You are on page 1of 14

Information Fusion 15 (2014) 518

Contents lists available at ScienceDirect

Information Fusion
journal homepage: www.elsevier.com/locate/inffus

A novel distance estimation approach for 3D localization in wireless


sensor network using multi dimensional scaling
Vijay K. Chaurasiya , Neeraj Jain, G.C. Nandi
Indian Institute of Information Technology, Allahabad, U.P. 211012, India

a r t i c l e i n f o a b s t r a c t

Article history: Node localization is very important in Wireless Sensor Network (WSN) and distance estimation between
Received 26 April 2010 pairs of nodes is the prerequisite for localization and thus the applicability of the reported events. The
Received in revised form 1 June 2013 paper proposes a novel distance estimation algorithm to estimate distances of each node to every other
Accepted 4 June 2013
node in the network. The main contribution of the paper is the denition of a dissimilarity matrix repre-
Available online 27 June 2013
senting the distance of each node to every other node in the network. MDS based localization algorithm is
used to determine coordinates of the node in a local coordinate system and Helmert Transformation is
Keywords:
used to convert the local coordinates of the node into a global coordinate system. The effect of various
Localization
WSN
parameters affecting the performance of proposed algorithm is also presented in the paper. Finally, the
MDS efciency of the proposed algorithm is established through the simulation results.
3D space 2013 Elsevier B.V. All rights reserved.
Distance matrix

1. Introduction nding local coordinates of a moving target with respect to the


sensors deployed for target tracking.
Localization refers to a process of nding location information Distance between the nodes is necessary for most of the locali-
of the objects in a given coordinate system. In order to gather more zation algorithms. One of the characteristic of radio communica-
information for a specic region of deployment, there is a pertinent tion [57] is the potential to partially determine the distance
need of the awareness regarding the location of the node in the between the sender and the receiver. Based on the distance estima-
concerned area. Many issues exist with the localization of the tion technique used, localization algorithms can be classied as
nodes in WSN, such as deployment in the three dimensional space Range Based Localization and Range Free Localization.
without any infrastructural support. Since sensor nodes are ran- The Range Based Localization means that the distance between
domly deployed in 3D space, manually embedding of location the nodes is estimated by using physical properties of communica-
information in each node is not feasible as in many applications, tion signal. The physical properties which are used to estimate the
nodes get deployed by the use of aircrafts. WSNs may have mobile distance between the two nodes are Received Signal Strength Indi-
nodes, which may move from one point to another in the area of cator (RSSI) [7], Time of Arrival (ToA) [5], Time Difference of Arrival
deployment, making manual updating of nodes location informa- (TDoA) [6] and Angle of Arrival (AoA) [8]. Estimation of location of
tion not viable. It is also a prominent issue that new nodes may join the nodes using the above listed techniques is known as range
or old nodes may be eliminated from the existing deployment based localization. Range Free Localization means that an indirect
making the manual localization of nodes difcult. Global Position- technique is used to determine the distance between the nodes
ing System (GPS) receivers can be used by the nodes to determine [9]. Hop count is an indirect technique of distance estimation in
their location in the area of deployment, but putting GPS receiver wireless sensor network which is also commonly used. In this case,
on each node is not feasible due to increased cost. Therefore we every node tries to determine shortest path to reach all other nodes
need an automated, cost effective and efcient localization algo- in the WSN. The weight assigned to the link is the number of hops
rithm for WSNs. Localization approaches in WSN can be classied from source to destination. Hop count can be multiplied by the
as node localization [13] which refers to the process of nding average transmission range of the node to convert into distance
coordinates of sensor nodes with respect to the given coordinate estimation.
system and target localization [4] which refers to the process of Going forward anchor nodes are the nodes which know their
location information. Based on the use anchor nodes, localization
algorithms can be classied into Anchor Based Localization and
Corresponding author. Tel.: +91 9415648188. Anchor Free Localization. The localization algorithm is called
E-mail address: vijay.chaurasiya@gmail.com (V.K. Chaurasiya). Anchor Based Localization when anchor nodes assists localization

1566-2535/$ - see front matter 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.inffus.2013.06.003
6 V.K. Chaurasiya et al. / Information Fusion 15 (2014) 518

