passive geolocalization of radio transmitters: algorithm and performance in narrowband context
TRANSCRIPT
Contents lists available at SciVerse ScienceDirect
Signal Processing
Signal Processing 92 (2012) 841–852
0165-16
doi:10.1
n Corr
avenue
E-m
Jonatha
ferreol@
(A. Fer
larzaba# E
journal homepage: www.elsevier.com/locate/sigpro
Passive geolocalization of radio transmitters: Algorithm andperformance in narrowband context
Jonathan Bosse a,b,#,n, Anne Ferreol a,b, Cecile Germond a, Pascal Larzabal b
a THALES Communications, 160 Boulevard de Valmy, 92704 Colombes, Franceb SATIE, ENS Cachan, CNRS, Universud 61 avenue du president Wilson, 94230 Cachan, France
a r t i c l e i n f o
Article history:
Received 27 September 2010
Received in revised form
12 July 2011
Accepted 9 September 2011Available online 24 September 2011
Keywords:
Emitter geolocalization
Array processing
AOA estimation
Cramer–Rao bounds
84/$ - see front matter & 2011 Elsevier B.V. A
016/j.sigpro.2011.09.008
esponding author at: SATIE, ENS Cachan, C
du president Wilson, 94230 Cachan, France.
ail addresses: [email protected],
[email protected] (J. Bosse),
satie.ens-cachan.fr, [email protected]
reol), [email protected] (C
[email protected] (P. Larzabal).
URASIP member.
a b s t r a c t
Passive localization commonly consists of a two steps strategy. In the first step,
intermediate parameters, often called measurements (such as angles of arrival (AOA),
times of arrival (TOA), etc.) are measured on several base stations equipped with sensor
arrays. In a second step, the transmitted intermediate parameters are then used to
estimate the position at a central processing unit. Such approach is suboptimal. To
overcome this limitation, one step algorithms were recently proposed. They exploit
simultaneously all received signals of all base stations seen as a global array in order to
provide the source positions directly. In this paper, we propose an original one step
algorithm called global MUSIC approach (GMA). The GMA offers better performance in
narrowband context. Moreover it does not require the use of filter banks in the
wideband signal context, contrary to the recently proposed direct position determina-
tion (DPD). GMA appears to outperform the DPD in wideband context. We also
investigate in this paper by means of a Cramer–Rao bound analysis the potential gain
achievable by a one step approach compared to a conventional two steps approach.
Finally, numerical results illustrate the improvement of the proposed method compared
to existing techniques in terms of location error and robustness to the time-bandwidth
product.
& 2011 Elsevier B.V. All rights reserved.
1. Introduction
In radio communication the geolocalization problem ofradio transmitters has always received great attention. Bymeans of multiple sensors base stations (seen as sub-arrays), forming what we call here a global array, theproblem lies in estimating the position of multiple radio
ll rights reserved.
NRS, Universud 61
roup.com
. Germond),
emitters in far field context. Traditionally [1,2], thelocalization algorithm consists of two steps illustrated inFig. 1.
In the first step, parameters (often called measure-ments) are independently estimated at each base station.Parameter estimation has always been an intensiveresearch theme in the array processing community. Dur-ing the last two decades it leads to a well-suited theore-tical framework (see [3,4]). Among the available studiedproblems the angle of arrival (AOA) estimation has beenwidely discussed [5]. Using the narrowband assumptionon a multi-sensor base station the AOA information iscontained in the phase difference between the sensors. Itappears to be particularly convenient for passive localiza-tion since it does not require any prior information on thetransmitted signals.
Fig. 1. Conventional two steps position determination scheme.
Fig. 2. One step position determination scheme.
J. Bosse et al. / Signal Processing 92 (2012) 841–852842
In the second step, the geolocalization is computedbased on the set of measurements available from step 1[6,7] and transmitted to a central processing unit (CPU).Previous studies aimed to quantify and reduce the biasand the variance of this second step [8,9].
However, such two steps approach suffers from limita-tions [1]. First, solving the problem by means of a multiplestep strategy is often suboptimal. As the first step does nottake into account the fact that the received signals at thestations come from the same emitters this step is sub-optimal. Secondly, in passive multi-emitter context, theproblem is ambiguous since one has to identify among allavailable measurements, which subset of measurementscharacterizes each source. This problem is sometimescalled ‘‘data association’’ [10]. Lastly, the performanceand the number of resolvable emitters at each station areknown to be locally limited by the number of sensors ofthe considered station. These are the reasons why anapproach based on a one step strategy, gathering thereceived signals of all available stations, that directlyprovides the emitter location in LOS context, as illustratedin Fig. 2, appears of great interest. Such an approachassumes that each station is able to transfer all the signalsto a central processing unit.
Fewer works provide one step procedures that directlyestimate the position of emitters using the signal col-lected by all multi-sensor stations. Only [11–14] investi-gate recently such one step procedures. Studies [14,15]deal with the global navigation satellite system (GNSS),another [13] focuses on moving targets proposing alocation algorithm based on frequency Doppler shiftsexploitation. To the best of our knowledge, only Weissand Amar [11,12] treat the passive localization problem ina one step approach called direct position determination(DPD) exploiting the information of the steering vector(angular response) of each sensor. Based on filter banks,they proposed a unique criterion gathering all signals ofall stations firstly in the presence of a single emitter [11]and then proposed a MUSIC [16] approach of the problemin order to treat the multiple emitters case [12]. Theirapproach relies on an incoherent sum of criteria obtainedfor each frequency of the filter bank. It suggests that thisstrategy is not optimal. They proved [17] the theoretical
gain of such techniques in terms of the number ofresolvable emitters. They also provide a performanceanalysis [12] under known waveforms or random Gaus-sian signal assumptions. To the best of our knowledge, theperformance study for deterministic signals was neverprovided for direct position estimation. We fill the lack inthis paper.
Solving the problem by means of a one step strategyleads to estimate the emitters location but it also impliesto estimate the complex gain of the signal received in allstations, seen as a nuisance parameter [11], in order toaccount for the possible path attenuation, which is dif-ferent for each station. It underlines the need of investi-gating the theoretical gain of such a one step strategycompared to a conventional two steps strategy.
