Localization based on Signal Strength

Using Kalman Approach

P.T.V.Bhuvaneswari1, V.Vaidehi1, M.AgnesSaranya1
1
Madras Institute of Technology, Anna University,
Chennai, India
ptvbmit@annauniv.edu, vaidehi@annauniv.edu,
saran_1131@yahoo.co.in

Abstract:- This paper proposes a distributed localization algorithm

based on Received Signal Strength (RSS) that consists of two
phases, distance estimation phase and coordinate estimation phase.
In distance estimation phase the distance of the unknown node is
computed based on the RSS measurements. The effect of path loss
and attenuation in the wireless medium are also considered in the
proposed algorithm. The distance estimation uses log normal
shadowing path loss model and ITU indoor attenuation model. The
distance error is minimized using one-dimensional Kalman filter.
The number of iterations in Kalman filter is limited using Cramer
Rao Bound (CRB) value. In the second phase the coordinates of the
unknown node is estimated by lateration technique. The accuracy in
lateration is improved by min-max algorithm. The RSS value is
experimentally obtained in real-time indoor environment using
zigbee series 1 RF module. The proposed algorithm is simulated
and analyzed in MATLAB version 7. From the simulation results it
is found that the proposed localization algorithm performs more
efficient in terms of computational cost and accuracy.

Keywords: Received signal strength, log normal shadowing model,

ITU model, one-dimension Kalman estimator, Cramer Rao bound,
min-max algorithm.

1 Introduction:
Localization in Wireless sensor networks makes the
sensed data more meaningful. It also supports in designing
network layer services, such as topology control, routing,
and clustering.
Using nodes equipped with GPS (Global
Positioning System) receiver is costly and not energy
efficient [4], [5], [8]. Hence there is a need for an efficient
distributed location finding algorithm to determine node’s
positions using distance measurements between neighboring
nodes. The determination of location information requires
two types of nodes namely, the beacon node and the
unknown node. The beacon node is a node that is aware of
its location while unknown node is the one whose locations
is to be determined.
The distance measurement between the beacon
nodes and unknown nodes can be done by three techniques
[4]. In all the three techniques, the location information of a
set of beacons and some pair-wise distances between the
beacons and unknown node are needed. Generally to
determine the two dimensional coordinates of a node,
minimum three numbers of reference coordinates are
required [4]. In this paper, an efficient distributed
localization algorithm based on RSS measurement is
developed which provides an accurate and cost-effective
solution.
The rest of the paper is organized as follows.
Section 2 presents the proposed localization algorithm with
relevant mathematical formulations. Section 3 presents the
simulation results and performance analysis of the proposed
algorithm. Finally, section 4 concludes the paper with future
work.

2 Proposed Location Finding Algorithm:

Consider a sensor network comprising of Sn nodes
located at the coordinates (xn, yn), where n represents the
total number of nodes in the network. Assume that there are
Sk anchor nodes whose positions are known (eg. obtained via
GPS or some other “absolute” reference) and that the
remaining Su =Sn – Sk nodes are located at unknown
positions. Let  1,  2 be the coordinates of the nodes whose
positions are known and unknown respectively.
 1 = {(xk,yk)} , k = S1,S2,….,Sk
 2 = {(xu,yu)} , u = Sk+1,…., Sn (1)
The localization problem considered is the
estimation of the coordinates  2 given the coordinates of
the anchor node  1 and many pair-wise range measurement
{ γ ku(t)} taken over time t=1,2…,N where N represents
number of samples. Table 1 presents the list of symbols
along with descriptions used in this paper. The proposed
location estimation algorithm is shown in Figure 1.
2.1 Range Measurement and Statistical Modeling:
The Received Signal Strength values between
anchor node and unknown node are measured in indoor
environment. RSS measured values consist of more error
components due to multi-path propagation, reflection, and
fading effects. To enhance accuracy in the range
measurement, more samples are taken at different time
instances for different distances. The commercially available
Zigbee series 1 RF module supports 16 numbers of channels
in the 2.4 GHz ISM band. The channels considered for the
range measurement are B, C, D, E and F [3].
Statistical modeling is done to evaluate the quality
of channel. The various statistical parameters that are
evaluated are maximum, minimum, mean, median, standard
deviation, lower quantile and upper quantile [1]. The channel
that results with low standard deviation compared to other
channels is chosen as the best channel.

