesgroup.com
We propose in this paper to prove and illustrate the benefit
of such a method, called in the sequel multi frequency and
bearing target motion analysis (MFBTMA).
The paper is articulated as follows:
In section 2, the architecture of a narrow band sonar
system is recalled. Then in section 3, the specific block
structure of Fisher Information Matrix (FIM) in the case of
single frequency line is presented. This structure is
generalized to the case of MFBTMA. It is then used in
subsection 4.3 to demonstrate the TMA accuracy
improvement. Finally, in section 5, real data collected at-
sea from a towed array sonar illustrate this benefit.
2 Narrow band sonar system.
We recall shortly the treatments necessary to the
construction of the measurement lines. In sonar
narrowband signal processing, the typical architecture of
the processing chain is the following.
First, different bands in the frequency domain are defined
by an appropriate resolution cell and a center frequency.
The received signal is sampled and filtered in these bands
before applying the fast Fourier transform (FFT) and beam
forming. Then, an interpolation algorithm returns a fine
angle of arrival estimate [6]. An ad-hoc threshold is used to
declare the detection. Finally, automatic detection and
tracking (ADT) is performed to associate elementary
detections from one scan time to another. Many methods
are available, from the simplest one (PDA, NNA [7]) to the
most recent ones (MHT [8], Multi-Dimensional
Assignment [9], [10]).
A single target can be detected and tracked by several
bearing and frequency lines. That is the reason why the
need of automatic track-to-track association is important. It
consists in clustering all detected track into a partition of
mutually disjoint subsets. A human operator would not be
efficient at manually detecting and extracting significant
information related to every single potential threat. For
example, in the case of a large surveillance acoustic
system, sonar operators have to handle more than thousand
tracked lines, which are radiated from numerous merchant
ships and a few from warships. Thus, automatic track-to-
track association algorithms ([11], [12]) are an important β m (tk ) = β ( X S , tk ) + ε β (tk )
requirement for modern sonar systems.  for ( k = 1,...., K )
 f m (tk ) = f (YS , t k ) + ε f (tk )
where ε β (t k ) et ε f (t k ) are the additive noise on
bearings and frequencies, respectively. In the sequel, these
3 Frequency and bearing target noise are assumed independent, 0-mean Gaussian and their
motion analysis (FBTMA) standard deviations are σ β (t k ) et σ f (t k ) , respectively,
assumed known (or previously estimated).
3.1 Notation and hypotheses. Remarks :
a) The azimuths are defined by X S only while the
The observer (O) and the single source (S) are assumed
traveling in the common plane at constant velocity and frequency must be defined by YS .
constant heading. Moreover, the source emits a single tone
b) In a TMA context, f 0 has to be considered as a
with constant and unknown frequency f 0 . The state
nuisance parameter, in the sense that we must
vector YS is hence defined as follows: estimate it but we are concerned essentially by the

[
YS = f 0 , x S ( t * ), y S ( t * ), v xS , v yS ] T
(1)
vector X S .

* *
where f 0 , xS ( t ), y S ( t ), vxS , v yS are respectively the
3.2 The FIM in the FBTMA
the unknown emitted frequency, the components of the
position at time t * (the chosen reference time ) and the The Fisher information matrix (FIM) about YS takes the
components of the velocity of the source. For commodity following form:
reason, we denote the last four components of YS by
F (YS βm , f m ) = ∑
K
1
∇YS β( X S , t k )∇YTS β( X S , t k )
[
X S = xS ( t ), y S ( t ), v xS , v yS
* *
] T
. k =1 σ (t k )
2
β

