Algorithms for Location Estimation

Based on RSSI Sampling

Charalampos Papamanthou, Franco P. Preparata, and Roberto Tamassia

Department of Computer Science and Center for Geometric Computing

Brown University
{cpap,franco,rt}@cs.brown.edu

Abstract. In this paper, we re-examine the RSSI measurement model

for location estimation and provide the first detailed formulation of the
probability distribution of the position of a sensor node. We also show
how to use this probabilistic model to efficiently compute a good esti-
mation of the position of the sensor node by sampling multiple readings
from the beacons (where we do not merely use the mean of the samples)
and then minimizing a function with an acceptable computational effort.
The results of the simulation of our method in TOSSIM indicate that
the location of the sensor node can be computed in a small amount of
time and that the quality of the solution is competitive with previous
approaches.

1 Introduction
Estimating the location of a roaming sensor is a fundamental task for most
sensor networks applications. For example, if a sensor network has been deployed
to provide protection against fire (in this case, sensor nodes report a sudden
increase in temperature), we want to know the location of the sensor that triggers
an alert so that action can be taken accordingly. Additionally, some routing
protocols for sensor networks, such as geographical routing [15, 44], make routing
decisions based on the knowledge of the locations of the sensor nodes. Common
location estimation protocols that are widely adopted in practice assume that
there are some fixed nodes (base stations) that know their location which are
called beacons. These nodes send a signal to the sensor nodes that want to
determine their location. According to the intensity (or for example the angle)
of this signal, the sensor node can have an estimate of the distance between them
and the beacons.
After performing a certain number of such measurements for different bea-
cons, the sensor node has to combine all this information (for RSSI (Received
Signal Strength Indicator), this information is the power of each individual sig-
nal and the coordinates of the corresponding transmitter) in order to estimate
its location. However one could ask the following question: Why cannot we use a
Geographic Positioning System (GPS) to efficiently achieve the task of localiza-
tion? The answer is that a GPS requires a strong computing platform which is
not available in sensor networks. Sensor nodes are typically very low-computing

S. Fekete (Ed.): ALGOSENSORS 2008, LNCS 5389, pp. 72–86, 2008.


c Springer-Verlag Berlin Heidelberg 2008
 Algorithms for Location Estimation Based on RSSI Sampling 73

power units that can efficiently perform only basic arithmetic operations; Re-
quiring the execution of complex arithmetic operations on a sensor node would
entail a quick depletion of its battery which is not desirable for most practical
applications. Finally, the localization problem gets even more difficult because
the available power on the sensor node is limited: therefore no accurate mea-
surements of the signal can be made (since an accurate measurement requires
more computing power) which means that the measurements are prone to er-
rors. This is something that should also be taken into consideration and treated
accordingly.
Therefore, any location estimation algorithm should have the following re-
quirements:
1. The sensor node should avoid complex and time consuming computations,
which would deplete its energy supply (typically a low-cost battery) rapidly;
2. The computations should take into consideration the error in the measure-
ments, which can be large.

1.1 Related Work and Observations

There are several proposed location estimation protocols for sensor networks,
see, e.g., [5, 7, 8, 11, 23, 24, 25, 26, 27]. All these protocols use the same model,
where some nodes know their location (either because they are fixed or by using
GPS) and are called beacons or anchor nodes, and some other nodes, called sen-
sor nodes, estimate their location using the information they receive from the
beacons. This information consists of the beacons’ coordinates and of features
of the beacon signal, such as the received signal strength indicator (RSSI) or the
time difference of arrival (TDoA). Also, other protocols (e.g., [26]) are based on
the capability of the nodes to sense the angle from which a signal is received.
Recently [38] presented a solution for aerial localization and [22] proposed a
solution where the localization is based on adopting slightly different periodic
signal frequencies (interferometric positioning). This solution [22] is very com-
petitive and achieves very good precision. However the computations used are
very intensive.
Several previous approaches use computationally demanding methods, such
as convex optimization [8], systems of complex equations [26], minimum mean
square error (MMSE ) methods [6, 33], and Kalman filters [34]. In these ap-
proaches, the measurement model is not adequately analyzed and the error is
assumed to be small, which is not the case in most real applications of sensor
networks.
Other approaches, notably [17, 30, 36], estimate the location of a node using
the RSSI method (analyzed in [31]), which is the most realistic model for sen-
sor network communication. In [17], the authors evaluate the ability of a sensor
network to estimate the location of sensor nodes. They assume that the location
of the sensor node is known and develop arguments concerning the probability
that the network will detect this location. They use the RSSI error model to
analyze the problem of evaluating the ability of the sensor network to locate a
74 C. Papamanthou, F.P. Preparata, and R. Tamassia

