You are on page 1of 6

Anchor-Free Localization in Rapidly-Deployed

Wireless Sensor Networks


Hui Qu Stephen B. Wicker
School of Electrical and Computer Engineering School of Electrical and Computer Engineering
Cornell University Cornell University
Ithaca, New York 14853 Ithaca, New York 14853
Email: hq24gcornell.edu Email: wickergece.cornell.edu

Abstract- In this paper we consider a wireless sensor anchor density to work well. The APS proposed by
network consisting of a single sink node at the center of a Niculescu and Nath [5] uses both distance-free and
field of randomly distributed sensors. An anchor-free node distance-aware approaches to obtain distance estimates
localization algorithm is proposed in which the sink node to anchors: in their DV-hop method each node finds its
imparts radial location information through the phased- distance estimate to an anchor based on the hop-counts of
array transmission of a series of beacons. The individual
sensors use knowledge of received beacons and information the shortest connecting path; in their DV-distance method
from nearest neighbors to identify the sub-sectors in each node sets its distance estimate to an anchor as the
which they reside. This location information can then be total length of all the links in its shortest connecting path.
used as the basis for an emergent routing algorithm. We DV-hop is less sensitive to measurement noise, but has
present the localization algorithm, deduce the partition poorer performance compared with DV-distance under
error analytically, and then provide simulation results. The moderate measurement noise.
results indicate that the localization performance improves Unlike the above, our proposed algorithm is anchor-
with node density, and is robust against RSS measurement free. In this paper we present a low-cost localization
noise.
algorithm for randomly distributed sensors with a single,
I. INTRODUCTION centrally-located data sink (variations on this topology
are easily accommodated, but are not considered here
Sensor location is an important element in habitat due to length limitations). The proposed algorithm uses
monitoring, battle-field surveillance, search and rescue, a series of beacons generated by the data sink as well
and target tracking, but such knowledge often comes at as localized peer-to-peer transmissions as the basis for
the cost of pre-deployment planning or the use of GPS or sensor localization.
similar relatively expensive technologies. We also note Section II of this paper lists network assumptions
that location awareness plays a significant role in many made in this project. Section III describes our localiza-
energy-preserving routing protocols designed for sensor tion algorithm in detail. Section IV provides an error
networks, such as GEDIR [1], GeoCast [2], and LAR analysis for this algorithm, while section V presents
[3], etc. simulation results.
Several localization methods have been proposed,
most of which assume the existence of several location- II. NETWORK ASSUMPTIONS
aware nodes, or "anchors". These techniques can be Consider a circular coverage area with radius R and
divided into two different categories: distance-free and a data sink (base station, BS) at its center. During
distance-aware. An example for the former is the sys- deployment, N sensor nodes are randomly distributed
tem proposed by Buluse [4], which uses connectivity in the circular area (without prior planning), creating a
information, employs a grid of anchors with known 2-D uniform distribution of sensor nodes.
locations, and lets each unknown node set its location to The BS uses phased array antennas to broadcast M
the centroid of its neighboring anchors. The localization beacons with limited beamwidth, each beacon associated
accuracy is about one-third of the separation distance with an angular range in the coverage area. The beacons
between anchors, and therefore this method needs high thus divide the roughly circular coverage area into M

1-4244-0507-6/06/$20.00 ©2006 IEEE 627


beacon sectors. The beacon transmissions are scheduled
in sequential time slots and distinguished by beacon
sector numbers included in the transmission.
After the transmission of all the beacons, the BS
broadcasts strong omnidirectional signals which enable
all the nodes to estimate their distances to the data
sink using received signal strength (RSS) measurements.
The RSS measurements may be noisy and therefore the
distance estimates might be imprecise, but this will only
have negligible effect on our localization algorithm's
performance as we will show in the simulation section.
If several nodes are mobile, the BS needs to perform
directional and omnidirectional broadcasts periodically,
allowing nodes to update their locations. A node's
movement might change its own and other nodes'
neighbor list, but only the nodes that have moved need
to carry out the localization algorithm again with the
changed neighbor list. We will show in future work that
mobility actually helps nodes locate themselves more
accurately.