So [
YS = f 0 , X ST ] T
. (2) +∑
K
1
∇YS f (YS , t k )∇YTS f (YS , t k ) (5)
k =1 σ (t k )
2
The noise-free azimuths are defined by f
 x (t ) − xO (t k )  Due to the special structure of the state vector (2), the FIM
β ( X S ,t k ) = atan  S k  (3)
 y S (t k ) − yO (t k )
(5) can be partitioned into four blocks:
 γ0 c0T 
where xO (t k ) and yO (t k ) are the components of F (YS β m , f m ) =   (6)
observer’s position at time tk . c0
 F ( X S βm ) + F ( X S f m )
The noise-free Doppler-shifted frequency is then with:
[
f ( t ) = f 0 1 − v(t ) / C (4) ]
F (X S β m ) = ∑
K
1
where v( t ) is the radial velocity and C is the sound ∇ X S β ( X S , t k )∇ XT S β ( X S , t k )
speed in the medium (about 1500 m/s in water). The radial k =1 σ (t2
βk )
velocity must be expressed in terms of X S : which is nothing else but the usual FIM about the state
vector X S given the bearing measurements.
( )
ν ( X S , t ) = v x S − v xO sin [ β ( X S , t )] The FIM about X S given the frequency measurements is
+ (v yS − v yO )cos [β ( X S , t )] the following (4× 4) matrix
v xO and v xO are the components of observer’s
F (X S f m ) = ∑ 2
where K
1
∇ X S f (YS , t k )∇ XT S f (YS , t k ) (7)
speed vector. k =1 σ f (t k )
.
Under those classical assumptions, the so-called
frequencies and bearings target motion analysis (FBTMA) The last two blocks of (5) are the vector (4×1) ,
consists in estimating YS from a collection of couples of K
1  ∂ f (YS , t k )
c0 = ∑  ∇ X f (YS , t k ) (8),
k =1 σ (t k ) 
measurements 2
∂ f0  S
f

510
and the scalar b) Again, the N unknown frequencies
1  ∂ f (YS , t k ) 
2
K f 0 , f1 ,m , f N −1 are nuisance parameters.
γ0 = ∑   (9).
k =1 σ (t k )  ∂ f0
2 c) We insist on the fact that the bearing rate is not
f  equal to zero and that the measurements are
unbiased.
We are going to exploit this decomposition of the FIM to d) The numbering of the frequencies is arbitrary.
prove a remarkable property about the accuracy of the e) The association between traks is supposed error-
TMA when the number of received frequencies increases. free.
4 Multi Frequency and bearing target
motion analysis (MFBTMA) 4.2 Expression of the FIM in a block
structure
4.1 Notations. As previously (16), the FIM of MBFTMA can be
In this section, the source is supposed to emit N pure partitioned as:
stable and unknown frequencies f 0 , f1 ,  , f N −1 .
(
F YS βm , f 0,m ,m, f N −1,m = )
The state vector (1) must hence be augmented as follows:
[
YS = f 0 , f1 , , f p , , f N −1 , X ST ] T
. The measurements
 γ0 0 0 c0T 
collected by the observer is then { βm , f 0 ,m ,m , f N −1,m }. 
0 r 0


o
 
The N measured frequencies are given by  0 0 γ N −1 c TN −1 
f p ,m (t k ) = f p (Y S, t k ) + ε f p (t k ), 
( )
c0 m c N −1 F (X S βm ) + F X S f 0,m + m + F X S f N −1,m ( )


for p = 0,1,m , N − 1 , with, as previously (4)
f p ( Ys , t k ) = f p [1 − v( X s ,t k ) / C ] .
The vectors c p and the scalars γ p are similarly defined
The variance of ε f p (tk ) is denoted σ f p (tk ) .
than c0 (8) and γ 0 (9) respectively.
We are interested by the inverse of the submatrix
The fundamental question of the MFBTMA is the
following: ( )
F ( X S βm ) + F X S f 0 ,m + m + F X S f N −1,m ( ) corres-
Does the MFBTMA potentially improve the accuracy of ponding to the position and the velocity of the source.
the estimate of X S ? In terms of Cramèr-Rao lower bound We are going to use the following classic result of linear
algebra (see [13]):
(CRLB), the question is
(
BCR X S βm , f 0 ,m ,m, f N −1,m , f N ,m ) Let a non-singular matrix A and B its inverse, both
partitioned in four blocks
(
≤ BCR X S βm , f 0 ,m ) ?
 A11 ( p × p ) A12 ( p × q ) 



A= 
Remarks:  A21 (q × p ) A22 (q × q )
a) We implicitly assume that there is one single
 B11 ( p × p ) B12 ( p × q ) 
bearing measurement, the same one for all the
frequencies at each sampling time. In reality, a B= .
measured frequency is always linked to a  B21 (q × p ) B22 (q × q )
measured bearing. Hence, we should have as
many azimuths as frequencies. Because all these −1
azimuths are relative to the same line of sight, then B22 = A22 − A21 A11−1 A12 provided A11 is
they are averaged and the result is a more accurate non-singular.
single measured bearing. In order to make a fair
comparison, the standard deviation of the bearing We then identify
has been kept at the same value. So, the impact of
the extra frequency lines will be objectively
judged.