algorithm [10,11] to determine the location of other nodes in the localization algorithm for locating sensor nodes in the 2D space.
network. Anchor Free Localization if anchor nodes are not used The conventional centroid localization has been further enhanced
in the process of localization [12]. by Chen et al. [18], they have proposed centroid theorem of coor-
Based on the computation involved in the localization algo- dinate tetrahedron in the volume-coordinate system to enhance
rithm, it can be classied as Centralized Localization Algorithm localization accuracy for the application of algorithm in 3D space
and Distributed Localization Algorithm. If a central computing unit of WSN. Both the algorithms proposed uses a large number of an-
gets all the information from the network and computes the loca- chor nodes and their localization accuracy is inversely proportional
tion information of all the nodes then it is known as centralized to the number of anchor nodes used.
localization algorithm [13]. This approach increases communica- Ji and Zha [14] have proposed the use of MDS for localization of
tion overhead but reduces computation overhead of individual nodes in WSN. They have considered the issue of anisotropic topol-
nodes. In Distributed Localization Algorithm, every node gathers ogy and complex terrain of sensor networks. They have proposed a
information from its neighbors and computes its location locally distributed localization algorithm based on MDS where every node
[14]. This approach limits the overhead imposed on the network creates a local map of the network and these maps are stitched to-
in terms of communication but increases computational overhead. gether to form a global map of the network topology. Further local
But the increased computational overhead results in depletion of optimization is used to reduce localization error. They are running
battery power of individual nodes. a distributed and iterative algorithm on every node which results
This work addresses the problem of node localization that use in more energy consumption during the localization phase. Shang
range based approach to compute distance between the nodes. An- et al. [19] proposed MDS MAP algorithm which uses classical
chor nodes are used in the algorithm to convert local coordinates of MDS to generate the map nodes used in WSN which was further
the nodes into global coordinates. Centralized computation is used extended to MDS-MAP(P) [20] and MDS-MAP(R). MDS-MAP(P)
to determine the location of the individual nodes because localiza- [21] is a distributed algorithm that uses patches of relative map
tion algorithms are computation intensive and nodes have limited of individual nodes which are combined together to determine
computation facility available on board. By adopting centralized ap- the original positions of the nodes. All these algorithms are distrib-
proach, proposed algorithm is reducing computation overhead of uted algorithms where individual nodes are weighed down for
individual nodes. Therefore the paper proposes a centralized local- location computation.
ization algorithm which is based on MDS. MDS is a statistical tech- The accuracy of MDS based localization depends upon the ef-
nique for arranging the data items in an n-dimensional space based ciency of distance estimation between all pair of nodes. Since sen-
on dissimilarities between the data items. Many versions of MDS sor nodes are having limited communication range, it is not
[15] are available which are appropriate for different requirements possible to estimate distance between all the pair of nodes by using
of data analysis. Metric and non-metric MDS are two forms of MDS radio resources. To estimate the distance between all pair of nodes
algorithm. MDS can further be classied as Classical MDS and an indirect approach is required which can produce accurate dis-
Weighted MDS. Classical MDS uses dissimilarity matrix as input tance information. This paper proposes a novel approach for dis-
whereas Weighted MDS uses different weights in the dissimilarity tance estimation between all pair of nodes in the network.
matrix for every dimension. The choice of MDS for data analysis de-
pends upon the dimensionality of the space in which data items
need to be arranged and the accuracy in the measurement of dis- 3. Localization process
similarity between the data items. Since the paper proposes the
use of Euclidian distance as dissimilarity parameter, classical MDS This section describes the particulars of the proposed algorithm.
can be used in our proposed algorithm. Other signicant reasons Consider the scenario of a large set of sensor nodes deployed in an
of using classical MDS is that our proposed distance estimation area that requires continuous monitoring. It is assumed that sensor
algorithm is capable of generating fairly accurate distance matrix nodes are deployed in a 3D space and nodes are static. It is well
and classical MDS can be efciently applied on large matrix because known that network is having two types of node, i.e. normal nodes
it is a closed-form solution of data analysis. Since sensor nodes are and anchor nodes.
having limited communication and battery power, these algorithms Normal nodes are homogeneous, i.e. they have identical sens-
are not useful particularly for localization of nodes in WSN. ing, processing and communication capabilities and the same ini-
tial battery power. Anchor nodes are equipped with GPS receiver,
where links are bidirectional and broadcast type and network is
2. Related work adopting a at topology.
The localization algorithm is divided into following steps:
Many localization algorithms had been proposed to locate
nodes in WSN and describe use of MDS for 2D localization with  Distance estimation: the algorithm estimates distance
the help of few anchor nodes. Pei et al. [16] proposed an algorithm between all pair of nodes in the WSN.
in which anchor nodes are dividing whole area into regions with  Location estimation: the algorithm estimates local coordinate
their regional ids and locate position of nodes accordingly. It also of the nodes in the network using MDS.
applies an optimization technique to reduce localization error.  Coordinate transformation: this step uses the local coordi-
The limitation of the algorithm is that it requires direct communi- nates of the nodes and transforms them into a global coordi-
cation among the anchor nodes. Yu and Wang [17] proposed a nate system.
Hierarchical MDS (HMDS) based on 2D localization algorithm
which has three steps named as creation of cluster, intra-cluster After completion of these three steps, every node knows its
localization and merging of local coordinates achieved in intra- location with respect to global coordinate system dened by the
cluster localization phase. The main drawback of this algorithm anchor nodes in 3D space. To facilitate localization process, every
is that if there are many disjoint clusters exist in the system then node needs to determine its distance from its neighbors and send
it will be impossible to map local coordinate system into global this information to a central server. The central server is responsi-
coordinate system. The error in the shortest path based distance ble to run distance estimation process, MDS to determine local
estimation is high which leads to high localization error. Bulusu coordinate of individual nodes and synchronize transformation
et al. [12] have proposed anchor based conventional centroid process to map local coordinates into global coordinate. Finally,
V.K. Chaurasiya et al. / Information Fusion 15 (2014) 518 7