III. ANCHOR-FREE LOCALIZATION ALGORITHM


One-hop neighbor information is used to create a
neighbor list in each node as follows:
1. After deployment and receiving the BS's broadcasts,
all the nodes use full transmission power to broadcast
their existence using a message MSG as described below:
node number i distance to the BS ri current sub-
sector number mi total number of sectors Mi flag
fi 1 estimated angle foj 1 energy level indicator 1i
Each node i sets its ri using RSS measurement, sets
its current sub-sector number mi to the beacon sector
number received, and sets total number of sectors Mi
to M. The flag fi indicates whether additional partition
rounds are needed for node i, and is set to 1 for all
7..
-L-1

k A Fj

.
I.
. |!
When a node first receives a MSG from its one-hop
neighbor, it will make an entry for this neighbor in its
neighbor list. The entry is simply the received MSG
of that neighbor, combined with the distance from that
neighbor estimated using RSS. The entries are updated
as new MSG messages are received, and therefore kept
up to date.
The flow chart in Fig.1 is the detailed description
of our anchor-free localization method carried out at
each node i. As is shown in the flow chart, during

14
1.

,,1
f
-Cs .

t gS10

_E
.X
;

_00?g

Fig. 1. Flow Chart of the Anchor-Free Localization Algorithm

the localization process, each node i waits for Tupdate


to update entries for all its neighbors before starting a
Xi
)l

it
_.r

-
I
I

nodes at startup. The estimated angle is set as fo = further partitioning process. Then it traverses its neighbor
(mi + 0.5) M4 at the beginning, and it is used only when list, applying the checking conditions on each entry and
fi is set to 0. The energy level indicator 1i is the ratio of increasing its number of left neighbors (nL) or number of
the available energy over the full available energy, and right neighbors (nR) accordingly. After checking all the
is designed for energy aware routing. entries, it determines itself to be in the left (larger angle)
2. Each node carries out the localization algorithm and sub-sector if nL > nR, or in the right (smaller angle) sub-
updates its own sub-sector number mi, total number of sector if nL < nR. In addition, if nL = nR, it announces
sectors Mi, flag fi, estimated angle fo, and energy level itself to be on the common edge of these two sub-
1j, as well as uses full transmission power to broadcast sectors (middle line of the parent-sector) and terminates
a new MSG if its content does change whenever a localization by setting fi to 0. Otherwise, if it finds itself
partition round ends. The localization process of the in a sub-sector with maximal location inaccuracy no
whole network terminates when all the nodes have their larger than the preset threshold F, it announces itself to
flags set to 0. be on the middle line of current sub-sector and terminates

628
localization by setting fi to 0. Define
It is clear that our localization algorithm does not rLR = min(rtrans,ri * sin("IT))
depend on the exact distances between neighbors, and
therefore we claim it to be partial-distance-aware. As it dL ri- sin(47 (mi+ 1)- Oi)
is a localized algorithm, it scales well. dR ri sin(Qi - * mi)

IV. ERROR ANALYSIS FOR THE LOCALIZATION arccosdL) dL < rLR


ALGORITHM (L ={
O dL > rLR
:
Under the assumption of 2-D uniform distribution arccos(R) : dR < rLR
(R ={ dR > rLR
of sensor nodes, we can study the error probability of O :
our localization algorithm analytically. The localization
process at each node i is basically partitioning current XL max(O, ri + rLR sin 1L R)
sector (sub-sector) and then deciding itself to be in XR max(O, ri + rLR sin (R R)
the left part, in the right part, or on the middle line.
7T-r2
rLR iD2 L 12
According to the algorithm described in the above flow SL T TrLR * sin(21L)
chart, given node i' location xi = (ri, Oi), we have: SR 2T
FrLR R (DR
i
2 -
12r2LR sin (2(R)
4r
P1 = P {node i in the left part} Where rtrans is the maximal transmission range of all the
nodes.
P{NL-NR> 1} Then the areas of the left and right regions being
F(N- 1) /2] -1 considered are:
=
2E S .0C/_1pnR(l
R I R) N-1-nR
nR=o AL =SL * (2 -( rLRXLjjsinf(L )2)
N-I-nR
AR= SR- (2 ( rLR2 XR 2
j
N X, CN-1-nRPL( pL) N-1-nR-nL(j) sin (DR
nL=nR+l
For 2-D uniform distribution, we have:
P2 =P {node i in the right part p = AL PR ARs
= P{NR-NL>I} Where S is the area of the circular network.
F(N-1) 2] -1 In the first partition round, there is an error if node i
p N-1-nL
=
20,
nL=o
5 CZ_
C- 1pnL(l
P PU announces itself to be on the middle line while its actual
N-1-nL
location's distance from the estimated location is larger
X5 C-1-n LPR( PR) 1--nL-nlR(2) is not in the sub-sector that it determines itself to be in.
than the preset threshold F. Error also arises if node i
nR=nL+l
Combining all these considerations, we deduce the error
P3 P {node i on the middle line} probability of the first partition as follows:
P{NL -NR = O} For any node i in the circular area with location xi
L(N-1) 2j (ri, Qi), define its error indicator as:
X,R0 -1PR )R)N-1-nR { I : mistake
=