sensor node. However, they do not describe how their algorithms can be imple-
mented on a sensor node to estimate its own location. Moreover, their method
does not take into account the basic parameters of the RSSI model (standard
deviation and path loss exponent) and thus gives incorrect results.
In this paper, we formulate the correct probability distribution of the position
of a sensor node based on one reading produced with the RSSI model. Due to
the errors implicit in the RSSI model, it is unrealistic to try to compute a good
estimation of the location of the sensor node based only on a single measurement
(or even few measurements) from each beacon. Such an approach would be
so inaccurate to make the estimate practically worthless. Notwithstanding this
difficulty, we show that a reliable estimation of the location can be achieved by
processing a reasonably small number of readings of the signals.
Especially for indoor positioning systems, this is an assumption that has been
extensively used. For example, in [13, 14], the position estimation is based on a
location fingerprint t = [t1 t2 . . . tN ], where N is the number of beacons and ti
(i = 1, . . . , N ) is the mean value of the received signal strength over a certain
time window. Also, in [3, 9, 20, 42], experiments with various sample sizes are
presented where the samples are used to compute certain features of the signal
strength such as the standard deviation and the path loss exponent. Finally, [12]
presents simulations that use various number of samples, where more than 50
samples to filter out the errors in the probability distribution are used.
We show that using only the mean of the measurements is not a correct
procedure, due to the lognormal distribution of the distance from the beacon (see
Theorem 3). Instead of using directly the mean value, we use another value that
is adequate according to the specific underlying probability distribution of the
distance. The number of samples that are used vary from 20 to 60 and obviously
the accuracy of the computed location grows with the number of samples. Finally,
once the sampling has been performed, we show how to seek the minimum of a
function that approximates the actual location with small computational effort.

1.2 Our Contributions

The main contributions of this paper are as follows:
1. We evaluate the probability that a sensor node lies within a certain region,
given that the power received from the beacons is modeled with RSSI. To the
best of our knowledge, this is the first detailed formulation of the probability
distribution of the position of a sensor node. We show that unlike the normal
distribution of the received power, the probability distribution of the actual
position is lognormal. Thus, we give evidence to the role of the parameters
σ and n in the probability distribution of the actual distance, where σ is the
standard deviation of the normal variable that models the power received
by the sensor and n is a parameter, called path loss exponent, that depends
on the transmission medium. In previous approaches [17], the probability
distributions used did not exhibit dependency on these two variables.
2. We present a method for estimating the location of a node from multiple
sample power readings from the beacons. Our method computes the expected
 Algorithms for Location Estimation Based on RSSI Sampling 75

value of the received power and combines it with the mean and the standard
deviation of the sample readings using a steepest descent approach [37]. We
show that our method is simple and efficient and provides a good estimation
of the position. Note that using multiple sample readings is necessary for a
reliable location estimation since the probability distribution of the location
for a single sample implies that the domain within which the sensor lies with
high probability has large area.
3. We describe an implementation of our location estimation algorithm that is
suitable for execution on standard sensor hardware and we analyze the re-
sults of an extensive simulation of the execution of the algorithm in TOSSIM
[10, 18]. Our simulation shows that our method has accuracy that is compa-
rable to or better than that of previous methods.

1.3 Organization of the Paper

The rest of this paper is organized as follows. In Section 2, we overview the
RSSI model, give formulas for the probability distribution of the position of a
sensor node due to power measurements, and show how to estimate the actual
distance given a set of sample power readings. We develop an efficient algorithm
for location estimation and analyze its running time in Section 3. Finally, in
Section 4, we report the results of the simulation and present a comparison (in
terms of localization error) of our method with previous approaches. Concluding
remarks are in Section 5.

2 Theoretical Framework
This section provides a formal probabilistic framework for estimation of the
position of a sensor node from power measurements.

2.1 RSSI Model

Suppose we are given a region of the plane with k beacon nodes b1 , b2 , . . . , bk
(nodes of known location). The coordinates of the beacons are (xi , yi ) for i =
1, . . . , k. The beacons transmit information about their location with a signal of
normalized intensity to a sensor node s that does not know its location. Based
on the locations of the beacons and the estimated distances from the beacons
(computed from the received signals), the sensor is to compute its actual location.
Among the several models proposed for estimating the distance between a
beacon and a sensor node, the most realistic and commonly used one is the
received signal strength indicator model (RSSI) [31]. In this model, the beacon
broadcasts signal to all sensors and the sensors can estimate the distance between
them and the beacons on the basis of the strength of the signals they receive.
Let bi be a beacon located at (xi , yi ) and s a sensor node located at (x, y).
We define the relative error i pertaining to bi as follows. Suppose that s reads
a distance rˆi , while the actual distance is ri . The relative error is
76 C. Papamanthou, F.P. Preparata, and R. Tamassia

rˆi
i = − 1 ∈ [−1, +∞). (1)
ri
The commonly accepted transmission model [31] expresses the received power
pi (in dBm) as  
ri
pi = p0 + 10n log (2)
r0
where p0 is the received power in dBm at a reference distance r0 and n is the
path loss exponent which is a constant depending on the transmission medium
(indoors, outdoors) and ranges typically from 2 to 4. In some environments, such
as buildings, stadiums and other indoor environments, the path loss exponent
can reach values in the range of 4 to 6. On the other hand, a waveguide type of
propagation may occur in tunnels, where the path loss exponent drops below 2.
We recall that if the received power in mW at a point k is Pk , and Pk is
the received power at some reference point k  (again in mW), then the received
power pk in dBm at point k is defined as

pk = 10 log (Pk /Pk ) . (3)

