You are on page 1of 5

The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’07)

A JOINT ESTIMATION OF TARGET LOCATION AND CHANNEL MODEL


PARAMETERS IN AN IEEE 802.15.4-BASED WIRELESS SENSOR NETWORK
Radim Zemek†, Shinsuke Hara††, Kentaro Yanagihara‡ and Ken-ichi Kitayama†
†Graduate School of Engineering, Osaka University, Osaka, Japan
††Graduate School of Engineering, Osaka City University, Osaka, Japan
‡Corporate R&D Center, Oki Electric Industry Co., Ltd., Osaka, Japan

A BSTRACT curacy of this method is superior to the accuracy of trilatera-


tion [7].
Many target localization techniques based on received signal The RSSI ranging technique is based on the principle of sig-
strength indicator (RSSI) assume a prior knowledge of a chan- nal power attenuation with transmission distance. The power
nel model and its parameters for an area where a target node attenuation is described by a path loss model which is charac-
is localized. This is a limiting factor since an intensive pre- terized by two parameters, as presented in Section II. Several
measurement campaign is required to determine the model and methods to estimate a node location based on RSSI have been
its parameters. proposed [8–12]. However, their main disadvantage is that the
Basing on channel measurement campaigns in different ar- parameters need to be measured prior to using this technique
eas we confirmed that the variation in the RSSIs of the IEEE as the values of the parameters are particular to every envi-
802.15.4 signal can be described by a two-layer model. We ronment. Incorrect path loss model parameters can drastically
therefore propose a novel target localization method with no affect the location estimation accuracy [8], [13].
prior knowledge of model parameter values. The method is To overcome the disadvantage of the RSSI ranging tech-
based on a joint maximum likelihood target localization and nique, we propose a novel joint target location and channel
channel model parameters estimation. The proposed method model parameters estimation method based on ML estimation.
was experimentally verified in real environments and the re- The advantage of our proposed method is that the target node
sults show that the location estimation accuracy outperforms localization based on RSSI does not require pre-measurement
the accuracy of the conventional method where the values of campaign to obtain the channel model parameters. The per-
the channel model parameters are known in advance. formance of the proposed method was verified by target node
Keywords - target localization, channel parameter estima- localization experiments conducted in a corridor of a shopping
tion, RSSI, IEEE 802.15.4 center and in an office room. The results show that the location
estimation accuracy of the proposed method is higher than the
I. I NTRODUCTION accuracy of a conventional method which uses a prior knowl-
edge of the path loss model parameters.
Node localization using wireless sensor networks (WSNs) is
a very attractive application of WSNs and the research in this II. P ROBLEM S TATEMENT AND C ONVENTIONAL
area is very active. Application areas range from interactive L OCALIZATION M ETHOD
museums and customer observation to object detection and ve-
hicle tracking [1]. In addition, node localization is important Localization method based on RSSI utilizes the principle of
in a WSN setup phase, and knowledge of the node locations is signal power attenuation with transmission distance. The aver-
essential to effective routing [2]. aged received signal power P at distance d is expressed by a
A ”target node” location is estimated through ”anchor path loss model
nodes” whose locations are known. To estimate a target node
location in a plane, it is necessary to measure the distances be- P (d) = α · d−β (mW) (1)
tween a target node and at least three anchor nodes. Several
where α and β are the path loss model parameters. The param-
ranging techniques exist to measure the distance, such as time
eter α is a constant proportional to measured signal power at a
of arrival (TOA), time difference of arrival (TDOA), angle of
certain distance, usually 1 meter, from transmitter. Parameter β
arrival (AOA), and received signal strength indicator (RSSI).
is a path loss attenuation factor. We have conducted measure-
The details of these techniques have been described elsewhere
ment campaigns of the two parameters in various environments
[3]. We selected the RSSI technique because of its simplic-
both with and without human presence. The values of the pa-
ity and inexpensive implementation. Moreover, the nodes used
rameters vary between environments and conditions as shown
in the experiments were based on the IEEE 802.15.4 standard
in Tab. 1.
which supports the RSSI technique as a link quality measure-
Conventional ML estimation method assumes a prior knowl-
ment [4]. edge of channel model and its parameter values in the tar-
The measured distances between the target and anchor nodes get node environment (location estimation area). To be able
are used in a localization estimation method such as trilatera- to use ML estimation, we have to construct a joint condi-
tion [5] or maximum likelihood (ML) estimation [6]. We chose tional probability density function (PDF) p(P|d). Vector P =
ML estimation over trilateration, because the estimation ac- [P11 , . . . , PM
N
] is composed of k-th measured received signal

