IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO.

5, MAY 2015 2037

RSS-Based Localization in Wireless Sensor

Networks Using Convex Relaxation: Noncooperative
and Cooperative Schemes
Slavisa Tomic, Marko Beko, and Rui Dinis, Member, IEEE

AbstractIn this paper, we propose new approaches based second-order cone programming (SOCP) problem, semidefinite
on convex optimization to address the received signal strength programming (SDP) problem, wireless localization, wireless sen-
(RSS)-based noncooperative and cooperative localization prob- sor network (WSN).
lems in wireless sensor networks (WSNs). By using an array of
passive anchor nodes, we collect the noisy RSS measurements from I. I NTRODUCTION
radiating source nodes in WSNs, which we use to estimate the
source positions. We derive the maximum likelihood (ML) esti-
mator, since the ML-based solutions have particular importance
due to their asymptotically optimal performance. However, the
W IRELESS SENSOR networks (WSNs) comprise a num-
ber of sensor nodes, which can, in general, be clas-
sified as anchor and source (target) nodes [1]. The locations
ML estimator requires the minimization of a nonconvex objective
function that may have multiple local optima, thus making the of the anchor nodes are known, whereas the locations of the
search for the globally optimal solution hard. To overcome this source nodes are yet to be determined. WSNs have application
difficulty, we derive a new nonconvex estimator, which tightly in various areas such as target tracking, intrusion detection,
approximates the ML estimator for small noise. Then, the new energy-efficient routing, monitoring, underground, deep water,
estimator is relaxed by applying efficient convex relaxations that
are based on second-order cone programming and semidefinite outer space explorations, etc. [2]. In WSNs, sensor nodes are
programming in the case of noncooperative and cooperative local- deployed over a monitored region to acquire some physical
ization, respectively, for both cases of known and unknown source data about the environment, such as temperature, pressure,
transmit power. We also show that our approaches work well in humidity, wind speed, etc. The collected information, together
the case when the source transmit power and the path loss expo- with the objects position information, enables us to develop
nent are simultaneously unknown at the anchor nodes. Moreover,
we show that the generalization of the new approaches for the intelligent systems. Such systems offer improved safety and
localization problem in indoor environments is straightforward. efficiency in everyday life, since each individual device in
Simulation results show that the proposed approaches significantly the network can respond faster and better to the changes
improve the localization accuracy, reducing the estimation error in the environment (e.g., location-aware vehicles and asset
between 15% and 20% on average, compared with the existing management in warehouses) [3]. The idea of wireless posi-
approaches.
tioning was initially conceived for cellular networks, since
Index TermsCentralized localization, cooperative localiza- it invokes many innovative applications and services for its
tion, noncooperative localization, received signal strength (RSS), users. Nowadays, rapid increase in heterogeneous smart de-
vices (mobile phones and tablets), which offer self-sustained
Manuscript received December 17, 2013; revised April 23, 2014 and applications and seamless interfaces to various wireless net-
June 19, 2014; accepted June 26, 2014. Date of publication July 1, 2014;
date of current version May 12, 2015. This work was supported in part by works, is pushing the role of the location information to
the Fundao para a Cincia e a Tecnologia under Projects PEst-OE/EEI/ become a crucial component for mobile context-aware ap-
UI0066/2014, EXPL/EEI-TEL/0969/2013-MANY2COMWIN and EXPL/ plications [1].
EEI-TEL/1582/2013-GLANC, PEst-OE/EEI/LA0008/2013 (IT pluriannual
founding and HETNET), PEst-OE/EEI/UI0066/2011 (UNINOVA pluriannual Location information in WSNs is usually obtained by range-
founding), EnAcoMIMOCo EXPL/EEI-TEL/2408/2013, and ADIN PTDC/ based or range-free measurements. In this paper, we focus only
EEI-TEL/2990/2012, and Grant SFRH/BD/91126/2012 and the Cincia 2008 on the former, since they provide higher estimation accuracy
Post-Doctoral Research grant. This paper was presented in part at the
IEEE Workshop on Signal Processing Advances in Wireless Communications in general. In the range-based localization process, the main
(SPAWC 2013) Darmstadt, Germany, June 1619, 2013. The review of this concern is to accomplish good estimation accuracy from in-
paper was coordinated by Prof. G. Mao. accurate position-bearing measurements collected inside the
S. Tomic is with the Institute for Systems and Robotics (ISR), Instituto
Superior Tcnico, Universidade de Lisboa, 1049-001 Lisbon, Portugal (e-mail: network. To obtain the information of interest, it is necessary
stomic@isr.ist.utl.pt). to enable node communication, which can be noncooperative
M. Beko is with the Universidade Lusfona de Humanidades e Tecnologias,
1749-024 Lisbon, Portugal, and also with UNINOVA, 2829-516 Caparica,
or cooperative. The former approach allows source nodes to
Portugal (e-mail: beko.marko@ulusofona.pt; mbeko@uninova.pt). communicate only with the anchor nodes, whereas the latter
R. Dinis is with the Instituto de Telecomunicaes, 1049-001 Lisbon, approach allows source nodes to communicate with all nodes
Portugal; with the Centre of Technology and Systems (CTS), UNINOVA,
2829-516 Caparica, Portugal; and also with the Departamento de Engenharia
inside their communication range, whether they are anchor
Electrotcnica, Faculdade de Cincias e Tecnologia (FCT), Universidade Nova or source nodes [1]. Algorithmically, both approaches can
de Lisboa, 2829-516 Caparica, Portugal (e-mail: rdinis@fct.unl.pt). be executed in a distributed (self-positioning) or centralized
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
(network-centric positioning) mode. Although the distributed
Digital Object Identifier 10.1109/TVT.2014.2334397 approach has, in general, low complexity and high scalability,

0018-9545 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
2038 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 5, MAY 2015