511
 γ0 0 0  4.3 Fundamental property of the MFBTMA
  Now we are able to answer the question about the
A11 to  0 r 0  ,
  improvement of the accuracy concerning X S only when
 0 0 γ N −1  we measure an extra frequency, in a recursive way:
A22 to (
BCR X S β m , f 0 ,m ,m , f N −1,m , f N ,m )
[F (X S ( )
β m ) + F X S f 0, m + m + F X S f N −1, m ( )] (
≤ BCR X S β m , f 0 ,m ,m , f N −1,m ? )
, or equivalently
 c0T  (
BCR X S β m , f 0 ,m ,m , f N −1,m , f N ,m )−1

 
A12 to  o  and (
≥ BCR X S β m , f 0 ,m ,m , f N −1,m )
−1
?
c TN −1 
 
A21 to [c0 m c N −1 ] . We can re-write (11) as follows
(
BCR X S β m , f 0 ,m ,m , f N −1,m , f N ,m )−1
=

(X ) ( )
Because the bearing rate is not equal to zero, the frequency −1
rate are non null too and as a consequence the matrix A11 BCR S β m , f 0 ,m ,m , f N −1,m + F X S f N ,m − γ N−1 c N c NT .
is non singular. The recalled result can be applied. The answer is unambiguous if the matrix
We get (
B22 = BCR X S βm , f 0 ,m ,m , f N −1,m given by ) ( ) −1
F X S f N ,m − γ c N c is non negative.
N
T
N

(
BCR X S β m , f 0 ,m ,m , f N −1,m )
−1
=
This
γ N
property
c T
is satisfied since the matrix

[F (X ( )
βm ) + F X S f 0 ,m + m + F X S f N −1,m ( )]
N
  is the Fisher information matrix
S
−1
c N
 (
F X S f N ,m  )
γ0 0

0 

 c0T 
  ( )
F YS f N ,m , if we measure f N ,m only. Hence, this
− [c0 m c N −1 ] 0 r 0   o . matrix is a symmetric non negative matrix.
    The property (10) is then rigorously proven.
 0 0 γ N −1  c TN −1 
 
4.4 Algorithm aspects
We show readily that
−1 The MFBTMA can be obtained via several estimators:
  γ0 0 0   c0T  
 - The recursive estimators given by the pseudo
   
[c m c ] 0 r 0   o   = linear filters [4], or by the extended Kalman filter
N −1
 0    
[3] which all suffer from the same instability than
  0 0 γ N −1  c N −1  
T in BOTMA, especially when the source is far
  from the CPA (Closest Point of Approach). The
N −1
particular filters [15] could be used for a recursive
∑γ
p =0
−1
p c p c pT . estimation.
- The batch estimators, for example those given by
the so-called MIV (for modified instrumental
As a consequence, we get method) [2] [17], or the least squares estimator
equivalent to the maximum likelihood estimator

( )
in the case of additive and Gaussian noise.
−1
BCR X S βm , f0,m ,m, f N−1,m = We suggest retaining the latter estimator which consists in
minimizing the following quadratic criterion
[F(X β ) + F(X
S m S )
f0,m +m+ F X S f N−1,m ( )].
N −1
− ∑γ−p1 c pcpT (11).
p=0
This first result is will help us to prove the fundamental
property of the MFBTMA in the coming section.

512

 β (t ) − β ( X S , t k ) 
2
K
C (YS ) = ∑  m k  +

k =1  σ β (t k ) 
 f p ,m (t k ) − f p (YS , t k ) 
2
N −1 K
∑∑ 
p = 0 k =1  σ f p (t k )
 .
 