In this paper, we consider a global array composed ofseveral multi-sensor base stations, each one able totransfer the data to a CPU, and we focus on passive (ornon-cooperative) localization problem of multiple trans-mitted signals in LOS context, with no prior informationon the signals. Based on preliminary work [18], wepropose a one step location algorithm called global MUSICapproach (GMA) that does not require the use of filterbanks in wideband signal context. We explore moresoundly the performance improvement in terms of loca-tion error of such a one step approach using deterministicCramer–Rao bounds (CRBs) comparisons. In Section 2, weformulate the problem and study the case of narrowbandsignals for the global array. Then in Section 3 the MUSICapproach called GMA is derived. Section 3 also providesthe gradient and the Hessian required for a practicalimplementation based on a Newton algorithm. The Cra-mer–Rao bound expression of the conventional two stepsand the one step approach for deterministic signals aregiven in Section 4. Extension of the GMA to the widebandsignal case is discussed in Section 5. Simulations illustratein Section 6 the gain of the proposed approach comparedto existing techniques.
2. Signal model
We focus on the problem of locating multiple LOSemitters on L separated base stations each composed of Nl
(1r lrL) sensors as illustrated on Fig. 3. Let us denoteN¼
PlNl the total number of sensors. Let xlðtÞ be the
Fig. 3. Propagation from an emitter located at pm to multiple multi-
sensor base station located in Sl.
J. Bosse et al. / Signal Processing 92 (2012) 841–852 843
observation vector of the l-th station and let us considerM emitters, whose unknown signals are sm(t) (1rmrMÞ.Assuming classically that the impinging signals are nar-rowband for the base station, we have then:
xlðtÞ ¼XM
m ¼ 1
rl,malðylðpmÞÞe�j2pf 0tlðpmÞsmðt�tlðpmÞÞþnlðtÞ,
ð1Þ
where f0 is the carrier frequency, pm denotes the D� 1position of the m-th emitter (if D¼2, pm ¼ ½xm,ym�
T and ifD¼3, pm ¼ ½xm,ym,zm�
T , where ð�ÞT denotes the transposeoperator). rl,m is an unknown complex parameter stand-ing for the channel attenuation. We denote alðylðpÞÞ thesteering vector of the array at station l depending on theAOA yl, seen here as a function of the position of thesource p. Since we assume that the wavefront is planar ateach station the only parameter that can be estimated at aparticular station is the AOA. This is why a multiple basestation system is required in order to localize an emitter.For notational convenience we will write in the sequelalðpÞ9alðylðpÞÞ usually obtained by a calibration process inpractice and whose theoretical expression on n-th sensoris
ðalðpÞÞn ¼ e�j2pf 0dl,nðpÞ, ð2Þ
where dl,nðpÞ is the time delay between the n-th sensor ofthe base station l and the center of phases of the station.We define tlðpmÞ as the relative time delay of the m-thsignal on the l-th station, the origin being arbitrarilychosen on the first station. So, we define
tlðpmÞ9Jpm�pðlÞJ�Jpm�pð1ÞJ
c, ð3Þ
where pðlÞ ð1r lrLÞ is the position vector of the l-thstation. The additive noise vector nlðtÞ is white Gaussianwith a covariance matrix equal to s2INl
, where IN is theN�N identity matrix. We also assume that all nlðtÞ
ð1r lrL,1rtrTÞ are statistically independent. All sig-nals are supposed temporally independent.
Note that there is an ambiguity in (1) between thephase of rl,m and e�j2pf 0tlðpmÞ, so that both cannot beestimated separately. Since in practice the steering vec-tors of the stations are not calibrated with a commonphase reference and since the impinging signals at eachstation are not supposed to have the same amplitude andphase, we will include the term e�j2pf 0tlðpmÞ into rl,m.
We build the following stacked observation vector ofthe global array:
xðtÞ ¼
x1ðtÞ
^
xLðtÞ
264
375: ð4Þ
We assume here that the impinging signals (complexenvelops) are narrowband for the global array. That is tosay the time-bandwidth product (denoted tm � Bm) isclose to zero. Here tm stands for the time delay ofpropagation across the sensor network (i.e. time delay ofarrival between the two most separated sensors) and theemitter bandwidth Bm. We have
Bm �maxi,j
9tiðpmÞ�tjðpmÞ951: ð5Þ
We can write the signals sm(t) as smðtÞ ¼ amðtÞej2pf mt ,where amðtÞ is the complex envelop, f0 the carrier fre-quency and fm an eventual carrier frequency residual.Assuming (5) amðt�tlðpmÞÞ � amðtÞ so that we can write
smðt�tlðpmÞÞ � smðtÞe�j2pf mtlðpmÞ: ð6Þ
The term e�j2pf mtlðpmÞ can be also included into rl,m. Let usremark that since the base stations are not assumed to lieclose to each other, this assumption is more restrictivethan the classical narrowband assumption performed ateach station that only requires
Bm �maxi,j
9dl,iðpmÞ�dl,jðpmÞ951: ð7Þ
The global observation (4) writes
xðtÞ ¼XM
m ¼ 1
uðpm,wm,/mÞsmðtÞþnðtÞ, ð8Þ
where we introduce the following stacked steeringvector:
uðpm,wm,/mÞ ¼1ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP
l
Nl
Nc2
l,m
rc1,mejf1,m a1ðpmÞ
c2,mejf2,m a2ðpmÞ
^
cL,mejfL,m aLðpmÞ
266664
377775, ð9Þ
where
wm ¼ ½c1,m � � � cL,m�T ð10Þ
and
/m ¼ ½f1,m � � � fL,m�T ð11Þ
are unknown real deterministic parameter vectors suchthat
rl,m ¼cl,mejfl,mffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP
l
Nl
Nc2
l,m
r : ð12Þ
The steering vector uðpm,wm,/mÞ has now a constantnorm equal to the number of sensors N as encounteredin classical AOA estimation problem. We choose withoutloss of generality c1,m ¼ 1 and f1,m ¼ 0 (relative nuisanceamplitudes and phases origin). We also define
nðtÞ ¼ ½nT1ðtÞ � � � nT
L ðtÞ�T : ð13Þ
J. Bosse et al. / Signal Processing 92 (2012) 841–852844
The covariance matrix of the additive noise nðtÞ is s2IN ,assuming that the noise vectors are spatially white. Wedefine the following signal to noise ratio for the m-thsource:
SNRm ¼ 10 log10
E½9smðtÞ92�
s2
!: ð14Þ
We can write
xðtÞ ¼XM
m ¼ 1
uðpm,wm,/mÞsmðtÞþnðtÞ, ð15Þ
or, using a more compact expression:
xðtÞ ¼ AðP,W,UÞsðtÞþnðtÞ, ð16Þ
where
P¼ ½p1 � � � pM �, ð17Þ
W¼ ½w1 � � � wM�, ð18Þ
U¼ ½/1 � � � /M�, ð19Þ
AðP,W,UÞ ¼ ½uðp1,w1,/1Þ � � � uðpM ,wM ,/MÞ�, ð20Þ
sðtÞ ¼ ½s1ðtÞ � � � sMðtÞ�T : ð21Þ
3. A global MUSIC approach (GMA)
3.1. Algorithm
In this section we propose a one step approach calledglobal MUSIC approach (GMA), in contrast to a conven-tional two steps approach where all measurements areindependently estimated on each station. As it can beseen from Eq. (15) the number of parameters to beestimated quickly grows with the number of stations. Inorder to avoid optimizations in a possible huge dimen-sion, a MUSIC [16] approach is preferred compared to amaximum likelihood. The MUSIC algorithm exploits thestructure of the following covariance matrix:
R¼ E½R�, ð22Þ
where E½�� stands for the mathematical expectation and
R ¼1
T
XT
t ¼ 1
xðtÞxHðtÞ: ð23Þ
The signal model (15) is well suited to the use of a MUSICalgorithm applied on the estimated covariance matrix R ,knowing the steering vector uðp,w,/Þ.