it is sensitive to error propagation and may require long con- The grid search methods are time consuming and require a
vergence time. Therefore, in this paper, we concentrate only on huge amount of memory when the number of unknown pa-
the centralized approach. To establish an accurate estimate of rameters is too large. Linear estimators are very efficient in
the source position, the processor must have prior knowledge the sense of time consumption, but they are derived based
about the anchor nodes positions. In cellular networks, the base on many approximations that may affect their performance,
stations can be seen as anchor nodes, and the mobile stations particularly in the case when the noise is large [14]. In convex
can be seen as source nodes. relaxation techniques such as those in [12][18], the difficulties
Depending on the available hardware, current distance- in the ML problem are overcome by transforming the original
based localization techniques exploit different measurements nonconvex and nonlinear problem into a convex problem. The
of the radio signal transmitted between nodes [received signal advantage of this approach is that convergence to the globally
strength (RSS), time-difference-of-arrival (TDOA), time-of- optimal solution is guaranteed. However, due to the use of
arrival (TOA), roundtrip-time (RTT), or angle-of-arrival (AOA) relaxation techniques, the solution of a convex problem does
measurements]. The tradeoff between the localization accu- not necessarily correspond to the solution of the original ML
racy and the implementation complexity of each technique problem [20].
is a very important factor when deciding which method to In [10], different weighting schemes for the multidimen-
employ. For example, localization based on TOA or TDOA sional scaling (MDS) formulation were presented and com-
[including the Global Positioning System (GPS)] gives high pared. It was shown that the solution of the MDS can be
estimation accuracy, but it requires a very complex process used as the initial value for iterative algorithms, which then
of timing and synchronization, thus making the localization converge faster and attain higher accuracy when compared with
cost expensive [4]. Although less accurate than the localization random initial values. Convex semidefinite programming (SDP)
using TOA, TDOA, or AOA information, localization based estimators were proposed in [12] to address the nonconvexity
on the RSS measurements requires no specialized hardware, of the ML estimator, for both noncooperative and coopera-
less processing, and communication (and, consequently, lower tive localization problems with known source transmit power,
energy), thus making it an attractive low-cost solution for i.e., PT . Ouyang et al. in [12] reformulated the localization
the localization problem [2], [5]. Another attractive low-cost problem by eliminating the logarithms in the ML formulation
approach might be exploiting RTT measurements, which are and approaching the localization problem as a minimax opti-
easily obtained in wireless local area network (WLAN) sys- mization problem, which is then relaxed as SDP. Although the
tems by using a simple device such as a printed circuit board approach described in [12] provides good estimation results,
[6]. Although RTT systems circumvent the problem of clock particularly for the case of cooperative localization, it has high
synchronization between nodes, the major drawback of this computational complexity, which might restrict its applica-
approach is the need for double signal transmission to perform tion in large-scale WSNs. In [13], the RSS-based localization
a single measurement [7]. Recently, hybrid methods that fuse problem for known PT was formulated as the weighted least
two measurements of the radio signal (e.g., RSS-RTT) attracted squares (WLS) problem, based on the unscented transformation
considerable attention in the research society [6], [8], [9]. These (UT). It was shown that for the cooperative localization, the
methods try to improve the estimation accuracy in node posi- WLS formulation can be relaxed as a mixed semidefinite and
tions by exploiting the benefits of the combined measurements, second-order cone programming problem (SD/SOCP), whereas
together with minimizing their drawbacks. In this paper, we for the noncooperative localization, the WLS problem can be
focus on providing a good localization accuracy by using RSS solved by the bisection method. In [16], Wang et al. addressed
measurements exclusively. the noncooperative RSS localization problem for the case of
unknown PT and the path loss exponent (PLE). For the case
of unknown PT , based on the UT, a WLS formulation of the
A. Related Work
problem is derived, which was solved by the bisection method.
Source localization based on the RSS measurements has When both PT and PLE are not known, an alternating esti-
recently attracted much attention in the wireless communica- mation procedure is introduced. However, both [13] and [16]
tions community [10][18]. The most popular estimator used have the assumption of perfect knowledge of the noise stan-
in practice is the maximum likelihood (ML) estimator, since dard deviation (STD). This might not be the case in practice,
it is asymptotically efficient (for large enough data records) particularly in low-cost systems such as RSS where calibration
[19]. However, solving the ML estimator of the RSS-based is avoided due to maintaining low system costs [2], [5]. In
localization problem is a very difficult task, because it is [18], Vaghefi et al. addressed the RSS cooperative localization
highly nonlinear and noncovex [5]; hence, it may have multiple problem for unknown PT . The case where the source nodes
local optima. In this case, search for the globally optimal have different PT (e.g., due to different antenna gains) was
solution is very hard via iterative algorithms, since they may considered in [18]. The authors solved the localization problem
converge to a local minimum or a saddle point resulting in a by applying an SDP relaxation technique and converting the
large estimation error. To overcome this difficulty and possibly original ML problem into a convex problem. Furthermore, in
provide a good initial point (close to the global minimum) [18], the authors examined the effect of imperfect knowledge
for the iterative algorithms, approaches such as grid search of the PLE on the performance of the SDP algorithm and used
methods, linear estimators, and convex relaxation techniques an iterative procedure to solve the problem when PT and PLE
have been introduced to address the ML problem [10][18]. are simultaneously unknown.
TOMIC et al.: RSS-BASED LOCALIZATION IN WIRELESS SENSOR NETWORKS USING CONVEX RELAXATION 2039

B. Contributions II. N ONCOOPERATIVE L OCALIZATION V IA

S ECOND -O RDER C ONE P ROGRAMMING R ELAXATION
In this paper, the RSS-based source localization problem for
both noncooperative and cooperative scenarios is considered. Let us consider a WSN with N anchors and one source,
Instead of solving the ML problem, which is highly nonconvex where the locations of the anchors are respectively denoted
and computationally exhausting to solve globally, we propose a by s1 , s2 , . . . , sN , and the location of the unknown source is
suboptimal approach that provides an efficient solution. Hence, denoted by x. Without loss of generality, this paper focuses on
we introduce a new nonconvex least squares (LS) estimator that the 2-D scenario, i.e., x, s1 , s2 , . . . , sN R2 (the extension for
tightly approximates the ML estimator for small noise. This a 3-D scenario is straightforward). For the sake of simplicity, we
estimator represents a smoother and simpler localization prob- assume that all anchors are equipped with omnidirectional an-
lem in comparison to the ML problem. Applying appropriate tennas and connected to the source. Furthermore, it is assumed
convex relaxations to the derived nonconvex estimator, novel that the anchor positions are known. Under the lognormal
SOCP and novel mixed SDP/SOCP estimators are proposed for shadowing and log-distance path loss model, the path loss
noncooperative and cooperative localization cases, respectively. between the ith anchor and the unknown source, i.e., Li , can
The proposed approach offers an advantage over the existing be modeled according to the following radio propagation path
approaches as it allows straightforward adaptation to different loss model (in decibels) [22][25], i.e.,
scenarios of the RSS localization problem, thereby signifi-
x si 
cantly reducing the estimation error. In both noncooperative Li = L0 + 10 log10 + vi , i = 1, . . . , N (1)
and cooperative scenarios, we first consider the simplest case d0
of the localization problem where PT is known at the anchor where L0 denotes the path loss value at a short reference
nodes. Next, we consider a more realistic scenario in which we distance d0 (x si  d0 ), is the PLE, and vi is the
assume that PT is an unknown parameter, and we generalize lognormal shadowing term modeled as a zero-mean Gaussian
our approaches for this setting. Finally, we investigate the most random variable with variance i2 , i.e., vi N (0, i2 ). The
challenging scenario of the localization problem when PT and model has been validated by a variety of measurement results
PLE are simultaneously unknown at the anchor nodes. In this [23][28].
case, for the noncooperative localization, we apply an iterative From the relationship Li (dB) = 10 log10 (PT /Pi ), where
procedure based on the proposed SOCP method to estimate Pi is the RSS measured by the ith anchor, and PT is the
all unknown parameters. We also provide details about the transmission power of the unknown source, it is easy to see
computational complexity of the considered algorithms. that the localization problem can be formulated by the path
In contrast to [12] and [13] where the authors considered the loss instead of the RSS. Hence, as in [12] and [13], the path-
localization problem for the case when PT is known, here, we loss-based approach is adopted in this paper. Based on the
address a more challenging scenario when both PT and PLE are measurements in (1), the ML estimator is found by solving the
not known. In [13] and [16], Wang and Yang and Wang et al., nonlinear and nonconvex LS problem, i.e.,
respectively, assumed that accurate knowledge of the noise
STD is available, which might not be a valid assumption in N  2
1 x si 
some practical scenarios. Hence, we consider a more realistic = arg min
x (L i L 0 ) 10 log 10 .
x 2
i=1 i
d0
scenario in which the noise STD is not available. In contrast to
(2)
[18] where an SDP estimator is derived for the case of unknown
PT , we derive our estimators by using SOCP relaxation for the To solve (2), recursive methods such as Newtons method, com-
noncooperative case and mixed SDP/SOCP relaxation for the bined with the gradient descent method, are often used [19].
cooperative case. However, the objective function may have many local optima,
The remainder of this paper is structured as follows. In and local search methods may easily get trapped in a local
Section II, the RSS measurement model for locating a single optimum. Hence, in this paper, we employ convex relaxation
source node is introduced, the source localization problem is to address the nonconvexity of the localization problem.
formulated for the case of known source transmit power, and The remainder of this section is organized as follows.
the development of the proposed SOCP estimator is presented. Section II-A deals with the case where PT is known, whereas
We then extend this approach for the case where PT , and PT Section II-B deals with the case where PT is considered to
and PLE are simultaneously unknown. Section III introduces be an unknown parameter that needs to be estimated. Finally,
the RSS measurement model for the cooperative localization Section II-C addresses the more general problem when PT and
where multiple source nodes are simultaneously located. We PLE are simultaneously unknown.
give the formulation of the cooperative localization problem
and provide details on the development of the proposed SDP es-
A. Noncooperative Scenario With Known PT
timators for both cases of known and unknown source transmit
power. The complexity analysis is summarized in Section IV. The source might be designed to measure and report its own
In Section V, we provide both the complexity and simulation calibration data to the anchors, in which case it is reasonable to
results to compare the performance of the newly proposed assume that the source transmission power is known [5]. This
estimators with the existing estimators. Finally, in Section VI, corresponds to the case when the reference path loss L0 , which
we summarize the main conclusions. depends on PT [22], is known.
2040 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 5, MAY 2015

For the sake of simplicity, in the rest of this paper, we assume Here, a more realistic and challenging scenario where the
that i2 = 2 , for i = 1, . . . , N . When the noise is sufficiently anchor nodes are not aware of the source transmission power
small, from (1), we get is considered; thus, L0 is assumed to be unknown and has to
be estimated. The joint ML estimation of x and L0 can be
i x si  d0 (3) formulated as
 2
where i = 10(L0 Li )/10 . One way for estimating the source N
1 AT si 
= arg min
(L i l T
) 10 log
location x is via the minimization of the LS criterion. 2 10
d0
=[x;L0 ] i=1
Thus, according to (3), the LS estimation problem can be (7)
formulated as1
where l = [021 ; 1], and A = [I 2 ; 012 ].