The measured power, however, differs from that given in Equation (2); due to
channel fading (variation of the received signal power caused by changes in trans-
mission medium or path), the measured power is pˆi = pi +x. The random variable
x represents the medium-scale channel fading and is typically modelled as Gaus-
sian zero-mean with variance σ 2 (in dBm). Typically, σ is as low as 4 and as
high as 12 (this implies that the error may be large). Inserting pˆi and rˆi into
(2), we get  
rˆi
pˆi = p0 + 10n log (4)
r0
where now the measured power pˆi in dBm relates to the measured distance rˆi by
the sensor. By combining the above equations, we get that the relation between
the measured distance and the actual distance is
x
rˆi = ri 10 10n (5)

which gives
x
i = 10 10n − 1. (6)

2.2 Probability Distributions

In this section, we present the probability distribution of the position of the
sensor node based on measurements of one or more beacons. We denote ex with
exp(x).
Theorem 1. Let bi be a beacon node located at (xi , yi ) sending information to
a sensor node under the RSSI model with standard deviation σ and path loss
exponent n. Let rˆi be the measured distance from beacon node bi at the sensor
 Algorithms for Location Estimation Based on RSSI Sampling 77

node. We have that the probability density function of the actual position (x, y)
of the sensor node is given by
  2 
10n exp − 10n log √ r
ˆ
2
i
2
/2σ 2
(x−xi ) +(y−yi )
(i)
PX,Y (x, y) = √ .
2πσ 2π ln(10)((x − xi )2 + (y − yi )2 )

To simplify the notation, we denote the probability distribution due to beacon

bi with
(i)
Φbi (x, y) = PX,Y (x, y).
The previous argument can be extended to a finite set of beacons B = {b1 ,
. . . , bk } yielding the following theorem:

Theorem 2. Let B = {b1 , b2 , . . . , bk } be a set of beacons sending information

to a sensor node under the RSSI model with standard deviation σ and path loss
exponent n. If the measured distance from beacon node bi at the sensor node is
rˆi (i = 1, . . . , k), then the probability density function (due to all the beacons in
B) of the actual position (x, y) of the sensor node is given by
k
Φbi (x, y)
Φ(B) (x, y) =  +∞  +∞ i=1
k
−∞ −∞ Φ
i=1 bi (x, y) dydx

where Φbi (x, y) is the probability distribution due to beacon bi , as defined in

Theorem 1.

As we will see later, if we use only one measurement, we may end up with a
density function having more than one maximum. Moreover, the computation
of this maximum on a sensor node is a difficult task since there is no simple
analytical expression for the maximum of the probability distribution. Also,
there is no suitable analytical expression for the integrals needed to compute the
probability for a certain region. Due to space limitations, the proofs of the two
theorems can be found in the full version of the paper.

2.3 Samples of Measurements

In the previous sections, we have examined the probability distribution of the
sensor’s position based on a single measurement. This setting, however, can give
rise to unacceptable errors for the values of σ (σ = 4, 6, 8) reported in the liter-
ature [31]. A consequence of this situation is that we may be unable to define a
disk containing the sensor’s location with an acceptable degree γ of confidence
(say, γ > 0.9). Additionally, if our practice is based on only one reading, there
is no way for the sensor to estimate the standard deviation σ (this is actually
the standard deviation of the Gaussian random variable x introduced in Section
2.1) of each beacon. This parameter σ—conveniently assumed to be known in
78 C. Papamanthou, F.P. Preparata, and R. Tamassia

Section 2.1—must be estimated in practice as it is needed in the computations

(see below).
To overcome these difficulties, we show in this section how we can obtain a
good estimate of the location of the sensor node based on small number of read-
ings. Especially for indoor positioning systems, this is an assumption that has
been extensively used. For example, in [13, 14], the position estimation is based
on a location fingerprint t = [t1 t2 . . . tN ], where N is the number of beacons and
ti (i = 1, . . . , N ) is the mean value of the received signal strength from the i-th
beacon over a certain time window. Note that ti denotes the measured power
pˆi that appears in Equation 4. Hence, the “measured power” location finger-
print t can be transformed to a location fingerprint r of “measured radii” by
using Equation 4, since the reference values p0 and d0 are known. The number
of samples that are used vary from 20 to 60 and obviously this number affects
the accuracy of the computed location. Also, in [3, 9, 20, 42], experiments with
various sample sizes are presented where the samples are used to compute cer-
tain features of the signal strength such as the standard deviation and the path
loss exponent.
Suppose now that we use a sample of k readings from beacon bi . We have a
sequence of radii rˆi1 , rˆi2 , . . . , rˆik . Let r¯i , s¯