According to (16) under the deterministic signalassumption:
R¼AðP,W,UÞRsAHðP,W,UÞþs2I, ð24Þ
Rs ¼1
T
XT
t ¼ 1
sðtÞsHðtÞ: ð25Þ
Following the MUSIC approach, assuming that the size ofthe signal subspace is K¼M in this context, the para-meters fP,W,Ug are estimated by finding the M zeros of
the following objective function:
Cðp,w,/Þ ¼uHðp,w,/ÞPnuðp,w,/Þ
uHðp,w,/Þuðp,w,/Þ, ð26Þ
where Pn ¼ I�UsUH
s is the estimated noise projector, andthe N�K matrix Us is the signal subspace, consisting ofthe K eigenvectors of the matrix corresponding to the K
largest eigenvalues of R . The optimization of (26) wouldlead to a 2ðL�1ÞþD dimensional search, whereas w and /
are nuisance parameters. Note that
uðpm,wm,/mÞ ¼UðpmÞqðwm,/mÞ, ð27Þ
where
UðpmÞ ¼ diagða1ðpmÞ, . . . ,aLðpmÞÞ, ð28Þ
qðwm,/mÞ ¼1ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP
l
Nl
Nc2
l,m
r1
c2,mejf2,m
^
cL,mejfL,m
266664
377775, ð29Þ
and diagðvÞ is a diagonal matrix built with the entries inthe vector v. Since finding the minimum of the quadraticform ratio (26) is equivalent to searching the minimaleigenvalue of the matrix ½UH
ðpÞUðpÞ��1UHðpÞPnUðpÞ [19],
criterion (26) can be simplified by computing the follow-ing criterion:
C2ðpÞ ¼ lminf½UHðpÞUðpÞ��1UH
ðpÞPnUðpÞg, ð30Þ
where lminfAg is the minimal eigenvalue of the matrix A.For the sake of computational cost and to avoid the fullnumerical optimization of (30) the following equivalentcriterion is preferred [20]:
CrðpÞ ¼9UHðpÞPnUðpÞ9
9UHðpÞUðpÞ9
, ð31Þ
where 9A9 stands for the determinant of the matrix A.Finally, the estimated positions are deduced from the M
zeros of the cost function (31).
3.2. Practical implementation
A first possible implementation could be a grid searchperformed with the criterion CrðpÞ. The positions pm
correspond to the M lowest minima. But we can alsooptimize (31) by using a Newton algorithm, since theobjective function (31) is particularly well suited to obtainclosed-form expressions for the gradient and the Hessiancontrary to (30).
The Newton algorithm can be initialized with p0m
ð1rmrMÞ, obtained by a conventional triangulation,that can be performed, for example, through two AOAestimation algorithms on two stations. In order to miti-gate the ambiguities, among the M2 possible positions(deduced from the two sets of the M estimated AOAs ateach station), the M most appropriate positions are thoseproviding the M lowest values when inserted in thecriterion (31). A grid search can also be employed.
At each step i we obtain
piþ1m ¼ pi
m�H�1ðpi
mÞ=ðpimÞ, ð32Þ
J. Bosse et al. / Signal Processing 92 (2012) 841–852 845
where Hð�Þ and =ð�Þ are the Hessian and the gradientrespectively of the criterion (31). Since UH
ðpÞUðpÞ doesnot depend on p, well known mathematical formulasprovide [3]:
½=ðpÞ�k ¼9M9
9UHðpÞUðpÞ9
Tr½M�1Mk�, ð33Þ
½HðpÞ�kl ¼9M9
9UHðpÞUðpÞ9
ðTr½M�1Ml�Tr½M�1Mk�
�Tr½M�1MkM�1Ml�þTr½M�1Mkl�Þ, ð34Þ
where
MðpÞ ¼UHðpÞPnUðpÞ, ð35Þ
and with Mk ¼ @M=@pk and Mkl ¼ @2M=@pk@pl.
4. Deterministic Cramer–Rao bounds
By means of the Cramer Rao bound we provide in thissection new insights in the theoretical gain provided bythe global one step approach in narrowband context forthe global array compared to a conventional two stepsapproach. In this approach all signals of all stations haveto be transmitted to a CPU. Due to this stronger practicalrequirement, it appears of fundamental interest to quan-tify when and in what manner the global approach out-performs the conventional one. We here provide theexpression of the deterministic CRB of the model (16)called global CRB (GCRB), assuming the noise to beGaussian circular distributed. Then we also provide thedeterministic CRB for a conventional approach calledCCRB based on observation vectors received on eachstation separately seen as sub-blocks of the vector xðtÞ.