N
2 In (3), we assumed that PT , i.e., L0 is known. Assuming that
= arg min
x (i x si  d0 ) . (4)
x
i=1
L0 is unknown, we can rewrite (3) as

Although the problem in (4) is nonconvex, when i = d0 , for i x si  d0 (8)

i = 1, . . . , N , it can be accurately solved by the SCLP method
presented in [30].2 In the following text, we will present a novel where i = 10Li /10 , and = 10L0 /10 . By following a
approach to solve the problem defined in (4). procedure similar to that in Section II-A, we obtain the SOCP
Defining auxiliary variables z = [z1 , . . . , zN ]T , where zi = problem, i.e.,
i gi d0 , and gi = x si , from (4), we get
minimize t
x,g,z,,t
minimize z2
x,g,z
subject to
subject to
[2z; t 1] t + 1, x si  gi
gi = x si , zi = i gi d0 , i = 1, . . . , N. (5) zi = i gi d0 , i = 1, . . . , N. (9)
Introducing an epigraph variable t and relaxing the nonconvex Although the approach in (9) efficiently solves (7), we can
constraint gi = x si  as gi x si  yield the following further improve its performance. To do so, we will exploit the
SOCP problem:  0 , which we get by solving (9), and solve
estimate of L0 , i.e., L
another SOCP problem. This SOCP approach will be described
minimize t
x,g,z,t in the following text.
Introducing auxiliary variables ri = x si  and i = ri2 ,
subject to
expanding (4), and dropping the term d20 , which has no effect
[2z; t 1] t + 1, x si  gi , on the minimization, yield

zi = i gi d0 , i = 1, . . . , N. (6) 
N
 
minimize i2 i 2d0
i ri
x,,r
Problem (6) can be efficiently solved by [21], and we will refer i=1

to it as SOCP1 in the following text.3 subject to i = ri2 , ri = x si , i = 1, . . . , N (10)

where 0 )/10 . One can relax (10) to a convex

i = 10(Li L
B. Noncooperative Scenario With Unknown PT
optimization problem as follows. The nonconvex constraint
The assumption that the anchors know the actual source i = ri2 will be replaced by the second-order cone constraint
transmission power may be too strong in practice since it would (SOCC) ri2 i . In fact, the inequality constraint ri2 i will
require additional hardware in both source and anchors [5]. be satisfied as an equality since i and ri will decrease and
increase in the minimization, respectively. Furthermore, define
1 A justification for dropping the shadowing term in the propagation an auxiliary variable y = x2 . The constraint y = x2 is
path loss model is provided in the following text. We can rewrite (1) as relaxed to a convex constraint y x2 , which is evidently an
(Li L0 )/10 = log10 (x si /d0 ) + (vi /10), which corresponds to SOCC. With the use of all developed constraints, problem (10)
i x si  = d0 10vi /10 . For sufficiently small noise, the first-order Taylor
series expansion to the right-hand side of the previous expression is given by
is approximated as a convex, SOCP, optimization problem, i.e.,
i x si  = d0 (1 + (ln 10/10)vi ), i.e., i x si  = d0 + i , where
i = d0 (ln 10/10)vi is the zero-mean Gaussian random variable with the 
N
 
variance d20 ((ln 10)2 /100 2 ) 2 . Clearly, the corresponding LS estimator is minimize i2 i 2d0
i ri
x,,r,y
given by (4). The same has been done in [18]. i=1
2 It is possible to generalize the SCLP method to the weighted case, i.e., to
the case when i = d0 for some i. However, the algorithm in [30] yields a subject to
meaningless solution. This is due to the fact that the almost convexity property
of the resulting constraints is not preserved.
3 It is worth noting that the SOCP1 approach can be modified to solve the [2x; y 1] y + 1, [2ri ; i 1] i + 1
localization problem in a distributed fashion. i = y 2sTi x + si 2 , i = 1, . . . , N. (11)
TOMIC et al.: RSS-BASED LOCALIZATION IN WIRELESS SENSOR NETWORKS USING CONVEX RELAXATION 2041

In summary, the proposed procedure for solving (7) is given as objective function, i.e., fk . If |fk fk1 |/fk1 < 
follows. ( is a small positive number) or k > Kmax (Kmax
Step 1) Solve (9) to obtain the initial estimate of x, x . is the maximum number of iteration) stop; otherwise
Step 2) Use x  
to compute the ML estimate of L0 , L0 , from let k = k + 1 and go to step 2).
(7) as We refer to the given described iterative procedure as
N 

SOCP3 in this paper.
i=1 Li 10 log10 x ds i

 0 =
L
0
. (12)
N III. C OOPERATIVE L OCALIZATION V IA S EMIDEFINITE
P ROGRAMMING R ELAXATION
Step 3) Use L   to solve the SOCP in (11) and obtain the
0 Consider now a WSN with N anchors and M source nodes,
new source position estimate x, x  . Compute the where, as before, the locations of the anchors s1 , s2 , . . . , sN
ML estimate of L0 , L  , from (12), by using x  .
0 are known, and the locations of the source nodes are
The main reason for applying this simple procedure is that x1 , x2 , . . . , xM (where xi , sj R2 , for i = 1, . . . , M and j =
we observed in our simulations that after solving (9), we obtain 1, . . . , N ). Due to the limited communication range, i.e., R, or
an excellent ML estimation of L0 , i.e., L   , which is very close
0 other physical limitations, only some source nodes can directly
to the true value of L0 . This motivated us to take advantage of connect to the anchor nodes, making the information gathered
this estimated value and solve another SOCP problem (11), as if inside the network insufficient to perform a good estimation. To
PT , i.e., L0 is known. In Section V-A2, we will see remarkable overcome this problem, node cooperation is required. Node co-
improvements in the estimation accuracy of both x and L0 operation allows direct communication between any two nodes
by employing the given procedure. We denote this three-step in the WSN, which are within the communication range of
procedure as SOCP2. each other. This means that the source nodes also perform RSS
measurements, i.e., they play a role of the anchor nodes, which
C. Noncooperative Scenario With Unknown PT and number is scarce, to acquire adequate amount of information.
After the RSS measurements are collected, all locations of
Signal attenuation may be caused by many effects, such the source nodes are simultaneously estimated. This kind of
as multipath fading, diffraction, reflection, environment, and localization is called cooperative localization [1].
weather condition characteristics. Thus, it is reasonable to For ease of expression, matrix X is built such that it con-
assume that PLE, i.e., , is not known at the anchor nodes. Here, tains the positions of all sources, i.e., X = [x1 , x2 , . . . , xM ]
we investigate the case where PT and are simultaneously (X R2M ). Furthermore, sets A = {(i, j) : xi sj 
unknown at the anchor nodes. The joint ML estimation of x, R, i = 1, . . . , M ,j = 1, . . . , N }, and B = {(i, k) : xi
L0 , and is written as xk  R, i, k = 1, . . . , M, i = k} denote the existence of the
N  2 source/anchor and the source/source connections, respectively.
1 C T si 
2 (Li h )10g log10
T T For the sake of simplicity, we assume that all source nodes
= arg min
=[x;L0 ;] i=1 i d0 radiate with the same power PT , i.e., L0 and R are the same
(13) for all source nodes.
where h = [021 ; 1; 0], g = [031 ; 1], and C = [I 2 ; 022 ]. According to the radio propagation path loss model in [22],
Problem (13) is nonconvex and has no closed-form solution. the RSS measurement for the cooperative localization can be
To tackle (13), we employ a standard alternating procedure formulated as
explained as follows (see also [16] and the references therein). xi sj 
LA ij = L0 + 10 log10 + vij , (i, j) A
Step 1) Instead of blind estimation, use empirical values, d0
e.g., [22], and set the initial estimate of , 0 xi xk 
LBik = L0 + 10 log10 + wik , (i, k) B (14)
[min , max ], and solve (9) to find the initial estimate d0
of x, x 0 . Use 0 and x 0 to calculate the ML estimate where vij and wik are the lognormal shadowing terms mod-
 0
of L0 , L0 . Compute the value of the objective func- eled as zero-mean Gaussian random variables with variances
tion, i.e., f0 , by plugging {  0 , 0 } into (13). Set
x0 , L v2ij and w
2
, i.e., vij N (0, v2ij ) and wik N (0, w 2
).
0 ik ik
k = 1. We assume that the source/source path loss measurements are
Step 2) Use x k1 and L  k1 to find the ML estimate of , symmetric, i.e., LBik = LBki for i = k.
0
 , as
k
As in (1), the path loss approach and the measurements in
N
(14) lead to the ML estimator, corresponding to solving the
10 log10 x d0si  Li L
k1
 k1
i=1 0 nonlinear and nonconvex LS problem, i.e.,
k =
N
2 .  2
10 log xk1 si 
  1  A  xisj 
i=1 10 d0
X=arg min L L 0 10 log 10
ij
X 2 d0
(i,j):(i,j)A vij
k [min , max ] go to step 3); else stop.
If  2
 1  B  xi xk 
Step 3) Use k and L  k1 to solve (11), obtain the estimate of + Lik L0 10 log10 .
0 2
x, x k , and update the ML estimate of L0 , L  k . Plug (i,k):(i,k)B
w ik
d0
0
k k
{
x , L0 , } into (13) and compute the value of the
k
(15)
2042 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 5, MAY 2015

