Professional Documents
Culture Documents
org
ISSN 1751-9675
Abstract: A common technique for source localisation is to utilise the time-of-arrival (TOA) measurements
between the source and several spatially separated sensors. The TOA information defines a set of circular
equations from which the source position can be calculated with the knowledge of the sensor positions.
Apart from nonlinear optimisation, least squares calibration (LSC) and linear least squares (LLS) are two
computationally simple positioning alternatives which reorganise the circular equations into a unique and non-
unique set of linear equations, respectively. As the LSC and LLS algorithms employ standard least squares (LS),
an obvious improvement is to utilise weighted LS estimation. In the paper, it is proved that the best linear
unbiased estimator (BLUE) version of the LLS algorithm will give identical estimation performance as long as
the linear equations correspond to the independent set. The equivalence of the BLUE-LLS approach and the
BLUE variant of the LSC method is analysed. Simulation results are also included to show the comparative
performance of the BLUE-LSC, BLUE-LLS, LSC, LLS and constrained weighted LSC methods with Cramér– Rao
lower bound.
156 IET Signal Process., 2008, Vol. 2, No. 2, pp. 156– 162
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-spr:20070190
www.ietdl.org
estimation performance as long as the (M 2 1) linear parameters to be estimated linear in the data. It is suitable for
equations correspond to the independent set [10]. (ii) The practical implementation as only the mean and covariance of
covariance matrices of the position estimates in the BLUE- the data are required and complete knowledge of the
LLS scheme with the independent set and the BLUE probability density function is not necessary. The BLUE
version of the LSC algorithm are identical. By comparing version of the LSC estimator is derived as follows.
with Cramér – Rao lower bound (CRLB) for TOA-based
localisation [14], it is then shown that they are suboptimal Squaring both sides of (1), we have [9]
estimators, and this result is different from the iterative
BLUE estimator of [13], which gives maximum likelihood 1 2
xi x þ yi y 0:5R ¼ x þ yi2 ri2 þ mi ,
estimation performance. (iii) Among the BLUE-LLS and 2 i (2)
BLUE-LSC algorithms, the latter is preferable as it i ¼ 1, 2, . . . , M
involves lower computational complexity. Note that the
research results can also be applied to source localisation where mi ¼ n2i =2 þ di ni and R ¼ x 2 þ y 2 is the introduced
systems with received signal strength [2] measurements as variable to reorganise (1) into a set of linear equations in x,
they employ the same trilateration concept where the y and R. To facilitate the development, we express (2) in
propagation path losses from the source to the sensors are matrix form
measured to give their distances.
Au þ p ¼ b (3)
The organisation of this paper is as follows. In Section 2,
where
we first develop the weighted versions of the LSC and
LLS methods based on BLUE. The equivalences between 2 3
x1 y1 0:5
various forms of the BLUE-LLS solutions within the 6 . .. .. 7
independent set and the BLUE-LSC estimate are then A¼6 4 .
. . . 75
proved. Furthermore, their suboptimality and computational xM yM 0:5
requirement will be discussed. Simulation results are included 2 3
in Section 3 to evaluate the estimation performance of the x
BLUE-LSC and BLUE-LLS algorithms by comparing with 6 7
u¼4y5
the LSC, LLS and constrained weighted LSC [14] methods R
as well as verify our theoretical development. Finally, 2 3
conclusions are drawn in Section 4. m1
6 . 7
p ¼ 4 .. 5
2 BLUE-based positioning mM
In this Section, we first present the signal model for TOA-
based localisation. The BLUE-LSC and BLUE-LLS and
algorithms are then devised from the LSC and LLS 2 3
x21 þ y12 r12
formulations, respectively. Their relationship, estimation
16 .. 7
performance and computational complexity are also provided. b¼ 64 . 7
5
2
2 2 2
Let (x, y) and (xi , yi), i ¼ 1, 2, . . . , M, be the unknown xM þ yM rM
source position and the known coordinates of the ith
sensor, respectively. With known signal propagation speed, For sufficiently small noise conditions,
the range measurements between the source and sensors are p ’ [ 2 d1n1 . . . 2dMnM]T and Efr 2i g ’ d 2i , i ¼ 1, 2, . . . , M,
straightforwardly determined from the corresponding TOA where T denotes transpose operation and E is the expectation
measurements, which are modelled as operator. Hence we have Efbg ’ Au, which corresponds to
the linear unbiased data model. Using the information that p
ri ¼ di þ ni , i ¼ 1, 2, . . . , M (1) is approximately zero-mean and its covariance matrix, denoted
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi by Cp , is a diagonal matrix of the form
where di ¼ (x xi )2 þ (y yi )2 is the noise-free range and
2 2 2 3
ni is the noise in ri . For simplicity, we assume line-of-sight d1 s1 0 0
propagation between the source and all sensors such that 6 0
6 d22 s22 0 7 7
each ni is a zero-mean white process with known variance Cp ’ 6 . .. .. .. 7 (4)
4 .. . 5
si2 [14]. . .