4.1. CRB of the global observation (GCRB) in the
narrowband context
Let us denote n the following unknown parametervector in the global approach, where the term s2 isdiscarded from the unknown parameter g:
n¼ ½gT sTr ð1Þ sT
c ð1Þ . . . sTr ðTÞ sT
c ðTÞ�, ð36Þ
where
g¼ ½P1 . . . PD W2 . . . WL U2 . . . UL�T ð37Þ
and srðtÞ and scðtÞ are respectively the real and imaginarypart of sðtÞ. Pd denotes the d-th row of the D�M matrix P,Wl the l-th row of the ðL�1Þ �M matrix W and Ul the l-throw of the ðL�1Þ �M matrix U. The log-likelihood of theproblem writes, up to an additive irrelevant constant
Lðx9nÞ ¼ �1
s2
XT
t ¼ 1
½xðtÞ�AðgÞsðtÞ�H½xðtÞ�AðgÞsðtÞ�, ð38Þ
where x¼ ½xT ð1Þ . . . xT ðTÞ�T . The Fisher informationmatrix (FIM) is [3]
JðnÞ ¼ E@Lðx9nÞ@n
� �@Lðx9nÞ@n
� �T( )
: ð39Þ
We consider the noise as Gaussian, circular. Using [21] asin [22] we finally get the block of the CRB from the invert
of J, only function of the parameter vector g:
GCRB�1ðgÞ ¼
2
s2
XT
t ¼ 1
RefSHðtÞDHPADSðtÞg, ð40Þ
PA ¼ I�AðAHAÞ�1AH , ð41Þ
SðtÞ ¼ IðDþ2ðL�1ÞÞ � diagðsðtÞÞ, ð42Þ
D¼ ½DP DW DU�, ð43Þ
DP ¼ ½DP1. . . DPD
�, ð44Þ
DPd¼
@uðp1,w1,/1Þ
@pd,1. . .
@uðpM ,wM ,/MÞ
@pd,M
� �, ð45Þ
DW ¼ ½Dw2. . . DwL
�, ð46Þ
Dwl¼
@uðp1,w1,/1Þ
@cl,1
. . .@uðpM ,wM ,/MÞ
@cl,M
" #, ð47Þ
DU ¼ ½D/2. . . D/L
�, ð48Þ
D/l¼
@uðp1,w1,/1Þ
@fl,1
. . .@uðpM ,wM ,/MÞ
@fl,M
" #, ð49Þ
where � denotes the Kronecker product. The result (40)can be straightforwardly expressed in the more compactform
GCRB�1ðgÞ ¼
2T
s2RefDHPAD� ðR 0sÞ
Tg, ð50Þ
where
R 0s ¼ 1K�K �1
T
XT
t ¼ 1
sðtÞsHðtÞ, ð51Þ
with � denoting the Hadamard product, andK ¼Dþ2ðL�1Þ and 1K�K being the matrix of dimensionK�K with all elements equal to 1. The expressions for thederivatives of the steering vectors can be found inAppendix A.
Finding an analytical expression of the GCRB depend-ing only on p is complicated in the general case, never-theless we provide in the sequel the calculation in theparticular case of one source and two stations.
4.2. CRB of the conventional observation (CCRB)
Conventional methods does not exploit the extendedstaked vector, they do not exploit any relation betweenthe signals received on each station. Consequently, theyrely on the following model, seen as a set of reparame-trized sub-blocks of the global observation xðtÞ consideredindependently of each other:
x1ðtÞ ¼XM
m ¼ 1
a1ðy1ðpmÞÞs1,mðtÞþn1ðtÞ,
^
xLðtÞ ¼XM
m ¼ 1
aLðyLðpmÞÞsL,mðtÞþnLðtÞ,
8>>>>>>><>>>>>>>:
ð52Þ
J. Bosse et al. / Signal Processing 92 (2012) 841–852846
where
sl,mðtÞ ¼
ffiffiffiffiNp
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPlc
2l,mNl
q clejfm smðtÞ: ð53Þ
The noise vectors nlðtÞ ð1r lrLÞ are assumed Gaussiancircular. The unknown parameters are
ptot ¼ ½P1 . . . PD�T , ð54Þ
s¼ ½sT1 . . . sT
L �T , ð55Þ
sl ¼ ½sTl ð1Þ . . . sT
l ðTÞ�T , ð56Þ
slðtÞ ¼ ½sTl,1ðtÞ . . . sT
l,MðtÞ�T , ð57Þ
so that the unknown parameter vector is
g¼ ½pTtot sT
r sTc �
T : ð58Þ
Where, again, sr and sc denotes the real and imaginarypart of s, respectively. Let us denote JðgÞ the correspond-ing Fisher matrix. We now consider the following variablechange:
g : g¼
ptot
sr
sc
0B@
1CA/n¼
htot
sr
sc
0B@
1CA, ð59Þ
where
htot ¼ ½hT1 . . . hT
L �T , ð60Þ
hl ¼ ½yl,1 . . . yl,M�T , ð61Þ
with yl,m ¼ ylðpmÞ. Consequently, the reparametrizedmodel is
x1ðtÞ ¼XM
m ¼ 1
a1ðy1,mÞs1,mðtÞþn1ðtÞ,
^
xLðtÞ ¼XM
m ¼ 1
aLðyL,mÞsL,mðtÞþnLðtÞ:
8>>>>>>><>>>>>>>:
ð62Þ
We denote JðnÞ the corresponding FIM. We have [23]
JðgÞ ¼@nT
@gJðnÞ
@n
@gT: ð63Þ
Let us denote
JðnÞ ¼Jyy Jys
Jsy Jss
" #: ð64Þ
Since
@nT
@g¼
@pTtot
@htot0
0 I2LMT
" #, ð65Þ
using the matrix inversion lemma [3], we have
CCRB�1ðptotÞ ¼
@hTtot
@ptot
Jyy@htot
@pTtot
�@hT
tot
@ptot
JysJ�1ss Jsy
@htot
@pTtot
, ð66Þ
CCRB�1ðptotÞ ¼
@hTtot
@ptot
ðJyy�JysJ�1ss JsyÞ
@htot
@pTtot
: ð67Þ
And since Jyy�JysJ�1ss Jsy ¼ CRB�1
ðhtotÞ, thanks to the matrixinversion lemma again, we have
CCRB�1ðptotÞ ¼
@hTtot
@ptot
CRB�1ðhtotÞ
@htot
@pTtot
, ð68Þ
with
CRB�1ðhtotÞ ¼
CRB�1ðh1Þ 0 0
0 & 0
0 0 CRB�1ðhLÞ
264
375: ð69Þ
The term CRB�1ðhlÞ is the deterministic CRB on AOA [24]
on the l-th station:
CRB�1ðhlÞ ¼
2T
s2RefDHPlD� R
T
s,lg, ð70Þ
Rs,l ¼1
T
XT
t ¼ 1
sHl ðtÞslðtÞ ð71Þ
D¼@alðyl,1Þ
@yl,1. . .
@alðyl,MÞ
@yl,M
� �, ð72Þ
Pl ¼ I�AlðAHl AlÞ
�1AHl , ð73Þ
Al ¼ ½alðyl,1Þ . . . alðyl,MÞ�: ð74Þ
Since the global approach exploits more information fromthe observations by taking into account correlationbetween the impinging signals on each station, it isexpected that the GCRB is smaller than the CCRB. Never-theless, each bound reflects different technological meansso that their comparison allows us to quantify thepotential gain of a global one step approach comparedto a conventional approach.