The problem defined in (15) is nonconvex and nonlinear and, as The given problem is SDP (more precisely, it is mixed
far as we know, has no closed-form solution. As before, from SDP/SOCP), which can be readily solved by CVX [21]. If
(15), we distinguish two different cases. In the first case, PT rank(Y ) = 1, then constraint Y
yy T is satisfied as an equal-
(i.e., L0 ) is assumed to be known, whereas in the second case, ity [20]. Note that we applied the Schur complement to rewrite
PT is considered to be an unknown parameter that has to be Y
yy T into a semidefinite cone constraint form [20]. Note
estimated. In Sections III-A and III-B, a centralized localiza- also that, in huge contrast to the existing SDP-based approaches
tion algorithm based on SDP relaxation will be presented to [12], [13], [16], [18], which consider the unknown parame-
address problem (15) for the case of known and unknown PT , ters as a matrix, here, we consider them as a vector. In the
respectively. remainder of this paper, we will denote the given approach
as SDP1.
A. Cooperative Scenario With Known PT
For the sake of simplicity, we assume that v2ij = w
2
= 2 B. Cooperative Scenario With Unknown PT
ik
in the remainder of this paper. The following procedures similar The joint ML estimation of X and L0 is given by
to those in the noncooperative localization problem, a convex
estimator for cooperative localization is derived by applying  1
=
arg min
semidefinite relaxation to the nonconvex problem (15). =[xT T T 2
1 ,...,xM ,L0 ] (i,j):(i,j)A
As in Section II-A, we can approximate (14) as
  T 
A 2 B 2    Ai s j  2
ij xi sj 2 d20 , ik xi xk 2 d20 (16) LA q 10 log10
T
ij
d0
A (L0 LA
ij )/10 B (L0 LB )/10
where = 10
ij , and = 10 ik . ik
  T 
From (16), the following LS minimization problem is derived:  1  B  Ai ATk  2
+ Lik q 10 log10
T
 2
2 2 d0
A
minimize ij xi sj 2 d20 (i,k):(i,k)B
X
(i,j):(i,j)A

2 (20)
 2
B
+ ik xi xk 2 d20 . (17)
(i,k):(i,k)A
where q = [02M 1 ; 1], Ai = [r 2i1 , r 2i ], and r i represents
the ith column of the identity matrix I 2M +1 .
Next, define vector y = vec(X), where vec(X) denotes the To solve (20), we will use a similar idea as in the SDP1
column-wise vectorization of X. Then, (17) can be written as approach. We can rewrite (16) as
 2 2
2 A B
minimize A  T
ij E i y sj  d20 ij xi sj 2 d20 , ik xi xk 2 d20 (21)
y
(i,j):(i,j)A A
  T 
2 A
where ij = 10Lij /5 , ik
B B
= 10Lik /5 , and = 10L0 /5 .
+ B 2
ik E i y E Tk y 2 d20 (18) Following the steps highlighted in Section III-A, the following
(i,k):(i,k)B convex problem is obtained:
where E i = [e2i1 , e2i ], and ei represents the ith column minimize t
of the identity matrix I 2M . Introducing an epigraph vari- y,Y ,z,t,
A B T
able t and auxiliary variables Y = yy T and z = [zij , zik ] ,
2 2
subject to
A A
where zij = ij E Ti y sj  d20 , for (i, j) A, and zik
B
=   T  
A A
B
ik
2 2
E Ti y E Tk y d20 , for (i, k) B, together with the zij = ij tr E i Y E i 2sTj E Ti y+sj 2 d20
convex relaxation Y
yy T , the following convex epigraph
form is obtained from (18): for (i, j) A
B B
  T   
zik = ik tr E i Y E i 2tr E Ti Y E k
minimize t
y,Y ,z,t  
+tr E Tk Y E k d20
subject to
  T2   for (i, k) B
A A
zij = ij tr E i Y E i 2sTj E Ti y+sj 2 d20
for (i, j) A [2z; t 1] t + 1, [Y y; y T 1]
02M +1 . (22)
B B 2
    
zik = ik tr E Ti Y E i 2tr E Ti Y E k Although the approach in (22) efficiently solves (20), we can
 
+ tr E Tk Y E k d20 further improve its performance. As in Section II-B, we propose
for (i, k) B a simple three-step procedure as follows.
[2z; t 1] t + 1, [Y y; y T 1]
02M +1 . (19) .
Step 1) Solve (22) to obtain the initial estimate of y, y
TOMIC et al.: RSS-BASED LOCALIZATION IN WIRELESS SENSOR NETWORKS USING CONVEX RELAXATION 2043

Step 2) Use y  to compute the ML estimate of L0 , L   , from interpret the cooperative localization problem defined in [12] in
0
(20) as dual form, we get ((M + 2)(M + 3)/2) + K variables, which
  correspond to ((M + 2)(M + 3)/2) + K equality constraints
 E Ti y  sj 
LAij 10 log 10 d0
in the primal form. In contrast, if we interpret the same problem
 = (i,j):(i,j)A in the primal form, we get 4K + 3 equality constraints, cor-
L 0
|A| + |B| responding to the same number of variables in the dual form.
  While performing the simulations, we have experienced that
 E Ti y  E Tk y  
LBik 10 log10 d0
the latter interpretation is computationally more efficient; thus,
(i,k):(i,k)B the complexity analysis is performed based on the primal form
+ (23)
|A| + |B| representation only.
To provide a more complete overview of the algorithms
where |A| and |B| represent the cardinalities of sets performance to the reader, we present the complexity results
A and B. together with the simulations results in the following section.
Step 3) Use L  to solve (19) and obtain the new estimate of
0
y, y 
. Compute the ML estimate of L0 , L   , from
0
 V. P ERFORMANCE R ESULTS
(23), by using y .
We will refer to the given three-step procedure as SDP2. Here, we present a set of performance results to compare
the proposed approaches with the existing approaches, for both
noncooperative and cooperative localization with known and
IV. C OMPLEXITY A NALYSIS unknown PT . All of the presented algorithms were solved by
The tradeoff between the estimation accuracy and the com- using the MATLAB package CVX [21], where the solver is
putational complexity is one of the most important features of SeDuMi [36].
any method since it defines its applicability potential. This is
the reason why, apart from the performance, we are interested A. Noncooperative Localization
in comparing the complexity of the proposed and the existing
approaches. To generate the RSS measurements, the propagation model
The formula for computing the worst case complexity of (1) is used. Extensive simulations have been carried out to
mixed SD/SOCP [32], given in the following equation, is used compare the performance of the proposed methods in Section II
to analyze the complexities of the considered algorithms in this with that of the existing methods for the cases of known and
paper: unknown source transmit power. Unless stated otherwise, in
  N all simulations presented here, the number of Monte Carlo
sd
3

Nsd 
N soc
runs is Mc = 10 000, the PLE is = 3, the reference distance
sd2
O L m nsd
i + m 2
ni + m 2
nsoc
i d0 = 1 m, and the path loss L0 = 40 dB. The anchor nodes
i=1 i=1 i=1
 are uniformly distributed at a circle with the center at the