server sends the estimated global coordinate of nodes back to the row jth column in the matrix D(X) (it means that the
individual nodes in the network. distance between node i and node j is not known).
10  The algorithm nds all the set of 3 nodes which are
4. Elaboration of proposed localization process common neighbors of both i and j and put it in a set CN.
11  Dij = 0; C = 0;
This section presents the detail of each of the three steps in- 12  WhileCN! ;
volved in the proposed localization process. The input to MDS algo- 13 repeat
rithm is a dissimilarity matrix representing the distance between  C++;
every two nodes present in the network. Suppose there are n sen-  If (3 common nodes are not in a line)
sor nodes in the network deployed in a 3D space. The dissimilarity 14  If (4 nodes are not in a plane)
between the points is represented by the distance between the  Dene a plane by using 3 common nodes in a
points. The distance between point i and j can be represented as local coordinate system in 3D.
dij then the distance between all the points in the network can  Estimate the coordinate of node i and node j
be represented by a n  n matrix D(X) as using trilateration.
2 3  Estimate the Euclidian distance between node i
0 d12 d13 . . . d1n
6 7 and node j using their estimated coordinate.
d
6 21 0 d23 . . . d2n 7 15  Set: Dij + = Dij;
6 7
6 . . . d3n 7
DX 6 d31 d32 0 7 1 16 end
6 . .. .. .. .. 7
6 . 7 17 //Put estimated distance in matrix D(X)
4 . . . . . 5 D
18  Dij Cij ;
dn1 dn2 dn3 ... 0
19  put Dij in D(X)b for corresponding i and j.
Therefore to apply MDS the algorithm needs distance information 20 end
for every pair of node present in WSN. 21 //after this loop there may be some i, j for which D(X)i,j = -
1.
4.1. Distance estimation algorithm 22 //to estimate the distance between these nodes following
steps are carried //out by the algorithm.
In wireless sensor network, distances between the two nodes 23 Scan each row of distance matrix D(X) and if it nds value
can be estimated if the they are neighbors. Two nodes are said to 1 for an element present at ith row jth column in the
be neighbor of each other if the distances between the nodes is less matrix D(X) (it means that the distance between node i
than or equal to the communication range. For example suppose d and node j is still not known).
is the distance between the two nodes and r is their communica- 24 Find all node k such that D(X)i,k and D(X)k,j is known.
tion range then they are the neighbors of each other if d 6 r. The 25 Set: p = {total number of common node k}
distance between the two neighboring nodes in wireless commu- 26 For all k compute dl = D(X)i,k + D(X)k,j; for l = 1, 2, . . . , p.
nication can be estimated using RSSI, ToA, TDoA or AoA techniques. 27 Set: D(X)i,j = min(dl)
The issue in distance estimation in WSN is that how to determine 28 Put: D(X)i,j into D(X)
distances between all pair of nodes which are not neighbors. RSSI is 29 end
a technique which may produce relatively less accurate distance
estimation among the neighbors but does not need additional
hardware or software, thus least cumbersome. In this paper, the The algorithm is able to estimate distance between each pair of
use of RSSI is proposed to produce distance estimation between nodes in the network. In the rst iteration, the algorithm estimates
the neighbors. the distance between any two nodes i and j by selecting three
The distance estimation algorithm runs at the central server nodes common to both i and j and try to estimate the distance be-
with sufciently high computational power and memory. It uses tween them by applying trilateration. After completion of these
mere connectivity information provided by the nodes to compute iterations, there may be some pair of nodes whose distance is still
distance between every pair of nodes in WSN. Trilateration [22] unknown. To estimate the distances for each such pair, algorithm
is applied on the neighbors distance information which is gener- nds the set of common nodes k, between i and j, such that the dis-
ated in the rst step to estimate the distance between the each pair tance between the pair i, k and k, j exists in distance matrix D(X).
of nodes in the network. The algorithm is using a matrix represent- Now for every node k, the algorithm estimates the distance be-
ing distance information of the neighbors and 1 for the other tween node i and j, by adding the distance between pairs i, k and
pairs. The algorithm has following steps: k, j (i.e. D(X)i,k + D(X)k,j. In the next step, the algorithm determines
the minimum distance estimated between nodes i and j. Finally,
1 // Initialization
the algorithm puts the minimum distance determined between
2  Every node estimates distance to its neighbors and
nodes i and j into the distance matrix D(X). This process ensures
send it to the central server.
the availability of distance information for all the pairs of nodes
3  Central sever puts all the distances received by the
i, j exists in the WSN.
nodes in the matrix D(X) and marks all unknown
The above algorithm tries to nd three common neighbors be-
distances by 1.
tween each pair of nodes where distance is still unknown and uses
4 start
these common nodes to nds distance in between the two. Using
5  Set: D(X)a = 0; D(X)b = D(X);
each group of three common nodes the algorithm estimates the
6  while(DXa ! DXb
distances between two nodes (whose distance is still unknown)
7 repeat and nally takes the average of these estimated distances to
8  D(X)a = D(X)b; represent the nal distance in the distance matrix. Since we are
9  The algorithm scans each row of distance matrix using RSSI approach to nd distance between the neighbors,
D(X) and if it nds value 1 for an element present at ith including all groups of three common nodes in distance estimation
8 V.K. Chaurasiya et al. / Information Fusion 15 (2014) 518

process will result in larger distance estimation errors. Therefore to


limit distance estimation error in the algorithm, it is considering
only those three common nodes which are nearest and whose dis-
tance needs to be estimated. Finally the average of these estimated
distances are taken to nd the nal distance between two nodes.

4.1.1. Elaboration of distance estimation algorithm


Suppose node A and node E are not neighbors of each other,
therefore the distance between the two nodes cannot be deter-
mined based on the communication parameters. Suppose node A
is the ith node in the network and node E is the jth node in the net-
work then the distance matrix D(X) have the value 1 at D(X)i,j po-
sition and D(X)j,i position. This means that the distance between
node A and node E is not known as shown in Fig. 1.
Suppose nodes B, C and are neighbors of both A and E as shown
in Fig. 1.
Nodes A, B, C and D form a tetrahedron and nodes E, B, C and D
Fig. 3. Tetrahedron formed by node A, B, C and D.
form another tetrahedron as shown in Fig. 2. To estimate the dis-
tance between node A and node E, following steps are carried
out by the distance estimation algorithm. q
Area of triangle DBCDA ss  dbc s  dcd s  ddb ; 2

4.1.1.1. If (3 common nodes are not in a line). To test whether the 3 where
common nodes B, C and D are nonlinear, the algorithm calculate
dbc dcd ddb
the area of the triangle formed by the nodes B, C and D and checks s 3
whether the area is greater than zero or not. If the area of triangle 2
formed by the nodes B, C and D is greater than zero, it means points Since all the distances are known, the area of base triangle can be
are not linear. Fig. 3 shows the tetrahedron formed by nodes A, B, C estimated. If A > 0 proceed to next step.
and D where nodes B, C, D are the base of tetrahedron.
The area of base triangle can be obtained by using Herons for- 4.1.1.2. If (4 nodes are not in a plane). To test whether the 4 nodes
mula as follows: B, C, D and A are nonplaner, the algorithm considers a tetrahedron
form by nodes B, C, D and A as shown in Fig. 3. The algorithm cal-
culates the volume of the tetrahedron formed by the nodes
B, C, D and A and checks whether the volume is greater than zero
or not. If the volume of the tetrahedron formed by the nodes
B, C, D and A is greater than zero, it means points are not planer
and form a rigid tetrahedron in 3D space.
Since the distances between the vertices of a tetrahedron are
known, the volume of the tetrahedron can be computed by using
the CayleyMenger determinant [23] as follows:
 
0 1 1 1 1 

 2 2 2 
1 0 dab dac dad 
 
2  2 2
dbd 
2
288  V  1 dab 0 dbc 4
 
 1 d2 2
dbc 0 dcd 
2
 ac
 
 1 d2 2
dbd dcd
2
0 
ad

Fig. 1. Organization of node A, B, C, D and E in 3D space. or


v
 
u 0 1
u  1 1 1 
u  2 
u  1 0 dab dac dad 
2 2
u  
u
udet  1 d2 0 dbc dbd 
2 2
u  ab
 5
u  1 d2 d2 0 dcd 
u 2
u  ac bc
u  
t  1 d2 d2 d2 0 
ad bd cd
V
288
Since all the distances are known, the volume of tetrahedron
formed by node B, C, D and A can be estimated. If V > 0 proceed to
next step.