4.3. Analytic calculation for one source and two stations
We now examine the special case of two stations(L¼2) and one source (M¼1) for the planar geolocaliza-tion problem (D¼2). We have
xðtÞ ¼ uðp,c,fÞsðtÞþnðtÞ: ð75Þ
In this section, since there is only one source, the index m
is omitted and the parameters c2;1 and f2;1 are denotedas c and f instead since there is no possible ambiguity.Here we have
uðp,c,fÞ ¼ffiffiffiffiNp
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiN1þN2c
2q a1ðpÞ
cejfa2ðpÞ
" #: ð76Þ
Calling y1 and y2 the angles of arrival on the first and thesecond station respectively, we can write the observationson both stations used by a conventional approach as
x1ðtÞ ¼ a1ðy1Þs1ðtÞþn1ðtÞ, ð77Þ
x2ðtÞ ¼ a2ðy2Þs2ðtÞþn2ðtÞ, ð78Þ
where
s1ðtÞ ¼
ffiffiffiffiNp
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiN1þN2c
2q sðtÞ, ð79Þ
s2ðtÞ ¼
ffiffiffiffiNp
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiN1þN2c
2q cejfsðtÞ: ð80Þ
J. Bosse et al. / Signal Processing 92 (2012) 841–852 847
4.3.1. CCRB for two stations and one source
The CCRB reads
CCRB�1ðpÞ ¼G½CRB�1
ðy1,y2Þ�GT , ð81Þ
where
CRB�1ðy1,y2Þ ¼
CRB�1ðy1Þ 0
0 CRB�1ðy2Þ
" #, ð82Þ
G¼
@y1@x
@y2@x
@y1@y
@y2@y
24
35: ð83Þ
We have
CRB�1ðyiÞ ¼
2T
s2Re
@ai
@yi
H
PðaiÞ@ai
@yiRsi
( ), ð84Þ
with
Rsi¼
1
T
XT
t ¼ 1
9siðtÞ92, ð85Þ
and for any vector b
PðbÞ ¼ I�bðbHbÞ�1bH: ð86Þ
We finally get
ðCCRB�1Þij ¼
2T
s2Re
@a1
@pi
H
Pða1Þ@a1
@pj
Rs1
(þ@a2
@pi
H
Pða2Þ@a2
@pj
Rs2
):
ð87Þ
4.3.2. GCRB for two stations and one source
We prove in Appendix B that the CRB of the globalapproach is written as
ðGCRB�1Þij ¼
2TRs
s2Re
@u1
@pi
H
Pðu1Þ@u1
@pj
(þ@u2
@pi
H
Pðu2Þ@u2
@pj
),
ð88Þ
where
u1 ¼
ffiffiffiffiNp
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiN1þN2c
2q a1,
u2 ¼
ffiffiffiffiNp
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiN1þN2c
2q cejfa2:
We can easily deduce that PðuiÞ ¼PðaiÞ, @u1=@pi ¼ ðffiffiffiffiNp
=ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiN1þN2c
2q
Þ@a1=@pi and @u2=@pi ¼ ðffiffiffiffiNp
=ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiN1þN2c
2q
Þ
cejf@a2=@pi. Since by definition Rs1¼ ðN=ðN1þN2c
2ÞÞRs
and Rs2¼ ðN=ðN1þN2c
2ÞÞc2Rs we finally get
ðGCRB�1Þij ¼
2T
s2Re
@a1
@pi
H
Pða1Þ@a1
@pj
Rs1
(þ@a2
@pi
H
Pða2Þ@a2
@pj
Rs2
),
ð89Þ
which proves
GCRB�1ðpÞ ¼ CCRB�1
ðpÞ: ð90Þ
This means that in the narrowband signals case, in thespecial case of one source with two base stations noasymptotic gain of performance can be achieved with theproposed one step approach compared to a conventional
approach. Identity (90) is valid for one source and is dueto the particular form of the noise projector and to thefact that the FIM of the narrowband one step approachcontains some terms equal to zero. This study appears tobe more complicated in multiple sources context. This isthe reason why in the sequel we will calculate by meansof simulations the GCRB under some multi-emitter con-text in order to examine the gain of the one step narrow-band approach.
5. Wideband signal case
When the narrowband assumption is not verified onthe global array we can still write, according to (1), (4)and (9):
xðtÞ ¼XM
m ¼ 1
XL
l ¼ 1
0N1þ���þNl�1
alðpmÞ
0Nlþ 1þ���þNL
264
375smðt�tlÞþnðtÞ, ð91Þ
xðtÞ ¼XM
m ¼ 1
XL
l ¼ 1
uðpm,cl,mel,fl,melÞslmðtÞþnðtÞ, ð92Þ
where slmðtÞ ¼ smðt�tlÞ and el is the l-th column of theL� L identity matrix. We have rl,m ¼cl,mejfl,m . We define0K as a K � 1 vector composed of zeros.
In wideband signal context the model (92) is equiva-lent to (16) except that the size of the signal subspace K ofthe covariance matrix R is now equal to ML. The noiseprojector is Pn ¼ I�UsU
H
s where the columns of Us nowconsist of the K¼ML eigenvectors corresponding to the K
largest eigenvalues of R . Then, we can still optimize thecriterion (31) in order to estimate the position of theemitters. Eq. (91) shows that a wideband emitter can beseen as multiple components at same location and differ-ent nuisance parameters. The difference between thenarrowband and the wideband signal case lies in the sizeof the signal subspace embedded in R . In practice we donot know if the signals are narrowband or wideband onthe global array and both case can occur simultaneouslyin multiple emitter context. This is why, even if thenumber of sources is known, since the signal subspacecan be larger than the number of sources, we propose toestimate the signal subspace size. We use here a simpleempirical thresholding on the eigenvalues of R . Of coursethis stage can be refined with a more elaborate test but itis out of the scope of this paper. Note that [12] proposesto treat the case of wideband signals on the sensornetwork by increasing the size of the filter banks. As aconsequence, the number of independent snapshotsdecreases at the output of each frequency of the filter.Moreover, since the DPD criterion relies on an incoherentsum at the output of the filter banks, it appears to besuboptimal. In the proposed approach the number ofindependant snapshots is preserved.