The estimate ŶS is reached with the Gauss-Newton
procedure (or the Newton-Raphson one) demanding less
then twelve iterations. The initialization point must be
carefully chosen (or guessed): according to the situation,
we can benefit from prior information or we can use a
coarse estimate chosen among the nodes of an ad-hoc grid.
We recommend to employ the so-called pseudo-linear
estimator (given by the first iteration of the MIV), as
suggested in [2]. Concerning the frequencies, we propose
to used the mean value of each track as initial value of the
emitted frequency.
5 Experimental Results
5.1 Experimental scenario
The experiment was held in coastal shallow waters [14]
(see Fig. 2 for geographic plot). The range between the two
mobiles was between 12 and 18 km. The duration was ~60
minutes and the target reached the CPA at time ~40
minutes. We can see that the observer was not strictly in
straight line, but this maneuver was inefficient for the
bearing-only TMA; consequently, the frequency
measurements allowed one to perform the TMA.
Bottom depth is approximately 100 m. Acoustic
propagation was such that interferences between direct and
reflected paths occurred (see Fig. 1). Those interferences
caused measurement lines interruptions. More than ten
frequency lines were detected from the target on the so
called lofargram as depicted in Fig. 1. Among the set of
frequency lines, seven were chosen for the study: those
corresponding to the frequency lines at about 64 Hz, 66
Hz, 77 Hz, 84 Hz, 90 Hz, 97 Hz and 110 Hz, respectively
(they can be “guessed” in the Fig. 1).
Typical bearing and frequency accuracies are respectively
2 degrees and 0.1 Hz for the sonar measurements.
5.2 MFBTMA at sea results
Batch TMA processing is performed to estimate the target
state vector successively from overlapping time periods.
Fig. 3 displays an example of the time evolution of the
normalized residuals: the top of fig. 3 is concerned by
bearing residuals and the bottom by frequency residuals.
Obviously, the MBFTMA performs properly on real data
provided the assumptions are respected.
Fig.4 presents the true range history and the estimated
range history obtained by the MBFTMA with the seven
frequency lines, the estimation period being equal to 2
513
minutes (corresponding to the ticks displayed in Fig. 2).
The obtained results were exploitable after the 26th minute,
before the weak observer maneuver which occurred at the
32nd minute.
The relative error of the estimated range is less 20% at the
30th minute and less than 10% at the final instant.
5.3 Comparison between FBTMA and
MFBTMA at sea
With each frequency line, we ran the FBTMA with the
batch algorithm in the same previous conditions. We
obtained seven estimated range history displayed in Fig.5.
The result are disperse and the error can reach 30 to 60 %
at the 30th minute depending on the chosen frequency line.
The same report can be made at the final time.
For instance, with the frequency line around 90Hz or by
around 110Hz, the error in the estimated range can reach
20 % at the final time, whereas with these close to 97Hz,
the result is excellent. We can conclude that the accuracy
of the FBTMA depends hence on the chosen frequency
line and so not very reliable. This is likely due to some bias
of the measurement lines.
If these seven FBTMA results are averaged, we obtain the
dash line of the fig. 5 very close to the similar curve
displayed in fig. 4 concerning the MFBTMA.
The advantage of the MFBTMA on the FBTMA is that a
sole run of the ad-hoc algorithm provides a reliable
solution while the FBTMA demands as many runs as
frequency lines to get the same result. This point
disqualifies the use of the FBTMA in case of a large
number ( N ≥ 100 ) of frequency lines.
6 Conclusion
In passive narrow band sonar, it is usual to track targets
emitting several pure single tones. At each sampling time,
the available measurements are a bearing and a set of
frequencies. The crucial question we answered was to
know if it’s worth using all these frequencies in a TMA
function, since each frequency augments the dimension of