i denote the unbiased estimators of the

value E[rˆi ] and of the standard deviation Var(rˆi ) of the underlying distribution
of the measured radii rˆi (i = 1, . . . , 3) respectively. We have the following result
that relates estimates of the actual distance and the standard deviation with
reference to a beacon bi with features of the lognormal distribution:
Theorem 3. Suppose a sensor node reads k distance samples rˆi1 , rˆi2 , . . . , rˆik
from a beacon bi that is modeled with the RSSI of path loss exponent n and
standard deviation σi . If r¯i is the sample mean and s¯i is the sample standard
deviation then we have the following:
1. The estimate of the square of the actual distance ri2 from beacon bi is
r¯i 4
.
r¯i 2 + s¯i 2
2. The estimate of the square of the standard deviation σi2 is
  2 
100n2 s¯i
2 ln 1 + .
ln (10) r¯i
Note that the above theorem indicates that the quality of estimation of the
actual distance is heavily dependent on the estimation of the distribution of the
measured radii.

3 Location Estimation
In this section, we develop an algorithm for location estimation based on several
samples. This algorithm does not involve any complex calculations (such as
square roots) which is very important to consider when we develop algorithms
to be executed on sensor nodes, due to the sensor’s modest computing power.
 Algorithms for Location Estimation Based on RSSI Sampling 79

3.1 Algorithm
As we saw in the previous section, after completing the sampling procedure, we
derive estimates for r12 , r22 , r32 , given by Theorem 3. Our aim is to formulate a
function whose minimum will yield a good approximation of the sensor’s location.
This function should be convex and also its derivatives should not include roots.
Suppose now we have three beacons located at (x1 , y1 ), (x2 , y2 ), (x3 , y3 ). Let
f (x, y) be the function

3
f (x, y) = ((x − xi )2 + (y − yi )2 − ri2 )2 . (7)
i=1

Note that if all 3 circles intersect at the same point (x0 , y0 ), this function has
minimum 0 at (x0 , y0 ). Unfortunately, minimizing that function is not an easy
task, if we are restricted on the available primitives. Hence we are going to use
methods that are based on the gradient of the function. The good feature about
such methods is that we can get to a point very close to the minimum in a small
number of computationally simple iterations. Indeed, let

∂f (x, y) 3
α(x, y) = =4 (x − xi )((x − xi )2 + (y − yi )2 − ri2 ) (8)
∂x i=1

and
∂f (x, y) 3
β(x, y) = =4 (y − yi )((x − xi )2 + (y − yi )2 − ri2 ) (9)
∂y i=1
be the partial derivatives of f . Note that the above expressions are computable
on a sensor node. The function z = f (x, y) describes a convex solid surface
with obvious definitions of “interior” and “exterior”. Initially, we make a guess
for our point (this is required by all steepest descent methods [37]). Suppose,
for uniformity, we choose as our initial point (x0 , y0 ) the centroid of the beacon
triangle. We compute the vector v which is orthogonal to the tangent plane T and
 T
pointing toward the exterior. Hence v = α(x0 , y0 ) β(x0 , y0 ) −1 . Let now P be
the vertical plane containing v applied to (x0 , y0 , f (x0 , y0 )). Since P is a vertical
plane, any normal vector w of P will have a zero z-component. Additionally, w is
 T
orthogonal to v and therefore may be chosen as w = −β(x0 , y0 ) α(x0 , y0 ) 0 .
We seek the vector q pointing towards the minimum of the function. Such vector
belongs to P and is orthogonal to v (q is orthogonal both to v and w), i.e.,
 T
q = α(x0 , y0 ) β(x0 , y0 ) α2 (x0 , y0 ) + β 2 (x0 , y0 ) .
Now we compute the intersection point (x0 , y0 ) of the line (with the surface
z = 0) passing by (x0 , y0 , f (x0 , y0 )) which is collinear with the direction q and
the xy-plane. The parametric equation of this line is (x, y, z) = (x0 + tqx , y0 +
tqy , f (x0 , y0 ) + tqz ) for all t ∈ R. The new point (x0 , y0 ) is then given by
 
  f (x0 , y0 )α(x0 , y0 ) f (x0 , y0 )β(x0 , y0 )
(x0 , y0 ) = x0 − 2 , y0 − 2 .
α (x0 , y0 ) + β 2 (x0 , y0 ) α (x0 , y0 ) + β 2 (x0 , y0 )
80 C. Papamanthou, F.P. Preparata, and R. Tamassia

The described process gives a new point (x0 , y0 ). This point is expectedly closer
to the point that corresponds to the minimum of f as we follow the direction
of the gradient as long as the products α(x0 , y0 )α(x0 , y0 ) > 0 and β(x0 , y0 )
β(x0 , y0 ) > 0. When this condition no longer holds, we have “overshot”; to
remedy, we backtrack to the previous point referred here as (x, y) and apply a
typical steepest descent method with very small rate λ. We therefore compute
our new point (x , y  ) by setting
(x , y  ) = (x − λα(x, y), y − λβ(x, y)). (10)
We continue this process until the gradients α(x, y), β(x, y) change sign. At that
point we stop and we report the final point as our estimation. Here we should
emphasize the fact that it is important to take samples of adequate size. Taking
samples implies a better behavior for f , meaning that there would be only one
minimum and therefore the algorithm will quickly converge to the minimum. As
far as the value of the variable λ is concerned, this variable is chosen to be small
enough and inversely proportional to the size of the grid since these features of
λ force the second repeat loop of the algorithm to converge quickly. This has
been observed in the experiments. For the experiments, the value of λ is equal
to 1000−m/100 .