2.2 Distance Calculation:

The distance of the unknown node with respect to
anchor node can be computed from the ensemble mean RSS
value of the best channel. The indoor environment
considered is destructive as it consists of thick and thin walls,
wooden and steel furniture, floorings, roofings etc. Further
the effect of interference from WLAN and Bluetooth devices
present in the environmental set up makes the medium more
destructive. So, the effect of path loss, attenuation and
interference are also included in calculating the distance
parameter [7]. The distance of unknown node is computed
using two models namely path loss log normal shadowing
and ITU indoor attenuation models. Then the distance error
which is the difference between the actual distance and the
calculated distance is computed.

2.2.1 Log normal shadowing path loss model:

The path loss between node k and node u is random
and distributed log normally which can be modeled as [8],
[10].
PL (dku’) [dB] = PL (dref) +10nplog (dku’/dref) +X  (2)
where PL (dref) is the ensemble path loss at a short reference
distance dref and X  is the zero mean Gaussian random
variable with standard deviation  and np is path-loss
exponent, typically lies between 2 and 4.
The ensemble mean received power is the
difference of the transmitted power and the path loss
component at the specific distance dku’ which is modeled as,
Pr (dku’)[dBm] =Pt (dBm) - PL(dku’)[dB] (3)
Substituting eqn (2) in (3),
Pr (dku’)[dBm] = [ Pt (dBm) - PL (dref) –
10nplog(dku’/dref) - X  ] (4)

where Pr (dku’) is the average RSSI value in dBm or received

power and Pt(dBm) is the power transmitted.
2
PL (dref) = - 10 log (GtGr  / (4π)2 dref 2 ) (5)
where Gt, Gr = Gain of the transmitter and receiver
antennas respectively

 = wavelength of the signal by c /f

c = Velocity of light in m/s
f =frequency of the signal in Hz.
The distance dku’ computed from log-normal
shadowing model can be expressed as
[(Pt (dBm) - (dref) - X  -
dku’ = dref (10 PL
Pr (dku’)[dBm] ) (10 np]
/ ) (6)

2.2.2 ITU indoor attenuation model:

The amount of attenuation is predominant in an
indoor environment. The range calibration can also be
computed by an attenuation model. The ITU indoor
attenuation model [11] is formally expressed as
PL (dref)[dB] = 20 log f + L log dref + Pf (frku ) – 28 (7)
where PL (dref )[dB] = ensemble mean path loss at the
reference distance in dB
L = distance power loss coefficient
frku = Number of floors between the node k
and node u
Pf ( frku ) = floor loss penetration factor.
The distance dku ’’ computed from ITU indoor attenuation
model can be obtained by substituting eqn. (7) in eqn. (6)
[(Pt (dBm) – 20 log f - L log dref - Pf (frku ) +
dku’’ = [dref (10
28- X  - Pr (dku’)[dBm] ) / (10 np)]
)] (8)
2.3 Distance Estimation:
The accuracy of the distance calculated by the two
models can be improved by minimizing the distance error
between the actual distance and the calculated distance. The
accurate distance can be estimated through one-dimensional
Kalman estimator [9]. The one-dimensional Kalman
estimator is modeled as average filter that combines, the
distance values calculated by model 1 and 2. The distance
values for the two models are obtained after statistical
modeling by which the best channel selection is done.
Let dku’ be the range obtained by log-normal
shadowing model (Model 1) and dku” be the range obtained
by ITU indoor attenuation model (Model 2). Let  1 and  2
be the standard deviations obtained from model1 and model
2 respectively. Both errors can be combined to get an
estimation using Kalman algorithm [9].
 1 = standard deviation for dku’ about the mean which is
given by
 1 = (((dku’–m)2 + (dku–m)2) /(nd -1))1/2 (9)
where m = (dku’ + dku)/2
nd = number of samples
 2 = standard deviation for dku” about the mean which is
given by
 2 =(((dku’’ –m1)2 +(dku –m1)2)/ (nd -1))1/2 (10)

where m1 = (dku’’ + dku ) / 2

If both models have same standard deviation, then
1 =  2. If path loss is predominant, then 1 > 2. Similarly,
if attenuation is high, then  1 <  2. The complexity of the
proposed algorithm increases if the error in the calculated
distance increases. The iterations can be limited by the
amount of errors. If  1 =  2 , the estimated distance will be
the average of dku’ and dku”. If  1 >> 2, then dku” can be
taken as the estimated distance. If  1 <<  2, then dku’ can
be taken as the estimated distance [9].
 In any case, the weighted average of the distances
calculated from both the models are used to estimate dku,
^
which may be called as d k u
^ d ku ’ d ku ’’ 1 1
d ku = (  ) ( 2  2) (11)
1 2 1 2
(Or)
^
d k u = (d ku ’2  d ku ’’1 ) (1   2 )
2 2 2 2
(12)