0 0 dM2 sM2
2.1 BLUE-LSC algorithm The BLUE for u based on (3), denoted by u^ , is then [12]
BLUE [12] is a linear estimator which is unbiased and has
minimum variance among all other linear estimators. To
employ the BLUE technique, we need to restrict the u^ ¼ (A T C 1 1 T 1
p A) A C p b (5)
IET Signal Process., 2008, Vol. 2, No. 2, pp. 156 – 162 157
doi: 10.1049/iet-spr:20070190 & The Institution of Engineering and Technology 2008
www.ietdl.org
where 21 represents matrix inverse. Note that the LSC where Cq is the covariance matrix for q which has the form of
estimate is given by (5) with the substitution of Cp ¼ IM
where IM is the M M identity matrix, without utilising 2 3
the mean and covariance of the data. Since fdig are d12 s12 þ d22 s22 d12 s12 d12 s12
2 2
unknown, they will be substituted by frig in practice. The 6
6 d1 s 1 d1 s12 þ d32 s32
2
2 2
d 1 s1 7
7
covariance matrix for u^ , denoted by Cu , is [12] Cq ’ 6 .. .. .. .. 7
4 . . . . 5
2 2 2 2 2 2 2 2
d1 s 1 d1 s1 d1 s1 þ dM sM
C u ’ (A T C 1
p A)
1
(6) (10)
where the variances for the estimates of x and y are given by With the use of matrix inversion lemma, its inverse can be
the (1, 1) and (2, 2) entries of Cu , respectively. It is worthy to computed as
mention that the same weighting matrix of C 1 p has been
proposed in [14], which can be considered as a constrained
weighted least squares calibration (CWLSC) algorithm 1 1 1 1
with utilising the constraint of x 2 þ y 2 ¼ R. We expect C 1
q ’ diag 2 2 , 2 2 , , 2 2 PM
d2 s2 d3 s3 dM s M i¼1 1=di2 si2
that the BLUE-LSC algorithm is inferior to the CWLSC 2 3
scheme as the parameter relationship in u is not exploited. 1 1 1
6 d24 s24 d22 d32 s22 s32 d22 dM
2
s22 sM2
7
6 7
6 1 1 1 7
2.2 BLUE-LLS algorithm 6 7
6 d 2d 2s 2s 2 d34 s34 2 2 2 2 7
d3 dM s 3 s M 7
6 3 2 3 2
On the other hand, subtracting the first equation of (2) from 6
6 .. .. .. ..