3.2 Complexity and Limitations

The most expensive part of the presented algorithm is the sampling procedure
and the computation of the estimates r¯i 2 and s¯i 2 . These steps take time O(k),
where k is the number of samples. There is an obvious trade-off between accuracy
and power consumption. Also, the computation executed on the sensor nodes
depends on the time the gradient methods take to converge, which is generally
small for a well-behaved function. For the other parts of the algorithm there are
closed formulas, so we can assume that they take time O(1). We can also see
that the exact number of multiplications needed by the presented program is
(7k + 8) + 10n1 + 4n2 , where is k is the size of the sample, n1 is the number of
iterations of the first repeat loop and n2 is the number of iterations of the second
repeat loop.
The size of the code of the program (ROM) written in NesC [10] is 47K
whereas the amount of memory (RAM) needed to execute this program in
TOSSIM [18] (see Section 4) is 637K. As far as the complexity of the closed
formulas computation is concerned it is realistic to assume that the involved in
closed-formula calculation can be executed on a sensor node (essentially floating
point operations). For example, there are micro-controllers, such as the ATM-
Mega128L [2] and MSP430 [39], which have very rich instruction sets. Finally, a
hardware multiplier allows floating-point arithmetic to be carried out [21].

4 Simulation
In this section, we present and analyze extensive simulation results of our
method. We have run our experiments with TOSSIM [10, 18], a widely used
simulator of the TinyOS operating system for sensor networks.
 Algorithms for Location Estimation Based on RSSI Sampling 81

We executed our simulations in a square of area m × m cells, where m =

50, 100, 200. The three beacons are placed in positions that form a well condi-
tioned triangle (well-conditioning is synonymous with the fact that the function
f (x, y) has a single global minimum). Namely, the first beacon is placed at
(0, 0), the second beacon is placed at (m, 0) and the third beacon is placed at
(m/2, 3m/4). The standard deviations of the three beacons σ1 , σ2 , σ3 are set to
4 and the path loss exponent n is set to 2. We also recall that we set the variable
λ that appears in Equation 10 equal to 1000− 100 , where m is the dimension of
m

the grid. Finally the measured distance is computed using Equation 5.

We measure the execution time of the algorithm implemented in NesC [10]
(that runs in TinyOS) and the average number of iterations of the repeat loops
over 1000 runs. We also determine the mean of the distance d between the actual
d
point and the computed point. We use the ratio m to evaluate the quality of the
solution computed by the algorithm. Note that this metric was proposed in [41].
The results obtained for different numbers of samples and different sizes of
grids are shown in Table 1 , where m is the dimension of the square region, k is
the number of samples, the time is counted in milliseconds (we count the exact
time that the simulated processor in TOSSIM takes to execute this program), n1
is the number of iterations of the first repeat loop, n2 is the number of iterations
of the second repeat loop and d is the mean of the distance between the actual
point and the computed point.
The simulation results with TOSSIM show that the sensor node can execute
the algorithm in a small amount of time. This time is proportional to the number
of samples we use each time, which indicates that the sampling procedure domi-
nates the execution time on the sensor node. Only up to six iterations (n1 + n2 )
are enough to compute an estimation of the actual point and the quality of
the estimation is dependent on the number of the samples. Additionally, note
that for various grid sizes, the algorithm has a uniform behavior since the ra-
d
tio m is similar for different sizes of the grid. Finally, in all cases, the solution
we get is better for larger sizes of samples. For example, though not practical
for power consumption issues, for 200 samples the estimation gets even better
(d/m = 0.041 for m = 200).

Table 1. Results of the simulation in TOSSIM for an m × m square grid (m =

50, 100, 200): execution time, average number of iterations of the program (n1 , n2 ),
d
localization error (d) and ratio m for various samples sizes over 1000 runs

m×m k time (ms) n1 n2 d d/m

50 × 50 20 0.140 4.045 1.081 5.018 0.1003
50 × 50 40 0.230 4.226 1.061 3.774 0.0745
100 × 100 20 0.130 3.670 1.142 9.986 0.0986
100 × 100 40 0.240 3.817 1.046 7.634 0.0763
200 × 200 20 0.120 3.640 2.736 19.977 0.0998
200 × 200 40 0.240 3.820 2.323 14.957 0.0747
82 C. Papamanthou, F.P. Preparata, and R. Tamassia