the state vector.
We have rigorously demonstrated that the more
frequencies are taken into account in the TMA, the more
accurate the target trajectory estimate (positions and
speeds) will be: it is a theoretical point obtained by the
computation of the CRLB.
The confrontation to at sea data confirms that fact: more
precisely, we observe shorter convergence of the algorithm
and better accuracy on the estimated target range (or
conversely a better accuracy within the same time).
The same approach can be conducted in the case of
frequency comb, i.e. harmonically separated frequency
lines.
7 References
[1] Nardone, S.C., Lindgren, A.G., Gong, K.F.(1984)
Fundamental Properties and Performance of Conventional
Bearings-Only Target Motion Analysis.
IEEE Transactions on Automatic Control, AC-29, 9 (Sept.
1984), pp 775-787.
[2] J.M. Passerieux, D.Pillon, P. Blanc-Benon, C. Jauffret,
“Target Motion analysis with Bearings and Frequency
Measurements via Intrumental Variable”, in Proc. of IEEE
ICASSP 1989, Glasgow (UK), pp 2645-2648.
[3] G.W. Johnson, A.O. Cohen, D.E. Ohlms, C.W. Shier "
Modified polar for ranging from Doppler and bearing
measurements" IEEE International Conference on
Acoustics Speech and Signal Processing, Boston (USA),
1983 , pp 907-910.
[4] P.A. Rosenqvist "Passive Doppler-Bearing Tracking
using Pseudo–linear estimator" IEEE J. of Ocean. Eng.,
Vol 20, N° 2 (April 1995), pp 114-118.
[5] Jauffret, C., Pillon, D. “Observability in Angle Only
Target Motion Analysis”, IEEE, Trans. on AES, Vol. 32,
No 4, pp 1290-1300, oct. 1996.
[6] Nielsen,R.O. "Sonar Signal Processing " Artech House
1993.
[7] Y. Bar Shalom, T. Fortmann, “Tracking and Data
Association”, Academic Press, Vol. 179, 1988.
[8] S.S. Blackman, “Multiple Target Tracking with Radar
Application”, Artech House, MA, 1986
[9] R. Popp, K. Pattipati, Y. Bar Shalom “m-Best S-D
Algorithm with Application to Multitarget Tracking”,
IEEE Transactions on AES, Vol. 37, No. 1 (Jan. 2001), pp
22-39.
Figure 1: Lofargram with interferences and chosen frequency lines (green markers)
514
[10] A. Poore, “Multidimensional Assignments and
Multitarget Tracking”, DIMACS series in Discrete
Mathematics and Theoretical Computer Science, Vol. 19
(1995), American Mathematical Society, pp 169-196.
[11] P. Blanc-Benon, D. Pillon, "Multi-tracks Association
for Underwater Passive Listening", Y.T. Chan Editor,
NATO-ASI series, Kluwer Academic, 1989, pp 473-477.
[12] S. Sitbon, P. Blanc-Benon, "Frequency line Data-
Fusion for a multi-octaves towed array ", in Proc. IEEE
ICASSP 1990, Albuquerque (NM), pp 2755-2758.
[13] Frank Ayres, JR. "Schaum’s Outline Series Theory
and Problems of Matrices", pp 55-57 – 1962.
[14] P.Blanc-Benon "TMA with/without Doppler
measurement: shallow water experiments" Conf. UDT 2-4
July 1996, pp 317- 321.
[15] B . Ristic, S. Arulampalam, N. Gordon "Beyond the
Kalman Filter – Particle Filter for tracking applications",
Norwood, MA : Artech house 2004.
[16] K. Becker "A General Approach to TMA
Observability from Angle and Frequency Measurements".
IEEE, Trans. on AES, Vol. 32, No 1, pp 487-494, jan.
1996.
[17] K.C. Ho, Y.T. Chan "An Asymptotically Unbiased
Estimator for Bearings-Only and Doppler-Bearing Target
Motion Analysis". IEEE, Trans. on SP, Vol. 54, No 3, pp
809-822, mar. 2006.
 Figure 2: Scenario used for the comparison

Bearing normalized residuals for successive TMA estimations

Frequency normalized residuals for successive TMA estimations

Figure 3: Normalized residuals vs. time in min. (vertical axis) of the estimates returned by the MFBTMA.

515
 Estimated Target Range versus Time (MFBTMA)
21
64, 66, 77, 84, 90, 97, 110 Hz
Ground Truth
18

15

R
A 12
N
G
E
9
(km)

0
20 25 30 35 40 45 50 55 60 65 70
Time (mn)

Figure 4: MFBTMA estimated range and real range between the 20th and the 56th minute.

Estimated Target Range versus Time (FBTMA)

21

18 64 Hz
66 Hz
77 Hz
15 84 Hz
R 90 Hz
A 97 Hz
N 12
110 Hz
G
E 7-averaged
9 Ground Truth
(km)

0
20 25 30 35 40 45 50 55 60 65 70
Time (mn)

Figure 5: Several Single FB TMA results and averaged result.

516