7
7
the remaining equations, R can be eliminated and we get 6 . . . . 7
6 7
(M 2 1) equations 6 7
6 7
4 1 1 1 5
1 2 2 2 2 2 2 2 2
4 4
(xi x1 )x þ ( yi y1 )y ¼ (x2i þ yi2 x21 y12 ri2 þ r12 ) dM d2 sM s2 dM d3 sM s3 dM sM
2
(11)
þ mi m1 , i ¼ 2, 3, . . . , M (7)
Expressing (7) in matrix form yields Note that the LLS estimate is given by (9) with the
substitution of Cq ¼ IM21 , without utilising the mean and
Gf þ q ¼ h (8)
covariance of the data. The estimator of (9) has minimum
where variance according to the data model of (8). It is worthy to
2 3 note that although the dependent variable of R is
x2 x1 y2 y1 eliminated in the LLS approach, estimation performance
6 .. .. 7
G¼64 . . 7
5
degradation occurs in the conversion of (2) to (7) or (8).
This is analogous to TOA and TDOA-based positioning
xM x1 yM y1 where the former estimation performance bound is lower
than that of the latter if the TDOAs are obtained from
x
f¼ substraction between the TOAs [15, 16]. The covariance
y ^ , denoted by Cf , is
2 3 matrix for f
m1 m2
6 .. 7
q¼64 . 7
5 C f ’ (GT C 1 1
(12)
q G)
m1 mM
and Although there are at most M(M 2 1) LLS equations can be
2 3 generated from (2), only (M 2 1) are independent [10]. In
x22 þ y22 x21 y12 r22 þ r12 fact, (7) is an example of the independent set of equations.
16 .. 7
h¼ 4 5 Although we can use up to M(M 2 1) dependent equations
2 .
2 2 2 2
xM þ yM x1 y1 rM þ r1 2 2 in the standard LLS algorithm, this is not possible for the
BLUE realisation because the corresponding noise
covariance matrix will be singular. In the following, we will
Following the development of the BLUE-LSC algorithm, prove that as long as the (M 2 1) equations belong to the
the BLUE-LLS estimate for f based on (8), denoted by independent set, the BLUE-LLS estimator performance
f^ , is
will agree with the covariance matrices given by (6) and
^ ¼ (GT C 1 G)1 GT C 1 h (12). Their suboptimality is then illustrated by contrasting
f q q (9) with the CRLB.
158 IET Signal Process., 2008, Vol. 2, No. 2, pp. 156– 162
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-spr:20070190
www.ietdl.org
J ¼ {eM,i eM, j : i, j ¼ 1, 2, . . . , M,
" #1
i = j and em,n is the nth column of I m } (13) T 1
LT AT C 1
p AL 0:5LT AT C 1
p 1M
(A C 1
p A) ¼
and 0:51TM C 1 0:251TM C 1
p AL p 1M
C 1=2
p 1M 1TM C p1=2 Exploiting G ¼ 2 KF and Cq ¼ KCp KT and with the use of
B¼ þ C p1=2 K T (K C p K T )1 K C p1=2 (18) (12) and (20), the inverse of Cf can be expressed as
tr(C 1
p )
T 1
!
where tr is the trace operator. Since rank(B) ¼ tr(B) and the T
C 1
p 1M 1M C p
traces of the first and second terms in (18) can be computed C 1
f ¼F C 1
p F (24)
tr(C 1
p )
as 1 and (M 2 1), respectively, we get rank(B) ¼ M.
Employing the full rank property of B as well as
idempotent property of B(IM 2 B) ¼ 0MM yield From (23) and (24), we have
T 1
B ¼ IM (19) T
C 1
p 1M 1M C p
D1 C 1
f ¼F F (25)
tr(C 1
p )
With the use of (19), we pre-multiply and post-multiply both
sides of (18) by C p1=2 to obtain
which is positive semidefinite. This implies that Cf 2 D is
C 1 T 1 also positive semidefinite, and thus the variances of the
p 1M 1M C p
K T (K C p K T )1 K ¼ C 1
p (20) BLUE-LLS estimates of x and y are greater than or equal
tr(C 1
p ) to their corresponding performance limits.