4.1.1.3. Dene a plane by using 3 common nodes in a local coordinate


system in 3D. The coordinates of common nodes B, C and D can be
dened in a plane by choosing node C as the origin of coordinate
system, where D is on positive X-axis and B is in XY plane as
Fig. 2. The two tetrahedrons imposed by base DBCD. shown in Fig. 4.
V.K. Chaurasiya et al. / Information Fusion 15 (2014) 518 9

The equation of rst sphere with center as node C and radius dac
can be written as
2
X 21 Y 21 Z 21 dac 10
The equation of second sphere with center as node D and radius
dad can be written as
2 2
X 1  dcd Y 21 Z 21 dad 11

Fig. 4. Coordinate of node B, C and D in a plane. The equation of third sphere with center as node B and radius
dab can be written as
2 2 2
X 1  dbc cos h Y 1  dbc sin h Z 21 dab 12
Since the distance between all the three nodes are known as
shown in the Fig. 4 by dbc, dcd, and dbd, the angle h can be estimated By solving three Eqs. (10)(12) for three unknowns X1, X2 and X3
by using the following Trigonometric equation gives following solution
! 2 2 2
2 2
dbc dcd  dbd
2 dac  dad dcd
h cos1 6 X1 13
2dbc dcd 2dcd

2 2 2 2
If h is known then the coordinate of nodes in 3D can be dened as dac  dab dbc cos h dbc sin h
Y1  dbc  cot h
follows. For simplicity and without loss of generality, it is assumed 2  dbc  sin h
!
that the origin be lie with node C, therefore the coordinate of C 2 2 2
dac  dad dcd
become:  14
2dcd
C 0; 0; 0 7
q
2
Node D is assumed to exist on positive X-axis, therefore the coordi- Z 1  dac  X 21  Y 21 15
nate of node D becomes
Similarly the coordinates of node E can be estimated as
D dcd ; 0; 0 8 2 2 2
dec  ded dcd
X2 16
Node D is assumed to exist on XY plane, therefore the coordinate of 2dcd
node B becomes
2 2 2 2
dec  deb dbc cos h dbc sin h
B dbc cos h; dbc sin h; 0 9 Y2  dbc  cot h
2  dbc  sin h
!
2 2 2
dec  ded dcd
 Estimate the coordinate of node A and node E.  17
2dcd

To estimate the coordinates of node A and node E, the algorithm q


2
is using trilateration method [23]. Z 2  dec  X 22  Y 22 18
Fig. 5 is showing the nodes in 3D space. To estimate the coordi-
The trilateration method produces two values for Z axis. One is
nates of node A, trilateration method says write the equation of
positive and the other one is negative. Since the plane by nodes B, C
three spheres with their center as nodes B, C and D with radius
and D is dened, such that the two nodes lie on the either side of
dab dac and dad respectively.
the node, for one node the algorithm is taking positive Z value of
the estimated Z coordinate and for another node it is taking the
negative value of Z coordinate. Therefore the coordinate of node
A can be taken as (X1, Y1, Z1) and coordinate of node E can be taken
as (X2, Y2, Z2).
The Euclidian distance between node A and node E with their
estimated coordinates (X1, Y1, Z1) and (X2, Y2, Z2) can be obtained
by the following equation
q
dae X 1  X 2 2 Y 1  Y 2 2 Z 1 Z 2 2 19

The process of distance estimation is repeated for 3 sets of com-


mon nodes that exists between node A and node E, which has their
distance information from A and E and also in between them. The
average of the estimated distances from each set of the three com-
mon nodes are calculated and inserted in the distance matrix D(X).

4.2. Location estimation

MDS is used in the algorithm to estimate the local coordinate of


nodes in WSN. When MDS is applied on a data set, a MDS-MAP is
produced. It is a map in n-dimensional space where the data items
Fig. 5. Nodes A, B, C, D and E in 3D space with their Cartesian coordinates. are placed inside the n-dimensional space. The choice of dissimilar-
10 V.K. Chaurasiya et al. / Information Fusion 15 (2014) 518

 
ity is important in MDS. For localization in WSN the dissimilarity is C X 211 X 222 X 233 . . . X 2nn 26
taken in the Euclidian distance between the nodes.
To use MDS for 3D localization in WSN, suppose the dissimilar- Then Eq. (25) can be rewritten as
ity between the nodes in WSN is represented by the distance D2 X CI0 IC 0 2  XX 0 27
between the nodes in 3D then the distance between any pair of
0
node i and j can be represented as where XX is known as scalar product matrix.
q Let the scalar product matrix XX0 = B. The aim of MDS is to
dij X i1  X j1 2 X i2  X j2 2 X i3  X j3 2
2
20 determine the scalar product matrix B based on the given matrix
of squared distances D2(X). To determine B, let us multiply
where (Xi1, Xi2, Xi3) is the Cartesian coordinate of node i and Eq. (27) by J (a centering matrix) and another factor of  12 in both
(Xj1, Xj2, Xj3) is the Cartesian coordinate of node j in 3D space. sides where J = I  n1110 . After multiplication, Eq. (27) become
Now the Eq. (20) can be rewritten as
1 1
X
3  J  D2  J  JCI0 IC 0 2  XX 0 J 28
2 2 2
dij X ia  X ja 2 21
a1
1 1
 JC00  0C 0 J JBJ
or 2 2
B 29
2
X
3
dij X 2ia X 2ja  2X ia X ja 22
By applying Eigen value decomposition on the matrix B we get
a1