Histogram for 1000 runs and k=50 (mean=3.02)

15

10
frequency

0
0 2 4 6 8 10 12 14 16 18 20
distance from actual position

Fig. 1. Histogram for 1000 runs and k = 50

We show in Figure 1 the probability distribution of the distance of the com-

puted point from the actual point derived for 1000 runs and for m = 50. In
particular, we show the plot of the distribution of the error of the estimation
(which we define as the distance of the computed point from the actual point)
for sample size k = 50. The mean of the error for these measurements is about 3
for k = 50. Also, we observe that the distribution of the error appears to follow
a lognormal distribution.
Finally, Table 2 compares location estimation algorithms. We present the
average localization error d, the area A of the field where the experiments are
executed and the ratio √dA . We use as a comparison measure the quantity √dA ,
which for our algorithm is bounded by 0.1 (this is what we get for the smallest
number of samples k = 20). However, for a number of samples k = 300 (though
not so practical) we can get even smaller values (for example for k = 300 and
for an area 1000 × 1000 we get √dA = 0.026).
From Table 2, we see that our method gives always better or as good results
as the results obtained by the existing methods (except for [38] where, however,
the time needed for localization ranges from 10 milliseconds to 2 minutes —
something that is also observed in [22]— where the precision is even better).
Also, if we slightly increase the number of the samples we use, we get very good
results and the ratio drops substantially (for example for k = 60 we get a ratio
 Algorithms for Location Estimation Based on RSSI Sampling 83

Table 2. Comparison of existing work. In each row, we display the bibliographic ref-
erence and the respective average localization error (d), the size of the area of the
experiments A, the ratio √dA and finally the number of samples used by each method.
Note that it is not always feasible to compare between different methods since the
settings used can be different. N.A. stands for “not applicable” and it means that
the certain method does not refer explicitly to the number of samples used or that the
sampling technique is not used.
√
reference error d simulation area A d/ A number of samples
[3] 3 22.5 × 45.5 0.090 20
[32] 3 16 × 40 0.118 N.A.
[4] 4 35 × 40 0.107 250
[29] 7.62 13.71 × 32 0.360 40
[1] 3 500 0.130 N.A.
[30] 6 60 × 60 0.100 N.A.
[5] 1.83 10 × 10 0.183 20
[35] 0.8 6×6 0.130 50
[19] 13 18751 0.094 N.A.
[43] 10 26 × 49 0.280 N.A.
[38] 3.5 60 × 120 0.058 N.A.
[16] 0.82 5×5 0.164 N.A.
our scheme 4.350 50 × 50  0.087 25
our scheme 3.020 50 × 50  0.064 50

√d = 0.06). Note that previous methods use more than three beacon nodes (see
A
for example [28] where O(m) beacons are placed in the area of localization for
an m × m grid).

5 Conclusions
In this paper, we have analyzed the RSSI model for location estimation in sensor
networks. Given a normal distribution for the error in dBm, we compute the
correct probability distribution of the sensor’s location and then we adopt this
probability distribution in a theoretical analysis of sampling the measurements
for location estimation. We finally give a simple algorithm that can be executed
on sensor nodes; its complexity, for a constant number of beacons, is proportional
to the size of the sample.
Location estimation in sensor networks presents several trade-offs. If higher
accuracy is desired, one has to deploy more beacons or use more samples. Using
a large number of beacons and samples causes significant energy consumption.
The energy-optimal case occurs when only three beacons are deployed and an
estimation of the actual point is based on the probability distribution computed
by taking into consideration only one measurement. This solution, however, gives
unacceptable errors. Additionally, performing computations with the exact prob-
ability distribution is unrealistic, since it involves complex formulas. Hence, were
we to depend on few measurements, off-line computed data must be stored as
84 C. Papamanthou, F.P. Preparata, and R. Tamassia

tables within the sensor, which immediately creates a storage problem. However,
one can use more samples, thus increasing energy consumption.

Acknowledgments

This research was supported by the U.S. National Science Foundation under
grants IIS–0324846 and CCF–0830149 and by the Center for Geometric Com-
puting and the Kanellakis Fellowship at Brown University. The views in this
paper do not necessarily reflect the views of the sponsors. We thank Goce Tra-
jcevski for useful discussions.