The above estimate can also be rewritten as

^
d k u = dku’ + K(dku ’’ – dku’) (13)

12
where K = is defined as the Kalman gain.
(12  2 2 )
^
The accuracy of the estimator d ku can be determined based
on Carmer Rao bound value [3].
The CRB value is given below,
2
CRB bound = (14)
nd
where  =(  1+  2)/2
The iterations of the estimator is repeated till
following condition is satisfied,
^
Cov ( d k u ) ≥ CRB bound (15)
The estimated distance for every iteration is
compared with the actual distance. The iterations are
repeated till the variance reaches the CRB bound. The value
obtained in each iteration of the proposed algorithm is given
in Table 4.

2.4 Co-ordinates Estimation:

Lateration is one of the most popular techniques for
node positioning in Wireless sensor network. The lateration
technique focused in this research is tri-lateration. As
suggested in the literature [2], [4] to determine the node’s
coordinates in 2-dimension, a minimum of three anchor
nodes are required. Assume that there are three anchor nodes
with known positions (xk ,yk) where k =1,2,3 a node at
unknown position (xu ,yu ) and estimated distance value
^
d k u where k = S1,S2,….,Sk, and u = Sk+1,….,Sn.
Using the model of multi-lateration, the 2-D
coordinates of the unknown node can be given as
 xu   yn  yn  1  y 1  y n 
 
  2     
 y u  =
 
 x n  1  x n  x n  x 1 
 (r 1  rn )
2 2
 ( x1  xn )
2 2
 ( y  yn 2 )
2 
 1

. . . . .  (16)
 
. . . . . 
 
  ( xn  12  xn 2 )  ( yn  12  yn 2 )
( rn  1  rn )
2 2

where (xu,yu) be the coordinates of the unknown node, (xi,yi),

i=1,2,….n be the coordinates of the known nodes,
ri,i=1,2,…n be the distance between the node i and node u.
The perfect intersection of circles in trilateration is not
possible as the distance measurements are not accurate.This
result in overlapping of circles and the location of the node
can be anywhere in the overlapping region. To resolve this
problem an efficient min-max bounding box concept is used
in the proposed algorithm.

2.5 Min-max Algorithm:

In the proposed location finding algorithm, the
anchor node’s radio ranges are modeled as circular cells [6].
By trilateration algorithm the estimation of coordinates can
definitely lie in the minimum area that is bounded by the
overlapping of variable sized cells. To simplify calculation
without considerably degrading the accuracy, the cells are
modeled as squares. Furthermore the radio ranges of the
anchor nodes can be mapped to radio coverage circles and
square Positioning Cells (PC). A square can be fully
constructed by determining the coordinates of its bottom left
and top right corners. The set of PCs used in coordinate
estimation is denoted by PCS.
In order to find the Final Bounding Box (FBB)
containing the unknown nodes, the intersection of PCs in
one PCS needs to be determined.
FBB is given by

FBB =  PCn (17)

Equation (18) can be solved by min-max algorithm.
Min-max relies on the fact that the intersection of all PCs
can be acquired by taking the maximum of all coordinate
minimums and the minimum of all maximums.
FBB = {(max (Xbl), max (Ybl), min (Xtr), min (Ytr)} (18)
where Xbl is the set of bottom left x-coordinates, Ybl is the set
of bottom left y coordinates. Xtr is the set of top right x-
coordinates and Ytr is the set of top right y-coordinates in
PCs contained in PCS. The method developed is general
enough to adapt to any measurements methods like ToA,
TDoA, RSS or proximity.