Suppose D2 X represents the matrix of dij , i.e. the square of the


2 B Q KQ 0 30
distances between every pair of nodes then for n points the above where Q is a square matrix and its columns are representing Eigen
equation can be written as vectors and K is a diagonal matrix representing Eigen values. For
2 2 2 2 2
3 three dimensional solutions three positive Eigen values are re-
d11 d12 d13 . . . d1n
6 2 7 quired. Suppose K+ represents the three positive Eigen values and
6d 2
d22
2
d23 . . . d2n 7
2
6 21 7 Q+ represents the corresponding Eigen vectors, then the coordinates
6 2 2 2 2 7
D2 X 6
6 d31 d32 d33 . . . d3n 7
7 23 of nodes in MDS can be obtained in the form of a matrix represented
6 . . 7 by X0 as follows:
6 .
4 . ... ... ..
. .
. 5
7
2 2 2 2 X 0 K  Q 31
dn1 dn2 dn3 . . . dnn
2 3
From Eqs. (20) and (21), D2(X) can be rewritten as X 11 X 12 X 13
6 X 21 X 22 X 23 7
2 3 6 7
X 21a X 21a X 21a . . . X 21a where X 0 6 .. .. .. 7
4 . . . 5
6 2 7
6 X 2a X 22a X 22a . . . X 22a 7 X n1 X n2 X n3
3 6
X 6
7
7
2
D X 6
2
6 X 3a X 23a X 23a . . . X 23a 7
7 is the matrix representing coordinates of the nodes in local
a1 6 . .. .. .. .. 7 coordinate.
6 . 7
4 . . . . . 5 System dened by MDS. The matrix X0 obtained above gives the
X 2na X 2na X 2na . . . X 2na coordinates of all the nodes in the three dimensional space. Since
2 2 3 the coordinates obtained using MDS are the coordinates in a local
X 1a X 22a X 23a . . . X 2na
6 2 7 coordinate system, the MDS-MAP generated by MDS is the scaled
6 X 1a X 22a X 23a . . . X 2na 7 map of the nodes. So to get the global coordinates of the nodes
X6 2
3 6 7
7 coordinate transformation and scaling is required to convert the
6 X 1a X 22a X 23a . . . X 2na 7
6 7 local map into global map.
a1 6 . .. .. .. .. 7
6 . 7
4 . . . . . 5
4.3. Coordinate transformation
X 21a X 22a X 23a . . . X 2na
2 3 To map the local coordinates into a global coordinate, a coordi-
X 1a X 1a X 1a X 2a X 1a X 3a . . . X 1a X na
6X X nate transformation technique is required. The localization algo-
6 2a 1a X 2a X 2a X 2a X 3a . . . X 2a X na 7
7
X3 6 7 rithm uses a seven parameter Helmert Transformation [24] to
6 X 3a X 1a X 3a X 2a X 3a X 3a . . . X 3a X na 7
2 6 7 24 transform local coordinate of the nodes into a global coordinate.
6 .
a1 6 .. .. .. .. 7 7 The seven parameters used by Helmert Transformation are as
4 .. . . . . 5
follows:
X na X 1a X na X 2a X na X 3a . . . X na X na
or  Three rotation parameter to rotate the local coordinate sys-
tem around the coordinates axes to align axis of local coor-
X
3
dinate system with the global coordinate system.
D2 X CI0 IC 0 2  X a X 0a 25
 Three translation parameters to move the points along three
a1
coordinate axes to map the origins of the two coordinate
2 3 systems.
X 11 X 12 X 13
6X  One scaling parameter to scale the local coordinate of the
6 21 X 22 X 23 7
7 nodes to represent their global coordinates.
Suppose X 6
6 .. .. .. 77
4 . . . 5
To determine transformation parameters, the algorithm is using
X n1 X n2 X n3
the approach by Horn et al. [25]. The process proposed by them to
where Xa is the ath column of matrix X and determine transformation parameters is as follows:
V.K. Chaurasiya et al. / Information Fusion 15 (2014) 518 11

Let Xloc,i is the local coordinate of node i obtained from multidi- Eigen vectors obtained from Eigen-decomposition of RTm  Rm , then
mensional scaling and Xglo,i is the global coordinate of node i. To the rotation matrix R can be obtained by following equation:
transform local coordinates into global coordinates, Helmert Trans- s
 
formation require at minimum four anchor nodes with known po- 1 1 1
R Rm  p t1 t1 p t2 t2 p t3 t3
T T T
40
sition. The accuracy of the transformation increases with the e1 e2 e3
increase in the number of anchor nodes [26]. Suppose there are n
The scaling factor can be computed as follows:
nodes in the network and let Xloc,i and Xglo,i represents the local
v
and global coordinate of the nodes respectively. uPn
u i1 kX 0glo;i k2
2
X 11 X 12 X 13
3 s t Pn 0 2
41
i1 kX loc;i k
6X X 22 X 23 7
6 21 7
X loc 6
6 .. .. .. 77 32 Since R and s are estimated, t can be estimated by the following
4 . . . 5 equation:
X n1 X n2 X n3
t X 0glo  sRX 0loc 42
and
For each point X in the local coordinate system, corresponding X0 is
2 3
X 11 X 12 X 13 computed in the global coordinate system using following
6X X 22 X 23 7 equation:
6 21 7
X glo 6
6 .. .. .. 77 33
4 . . . 5 X 0 sRX t 43
X n1 X n2 X n3 where s is the scaling parameter, R is the rotation matrix of 3  3
and t is the transformation matrix 3  1.
where Xloc,i = (Xi1, Xi2, Xi3) are the local coordinates of node i and
Xglo,i = (Xi1, Xi2, Xi3) is the global coordinates of node i. According to
5. Simulation results
Helmert Transformation, the coordinate of a node from its local
coordinate system can be transformed into a global coordinate sys-
The proposed localization algorithm is simulated in MATLAB
tem by using following equation:
7.6.0.324(R2008a) [27]. The input parameters required for the sim-
X 0 sRX t 34 ulation are volume of deployment, number of normal and anchor
nodes, communication range and error in distance estimation.
where X is the coordinate of a node in a local coordinate system and
X0 is the coordinate of the node in global coordinate system. To ap-
5.1. Simulation results for 4 anchor node and 15 normal nodes
ply Helmert Transformation, the algorithm needs to determine the
parameters t for translation, s for scaling and R for rotation, which
This section presents result of the simulation carried on a vol-
can be used to transform a point X in the local coordinate system ume of 50  50  5 m3 with 4 anchor node and 15 normal nodes.
to the equivalent point X0 in the global coordinate system. Horn
The communication range of the nodes and error in distance esti-
et al. [25] proposed an approach to determine these parameters. mation is taken as 30 m and 05% respectively. Fig. 6a shows true
According to Horn et al., the transformation parameters can be esti- arrangement of the nodes in a 3D space and Fig. 6b shows the
mated as follows. mapped arrangement of nodes in a local coordinate system ob-
Let X loc and X glo represents the centroid of nodes in local and tained after MDS.
global coordinate system respectively then X loc and X glo can be cal- In Fig. 6b the nodes are arranged in a three dimensional space
culated as according to their local coordinate obtained after MDS. Fig. 7 is
1X n showing the original location and estimated location of the nodes
X loc X loc;i 35 after coordinate transformation.
n i1
In Fig. 7, squares are used to represent estimated position of the
nodes and circles are used to represent true position of the nodes
1X n
in the three dimensional space. Fig. 8 represents the error in local-
X glo X gob;i 36
n i1 ization of each normal node in meters.