References
[1] Alippi, C., Vanini, G.: A RSSI-based and calibrated centralized localization tech-
nique for wireless sensor networks. In: Proc. IEEE Int. Conf. on Pervasive Com-
puting and Communications Workshops (PERCOMW), pp. 301–306 (2006)
[2] Atmel Corporation. ATM128 Datasheet, Revised 2461-09/03 (2003)
[3] Bahl, P., Padmanabhan, V.N.: RADAR: An in-building RF-based user location
and tracking system. In: Proc. IEEE Conf. on Computer Communications (IN-
FOCOM), pp. 775–784 (2000)
[4] Brunato, M., Battiti, R.: Statistical learning theory for location fingerprinting in
wireless LANs. Computer Networks 47(6), 825–845 (2005)
[5] Bulusu, N., Heidemann, J., Estrin, D.: GPS-less low cost outdoor localization for
very small devices. IEEE Personal Communications Magazine 7(5), 28–34 (2000)
[6] Capkun, S., Hubaux, J.-P.: Secure positioning of wireless devices with application
to sensor networks. In: Proc. IEEE Conf. on Computer Communications (INFO-
COM), pp. 1917–1928 (2005)
[7] Dil, B., Dulman, S., Havinga, P.: Range-based localization in mobile sensor net-
works. In: Römer, K., Karl, H., Mattern, F. (eds.) EWSN 2006. LNCS, vol. 3868,
pp. 164–179. Springer, Heidelberg (2006)
[8] Doherty, L., Pister, K.S.J., Ghaoui, L.E.: Convex optimization methods for sensor
node position estimation. In: Proc. IEEE Conf. on Computer Communications
(INFOCOM), pp. 1655–1663 (2001)
[9] Faria, D.B.: Modeling signal attenuation in IEEE 802.11 wireless LANs. vol. 1.
Technical Report TR-KP06-0118, Stanford University (2005)
[10] Gay, D., Levis, P., von Behren, R., Welsh, M., Brewer, E., Culler, D.: The nesC
language: A holistic approach to networked embedded systems. In: Proc. ACM
Conf. on Programming Language Design and Implementation (PLDI), pp. 1–11
(2003)
[11] He, T., Huang, C., Blum, B.M., Stankovic, J.A., Abdelzaher, T.: Range-free lo-
calization schemes for large scale sensor networks. In: Proc. of the Int. Conf. on
Mobile Computing and Networking (MOBICOM), pp. 81–95 (2003)
[12] Hu, L., Evans, D.: Localization for mobile sensor networks. In: Proc. of the Int.
Conf. on Mobile Computing and Networking (MOBICOM), pp. 45–57 (2004)
[13] Kaemarungsi, K., Krishnamurthy, P.: Modeling of indoor positioning systems
based on location fingerprinting. In: Proc. IEEE Conf. on Computer Commu-
nications (INFOCOM), pp. 1012–1022 (2004)
 Algorithms for Location Estimation Based on RSSI Sampling 85

[14] Kaemarungsi, K., Krishnamurthy, P.: Properties of indoor received signal strength
for WLAN location fingerprinting. In: Proc. Int. Conf. on Mobile and Ubiquitous
Systems (MOBIQUITOUS), pp. 14–23 (2004)
[15] Karp, B., Kung, H.T.: GPSR: Greedy perimeter stateless routing for wireless
networks. In: Proc. of the Int. Conf. on Mobile Computing and Networking (MO-
BICOM), pp. 243–254 (2000)
[16] Krohn, A., Hazas, M., Beigl, M.: Removing systematic error in node localisation
using scalable data fusion. In: Langendoen, K.G., Voigt, T. (eds.) EWSN 2007.
LNCS, vol. 4373, pp. 341–356. Springer, Heidelberg (2007)
[17] Kuo, S.-P., Tseng, Y.-C., Wu, F.-J., Lin, C.-Y.: A probabilistic signal-strength-
based evaluation methodology for sensor network deployment. In: Proc. Int. Conf.
on Advanced Information Networking and Applications (AINA), pp. 319–324
(2005)
[18] Levis, P., Lee, N., Welsh, M., Culler, D.: TOSSIM: accurate and scalable simula-
tion of entire TinyOS applications. In: Proc. Int. Conf. on Embedded Networked
Sensor Systems (SENSYS), pp. 126–137 (2003)
[19] Lorincz, K., Welsh, M.: MoteTrack: A robust, decentralized approach to RF-based
location tracking. Personal and Ubiquitous Computing 11(6), 489–503 (2007)
[20] Lymberopoulos, D., Lindsey, Q., Savvides, A.: An empirical characterization of ra-
dio signal strength variability in 3-D IEEE 802.15.4 networks using monopole an-
tennas. In: Römer, K., Karl, H., Mattern, F. (eds.) EWSN 2006. LNCS, vol. 3868,
pp. 326–341. Springer, Heidelberg (2006)
[21] Lynch, C., Reilly, F.O.: Processor choice for wireless sensor networks. In: Proc.
ACM Workshop on Real-World Wireless Sensor Networks (REALWSN), pp. 52–
68 (2005)
[22] Maróti, M., Völgyesi, P., Dóra, S., Kusý, B., Nádas, A., Lédeczi, Á., Balogh, G.,
Molnár, K.: Radio interferometric geolocation. In: Proc. Int. Conf. on Embedded
Networked Sensor Systems (SENSYS), pp. 1–12 (2005)
[23] Moore, D., Leonard, J.J., Rus, D., Teller, S.J.: Robust distributed network local-
ization with noisy range measurements. In: Proc. Int. Conf. on Embedded Net-
worked Sensor Systems (SENSYS), pp. 50–61 (2004)
[24] Nagpal, R., Shrobe, H.E., Bachrach, J.: Organizing a global coordinate system
from local information on an ad hoc sensor network. In: Zhao, F., Guibas, L.J.
(eds.) IPSN 2003. LNCS, vol. 2634, pp. 333–348. Springer, Heidelberg (2003)
[25] Nasipuri, A., Li, K.: A directionality based location discovery scheme for wireless
sensor networks. In: Proc. ACM Int. Workshop on Wireless Sensor Networks and
Applications (WSNA), pp. 105–111 (2002)
[26] Niculescu, D., Badrinath, B.R.: Ad hoc positioning system (APS) using AOA. In:
Proc. IEEE Conf. on Computer Communications (INFOCOM), pp. 1734–1743
(2003)
[27] Niculescu, D., Nath, B.: DV based positioning in ad hoc networks. Telecommuni-
cation Systems 22, 267–280 (2003)
[28] Ochi, H., Tagashira, S., Fujita, S.: A localization scheme for sensor networks based
on wireless communication with anchor groups. In: Proc. Int. Conf. on Parallel
and Distributed Systems (ICPADS), pp. 299–305 (2005)
[29] Prasithsangaree, P., Krishnamurthi, P., Chrysanthis, P.K.: On indoor position
location with wireless LANs. In: Proc. IEEE Int. Symposium on Personal, Indoor,
and Mobile Radio Communications (PIMRC), pp. 720–724 (2002)
[30] Ramadurai, V., Sichitiu, M.L.: Localization in wireless sensor networks: A proba-
bilistic approach. In: Proc. Int. Conf. on Wireless Networks (ICWN), pp. 275–281
(2003)
86 C. Papamanthou, F.P. Preparata, and R. Tamassia