N soc
2
origin, and the radius of the circle r = 20 m. A source is
+ nsoc
i +m 3
(24) randomly distributed inside the square region {(x, y)| B
i=1 x B, B y B}, where B will be defined below, and it is
where L is the iteration complexity of the algorithm; m is the able to communicate with all anchors. The performance metric
number of equality constraints; nsd soc is the root mean square error (RMSE), which is defined as
i and ni are the dimensions

of the ith semidefinite cone (SDC) and the ith second-order  Mc
 xi x  i 2
cone (SOC), respectively; and Nisd and Nisoc are the number RMSE = 
of SDC and SOC constraints, respectively. Equation (24) corre- i=1
Mc
sponds to the formula for computing the complexity of an SDP
for the case when we have no SOCCs (in which case, L is the where x  i denotes the estimate of the true source location, i.e.,
dimension of the SDP cone, given as a result of accumulating xi , in the ith Monte Carlo run for the specific noise realization.
all SDP cones), and vice versa (in which case, L is the squared The CramrRao lower bound (CRB) on the RMSE of any
root of the total number of SOCCs) [32]. unbiased estimator is employed as a performance benchmark
We investigated the worst case asymptotic complexity of (see Appendix A for more details).
the algorithms, i.e., we present only the dominating elements, 1) Known PT : Table I gives an overview of the considered
which are expressed as a function of N and M . Since the worst algorithms in this section, together with their complexities. In
case complexity is considered, we assumed that the network [12] and [13], the authors have considered both indoor and out-
is fully connected, i.e., the total number of connections in door localization scenarios. Our simulation results show that,
the network is K = |A| + |B|, where |A| = M N , and |B| = for the chosen scenario, the proposed approach exhibits just
M (M 1)/2. a marginal gain when compared with the existing approaches.
It should be pointed out that the algorithms mentioned here Thus, for the noncooperative scenario when PT is known, we
for solving the source localization are not uniquely defined in focus on indoor localization only.
a clear primal or dual form; thus, we can interpret them in the Indoor localization: In Appendix B, we give more details
form that is more suitable for the solver [35]. For example, if we about the indoor propagation model. The simulation results
2044 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 5, MAY 2015

TABLE I
S UMMARY OF THE C ONSIDERED A LGORITHMS IN S ECTION V-A1

Fig. 2. Simulation results for noncooperative localization when L0 is not

known: RMSE versus N when = 5 dB, B = 15 m, r = 20 m, L0 = 40 dB,
= 3, d0 = 1 m, and Mc = 10 000.

TABLE III
L0 E STIMATION A NALYSIS FOR THE SOCP2 A PPROACH

Fig. 1. Simulation results for noncooperative localization in indoor environ-

ment when L0 is known: RMSE versus when N = 9, U = 5 dB, = 2.4,
w = 4 dB, d0 = 1 m, L0 = 30 dB, and Mc = 50 000.

TABLE II
S UMMARY OF THE C ONSIDERED A LGORITHMS IN S ECTION V-A2

TABLE IV
S UMMARY OF THE C ONSIDERED A LGORITHMS IN S ECTION V-A.3

for indoor localization are presented in Fig. 1. The scenario

described in [13] is used to execute the comparison of the
performances. Fig. 1 clearly demonstrates the superiority of the three-step procedure. When N increases, we can obtain a better
proposed approach over the existing approaches for the whole estimate of PT , i.e., L0 , in the first step of our procedure, which
range of . To illustrate this fact, consider the cases when = then allows us to achieve high estimation accuracy in the second
1 dB and = 6 dB. For the former case, SOCP1 shows a gain step with respect to (w.r.t.) x (see Table III for more details).
of approximately 0.2 m when compared with the existing ap- As shown, after the second step of our procedure, we obtain
proaches. For the latter case, SOCP1 outperforms WLS-1, an excellent estimate of L0 , which we then employ in the third
and SDPRSS approaches by 0.3 and 0.5 m, respectively. step. From Table III, we observe that as N is increased, the
In summary, the proposed approach outperforms the existing estimation of L0 is improved. Although the estimation of L0 in
approaches in terms of the estimation accuracy with an average the second step is not perfect, the simulation results in Fig. 2
error reduction of about 15%, whereas in terms of the compu- confirm the robustness of our SOCP approach to the imperfect
tational complexity, it represents a solid alternative. knowledge of L0 .
2) Unknown PT : Table II gives an overview of the consid- In Fig. 2, we can see that the new approach is biased for
ered algorithms in this section, together with their complexities. small N . This is not surprising since it is known that RSS-based
Fig. 2 shows the comparison of the RMSE versus N for = algorithms are generally biased [2]. For the sake of complete-
5 dB, B = 15 m and r = 20 m. In Fig. 2, we observe that ness, we also compared the considered approaches  c in terms of
the estimation error decreases as N is increased, as expected. bias. The bias is defined as Bias = (1/Mc ) M i=1 (xi x
i )1 ,
Furthermore, we observe the superior RMSE performance of where  1 represents the l1 norm. In our simulations, we
the SOCP2 approach for all chosen N . Fig. 2 also shows have observed that all approaches are slightly biased. However,
that the performance margin between the proposed and the the bias error is less than 0.14 m for all estimators, which is a
existing approaches grows as N is increased, e.g., for N = 24, relatively small error in the case when the RMSE error is of the
SOCP2 outperforms the existing approaches by more than order of a few meters.
0.5 m, which is a relatively large gain considering that when 3) Unknown PT and : Table IV gives an overview of
N = 24, the amount of information gathered in the network is the considered algorithms in this section, together with their
significant. This can be explained, to some extent, by the fact complexities. The results in [38] imply that better estimation
that our approach has somewhat higher computational com- accuracy w.r.t. x is attained by choosing the initial value of
plexity than the existing approaches and the use of the proposed the PLE, i.e., 0 , to be greater than the true value of . This
TOMIC et al.: RSS-BASED LOCALIZATION IN WIRELESS SENSOR NETWORKS USING CONVEX RELAXATION 2045

TABLE V
0 [min , max ]
U NKNOWN PARAMETER E STIMATION A NALYSIS FOR A R ANDOM C HOICE OF

TABLE VI
S UMMARY OF THE C ONSIDERED A LGORITHMS IN S ECTION V-B1

B. Cooperative Localization

Fig. 3. Simulation results for noncooperative localization when L0 and

This section presents the simulation results for the cooper-
are not known: RMSE versus N when = 5 dB, B = 15 m, r = 20 m, ative localization problem. The performance of the proposed
L0 = 40 dB, d0 = 1 m, Kmax = 30, min = 2, max = 4,  = 103 , and approaches in Sections III-A and B will be compared with
Mc = 10 000.
the existing algorithms for the cases of known and unknown
PT . In all simulation results presented here, the number of
motivated us to investigate the influence of the choice of 0 on Monte Carlo runs is Mc = 10 000, the PLE is = 3, the
the estimation accuracy of the remaining unknown parameters. reference distance d0 = 1 m, and the path loss L0 = 40 dB.
Our analysis showed that the influence of the choice of 0 on Unless stated otherwise, the anchor nodes are fixed at the posi-
the estimation accuracy of the source positions is significant. tions (B, B), (0, B), (B, B), (B, 0), (B, B), (0, B),
We present some of these results in Table V. Note that the (B, B), (B, 0), and (0,0), where B will be defined below.
average number of iterations for an algorithm to converge is Source nodes are randomly deployed inside the convex hull
denoted by Av. iter. Generally, the best estimation accuracy of the anchor nodes, and due to the limited communication
w.r.t. x is achieved for 0 = max . However, when the key range, only some nodes can directly connect to the anchors. The
requirement is to estimate L0 or , then a random choice performance metric is the normalized RMSE (NRMSE), which
of 0 [min , max ] is by far the best choice. The price to is defined as
pay for this choice of 0 is somewhat slower convergence, in 
comparison to the choice 0 = max . 
 1  Mc M
xij x ij 
Fig. 3 shows the comparison of the RMSE versus N for = NRMSE = 
5 dB, B = 15 m, and r = 20 m. In this figure, we represent M i=1 j=1 Mc
a curve for the proposed approach for a random choice of
0 [min , max ]. Fig. 3 confirms the efficiency of the iterative where x  ij denotes the estimate of the true location of the jth
procedure. It shows that the proposed approach is robust to the source, xij , in the ith Monte Carlo run for the specific noise
imperfect knowledge of the PLE, since it suffers only a small realization.
deterioration of the performance when both PT and PLE are 1) Known PT : Table VI gives an overview of the con-
not known, compared with the case when only PT is unknown. sidered algorithms in this section, together with their com-
Furthermore, Fig. 3 shows the superior RMSE performance of plexities. We observe from Table VI that the proposed SDP
the proposed approach for all chosen N . We observe that the approach has somewhat higher computational complexity than
performance margin between SOCP3 and WLS-3 increases the SD/SOCP-1 approach and lower complexity than the
as N increases. To illustrate this fact, consider N = 9 and N = SDPRSS approach. Although somewhat more complex than
24. In the former case, the SOCP3 approach outperforms the SD/SOCP-1 approach, the proposed approach has the
WLS-3 by roughly 0.3 m, whereas in the latter case, it lowest running time, as one can see from Table VII(a). This
shows a gain of roughly 0.5 m. In Fig. 3, it is shown that is due to the fact that for cooperative localization, when PT
SOCP3 achieves the value RMSE = 4 m with four anchors is known, we solve only one SDP, whereas SD/SOCP-1
less than WLS-3, which can reduce the cost of the network involves solving two SDP problems.
implementation in practice. Note that this gain comes at a cost Fig. 4 shows the comparison of the NRMSE versus M ,
of increased complexity. for N = 9, R = 6 m, = 5 dB, and B = 15 m. We observe
2046 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 5, MAY 2015