nR=o t0 : no mistake
nR
X, CN- 0f1 pP(1 _ PL)N 1nR-nL (3) (i is a random variable determined by NL and NR, as
well as node i's location xi, therefore, we can denote (i
nL =nR
as i (xi, NL, NR). Then for 2-D uniform distribution, we
NL and NR are random variables denoting the number have:
of neighbors in the left and right regions being con-
sidered, while nL and nR are the values they take in ENL,NR [i (xi, NL I NR)] = ei(oi)
a particular realization. PL and PR are the probabilities P2 + P3 : node i in region I
that there is a node in the left and right regions con- P2 : node i in region II
sidered separately. According to our algorithm, they are P+ P3 : node i in region III (4)
determined in the following way: PI + P3 : node i in region IV

629
Denote the angle of the middle line as oj, and region I, Case 3: It made a mistake and determined itself to be
II, III and IV are defined as: in sub-sector 2. Then in the second partition round, it
checks all the nodes in the circle of radius rLR as shown
region I 2ri * sin0ii
2> F
in Fig.2, and counts the number of left neighbors nL
region II o<2ri sin 0i2 i < F and the number of right neighbors nR. Since it assumes
region III o < 2ri sin ',2 < F its current sub-sector number is 2, it will regard nodes
region IV : 2ri .sin 'i20i > F with sub-sector number 3 as left neighbors, and nodes
In a 2-D uniform distribution, a node i's location is with sub-sector number 1 as right neighbors. Since node
also a random variable, we denote it by Xi, and all the i's actual location is in sub-sector 3, and we have set
nodes in the network are identical. Therefore, we can rLR to be no greater than ri sin( 27 ), this circle will not
generate the error probability of first partition for any cover nodes out of sub-sector 2,3,4. Therefore we would
node by taking average of ei(Xi) over Xi: definitely have nR = 0, and have nL > 0 with high
probability. As a result, node i would probably determine
P{first partition error of any node} itself to be in the left part of sub-sector 2 after the second
= Exj[e(i (Xi)] partition round, and in this way, the error in its estimated
rR r2- location decreases.
= X y ~ei (ri,Ii) ri dOidri (5) Similar discussions apply for the subsequent
partitions. Actually, in the case of last partition having
The second partition is based on the results of the first made mistakes, our localization
algorithm manages
partition which have an error probability, therefore its to decrease the location estimation errors as partition
performance needs to be analyzed in different cases. To rounds increase, which will be shown in the simulation
make it clear, we give a concrete example as follows: section.
Assume M=12, as in Fig. 2, node i received the BS's