IET Signal Process., 2008, Vol. 2, No. 2, pp. 156 – 162 159
doi: 10.1049/iet-spr:20070190 & The Institution of Engineering and Technology 2008
www.ietdl.org
Cp , C q 3MM 0 3MM 0
MA
C21 21
p , Cq MM 0 0:5M2 þ 1:5MM 0
M 1A
ATC21 T 21
p , G Cq 2M þ 1M 0 2M2 4M þ 2M 0
2M2 6M þ 4A
ATC21 T 21
p A, G Cq G 3M þ 4M 3M þ 4M 3M 3M 3M 3M
5M 5A 5M 5A 3M 6A 3M 6A
b, h 4MM 4MM 4MM 4MM
2MA 2MA 3M 1A 3M 1A
ATC21 T 21
p b, G Cq h 3MM 3MM 2M 2M 2M 2M
3M 3A 3M 3A 2M 4A 2M 4A
Gaussian elimination 17M 17M 6M 6M
14A 14A 4A 4A
total 16M þ 22M 10M þ 21M 2:5M2 þ 9:5M þ 3M 9M þ 1M
10M þ 6A 10M þ 6A 2M2 þ 6M 4A 10M 7A
BLUE-LLS and LLS algorithms are provided in Table 1 positioning methods against average noise power when the
which clearly shows the calculation breakdown. Note that source is at (x, y) ¼ (1000, 2000) m. That is, the source is
the Gaussian elimination is employed for performing the located inside the square bounded by the sensor coordinates.
matrix inverse operation. In particular, we have assumed The MSPE is defined as E{(x x^ )2 þ (y y^ )2 } where x^
the covariance matrix of (10) which has a closed-form and y^ denote the estimates of x and y, respectively,
P and the
inverse in the BLUE-LLS method. Excluding the average noise power is given by s 2 ¼ 1=M M i¼1 si
2
, where
computationally extensive task of solving the Lagrange si2 is chosen such that all {si2 =di2 } are kept identical. It is
multiplier corresponding to the constraint of x 2 þ y 2 ¼ R, seen that the CWLSC scheme has the best estimation
the CWLSC method needs (16M þ 24) multiplications performance as its MSPE attains the CRLB when the
and (10M þ 7) additions. Although the BLUE-LSC and average noise power is less than 70 dB m2 where m is
BLUE-LLS methods are identical, we see that the former referenced to one metre or s ¼ 103.5 m. That is, the average
is preferable because it is more computationally attractive.
Furthermore, the computational requirement of the
BLUE-LSC estimator is comparable with those of the
LSC and LLS schemes and is significantly less than that of
the CWLSC method.
3 Numerical examples
Computer simulations have been conducted to evaluate the
performance of the BLUE-LSC and BLUE-LLS
algorithms by comparing with the LLS, LSC and
CWLSC [14] algorithms as well as CRLB. The BLUE-
LLS method is implemented using the form of (9). The
range errors fnig are zero-mean white Gaussian processes.
All results were averages of 1000 independent runs.
160 IET Signal Process., 2008, Vol. 2, No. 2, pp. 156– 162
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-spr:20070190
www.ietdl.org
Figure 2 Mean square position errors at (x, y) ¼ (7000, Figure 4 Mean square position errors against number of
8000) m receivers at (x, y) ¼ (1000, 2000) m
IET Signal Process., 2008, Vol. 2, No. 2, pp. 156 – 162 161
doi: 10.1049/iet-spr:20070190 & The Institution of Engineering and Technology 2008
www.ietdl.org
one is higher for a larger number of sensors. Again, (6) agrees [6] CAFFERY J.J. JR, STUBER G.L.: ‘Subscriber location in CDMA
well with the simulation results and the suboptimality as well cellular networks’, IEEE Trans. Veh. Technol., 1998, 47,
as equivalence of the BLUE-LSC and BLUE-LLS methods pp. 406– 416
are demonstrated.