TABLE VII by the SDP2 approach, for N = 9, M = 40, R = 6 m, =

AVERAGE RUNNING T IME OF THE C ONSIDERED A LGORITHMS FOR THE
C OOPERATIVE L OCALIZATION. N = 8, M = 20, AND R = 6 m. 5 dB, and B = 15 m. In Fig. 5, it is shown that better estimation
CPU: I NTEL C ORE 7-363QM 2.40 GH Z . (a) PT accuracy is achieved for source nodes with higher number of
IS K NOWN . (b) PT IS N OT K NOWN connections (neighbors), as anticipated.
Fig. 6 shows the NRMSE performance of the considered
approaches for different M , when N = 9, R = 6 m, = 5 dB,
and B = 15 m. We can see that the estimation accuracy im-
proves as M is increased, as expected. Fig. 6 confirms the
superiority of our approach, since it outperforms the existing
approaches for all choices of M , with an average gain of
about 0.5 m. When the information gathered by the network is
not enough (low M ), our approach outperforms SD/SOCP-2
and SDP-URSS by roughly 2 and 1 m, respectively. As M
increases, the performance margin between all approaches de-
creases, as anticipated, since the information inside the network
becomes sufficient to allow good performance for all estima-
tors. Finally, in Figs. 4 and 6, it is shown that the new approach
suffers only a marginal deterioration in the performance for the
scenario where PT is not known.
Fig. 7 shows the NRMSE performance of the considered ap-
proaches for different R, when N = 9, M = 15, = 5 dB, and
B = 15 m. As anticipated, the estimation error decreases when
R is increased. In Fig. 7, it is clear that the proposed approach
Fig. 4. Simulation results for cooperative localization when L0 is known:
NRMSE versus M when N = 9, = 5 dB, B = 15 m, L0 = 40 dB, = 3,
outperforms the existing approaches for all choices of R, with
d0 = 1 m, and Mc = 10 000. an average gain of more than 1 and 0.5 m compared with
SD/SOCP-2 and SDP-URSS, respectively. An important
TABLE VIII practical scenario would be the case where R is chosen to be
S UMMARY OF THE C ONSIDERED A LGORITHMS IN S ECTION V-B2
low, due to the need to preserve low energy consumption in the
network. In this case, our approach reduces the estimation error
by about 2 and 0.5 m, compared with SD/SOCP-2 and SDP-
URSS, respectively. Intuitively, as R is increased, all methods
are expected to perform good, since the information gathered
by the network becomes sufficient enough. However, one can
see in Fig. 7 that the performance gains between the proposed
approach and SDP-URSS increases with R; for R = 10 m,
the new approach outperforms SDP-URSS by almost 1 m.
This result further confirms the superiority of the proposed
that, as M increases, the estimation accuracy of all algorithms approach over SDP-URSS.
improves, which is intuitive since more information is collected Fig. 8 shows the NRMSE performance of the considered
inside the network. Furthermore, in Fig. 4, it is shown that approaches for different N , when M = 15, R = 6 m, =
the proposed approach outperforms the existing approaches for 5 dB, and B = 15 m. Both anchor and source nodes were
all values of M , with the biggest improvement for medium- randomly positioned inside a square region of length 2B. This
to-high M . The proposed approach outperforms on aver- scenario is of particular practical importance because a com-
age SD/SOCP-1 by roughly 0.5 m, although SD/SOCP-1 mon requirement for a network is to be flexible and adaptable
solves two SDP problems. Moreover, the new approach im- to different layouts. Fig. 8 shows that the estimation accuracy
proves the estimation accuracy on average by about 0.25 m in increases as N is increased, as predicted. One can see that our
comparison to SDPRSS , although SDPRSS is more compu- approach outperforms the existing approaches for all choices of
tationally complex. In short, albeit SDPRSS is more compu- N , with an average gain of about 1 and 0.5 m, compared with
tationally complex than the new approach and SD/SOCP-1 SD/SOCP-2 and SDP-URSS, respectively. Having less an-
solves two SDP problems, which increases its execution time, chor nodes in the network might reduce its implementation
the new approach outperforms both of them for all M . costs, since, e.g., they might be equipped with a GPS to
2) Unknown PT : Table VIII gives an overview of the con- determine their own positions. Fig. 8 shows that the proposed
sidered algorithms in this section, together with their complexi- approach needs, on average, one or two less anchor nodes to
ties. From Table VIII, it can be seen that the proposed approach achieve the same estimation accuracy as the state-of-the-art
has the highest computational complexity. This was confirmed approaches.4
in our simulations (see Table VII(b) for more details).
In Fig. 5, we present one possible network configuration and 4 The CRB was not presented in Figs. 48, since the Fisher information
the estimation accuracy of the source positions accomplished matrix (FIM) is singular for the chosen scenarios.
TOMIC et al.: RSS-BASED LOCALIZATION IN WIRELESS SENSOR NETWORKS USING CONVEX RELAXATION 2047

Fig. 5. Example of (a) a network configuration and (b) estimation accuracy results for the SDP2 approach. Black squares represent the locations of the anchor
nodes, blue circles represent the true locations of the source nodes, and red symbols X represent the estimated source locations.

Fig. 6. Simulation results for cooperative localization when L0 is not known: Fig. 8. Simulation results for cooperative localization when L0 is not known:
NRMSE versus M when N = 9, R = 6 m, = 5 dB, B = 15 m, L0 = NRMSE versus N when M = 15, R = 6 m, = 5 dB, B = 15 m, L0 =
40 dB, = 3, d0 = 1 m, and Mc = 10 000. 40 dB, = 3, d0 = 1 m, and Mc = 10 000. Anchors and sources are ran-
domly deployed inside the square region of length 2B.

Fig. 7. Simulation results for cooperative localization when L0 is not known:

NRMSE versus R when N = 9, M = 15, = 5 dB, B = 15 m, L0 = 40 dB, Fig. 9. Simulation results for cooperative localization when L0 is not known:
= 3, d0 = 1 m, and Mc = 10 000. CDFs of mean errors in source position estimation when N = 9, M =
15, R = 10 m, = 5 dB, B = 15 m, L0 = 40 dB, = 3, d0 = 1 m, and
Mc = 10 000.
Fig. 9 shows the CDF comparison of the mean error (ME)
in the source position estimation of the considered approaches,
for N = 9, M = 15, R = 10 m, = 5 dB, and B = 15 m. Monte Carlo run, respectively. Fig. 9 shows that the proposed
ME is defined as ME = M i=1 (xij x
ij /M ) (m), for j = approach outperforms the existing approaches for all range of
i,j denote the true and the es-
1, . . . , Mc , where xi,j and x ME, improving the estimation accuracy by more than 0.5 m, on
timated source positions of the ith source node in the jth average. We can see that the new method achieves ME 3 m
2048 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 5, MAY 2015

in 80% of the cases, whereas the existing methods attain the A PPENDIX A
same value of ME in less than 50% of the cases. C RAMER R AO B OUND D ERIVATION
CRB provides a lower bound on the variance of any unbiased
VI. C ONCLUSION estimator, which means that it is physically impossible to find
an unbiased estimator whose variance is less than the bound.
In this paper, we have addressed the RSS-based source
CRB offers us a benchmark against which we can compare
localization problem. Both noncooperative and cooperative lo-
the performance of any unbiased estimator. If the estimator
calization problems were investigated for both cases of known
attains the bound for all values of the unknown parameters,
and unknown source transmit power, i.e., PT .
we say that such estimator is the minimum variance unbiased
In the case of the noncooperative localization when PT is
estimator [19].
known, we proposed the novel SOCP-based approach, which T
has an excellent tradeoff between the performance and the Let = [xTk , L0 , ] , k = 1, . . . , M , denote the 2M + 2
computational complexity, when compared with the existing vector of all unknown parameters. According to [19], the vari-
ance of any unbiased estimator is lower bounded by var( i)
approaches. For the case where PT is not known, we introduced
the simple three-step procedure based on the SOCP relaxation. [J 1 ()]ii , where J () is the (2M + 2) (2M + 2)
The simulation results show that the proposed approach pro- FIM. The elements of the FIM are defined as [J ()]i,j =
vides not only an excellent estimation of the source positions E[ 2 ln p(L|)/ i j ], where i, j = 1, . . . , (2M + 2),
but also an excellent estimation of PT . This motivated us to and p(L|) is the joint conditional probability density function
exploit the estimate of PT and solve another SOCP problem of the observation vector L = [L1 , . . . , LK ], given .
as if PT was known. The price we have to pay for apply- The FIM is computed as
ing this procedure was solving the problem twice. However, 1  1 
the simulation results confirm its effectiveness and show a J () = f ij f Tij + f ik f Tik (25)
2 2
remarkable improvement of the estimation accuracy. A gain of (i,j):(i,j)A (i,k):(i,k)B

more than 15% was achieved, when comparing with existing

where
approaches, for the case where N is high. We concluded the
 T 
noncooperative localization problem by investigating the case D i sj  10g T d0 D i D Ti D i sj
where both PT and PLE are simultaneously unknown. By f ij = +10g log10 +   ,
d0 ln(10) D T sj 2
applying the standard iterative procedure, we showed that our i
 T 
method efficiently solves the most challenging scenario of the D i D Tk 
RSS localization problem and outperforms the existing method f ik = + 10g log10
d0
in terms of the estimation accuracy; the biggest gain is obtained
for high N , where an improvement in the localization accuracy 10g d0 D i D Ti D i D Tk D k D Ti
T
+ D k D Tk
+  T 
of about 15% is attained. Moreover, we have shown that our ln(10) D i D Tk 2
approach can be used to solve the noncooperative localization
problem in indoor environments. The new approach reduces the and = [02M 1 ; 1; 0], g = [0(2M +1)1 ; 1], D i = [p2i1 , p2i ],
estimation error by more than 15% when compared with state- where pi represents the ith column of the identity matrix
of-the-art approaches. I 2M +2 .
In the case of the cooperative localization when PT is known, Therefore, the CRB for the estimate of the source positions
we proposed the novel SDP-based approach. In huge contrast to is computed as
the existing SDP approaches that consider the cooperative lo- 


calization problem, in which the unknown variables are treated CRB = tr J 1 () 1:2M,1:2M (26)
as a matrix, here, we consider them as a vector. This approach
implies a slight increase in complexity in comparison to state- where [M ]a:b,c:d represents the submatrix of matrix M com-
of-the-art methods. However, the performance evaluation in posed of the rows a to b and columns c to d of M .
Section V justifies the use of such an approach. For the case
where PT is known, the proposed approach outperforms the ex-
isting approaches for all choices of M with the biggest margin A PPENDIX B
for medium-to-high M . In the case where PT is assumed to I NDOOR L OCALIZATION
be not known, the simple three-step procedure based on SDP
In practice, the attenuation in indoor environments is superior
relaxation is applied. We have investigated the influence of
than that in outdoor environments due to additional deteriorat-
different M and N , as well as different R on the estimation
ing caused by obstacles (such as walls, floors, and other objects)
accuracy. For all the scenarios presented in this paper, the
and multipath fading. Therefore, the propagation model (1)
new approach outperforms the state-of-the-art approaches with
is not suitable for indoor localization, and hence, we adopt a
an increase in accuracy between 15% and 20% on average.
different propagation model [13], i.e.,
Furthermore, the simulation results show that our approach
achieves ME 3 m in 80% of the cases, whereas the existing x si 
approaches accomplish the same accuracy in less than 50% of Li = L0 + Lw,i + 10 log10 + vi , i = 1, . . . , N
d0
the cases. (27)
TOMIC et al.: RSS-BASED LOCALIZATION IN WIRELESS SENSOR NETWORKS USING CONVEX RELAXATION 2049