y 5 / 4 V. SIMULATION RESULTS FOR THE LOCALIZATION


/ 3/ -
, ALGORITHM
In this section, we use Matlab simulations to test our
localization algorithm's performance. In all simulations,
we assume the sensor network is a circular area with
radius R = 100m, the BS is at the center, and the original
total number of sectors provided by the BS's beacons is
Fig. 2. Analysis of Partition M = 12. There are N nodes of 2-D uniform distribution
in the circular area, and the maximal transmission range
beacon and set its beacon sector number to 1 at the be- of each node is rtrans = 30m. The maximal allowed
ginning of localization process. After the first partition, location inaccuracy F is set to 5m, which is proved to
there are three possible results: provide satisfying localization results for our energy-
Case 1: It announced itself to be on the middle line preserving routing algorithm in design (not described
and no more partitions are needed. here due to length limitations).
Case 2: It determined itself to be in sub-sector 3
correctly. Then in the second partition, suppose there are A. Performance of Partitions
actually n1 and n2 nodes in the left and right considered We vary N between 100 and 1000 with step size 100,
region of node i separately, but the neighbors might have and the results are plotted in Fig.3. As we can see, the
wrong decisions of their current sub-sectors, as well as first partition error probability decreases as N increases,
wrong entries in node i's neighbor list. Therefore the which indicates that our localization algorithm's perfor-
numbers of nodes that node i determines to be in the left mance might improve with node density. As explained
and right considered region might be different from their in last section, for the second partition, formula (1)-(5)
actual values, and formula (1)-(5) do not apply. However, can only be used as an estimation tool if the node's first
as long as the error probability of the first partition partition is right (Case 2), and when they are applied, we
is low, these formula still give good error probability are assuming that the first partition error probability of
estimation. every other neighbor is also 0, which makes the second

630
1 20 (
0.15f rI.
i (8 + ) M (7)
Therefore w e Ad2 0.5. Similar explanation
-0 0.11
applies for the subsequent partitions. Fig.4 plots the
a)

X0.0 15
V-
o the first partition
O the second partition
the third partition
+ the fourth partition
0
E~~~~~~
10
0
100 200 300 400 500 600 700 800 900 1000
number of nodes E 0

Fig. 3. Theoretical Error Probability of the First Partition (F=5m)

partition equivalent to the first partition with original 0


0 10 15 20 25 30
simulation count
number of beacon sectors M = 24.
For the case that the node has made a wrong sub- Fig. 4. Location Error Comparison (N=800, F=5m)
sector decision in the first partition, we show that our
localization algorithm manages to decrease the location results for the 30 simulations. In each simulation, we
estimation error as partition rounds increase. We set N to record all the nodes that have made wrong sub-sector
800 and do simulation 30 times, in each realization of the decisions in the first partition round, and are going into
2-D uniform distribution, we track every node that has the second partition round in a set Q. We sum up their
made a wrong sub-sector decision in the first partition, location estimation errors after the first partition and use
and compare its location estimation error after the first a red circle to represent the average. Each simulation
partition (Ad,) with its location estimation error after runs till all the nodes have their flags set to 0 (at
the second and subsequent partitions (Ad2,..., Adk). most 4 partitions under our parameter setting), and after
The simulation results show that, for any node that every subsequent partition round, we sum up the location
has determined itself to be in the wrong sub-sector in the estimation errors for all the nodes in set Q, as well as
first partition round, as long as it carries out the second plot the average in the figure.
partition round, its location estimation error Ad2 would It is clear that the average localization error after
decrease to a little bit larger than 1 Ad,. And as long as the second partition round is around one half of the
it carries out the kth partition round, it will have Ad, average localization error after the first partition round.
Adk < 1. The explanation is that, those nodes making For the following partition rounds, the error decreasing
wrong sub-sector decisions tend to be around the middle capability seems not so obvious in this figure, that is
line. Assume a node i with actual angle Qi = (mi+ 0.5- because more and more nodes in set Q have set their
e) * M,, << 1 makes a wrong sub-sector decision and flags to 0 and stopped partitioning in the third and fourth
determines itself to be in sub-sector 2mi + 1 instead of partition rounds, therefore, the error decreasing of those
2mi, accordingly, it will have localization error: nodes which do carry out the partition is diminished after
taking the average over all the nodes in the set Q.
(2mi + I + 0-5) * 2M -
Ad, = 2ri * sin( O)
B. Performance of the Localization Algorithm
1 2wr
M First, we show that our localization algorithm's perfor-
rI (+) (6)
mance does improve with node density. The performance
Then in the next partition round, as we discussed in is evaluated using the following routine: after all the
the last section, it will determine itself to be in sub- nodes having their flags set to 0, we compare each node's
sector 2 (2mi + 1) with high probability, and therefore estimated location with its actual location, and if the
the localization error will change into: Then in the next partition
location round,
difference larger than F, we declare
as we isdiscussed he a localiza- perormanc

tion error. The ratio of the number of localization errors


Ad2 = 2ri sin(
(2(2mi + 1) + 0.5)4 2j4 0* Nerror over the total number of nodes N, which we name
2 I error rate and denote as Perror Nror N ,i h efrac