To shift all the points with respect to the centroid of their coordi-
nate system, i.e. for points in local coordinate system, shift the
points with respect to the centroid of local coordinate system and
for points in global coordinate system, shift the points with respect
to the centroid of global coordinate system. The new coordinate of
the nodes can be obtained by following equations:

X 0loc;i X loc;i  X loc 37

X 0glo;i X glo;i  X loc 38

The rotation matrix for the transformation can be calculated as


follows. Let Rm is a matrix dened as follows:
X
n
T
Rm X 0glo;i X 0loc;i 39
i1

The Eigen value decomposition of RTm  Rm , produce Eigen values and


Eigen vectors. Let e1, e2, e3 are the Eigen values and t1, t2, t3 the Fig. 6a. True arrangement of nodes in 3D space.
12 V.K. Chaurasiya et al. / Information Fusion 15 (2014) 518

5.2. Simulation 5 anchor node and 100 normal nodes

This section presents result of the simulation carried on a vol-


ume of 100  100  50 m3 with 5 anchor node and 100 normal
nodes. The communication range of the nodes and error in distance
estimation is taken as 30 m and 05% respectively. Fig. 9 shows
the position of the nodes in the 3D space. The true position of
the nodes and their estimated position are represented by circles
and squares respectively.
The results obtained shows good accuracy in localization. The
minimum error in localization is 1.52 m. Only seven nodes are
showing high error in localization which is varying from 22.10 to
11.43 m. High error in distance estimations is the reason for high
localization error of these nodes. The reason of high error in dis-
tance estimation of these nodes is that these nodes are isolated
in the network and have only one or two neighbors only.
Fig. 10 represents the error in localization of each normal node
Fig. 6b. Mapped arrangement of nodes in 3D space after MDS. in meters. In Fig. 9 the minimum error in distance estimation of a
normal node is 1.52 m and maximum error is a 22.1 m.

5.3. Simulation 5 anchor node and 500 normal nodes


In Fig. 8 the minimum error observed in distance estimation of a
normal node is 0.948481 m and maximum error observed is a This section presents result of the simulation carried on a vol-
19.47069 m. ume of 100  100  50 m3 with 5 anchor node and 500 normal
nodes. The communication range of the nodes and error in distance

Fig. 7. True and estimated position after localization for 4 anchor node and 15 normal nodes.

Fig. 8. Localization error of normal nodes.


V.K. Chaurasiya et al. / Information Fusion 15 (2014) 518 13

Fig. 9. True and estimated position after localization for 5 anchor node and 100 normal nodes.

Fig. 10. Localization error of normal nodes.

estimation is taken as 30 m and 05% respectively. Fig. 11 shows localization algorithm with respect to the error occurred in locali-
the position of the nodes in the 3D space. The true position of zation of individual nodes in the network. We have performed 10
the nodes and their estimated position are represented by circles independent simulation runs to come up with an average RSME va-
and squares respectively. lue and computed 99% condence interval for the observed mean.
Fig. 12 represents the error in localization of each normal node This section presents result of the simulation carried on a vol-
in meters. In Fig. 12 the minimum error in distance estimation of a ume of 50  50  10 m3 with variable number of anchor node
normal node is 0.25785 m and maximum error is 31.693 m. and hundred normal nodes. The communication range of the nodes
and error in distance estimation is taken as 25 m and 05% respec-
5.4. Effect of various parameters on the performance of proposed tively. The results shows the average values obtained from 10 inde-
algorithm pendent simulations performed for every case.
Fig. 13 represents average RMSE with condence interval for
It is evident from the simulation results that the proposed local- 99% condence level. Fig. 13 shows how localization error is vary-
ization algorithm is able to localize all the nodes in the network ing with the number of anchor nodes used in the localization pro-
with low error. This section is going to discuss the effect of various cess. It has been inferred from the Fig. 13 that the localization error
parameters on the accuracy of localization algorithm. is decreasing with the increase in the number of anchor nodes. But
We have analyzed the effect of the number of anchor nodes, if the number of anchor nodes goes beyond a maximum value (for
node density, communication range and execution time on the example in Fig. 13 error becomes almost constant if we increase
proposed algorithm. The result obtained from the analysis is pre- the number of anchor nodes beyond 8), the error does not decrease
sented in Figs. 1316. We have used root mean squared error further. The reason of this behavior is that our proposed algorithm
(RSME) to measure the efciency of the algorithm because RSME uses anchor nodes in the process of transforming local coordinates
is a measure of the differences between values predicted by a mod- of the nodes into global coordinates and further Helmert Transfor-
el or an estimator and the values actually observed from the thing mation requires a minimum of four nodes as to transform local
being modeled or estimated. It can quantify the efciency of the coordinated into the coordinate system of reference nodes.
14 V.K. Chaurasiya et al. / Information Fusion 15 (2014) 518

Fig. 11. True and estimated position after localization for 5 anchor node and 500 normal nodes.