where Lw,i is the path-loss term that represents the attenu- [9] U. Hatthasin, S. Thainimit, K. Vibhatavanij, N. Premasathian, and
ation caused by partitions and multipath fading. Similar to D. Worasawate, The use of RTOF and RSS for a one base station RFID
system, Int. J. Comput. Sci. Netw. Security, vol. 10, no. 7, pp. 184195,
the model in [13], we assume that Lw,i = nw,i w + ui w , Jul. 2010.
where nw,i represents the number of partitions that the signal [10] X. Li, Collaborative localization with received-signal strength in wireless
passes through, w is the partition attenuation factor, and ui = sensor networks, IEEE Trans. Veh. Technol., vol. 56, no. 6, pp. 3807
3817, Nov. 2007.
U sin(2t/tu ) is a random variable that models the varying [11] K. W. K. Lui, W. K. Ma, H. C. So, and F. K. W. Chan, Semi-definite
indoor environment. The samples of t were drawn from the programming algorithms for sensor network node localization with un-
uniform distribution on the interval [0, tu ]. w is the indicator certainties in anchor positions and/or propagation speed, IEEE Trans.
Signal Process., vol. 57, no. 2, pp. 752763, Feb. 2009.
function that indicates whether the signal passes through parti- [12] R. W. Ouyang, A. K. S. Wong, and C. T. Lea, Received signal strength-
tions or not, i.e., based wireless localization via semidefinite programming: Noncoopera-
tive and cooperative schemes, IEEE Trans. Veh. Technol., vol. 59, no. 3,
 pp. 13071318, Mar. 2010.
0, if nw,i = 0 [13] G. Wang and K. Yang, A new approach to sensor node localization using
w =
1, otherwise. RSS measurements in wireless sensor networks, IEEE Trans. Wireless
Commun., vol. 10, no. 5, pp. 13891395, May 2011.
[14] R. M. Vaghefi, M. R. Gholami, and E. G. Strom, RSS-based sensor
0,i = L0 + nw,i w , the follow-
Letting vi = ui w + vi and L localization with unknown transmit power, in Proc. IEEE ICASSP,
ing problem formulation is obtained: Apr. 2011, pp. 24802483.
[15] H. Chen, G. Wang, Z. Wang, H. C. So, and H. V. Poor, Non-line-of-sight
node localization based on semi-definite programming in wireless sensor
0,i + 10 log10 x si  + vi ,
Li = L i = 1, . . . , N. (28) networks, IEEE Trans. Wireless Commun., vol. 11, no. 1, pp. 108116,
d0 Jan. 2012.
[16] G. Wang, H. Chen, Y. Li, and M. Jin, On received-signal-strength based
localization with unknown transmit power and path loss exponent, IEEE
In [13], it was shown that (28) can be expressed in the similar Wireless Commun. Lett., vol. 1, no. 5, pp. 536539, Oct. 2012.
form as (1), using some approximations for vi . Applying these [17] S. Tomic, M. Beko, R. Dinis, and V. Lipovac, RSS-based localization in
wireless sensor networks using SOCP relaxation, in Proc. IEEE SPAWC,
approximations, the implementation of the described algo- Jun. 2013, pp. 749753.
rithms for the noncooperative localization when PT is known [18] R. M. Vaghefi, M. R. Gholami, R. M. Buehrer, and E. G. Strm, Coop-
is straightforward for the case of indoor localization. erative received signal strength-based sensor localization with unknown
transmit powers, IEEE Trans. Signal Process., vol. 61, no. 6, pp. 1389
1403, Mar. 2013.
[19] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation
ACKNOWLEDGMENT Theory. Upper Saddle River, NJ, USA: Prentice-Hall, 1993.
[20] S. Boyd and L. Vandenberghe, Convex Optimization. Cambridge, USA:
The authors would like to thank Dr. J. Lfberg from ETH Cambridge Univ. Press, 2004.
Zrich, Zrich, Switzerland, and Dr. G. Wang from Ningbo [21] M. Grant and S. Boyd, CVX: Matlab Software for Disciplined Con-
vex Programming, ver. 1.21, CVX Res., May Austin, TX, USA, 2010,
University, Ningbo, China, for a constructive discussion about [Online]. Available: http://cvxr.com/cvx
the conic optimization problems. They would also like to thank [22] T. S. Rappaport, Wireless Communications: Principles and Practice.
the associate editor, Prof. G. Mao, and the anonymous review- Upper Saddle River, NJ, USA: Prentice-Hall, 1996.
[23] M. L. Sichitiu and V. Ramadurai, Localization of wireless sensor
ers for their valuable comments and suggestions that improved networks with a mobile beacon, in Proc. IEEE MASS, Oct. 2004,
the quality of this paper. pp. 174183.
M. Beko is a collaborative member of INESC-INOV, In- [24] G. Mao, B. Fidan, and B. D. O. Anderson, Wireless sensor network
localization techniques, Comput. Netw., vol. 51, no. 10, pp. 25292553,
stituto Superior Tcnico, Universidade de Lisboa, Lisbon, Jul. 2007.
Portugal. [25] B. Bjar and S. Zazo, A practical approach for outdoor distributed tar-
get localization in wireless sensor networks, EURASIP J. Adv. Signal
Process., vol. 2012, no. 1, p. 95, May 2012.
[26] N. Patwari, A. Hero III, M. Perkins, N. S. Correal, and R. J. ODea,
R EFERENCES
Relative location estimation in wireless sensor networks, IEEE Trans.
[1] G. Destino, Positioning in wireless networks: Noncooperative and coop- Signal Process., vol. 51, no. 8, pp. 21372148, Aug. 2003.
erative algorithms, Ph.D. dissertation, Univ. Oulu, Oulu, Finland, 2012. [27] G. D. Durgin, SpaceTime Wireless Channels. Englewood Cliffs, NJ,
[2] N. Patwari, Location estimation in sensor networks, Ph.D. disseration, USA: Prentice-Hall, 2002.
Univ. Michigan, Ann Arbor, MI, USA, 2005. [28] N. Patwari, R. J. ODea, and Y. Wang, Relative location in wireless
[3] Z. Sahinoglu, S. Gezici, and I. Gvenc, Ultra-Wideband Positioning Sys- networks, in Proc. IEEE VTC, May 2001, pp. 11491153.
tems: Theoretical Limits, Ranging Algorithms, Protocols. Cambridge, [29] H. L. Van Trees, Detection, Estimation, Modulation Theory, Part IV:
U.K.: Cambridge Univ. Press, 2011. Optimum Array Processing. Hoboken, NJ, USA: Wiley, 2002.
[4] N. Bulusu, J. Heidemann, and D. Estrin, GPS-less low cost outdoor [30] P. Oguz-Ekim, J. Gomes, J. Xavier, and P. Oliveira, A convex relaxation
localization for very small devices, IEEE Pers. Commun. Mag., vol. 7, for approximate maximum-likelihood 2D source localization from range
no. 5, pp. 2834, Oct. 2000. measurements, in Proc. IEEE ICASSP, Dallas, TX, USA, Mar. 2010.
[5] N. Patwari et al., Locating the nodes: Cooperative localization in wireless [31] A. Beck, P. Stoica, and J. Li, Exact and approximate solutions of source
sensor networks, IEEE Signal Process. Mag., vol. 22, no. 4, pp. 5469, localization problems, IEEE Trans. Signal Process., vol. 56, no. 5,
Jul. 2005. pp. 17701778, May 2008.
[6] A. Bahillo et al., Hybrid RSS-RTT localization scheme for indoor [32] I. Plik and T. Terlaky, Interior point methods for nonlinear optimiza-
wireless networks, EURASIP J. Adv. Signal Process., vol. 2010, no. 1, tion, in Nonlinear Optimization, G. Di Pillo and F. Schoen, Eds., 1st ed.
p. 126 082, Mar. 2010. Berlin, Germany: Springer-Verlag, 2010.
[7] N. Alam and A. G. Dempster, Cooperative positioning for vehicular [33] J. F. Sturm, Implementation of interior point methods for mixed semidef-
networks: Facts and future, IEEE Trans. Intell. Transp. Syst., vol. 14, inite and second order cone optimization problems, Optim. Meth. Softw.,
no. 9, pp. 17081717, Dec. 2013. vol. 17, no. 6, pp. 11051154, Aug. 2002.
[8] T. Gdeke et al., A bi-modal ad-hoc localization scheme for wireless [34] M. S. Lobo, L. Vandenberghe, S. Boyd, and H. Lebret, Applications
networks based on RSS and ToF fusion, in Proc. 10th WPNC, Mar. 2013, of second-order cone programming, Linear Algebra Appl., vol. 284,
pp. 16. no. 13, pp. 193228, Nov. 1998.
2050 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 64, NO. 5, MAY 2015

[35] J. Lofberg, Dualize it: Software for automatic primal and dual conver- Rui Dinis (S96M00) received the Ph.D. degree
sions of conic programs, Optim. Meth. Softw., vol. 24, no. 3, pp. 313325, from the Instituto Superior Tcnico (IST), Techni-
Jun. 2009. cal University of Lisbon, Lisbon, Portugal, in 2001
[36] J. F. Sturm, Using SeDuMi 1.02, a MATLAB toolbox for optimization and the Habilitation degree in telecommunications
over symmetric cones, Optim. Meth. Softw., vol. 11, no. 14, pp. 625 from the Faculdade de Cincias e Tecnologia da
653, Jan. 1998. Universidade Nova de Lisboa (FCT-UNL), Caparica,
[37] C. Fortin and H. Wolkowicz, The trust region subproblem and semidef- Portugal, in 2010.
inite programming, Optim. Methods Softw., vol. 19, no. 1, pp. 4167, From 2001 to 2008, he was a Professor with IST.
Feb. 2004. He is currently an Associate Professor with FCT-
[38] N. Salman, M. Ghogho, and A. H. Kemp, On the joint estimation of UNL. In 2003, he was an Invited Professor with
the RSS-based location and path-loss exponent, IEEE Wireless Commun. Carleton University, Ottawa, ON, Canada. He was a
Lett., vol. 1, no. 1, pp. 3437, Feb. 2012. Researcher with Centro de Anlise e Processamento de Sinal, IST, from 1992
to 2005 and with Instituto de Sistemas e Robtica from 2005 to 2008. Since
2009, he has been a Researcher with Instituto de Telecomunicaes. He has
Slavisa Tomic received the M.S. degree in traf- been actively involved in several national and international research projects
fic engineering according to the postal traffic and in the broadband wireless communications area. His research interests include
telecommunications study program from the Univer- modulation, equalization, channel estimation, and synchronization.
sity of Novi Sad, Novi Sad, Serbia, in 2010. He is Dr. Dinis is an Editor of the IEEE T RANSACTIONS ON C OMMUNICATIONS
currently working toward the Ph.D. degree with the (Transmission SystemsFrequency-Domain Processing and Equalization) and
Department of Electrical and Computer Engineering, a Guest Editor for the Elsevier journal Physical Communication (Special Issue
Universidade Nova de Lisboa, Caparica, Portugal. on Broadband Single-Carrier Transmission Techniques).
His research interests include target localization in
wireless sensor networks and convex optimization.

Marko Beko was born in Belgrade, Serbia, on

November 11, 1977. He received the Dipl. Eng. de-
gree from the University of Belgrade, in 2001 and the
Ph.D. degree in electrical and computer engineering
from Instituto Superior Tcnico, Lisbon, Portugal,
in 2008.
He is currently an Assistant Professor with the
Universidade Lusfona de Humanidades e Tec-
nologias, Lisbon. He is also a Researcher with
the UNINOVA, Campus da FCT/UNL, Monte de
Caparica, Portugal. His current research interests
include signal processing for wireless communications and nonsmooth and
convex optimization.
Dr. Beko received the 2008 IBM Portugal Scientific Award.