[7] SMITH J.O., ABEL J.S.: ‘Closed-form least-squares source
4 Conclusion location estimation from range difference measurements’,
IEEE Trans. Acoust. Speech Signal Process., 1987, 35,
BLUE versions of the LSC and LLS TOA-based pp. 1661 – 1669
positioning algorithms have been examined. It is proved
that various realisations of the BLUE-LLS approach are [8] FRIEDLANDER B.: ‘A passive localization algorithm and its
indifferent as long as the equations, which correspond to accuracy analysis’, IEEE J. Ocean. Eng., 1987, 12,
the independent set are employed, and their estimation pp. 234– 245
performance is identical to that of the BLUE-LSC
algorithm. In spite of the suboptimality of the BLUE [9] CHEN J.C., HUDSON R.E., YAO K.: ‘Maximum-likelihood source
approach, its estimation accuracy can be close to CRLB localization and unknown sensor location estimation for
particularly when the source is located inside the region wideband signals in the near field’, IEEE Trans. Signal
bounded by sensor coordinates. Furthermore, the Process., 2002, 50, (8), pp. 1843– 1854
computational requirement of the BLUE-LSC algorithm is
similar to that of the standard LSC and LLS methods and [10] FENWICK A.J.: ‘Algorithms for position fixing using pulse
is significantly less than that of the constrained weighted arrival times’, IEE Proc., Radar Sonar Navig., 1999, 146,
LSC estimator, which provides optimal positioning (4), pp. 208– 212
accuracy for sufficiently small noise conditions.
[11] CAFFERY J.J. JR.: ‘A new approach to the geometry of TOA
5 Acknowledgments location’. Proc. IEEE VTC 2000-Fall, 2000, vol. 4,
pp. 1943 – 1949
The authors would like to thank the anonymous reviewers for
their careful reading and constructive comments, which [12] KAY S.M.: ‘Fundamentals of statistical signal processing:
improved the clarity of this paper. The work described in estimation theory’ (Prentice-Hall, Englewood Cliffs, NJ,
this paper was supported by a grant from the Research 1993)
Grants Council of the Hong Kong Special Administrative
Region, China (Project No. CityU 119606). [13] COUTTS S.D.: ‘3-D emitter localization using
inhomogeneous bistatic scattering’. Proc. Int. Conf.
6 References Acoustics, Speech, and Signal Processing, Phoenix, USA,
March 1999, vol. 3, pp. 1509– 1512
[1] TSUI J.B.: ‘Fundamentals of global positioning system
receivers’ (John Wiley & Sons, New York, Chichester, 2000) [14] CHEUNG K.W. , SO H.C., MA W.-K. , CHAN Y.T.: ‘Least squares
algorithms for time-of-arrival based mobile location’, IEEE
[2] CAFFERY J.J.: ‘Wireless location in CDMA cellular radio Trans. Signal Process., 2004, 52, (4), pp. 1121 – 1128
systems’ (Kluwer Academic, Boston, 2000)
[15] MC GUIRE M., PLATANIOTIS K.N.: ‘A comparison of
[3] GAY S.L., BENESTY J. (EDS.): ‘Acoustic signal processing for radiolocation for mobile terminals by distance
telecommunication’ (Kluwer Academic, Boston, London, 2000) measurements’. Proc. Int. Conf. Wireless Communications,
2000, pp. 1356 – 1359
[4] PATWARI N., ASH J.N., KYPEROUNTAS S., HERO III A.O., MOSES R.L.,
CORREAL N.S.:‘Locating the nodes: cooperative localization [16] URRUELA A., SALA J., RIBA J.: ‘Average performance analysis
in wireless sensor networks’, IEEE Signal Process. Mag., of circular and hyberbolic geolocation’, IEEE Trans. Veh.
2005, 2, (4), pp. 54 – 69 Technol., 2006, 55, (1), pp. 52– 66
[5] SPIRITO M.A.: ‘On the accuracy of cellular mobile station [17] ANDERSSON C.: ‘GPRS and 3G wireless applications:
location estimation’, IEEE Trans. Veh. Technol., 2001, 50, professional developer’s guide’ (Wiley, New York, 2001),
pp. 674– 685 p. 260
162 IET Signal Process., 2008, Vol. 2, No. 2, pp. 156– 162
& The Institution of Engineering and Technology 2008 doi: 10.1049/iet-spr:20070190