631
indicator. We increase N from 100 to 1000 with step size 100
100 with other parameters fixed. For each value that N
takes, we run simulation 30 times and take the average 50
of Perror, the results are plotted in Fig.5 as red stars. As
0

0.4r
( = 0
0.35T c= 0.1 -50
( = 0.2
0.3 = 0.3

0.25 -100_
-100 -50 0 50 100
X (m)
2 02
=
0.15 Fig. 6. Localization Result of One Simulation (N = 900, F=5m)
01

0.05

0
VI. CONCLUSIONS
200 400 600 800 1000
number of nodes In this paper, we proposed an anchor-free localization
algorithm for rapidly-deployed wireless sensor networks.
Fig. 5. Localization Performance of the Anchor-Free Algorithm Our localization algorithm scales well with the size
(F=5m)
of network. During the localization process, each node
broadcasts messages of a preset format and collects one-
expected, the localization error rate decreases as node hop neighbors' messages to keep its neighbor list up
density increases, therefore our localization algorithm to date. The BS originally divides the circular network
works better for high density networks. We also notice area into M beacon sectors, and then each node keeps
that the decrease becomes slower when node density is partitioning its current sector and locating itself in a
larger, and after N approaches 500, the localization error smaller sub-sector with the help of its neighbor list till
rate is kept below 10%. the localization termination condition is satisfied. Both
Simulations are used to study the effect of RSS analytical and experimental results confirmed that our
measurement noise. For any node, we assume its distance localization algorithm's performance improves with node
from the BS estimated using RSS measurement rmeasured density, and it works well for high-density networks. It
has the following relationship with the actual distance was also shown that our algorithm is robust against RSS
ractual: rmeasured = ractual (1 + 6), where E is a Gaussian measurement noise.
random variable with zero mean and standard deviation
(X. Again we increase N from 100 to 1000 with step size REFERENCES
100, for each value N takes, we try different values of u-, [1] I. Stojmenovic and X. Lin, "GEDIR: Loop-free location based
and for each value of u-, we run simulation 30 times and routing in wireless networks", in Proc. IASTED Int. Conf on
record the average of Perror, the results are also plotted Parallel and Distributed Computing and Systems, November 3-
6, 1999, Boston, MA, USA, pp. 1025-1028.
in Fig.5. [2] J. C. Navas and Tomasz Imielinski, "GeoCast Geographic -

The results show that our localization algorithm is addressing and routing", in Proc. MobiCom'97, September 26-
robust against RSS measurement noise. Even when the 30, 1997, Budapest, Hungary, pp. 66-76.
[3] Y B. Ko and N. H. Vaidya, "Location-aided routing(LAR) in
standard deviation u- is 0.2, the performance degradation mobile ad hoc networks", in Proc. Mobicom'98, October 1998,
is negligible. Larger u- would result in larger performance Dallas, TX, USA, pp. 66-75.
degradation, but as far as the node density is high, our [4] N.Bulusu, J.Heidemann, and D.Estrin, "GPS-less low cost out-
localization algorithm still gives acceptable performance. door localization for very small devices", IEEE Personal Com-
munications Magazine, October 2000, 7(5):pp. 28-34.
It should also be noted that most of the nodes that [5] D. Niculescu and B. Nath, "Ad Hoc Positioning Systems", in
make errors are at the periphery of the coverage area, as Proc. GlobeCom, November 2001.
shown in Fig.6. This is actually an advantage, because
in the location-based routing scheme designed for our
single sink network, these boundary nodes will not be
used as relay nodes, and therefore their localization
errors are less detrimental.

632

You might also like