Fig. 12. Localization error of normal nodes.

Fig. 13. Average root mean squared error vs. number of anchor nodes for 99% condence interval.
V.K. Chaurasiya et al. / Information Fusion 15 (2014) 518 15

Fig. 14. Average root mean squared error vs. node density for 99% condence interval.

Fig. 15. Average root mean squared error vs. communication range for 99% condence interval.

Fig. 16. Average execution time vs. number of nodes for 99% condence interval.
16 V.K. Chaurasiya et al. / Information Fusion 15 (2014) 518

Fig. 17. Comparision of proposed algorithm with conventional centroid algorithm and novel centroid algorithm for varying number of anchor nodes used in the simulation.

Fig. 18. Comparision of proposed algorithm with conventional centroid algorithm and novel centroid algorithm for varying communication range of the nodes used in the
simulation.

This section presents result of the simulation carried on a vol- This section presents result of the simulation carried on a vol-
ume of 50  50  10 m3 with 5 anchor node and variable number ume of 50  50  10 m3 with 5 anchor node and 100 normal nodes.
of normal nodes. The communication range of the nodes and error The communication range of the nodes and error in distance esti-
in distance estimation is taken as 25 m and 05% respectively. The mation is taken as variable and 05% respectively. The results
results shows the average values obtained from 10 independent shows the average values obtained from 10 independent simula-
simulations performed for every case. tions performed for every case. Fig. 15 presets the summary of
Fig. 14 represents average RMSE with condence interval for the results obtained after simulations of the algorithm for varing
99% condence level. Fig. 14 represents the summary of results ob- communication range of the nodes. Fig. 15 represents average
tained after simulations of the algorithm for different node densi- RMSE with condence interval for 99% condence level.
ties in the network. It can been inferred from Fig. 14 that the It can be inferred from Fig. 15 that the localization error in-
localization error is decreasing with the increase in node density. creases with the increase in communication range of the nodes.
The results obtained indicates that the node density is very impor- This is an obvious result because if the communication range of
tant parameter for the designed algorithm. Node density in the the node increases, their coverage area also increases which in turn
network increases with the increase in number of nodes in the net- increases the number of neighbors of the nodes. Also, increased
work per unit volume. Since error is decreased with the increase in communication range reduces the number of isolated nodes in
node density, sufcient node density should be used to achieve de- the network which limits the localization error of the loosely con-
sired accuracy in the localization process. nected nodes.
V.K. Chaurasiya et al. / Information Fusion 15 (2014) 518 17