[31] Rappaport, T.S., Rappaport, T.: Wireless Communications: Principles and Prac-
tice, 2nd edn. Prentice-Hall, Englewood Cliffs (2001)
[32] Roos, T., Myllymaki, P., Tirri, H., Misikangas, P., Sievanen, J.: A probabilistic
approach to WLAN user location estimation. International Journal of Wireless
Information Networks 9(3), 155–166 (2002)
[33] Savvides, A., Han, C.-C., Strivastava, M.B.: Dynamic fine-grained localization in
Ad-Hoc networks of sensors. In: Proc. of the Int. Conf. on Mobile Computing and
Networking (MOBICOM), pp. 166–179 (2001)
[34] Savvides, A., Park, H., Srivastava, M.B.: The bits and flops of the n-hop multi-
lateration primitive for node localization problems. In: Proc. ACM Int. Workshop
on Wireless Sensor Networks and Applications (WSNA), pp. 112–121 (2002)
[35] Shen, X., Wang, Z., Jiang, P., Lin, R., Sun, Y.: Connectivity and RSSI based
localization scheme for wireless sensor networks. In: Huang, D.-S., Zhang, X.-P.,
Huang, G.-B. (eds.) ICIC 2005. LNCS, vol. 3645, pp. 578–587. Springer, Heidel-
berg (2005)
[36] Sichitiu, M., Ramadurai, V.: Localization of wireless sensor networks with a mobile
beacon. In: Proc. IEEE Conf. on Mobile Ad-hoc and Sensor Systems (MASS), pp.
177–183 (2004)
[37] Snyman, J.A.: Practical Mathematical Optimization: An Introduction to Ba-
sic Optimization Theory and Classical and New Gradient-Based Algorithms.
Springer, Heidelberg (2005)
[38] Stoleru, R., Vicaire, P., He, T., Stankovic, J.A.: StarDust: a flexible architec-
ture for passive localization in wireless sensor networks. In: Proc. Int. Conf. on
Embedded Networked Sensor Systems (SENSYS), pp. 57–70 (2006)
[39] Texas Instruments. MSP430C13x1 Datasheet (Revised September 04, 2004)
[40] Wackerly, D., Mendenhall, W., Scheaffer, R.: Mathematical Statistics with Appli-
cations, 6th edn. Duxbury Advanced Series (2002)
[41] Whitehouse, K., Karlof, C., Culler, D.: A practical evaluation of radio signal
strength for ranging-based localization. In: ACM Mobile Computing and Com-
munications Review, pp. 41–52 (2007)
[42] Xiang, Z., Song, S., Chen, J., Wang, H., Huang, J., Gao, X.: A wireless LAN-based
indoor positioning technology. IBM J. Res. Dev. 48(5/6), 617–626 (2004)
[43] Yedavalli, K., Krishnamachari, B., Ravula, S., Srinivasan, B.: Ecolocation: a se-
quence based technique for RF localization in wireless sensor networks. In: Proc.
Int. Conf. on Information Processing in Sensor Networks (IPSN), p. 38 (2005)
[44] Yu, Y., Govindan, R., Estrin, D.: Geographical and energy aware routing: A re-
cursive data dissemination protocol for wireless sensor networks. Technical Report
UCLA/CSD-TR-01-0023, UCLA Computer Science Department (2001)