6. Simulations
In this section we consider four base stations A, B, C
and D whose coordinates are ðxA,yAÞ, ðxB,yBÞ, ðxC ,yCÞ andðxD,yDÞ, respectively. Each base station is equipped with a
−10 −5 0 5 10101
102
103
SNR [dB]R
MSE
[m]
Ψ = [1 1]
CMA (2 steps)GMA (1 step)CCRB=GCRB
−10 −5 0 5 10SNR [dB]
RM
SE [m
]Ψ = [1 2]
CMA (2 steps)GMA (1 step)CCRB = GCRB
101
102
103
104
Ψ = [1 5]
J. Bosse et al. / Signal Processing 92 (2012) 841–852848
three sensor uniform circular array whose radius is 0.5wavelength (NA ¼NB ¼NC ¼ND ¼ 3). The signals aredeterministic (generated by a Gaussian random process).The carrier frequency is f 0 ¼ 100 MHz. The scenariodescribing the position of the emitters and the basestations is illustrated in Fig. 4. The performance is studiedby means of the root mean square error (RMSE) indistance:
RMSE¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1
K
XK
k ¼ 1
ðx�xkÞ2þðy�ykÞ
2
vuut , ð93Þ
where K is the number of Monte-Carlo runs and ðxk,ykÞ
denotes the k-th estimation of the true position (x, y) ofthe emitter. The empirical RMSE is compared to thecorresponding CRB in distance CRB¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiCRBðxÞþCRBðyÞ
p.
We here compare the proposed global MUSICapproach (GMA) to a conventional two steps MUSICapproach (CMA). The GMA is computed thanks to theNewton method initialized by a grid search. The CMArelies on MUSIC algorithms implemented at each station,estimating the AOAs. When there are two stations thelocation is achieved by a simple geometrical relation.When there are more than two stations the location isthen obtained by a classical least-square approach [7]. Wecompare these algorithms to their corresponding Cramer–Rao bound called GCRB and CCRB, respectively.
6.1. Single narrowband emitter case
In Fig. 5 we compare the RMSE for the GMA with theCMA in the case of one source in S1, with only two basestations A and B. As we saw in Section 4 the CRB are equal.Since the stations are not supposed to receive the samepower of signal we consider three cases in Fig. 5: W¼ ½1 1�that is to say the stations observed signals of the samepower, W¼ ½1 2� corresponding to a difference of 6 dBbetween the two impinging signals and W¼ ½1 5� corre-sponding to a difference of 14 dB on both impingingsignals. In all cases we can see that the GMA outperformsthe conventional approach at low SNR. Moreover whenthe signals impinging on station have different
Fig. 4. S1ð300;300Þ and S2ð�300,�300Þ and four base stations A, B, C and
D in ð�1500;0Þ, ð1500;0Þ, ð0;1500Þ and ð0;1500Þ, respectively.
amplitudes the gain becomes more obvious. If there ismore than two stations, as illustrated in Fig. 6 the CRBsappear to be still equal and the proposed GMA still
−10 −5 0 5 10101
102
103
104
SNR [dB]
RM
SE [m
]
CMA (2 steps)GMA (1 step)CCRB=GCRB
Fig. 5. RMSE of one source in S1 in the presence of two stations A and B,
with different values of the attenuation between the two stations W,
T¼300, number of Monte-Carlo runs¼100.
−15 −10 −5 0 5100
101
102
103
SNR [dB]
RM
SE [m
]
CMA (2 steps)GMA (1 step)CCRBGCRB
Fig. 6. RMSE of one source in S1 in the presence of a second source in S2
with the four stations A, B, C and D, T¼300, number of Monte-Carlo
runs¼100.
−5 0 5 10 15101
102
103
SNR [dB]
RM
SE [m
]
CMA (2 steps)GMA (1 step)CCRBGCRB
Fig. 7. RMSE of one source in S1 in the presence of a second source in S2
with two stations A and B, T¼300, number of Monte-Carlo runs¼100.
−15 −10 −5 0 5 10 15100
101
102
103
SNR [dB]
RM
SE [m
]CMA (2 steps)GMA (1 step)CCRBGCRB
Fig. 8. RMSE of one source in S1 in the presence of a second source in
S2 with four base stations (A,B,C,D), T¼300, number of Monte-Carlo
runs¼100.
J. Bosse et al. / Signal Processing 92 (2012) 841–852 849
outperforms the conventional two steps approach CCMAat low SNR.
6.2. Multiple narrowband emitter case
Now we consider the presence of two sources S1 and S2
(as the scenario is symmetric we will only show the RMSEof S1) in Fig. 7. We first observe that the GCRB lies underthe CCRB demonstrating the potential gain of an algorithmthat exploits all information of the stations together. Wesee that both algorithms converge toward their corre-sponding CRB, suggesting the relevancy of the consideredalgorithms with regard to the corresponding models. Aswe can see the GMA still outperforms the CMA (a 5 dBgain for a 100 m error). The GCRB still appears to be arelevant approximation of the asymptotic variance ofthe GMA.
In Fig. 8 we compare the RMSE of the source S1 in thepresence of the source S2 using the four base stations. As wecan see the GMA still outperforms the CMA, for example fora 20 m error the gain is approximatively 10 dB.
6.3. Performance comparison between one step and two
steps approach for the narrowband signal case
In [25] thanks to the EXIP theorem [26], it is demon-strated that the performance of a two steps approach arealways inferior or equal to the performance of a one stepapproach. At best a two steps approach can be asympto-tically equivalent to the one step approach. The drasticperformance improvement of the one step approachcompared to the conventional two steps lies in thenumber of sources. This is confirmed by the simulations.Indeed: if there is one source both approaches areasymptotically equivalent whereas in multiple sourcescase, the one step approach appears to be always better.This is consistent with the relationship between the twocorresponding CRBs. For one source we first demonstratethat with two stations both CRB are equal. For an arbitrarynumber of stations (four in the simulations) this equality
appears to be still verified. It underlines the fact thatimprovements for the one step approach can still appearat low SNR in single emitter context. For an arbitrarynumber of sources, the provided simulations underlinethe fact that the CRB of the one step approach is smallerthan the conventional two steps CRB. This is why the onestep approach appears to be always better in such amultiple source context.
6.4. Wideband signal case
It is of interest to compare the GMA performance whenthe time-bandwidth product (t� B) is growing, that is tosay when the narrowband assumption cannot be verifiedon the global array. In Fig. 9 we vary the bandwidth of thesignal without changing the emitter location, so that thetime-bandwidth changes too. The GMA is equivalent to aparticular case of the DPD technique when the numberof FFT-bins J is equal to one in narrowband signal case.