3 Simulation Results:
The experimentation is done in indoor environment
using zigbee series 1 RF module and the associated X-CTU
software of MAXSTREAM. 20 samples of RSS values are
measured at 20 different time instances for a specific
distance. The experiment has been repeated for five different
channels (B, C, D, E and F) with five different frequencies.
Figure 2 illustrate the relationship between distance and RSS
measurement for channels B. It is seen that the Received
Signal Strength of the unknown node decreases as distance
between the anchor node and unknown node increases.
Similar relationship can be obtained for the remaining
channels.
Figure 3 shows the statistical modeling of channel
B in which the various statistical parameters like minimum,
maximum, mean, standard deviation, lower quantile and
upper quantile are computed with the measured RSS values.
Similar analysis can be performed for the other channels as
shown in Table 2. Here the standard deviation is considered
as the deciding factor to quantify the quality of the channel.
It is found that the channel E can be selected as the best
channel as it possesses low standard deviation.
The distance of the unknown node is calculated by
both log normal shadowing path loss model with path loss
exponent np = 4 and ITU indoor attenuation model. Figure 4
illustrates the calculated distance obtained by both the
models. The distance error is computed by both the models.
 Figure 5 shows distance estimation with and with
out one-dimensional Kalman filter. Table 3 presents the
results of Kalman estimated distance and the associated error
percentage. It can be seen that the error has been drastically
reduced using Kalman estimator. The virtual coordinates of
the unknown nodes are determined by trilateration method.
Table 5 shows the result obtained with and without min-max
bounding box algorithm. It is found that accuracy is
improved through min-max bounding box algorithm.
The computational complexity of the proposed
algorithm is estimated as given in Table 6. So, the total
computation is (9A+2D) + (2A+5D) I. The division
operations involved in equation 6, 8 and 11 can be converted
into multiplications by performing the division once and
saving the quotient in temporary memory location to reduce
the computational complexity. As division takes more time
than addition and multiplication, computation of the
proposed algorithm will be O((2A+I(5D)) where I is the
number of iteration which is limited by the CRB value.

4 Conclusion:
The existing solutions of localization are costly and power
consuming due to the usage of GPS receiver.. Further the
solutions are not accurate as they do not consider effect of
wireless medium characteristics. In this paper, a novel cost-
effective accurate distributed location finding algorithm is
proposed that provides solution for the above mentioned
problems. The proposed work has following features to get
an efficient accurate estimation of a node’s position. It uses
path loss and attenuation models to compute the distance of
the nodes accurately. The error is minimized with one-
dimensional Kalman algorithm in which two models are
integrated to get better results. Further the one-dimensional
Kalman is not computationally heavy. The proposed
algorithm can be periodically updated to address the
topological changes occuring in the network due to mobility
of the nodes or failure of the nodes.
 RSS Measurement Ref distance

Statistical modeling Distance Distan

Calibration ce

Coordinate Decision 1D
estimation Kalman
making estimator

Min- Anchor Cramer rao

max coordinates bound

Figure1: Block Diagram of Localization Algorithm

RSS vs Distance for channel B
85
80
75
70
65
60
55
RSS(-dBm)

50
45
40
35
30
25
20
15
10
5
0
0 1 2 3 4 5 6 7 8 9 10 11
Distance(m)

Figure 2: RSS vs Distance in Channel B

Table 6: Computational Complexity

I* -Index value of Iterations
Equ. Operations
No. Divisions Additions Loop
6 1 3 ---
8 1 6 ---
11 5 2 I*
 Table 1: List of symbols

Notation Description

Sn=Sk+Su Total number of nodes in the wireless sensor

network
Sk Number of anchor nodes
Su Number of unknown nodes
1 Coordinates of anchor nodes, {(xk,yk)}, k =
S1,S2,….,Sk
2 Coordinates of unknown nodes, {(xu,yu)}, u =
Sk+1,…., Sn
dku Actual distance between the anchor node and
the unknown node
dku’ Distance estimated by the log normal
shadowing model
dku’’ Distance estimated by the ITU indoor
^ attenuation model
d ku Distance estimated by one dimensional
Kalman estimator
np Path loss exponent
dref Short reference distance
PL (dref ) Ensemble path loss at the short reference
distance in dBm
Pr (dku’) Received power at distance dku’ in dBm
Pt (dBm) Transmitted power in dBm
Gt,Gr Antenna gain of transmitter and receiver
 Wavelength of signal in meter
f Frequency of signal in Hz
L Distance power loss coefficient
Pf ( frku ) Floor loss penetration factor
frku Number of floors between node k and node u
 1 Standard deviation of error by log normal
shadowing
 2 Standard deviation of error by ITU indoor
attenuation model
K Gain of the one dimensional Kalman filter
 Average standard deviation obtained from Log
normal shadowing and ITU model
 nd Number of samples used in one dimensional
Kalman filtering

PC Positioning cell
PCS Set of positing cells
Xbl Set of bottom left x-coordinates
Ybl Set of bottom left y-coordinates
Xtr Set of top right x-coordinates
Ytr Set of top right y-coordinates

Statistical Modeling-Channel B
85
80
75
70
65
60
55
RSS(-dBm)