This section presents result of the simulation carried on a vol- central server executes localization algorithm by taking a distance
ume of 50  50  10 m3 with 5 anchor node and variable number matrix containing distance information of the node with any set of
of normal nodes. The communication range of the nodes and error four other nodes whose location is already known. We have com-
in distance estimation is taken as 25 m and 05% respectively. The pared the proposed algorithm with Conventional Centroid Algo-
results shows the average values obtained from ten independent rithm and Novel Centroid Algorithm and the results show that
simulations performed for every case. the proposed algorithm outperforms both the algorithm in term
Fig. 16 presets the summary of the results obtained after simu- of localization error.
lations of the algorithm for varying total number of nodes in the
network with respect to the execution time taken by the algorithm.
For the simulation we have used a computer with processing speed 7. Limitations and future work
of 2 GHz, RAM 2 GB, Operating System Windows Vista, Applica-
tion Program Used MATLAB 7.6.0.324(R2008a). It can be inferred The proposed algorithm has some limitations. The rst limita-
from Fig. 16 that the execution time is increasing with the increase tion of the algorithm is that it needs a minimum node density
in number of nodes in the network. The results obtained indicates for successful execution to localize nodes in the network. It has
that the execution time is increasing exponentially with the in- been observed during simulations that the algorithm is unable to
crease in number of nodes in the network but the algorithm is scal- produce result if the node density is below a particular threshold.
able to be applied to a larger network of more than thousand The second one is that the execution time taken by the algorithm
nodes. depends on the number of nodes and increases exponentially with
the increase in the number of nodes in the network. The third is
5.5. Comparison with other algorithm that the algorithm considers that nodes are static in the network,
so it does not support mobile nodes. The above limitations are piv-
The proposed algorithm has been compared with Conventional otal for future research in this area and testing of the proposed
Centroid Algorithm and Novel Centroid Algorithm and following algorithm at various platforms need to discussed.
results have been obtained. The simulation is performed in a region
of 100  100  100 m3 with 100 normal nodes with a communica-
tion range of 25 m. References
Figs. 17 and 18 presents the results obtained after simulations
[1] D.A. Tran, T. Nguyen, Localization in wireless sensor networks based on
performed for the three algorithms for varying anchor nodes and support vector machines, IEEE Transactions on Parallel and Distributed
communication range of the nodes respectively. Fig. 17 shows that Systems 19 (7) (2008) 981994.
our proposed algorithm localizes the nodes with relatively low [2] R. Mautz, W. Ochieng, G. Brodin, 3D wireless network localization from
inconsistent distance observations, Ad Hoc & Sensor Networks 3 (2) (2007)
localization error. Also the performance of algorithm is almost con- 141170.
stant for large anchor nodes. Fig. 18 shows that our localization [3] R. Iyengar, B. Sikdar, Scalable and distributed GPS free positioning for sensor
algorithm performs better than the other proposed algorithm for networks, in: IEEE International Conference on Communications, 2003, ICC03,
pp. 338342.
varying communication range of the nodes. [4] Y. Zou, K. Chakrabarty, Energy-aware target localization in wireless sensor
networks, in: Proceedings of the First IEEE International Conference on
Pervasive Computing and Communications, 2003 (PerCom 2003), 2003, pp.
6. Conclusions and discussions 6067.
[5] L. Girod, D. Estrin, Robust range estimation using acoustic and multimodal
sensing, in: Proceedings 2001 IEEE/RSJ International Conference on Intelligent
The results presented in Section 5 show that the localization Robots and Systems, Expanding the Societal Role of Robotics in the Next
algorithm is able to generate location information of the nodes in Millennium (Cat. No.01CH37180), vol. 3, pp. 13121320.
wireless sensor networks. The algorithm produces Cartesian coor- [6] X. Cheng, A. Thaeler, TPS: A time-based positioning scheme for outdoor
wireless sensor networks, in: 23rd IEEE International Conference on Computer
dinates of the nodes in 3D space. The accuracy of localization is in Communications, 2004, INFOCOM 2004, vol. 00(C), pp. 26852696.
the acceptable range for a large set of monitoring and automation [7] T.S. Rappapport, Wireless Communications: Principles and Practice, second ed.,
applications and the algorithm works if sufcient number of nodes Prentice Hall, New Jersey, 1996. pp. 50143..
[8] D. Moore, J. Leonard, D. Rus, S. Teller, Robust distributed network localization
is present in the network. The algorithm has been tested for differ-
with noisy range measurements, in: Proceedings of the Second ACM
ent node densities. Further it has been also an observation that the Conference on Embedded Networked Sensor Systems (SenSys 04), Baltimore,
algorithm is unable to generate distance matrix if node density MD. November 35, 2004.
falls below a particular threshold. The observations from the anal- [9] T. He, C. Huang, B.M. Blum, J.A. Stankovic, T. Abdelzaher, Range-free
localization schemes for large scale sensor networks, in: Proceedings of the
ysis of simulation results indicates that (i) localization error is 9th Annual International Conference on Mobile Computing and Networking
decreasing with the increase in the number of anchor nodes, (ii) MobiCom 03, 2003, pp. 8195.
localization error is decreasing with the increase in node density, [10] N. Bulusu, J. Heidemann, D. Estrin, Adaptive beacon placement, in: 21st
International Conference on Distributed, Computing Systems (ICDCS-21),
(iii) localization error is decreasing with the increase in communi- 2001, April, pp. 489498.
cation range of the nodes in the network, and (iv) the execution [11] N. Bulusu, J. Heidemann, D. Estrin, U.S.C. Isi, Density adaptive algorithms for
time is increasing exponentially with the increase in number of beacon placement in wireless sensor networks, in: Proceedings of the Sixth
International Symposium on Communication Theory and Applications (ISCTA
nodes in the network. 01), 2001, pp. 114.
The proposed algorithm is using a centralized approach to solve [12] N. Bulusu, J. Heidemann, D. Estrin, GPS-less low-cost outdoor localization for
the localization problem in WSN. The centralized approach reduces very small devices, IEEE Personal Communications 7 (2000) 2834.
[13] Q. Zhang, J. Huang, J. Wang, C. Jin, A new centralized localization algorithm for
the computational overhead on the individual nodes and as a result wireless sensor network, in: Third International Conference on
increases their service time for the particular application. Nodes Communications and Networking, China Com 2008, 2008, 2007, pp. 625629.
need to send their local information to the central server only once, [14] X. Ji, H. Zha, Sensor positioning in wireless ad-hoc sensor networks using
multidimensional scaling, in: Twenty-Third Annual Joint Conference of the
therefore it can reduce communication overhead on the nodes for
IEEE Computer and Communications Societies, INFOCOM 2004, 2004, vol.
localization. The algorithm reduces both computational as well as 00(C), pp. 26522661.
communication overhead, resulting in power saving. Therefore, the [15] V. Isham, N. Keiding, T. Louis, N. Reid, Monographs on statistics and applied
designed localization algorithm is an energy efcient algorithm. probability (1960) 2005.
[16] Z. Pei, Z. Deng, S. Xu, X. Xu, Anchor-free localization method for mobile targets
For newly joined node, localization algorithm needs to be executed in coal mine wireless sensor networks, Sensors 9 (4) (2009) 28362850 (Basel,
at the same central sever to generate location information. The Switzerland).
18 V.K. Chaurasiya et al. / Information Fusion 15 (2014) 518

[17] G. Yu, S. Wang, A hierarchical MDS-based localization algorithm for wireless [22] H. Air, F. Academy, Efcient solution and performance analysis of 3-D position
sensor networks, Journal of Information Technology and Applications 2 (4) estimation by trilateration, IEEE Transactions on Aerospace and Electronic
(2008) 197202. Systems 32 (4) (1996) 12391248.
[18] H. Chen, P. Huang, M. Martins, H.C. So, K. Sezaki, Novel centroid localization [23] D. Michelucci, S. Foufou, Using CayleyMenger determinants for geometric
algorithm for three-dimensional wireless sensor networks, in: 4th constraint solving, in: Ninth ACM Symposium on Solid Modeling and
International Conference on Wireless Communications, Networking and Applications, 2004, pp. 285290.
Mobile, Computing, 2008, pp. 14. [24] G.a. Watson, Computing Helmert transformations, Journal of Computational
[19] Y. Shang, W. Ruml, Y. Zhang, M.P.J. Fromherz, Localization from mere and Applied Mathematics 197 (2) (2006) 387394.
connectivity, in: Proceedings of the 4th ACM International Symposium on [25] B.K.P. Horn, Closed-form solution of absolute orientation using unit
Mobile Ad Hoc Networking & Computing MobiHoc 03, 2003, pp. 202212. quaternions, Journal of the Optical Society America 5 (7) (1987) 1127
[20] W. Ruml, Improved MDS-based localization, IEEE Infocom 4 (2004) 26402651. 1135.
[21] Y. Shang, J. Meng, H. Shi, A new algorithm for relative localization in wireless [26] R.E. Deakin, 3D coordinate transformations, Surveying and Land Information
sensor networks, Journal of Parallel and Distributed Processing 00 (C) (2004) Systems 58 (1998) 223234 (Featherstone 1996).
2630. [27] MATLAB version 7.10.0. Natick, Massachusetts: The MathWorks Inc., 2010.

You might also like