10
20
30
40
50
60
τ × B
τ × B
RM
SE [m
]SNR = 5dB
GMADPD, J=1DPD, J=2DPD, J=4
10−4 10−2 100 102
10−4 10−2 100 102
0
1
2
3
4
5
Estim
ated
sign
al su
bspa
ce si
ze
Fig. 9. (a) RMSE of one source in ð300;500Þ with four base stations
(A,B,C,D), SNR¼5 dB, T¼300, number of Monte-Carlo runs¼200.
(b) Mean estimated signal subspace size used for the GMA algorithm.
J. Bosse et al. / Signal Processing 92 (2012) 841–852850
When the time-bandwidth product is growing the receivedsignals at each station become uncorrelated. As a conse-quence, the size of the signal subspace grows too as shownin Fig. 9(b) and should be estimated. Fig. 9(a) showsthat the proposed methods outperforms the DPDwhen the time-bandwidth product becomes large. Wesee in Fig. 9(b) that the performance breakdown and thedistinction between the GMA and the DPD with J¼1 occurswhen the estimated signal subspace goes from one to four.When t� B is high emitters are uncorrelated on eachstations so that the size of the signal subspace becomesequal to ML. Choosing a larger filter banks for the DPDreduces the number of independent snapshots at eachfrequency and remains outperformed by the proposedapproach.
7. Conclusion
Whereas conventional location algorithms rely on a twosteps strategy, in this paper we investigate the potential gainoffered by a one step approach. By means of the Cramer–Rao
bounds we underline the fact that in a narrowband contexta one step approach will outperform the conventionalapproach in multiple emitter context. Moreover, we providean algorithm (called GMA) that is shown to perform close tothe corresponding Cramer–Rao bound. In a narrowbandcontext, it outperforms the conventional two steps approach.In a wideband context for the global array, the proposedmethod outperforms the recently proposed DPD technique.Future work might focus on the wideband signal case inorder to propose a performance study.
Acknowledgments
This work was partially supported by the network ofexcellence in wireless communication NEWCOMþþ underthe Contract no. 216715 funded by the European commis-sion FP7-ICT-2007-1.
Appendix A. Steering vectors differential
The expression of the differential of the steeringvectors are
@uðpm,wm,/mÞ
@cl,m
¼�Nlcl,mP
l0Nl0c2l0 ,m
Vcðcl,mÞuðpm,wm,/mÞ, ðA:1Þ
where
Vcðcl,mÞ ¼ diagðvCðcl,mÞÞ, ðA:2Þ
½vCðcl,mÞ�i ¼
1, ial,
Pl0al
Nl0c2l0 ,m
Nlc2l,m
, i¼ l,
8>><>>: ðA:3Þ
@uðpm,wm,/mÞ
@cl,m
¼ Vfðfl,mÞuðpm,wm,/mÞ, ðA:4Þ
where
Vfðfl,mÞ ¼ diagðvfðfl,mÞÞ, ðA:5Þ
½vfðfl,mÞ�i ¼0, ial,
j, i¼ l:
(ðA:6Þ
Appendix B. Analytic calculation of the GCRB for onesource and two base stations
Let us first denote
u¼
ffiffiffiffiNp
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiN1þN2c
2q a1ðpÞ
cejfa2ðpÞ
" #¼
u1
u2
" #ðB:1Þ
and
PðuÞ ¼P1 �C
�CH P2
" #, ðB:2Þ
where Pi ¼ INi�uiu
Hi =N and C¼ u1uH
2 =N. To go furtherwe need the simple following results obtained by
J. Bosse et al. / Signal Processing 92 (2012) 841–852 851
straightforward calculations:
P1u1 ¼N2c
2
N1þN2c2
u1, ðB:3Þ
P2u2 ¼N1
N1þN2c2
u2, ðB:4Þ
CHu1 ¼N1
N1þN2c2
u2, ðB:5Þ
Cu2 ¼N2c
2
N1þN2c2
u1, ðB:6Þ
uH1 P1u1 ¼ uH
2 P2u2 ¼NN1N2c
2
ðN1þN2c2Þ2
, ðB:7Þ
the unknown parameter vector is
n¼ ½RefsgT ImfsgT pT c f�: ðB:8Þ
Denoting
g¼ ½pT c f�, ðB:9Þ
according to (50) the FIM is
ðJðgÞÞij ¼2T
s2Re
@u
@Zi
H
PðuÞ@u
@Zj
Rs
( ): ðB:10Þ
Let us consider
JðgÞ ¼2TRs
s2
Jxx Jxy Jxc Jxf
Jxy Jyy Jyc Jyf
Jxc Jyc Jcc JcfJxf Jyf Jcf Jff
266664
377775: ðB:11Þ
Let us first remark that the derivative of u with respect top1 ¼ x and p2 ¼ y can be written as follows:@u=@pi ¼ diagðvpi
Þu where
vpi¼�j2pf 0
@d1@pi
@d2@pi
24
35, ðB:12Þ
where di ¼ ½di,1ðpÞ . . . di,NiðpÞ�. We can also write @u=@c¼
Vcu and @u=@f¼Vfu where
Vc ¼1
N1þN2c2
�cN2 0
0 N1c
" #, ðB:13Þ
Vf ¼0 0
0 j
" #: ðB:14Þ
Straightforward calculation provides
PðuÞ@u
@c¼PðuÞVcu
¼
ffiffiffiffiNp
ðN1þN2c2Þ3=2
�N2cðN2c2þN1Þu1
N1ðN2cþ N1c Þu2
24
35:
We can deduce
@u
@f
H
PðuÞ@u
@c¼�j
ffiffiffiffiNp
ðN1þN2c2Þ3=2
N1 N2cþN1
c
� �Ju2J
2:
ðB:15Þ
We conclude
Re@u
@f
H
PðuÞ@u
@c
( )¼ 0: ðB:16Þ
In the same manner we observe that
Re@u
@pi
H
PðuÞ@u
@c
( )¼ 0: ðB:17Þ
So the FIM is now
JðgÞ ¼2TRs
s2
Jxx Jxy 0 Jxf
Jxy Jyy 0 Jyf
0 0 Jcc 0
Jxf Jyf 0 Jff
266664
377775: ðB:18Þ
Thanks to a well-known matrix inversion lemma we have
ðGCRBðpÞ�1Þij ¼
2TRs
s2Jpipj�
JpifJpjf
Jff
!: ðB:19Þ
Straightforward calculations lead to
Jff ¼@u
@f
H
PðuÞ@u
@f, ðB:20Þ
¼ uH2 P2u2, ðB:21Þ
¼NN1N2c
2
ðN1þN2c2Þ2
, ðB:22Þ
Jpif ¼ Re@u
@pi
H
PðuÞ@u
@f
( ), ðB:23Þ
¼ Re@u
@pi
H P1 �C
�CH P2
" #Vfu
( ), ðB:24Þ
¼1
N1þN2c2
Im N2c2@u1
@pi
H
u1�N1@u2
@pi
H
u2
( ): ðB:25Þ
So it leads to
JpifJpjf
Jff¼
N2c2
NN1Im
@u1
@pi
H
u1
( )Im
@u1
@pj
H
u1
( )
�1
NIm
@u1
@pi
H
u1
( )Im
@u2
@pj
H
u2
( )
�1
NIm
@u1
@pj
H
u1
( )Im
@u2
@pi