50
Min
45
40 Max
35
30 Mean
25 Standard
20 Deviation
15 Low Quantile
10
High Quantile
5
0
2 4 6 8 10
Distance(m)

Figure3. Statistical modeling of Channel B

Table 2: Standard deviation values of RSS measurements for

different channels

Dist B C D E F
(m) (2.404- (2.409- (2.414- (2.419- (2.424-
2.406) 2.411) 2.416) 2.421) 2.426)
GHZ GHz GHz GHz GHz
2 2.39 2.13 2.62 0.92 1.69
4 1.87 1.65 1.53 1.23 1.38
6 2.13 1.67 1.26 1.37 2.05
8 1.98 1.29 1.04 1.31 1.92
10 1.11 1.44 0.96 0.61 1.18
 Calculated distance vs Actual distance
11
10
9

Calculated distance(m)
8
7
6
5
4
3
Log Normal
2 Shadowing
1 ITU Model

0
2 4 6 8 10
Actual distance(m)

Figure 4. Calculated distance vs Actual distance

Kalman Filter - Estimated distance vs Actual distance

11
10
9
Estimated distance(m)

8
7
6
5
4
3
2 With Kalman
1 without Kalman
0
2 4 6 8 10
Actual distance(m)

Figure 5. Comparison of distance estimation

Table 4: Iteration algorithm of Kalman Filter

Act. CRB Iteration 1 Iteration 2 Final
dist. bound Est. Var Est. Var dist
(m) dis(m) (m) dis(m) (m) (m)
2 0.0553 2.0603 0.0603 2.0502 0.0502 2.0502

4 0.2607 3.8476 0.1524 - - 3.8476

6 0.7601 5.4471 0.5529 - - 5.4471

8 1.0751 7.6493 0.3507 - - 7.6493
10 1.3050 10.029 0.0298 - - 10.029
8
 Table 3: One-dimension Kalman estimator
Act Err % of Err % of % of
dist. (without error (With error imp. Acc.
(m) KF) with- KF) with KF
out
KF
2 0.4705 23.5 0.0502 2.51 89

4 1.0213 25.53 0.1524 3.81 85

6 1.7437 29.06 0.5529 9.21 68

8 2.0737 25.92 0.3507 4.38 83

10 2.2848 22.85 0.0298 0.29 98

Table 5: Estimation of Coordinates of the unknown node

Distance Coordinates Coordinates
between of the of the
anchor unknown unknown
Anchor node
node and node(xu,yu)
coordinates (xu,yu)
(x,y) unknown without with min-
node(m) min-max max
algorithm algorithm

A1 (2,1) 2.0502

A2 (5,4) 3.8476 (2.9,1.3) (3.3,1.6)

A3 (8,2) 5.4471

Acknowledgement:

The first three authors acknowledge Tata Consultancy

Services for funding this Project.

References
1. Abdalkarim Awad, Thorsten Frunzke, Falko Dressler.: Adaptive
distance estimation and localization in WSN using RSSI measures,
Proceedings of 10th EUROMICRO Conference on Digital System
Design - Architectures, Methods and Tools, pp. 471-478,(2007).
2. Avinash Srinivasan, Jie Wu.:A survey of secure localization in
wireless sensor networks, Encyclopedia of Wireless and Mobile
Communications, Florida Atlantic University,(2008)
3. Erik G. Larsson .:Cramer-Rao bound analysis of distributed
positioning in sensor networks, IEEE Signal Processing Letters,
Vol. 11,No. 3,pp.334-337,(2007).
4. Holger Karl and Andreas Willig.: Protocols and Architectures
for Wireless Sensor Networks, Wiley Publications,(2006).
5. Jonathan Bachrach and Christopher Taylor.: Localization in
sensor networks, Handbook of sensor networks: algorithms and
architectures, Edition I, vol.1, Wiley Publications, Chapter 9,(2005).
6. Kiran Yedavalli, Bhaskar Krishnamachari and Lakshmi
Venkatraman.: Fast/Fair Mobile Localization in Infrastructure
Wireless Sensor Networks, Mobile Computing and Communications
Review, Vol.11, No. 1, pp.29-40,(2007).
7. Neal Patwari and Alfred O. Hero.: Demonstrating Distributed
Signal Strength Location Estimation, SenSys’06, pp.353-354,(2006).
8. Prasan Kumar Sahoo, I-Shyan Hwang and Shi-Yao Lin.: A
Distributed Localization Scheme for Wireless Sensor Networks,
International Conference On Mobile Technology, Applications, And
Systems, Vol:6, Article No. 77, No: 7, pp.1031-1050,(2008).
9. Raul Rojas.: The Kalman Filter, Technical report, Freie
University of Berlin,(2003).
10. Theodore S.Rappaport.: Wireless Communications-Principles
and Practice, Edition II, Prentice-Hall of India Publications.
11. ITU Model for Indoor Attenuation, www.wikipedia.com,