H
u2
( )
þN1
NN2c2
Im@u2
@pi
H
u2
( )Im
@u2
@pj
H
u2
( ): ðB:26Þ
J. Bosse et al. / Signal Processing 92 (2012) 841–852852
We have then
Jpipj¼ Re
@uH
@pi
PðuÞ@u
@pj
( )
¼ Re@uH
1
@pi
I�u1uH
1
N
� �@u1
@pj
( )�
1
NRe
@u1
@pi
H
u1uH2
@u2
@pj
( )
�1
NRe
@u2
@pi
H
u2uH1
@u1
@pj
( )þRe
@uH2
@pi
I�u2uH
2
N
� �@u2
@pj
( ):
ðB:27Þ
Using the fact that for any complex z1 and z2 we haveRefz1z2g ¼ Refz1gRefz2g�Imfz1gImfz2g and the fact that
Re@ui
@pj
H
ui
( )¼ 0, ðB:28Þ
we finally get
ðGCRBðpÞ�1Þij ¼
2TRs
s2Re
@u1
@pi
H
Pðu1Þ@u1
@pj
( )
þRe@u2
@pi
H
Pðu2Þ@u2
@pj
( )!: ðB:29Þ
References
[1] A. Amar, A. Weiss, Direct position determination: a single-stepEmitter localization approach, in: T.E. Tuncer, B. Friedlander (Eds.),Classical and Modern Direction-of-Arrival Estimation, AcademicPress, 2009.
[2] M. Porretta, P. Nepa, G. Manara, F. Giannetti, Location, location,location, IEEE Vehicular Technology Magazine 3 (2) (2008) 20–29.
[3] H.L. Van Trees, Optimum Array Processing, John Wiley and Sons,2002.
[4] H. Krim, M. Viberg, Two decades of array signal processing research,IEEE Signal Processing Magazine (1996) 67–94.
[5] T. Tuncer, B. Friedlander, et al., Classical and Modern Direction-of-Arrival Estimation, Academic Press, Elsevier, 2009.
[6] W. Foy, Position–location solutions by Taylor-series estimation,IEEE Transactions on Aerospace and Electronic Systems 12 (2)(1976) 187–194.
[7] D. Torrieri, Statistical theory of passive location systems, IEEETransactions on Aerospace and Electronic Systems 20 (2) (1984)183–198.
[8] M. Gavish, A. Weiss, Performance analysis of bearing only targetlocation algorithm, IEEE Transactions on Aerospace and ElectronicSystems 28 (3) (1992) 817–828.
[9] K. Doganc-ay, D. Gray, Bias compensation for least-squares multi-pulse TDOA localization algorithms, in: 2nd International Conference
on Intelligent Sensors, Sensors Network and Information Processing(ISSNIP), 2005, pp. 51–56.
[10] M. Wax, T. Kailath, Decentralized processing in sensor arrays, IEEETransactions on Acoustics, Speech and Signal Processing 33 (4)(1985) 1123–1129.
[11] A. Weiss, Direct position determination of narrowband radiofre-quency transmitters, IEEE Signal Processing Letters 11 (5) (2004)513–516.
[12] A. Amar, A. Weiss, Direct position determination of multiple radiosignals, EURASIP Journal on Applied Signal Processing 1 (2005)
37–49.[13] A. Amar, A. Weiss, Localization of narrowband radio emitters based
on doppler frequency shifts, IEEE Transactions on Signal Processing
56 (11) (2008) 5500–5508.[14] P. Closas, C. Fernandez-Prades, J. Fernandez-Rubio, Maximum like-
lihood estimation of position in GNSS, IEEE Signal Processing
Letters 14 (5) (2007) 359–362.[15] P. Closas, C. Fernandez-Prades, J. Fernandez-Rubio, Cramer–Rao
bound analysis of positioning approaches in GNSS receivers, IEEETransactions on Signal Processing 57 (10) (2009) 3775–3786.
[16] R.-O. Schmidt, Multiple emitter location and signal parameter
estimation, IEEE Transactions on Antennas and Propagation 34 (3)(1986) 281–290.
[17] A. Amar, A. Weiss, Fundamental limitations on the number ofresolvable emitter using a geolocation system, IEEE Transactions onSignal Processing 55 (5) (2007) 2193–2202.
[18] J. Bosse, A. Ferreol, P. Larzabal, A multiple narrowband emittersgeolocation algorithm based on AOA estimation, in: Proceedings ofEUSIPCO 2010, Aalborg, 2010, pp. 676–680.
[19] F. Gantmacher, Matrix Theory, vols. I–II, Chelsea, New York, 1959.[20] A. Ferreol, E. Boyer, P. Larzabal, Low-cost algorithm for some
bearing estimation methods in presence of separable nuisanceparameters, Electronics Letters 40 (15) (2004) 966–967.
[21] P. Stoica, A. Nehorai, MUSIC, maximum likelihood and Cramer–Rao
bound, IEEE Transactions on Acoustics, Speech and Signal Proces-sing 37 (1989) 720–741.
[22] H. Cirpan, E. Cekli, Deterministic maximum likelihood approach for
localization of near-field sources, International Journal of Electro-nics and Communications (1) (2002) 1–10.
[23] S. Kay, Fundamentals of Statistical Signal Processing: EstimationTheory, Signal Processing Series, vol. 1, Prentice Hall, 1993.
[24] P. Stoica, A. Nehorai, MUSIC, maximum likelihood and Cramer–Rao
bound: further results and comparisons, IEEE Transactions onAcoustics, Speech and Signal Processing 38 (1990) 2140–2150.
[25] J. Bosse, A. Ferreol, P. Larzabal, Performance analysis of passivelocalization strategies: direct one step approach versus two stepsapproach, in: Proceedings of IEEE Statistical Signal Processing (SSP),
Nice, 2011, pp. 701–704.[26] P. Stoica, T. Soderstrom, On reparametrization of loss functions
used in estimation, and the invariance principle, Signal Processing
17 (1989) 383–387.