1-4244-1144-0/07/$25.002007
c IEEE
The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’07)

power at j-th anchor node Pjk (RSSIs), where M is the total -5


10 1.0

Probability Density Function p(P|P(d))


measured power P measured data
0.9

Received Signal Power P (mW)


number of anchor nodes and N is the total number of RSSI -6 -1.8907
P(d) = 2⋅10 d exp(-P/P(d))
0.8
measurements. Vector d = [d1 , . . . , dM ] describes the dis- -6
10
0.7
tances dj between the target node at [X, Y, Z] and j-th anchor 0.6
node at [xj , yj , zj ] and is defined as -7
10 0.5

 0.4

dj (X, Y, Z) = (X − xj )2 + (Y − yj )2 + (Z − zj )2 . (2) -8
10
0.3
0.2
0.1
We assume that the heights of target node Z is known, and -9
10 0.0
therefore the target node location is estimated in a plane. How- 1 10 100 0 2 4 6 8 10
Distance d (m) Normalised Received Signal Power P/P(d)
ever, the method can be easily extended to localization in three- (a) (b)
dimensional space.
The conditional PDF expresses the probability of receiving
Figure 1: Typical (a) power attenuation and (b) fading charac-
signal of power Pjk given the distance dj . For simplification,
teristics.
we assume that the received signal powers Pjk are temporally
and spatially independent of each other, and thus the joint con-
ditional PDF is written as Note that the ML estimation using this log-likelihood function
assumes a prior knowledge of the α and β model parameters

M 
N
for the given environment obtained by a pre-measurement cam-
p(P|d) = p(Pjk |dj ). (3)
j=1 k=1 paign.
The log-likelihood function L(d) have a maximum at ML
The variation in received signal power (RSSI) of the IEEE estimates X̂ and Ŷ , which represent the estimated target node
802.15.4 signal is given by the decreasing exponential func- location in a plane. The maximum occurs when the partial
tion [8]: derivates of the log-likelihood function L(d) with respect to
  X and Y are equal to 0
1 Pjk
p(Pjk |dj ) = exp − . (4)    
α · d−β
j α · d−β
j
∂L(d)
=0
∂L(d)
= 0. (5)
∂X X=X̂ ∂Y Y =Ŷ

Equations (1) and (4) show that the variation in RSSI of the The solution of these two nonlinear equations, obtained by
IEEE 802.15.4 signal can be described by a two-layer chan- Newton-Raphson method [14], is the estimated target node lo-
nel model defined by only α and β parameters. Namely, the cation X̂ and Ŷ solved with the prior knowledge of the channel
first layer describes the relationship between the average RSSI model parameter values.
and the distance expressed by (1), and the second layer de-
scribes the relationship between the variation in the instanta- III. P ROPOSED L OCALIZATION M ETHOD
neous RSSI and the distance expressed by (4). Figure 1(a)
shows a typical power attenuation characteristic; Fig. 1(b) We can see from Tab. 1 that the smallest correlation coeffi-
shows a fading of the signal as a function of Pjk /P (d) obtained cient ρ is equal to 0.77 which still means a good approxima-
tion. Therefore, in the following, we assume that the two-layer
in one of the measurement campaigns. Table 1 shows values of
channel model can be applied to the propagation characteris-
the correlation coefficient ρ between the measured data and the tics of the IEEE 802.15.4 signal for any unknown areas. This
two-layer channel model in various environments. assumption allows a parametric joint estimation of target nodes
The joint conditional PDF p(P|d) is called the likelihood locations and the channel model parameters using the two-
function l(d) if it is considered as a function of d. Because the layer channel model. We define a new joint conditional PDF
1 N
likelihood function evolves exponentially, it is more convenient p(P|d, α, β), where P = [P11 , . . . , POM ] represents a vector
to use log-likelihood function L(d) = ln l(d) = ln p(P|d). of k-th measured received signal power at j-th anchor node re-
ceived from i-th target node Pijk , O is the total number of target
nodes. Vector d = [d11 , . . . , dOM ] is composed of distances
dij between i-th target node and j-th anchor node defined as
Table 1: Model parameters and correlation coefficients.
Environment Model parameters 
dij = (Xi − xj )2 + (Yi − yj )2 + (Zi − zj )2 (6)
α β ρ
room 1with human presence 1.1 · 10−6 1.61 0.85 where [Xi , Yi , Zi ] and [xj , yj , zj ] represent the locations of i-
room 2with human presence 9.9 · 10−7 1.51 0.87 th target node and j-th anchor node, respectively. Again, we
room 3without human presence 1.4 · 10−5 3.51 0.90 assume that Zi is known and therefore a target node location is
room 4without human presence 8.0 · 10−7 1.16 0.78 estimated in a plane.
room 4with human presence 1.4 · 10−6 1.63 0.84 We assume that the received signal powers Pijk are indepen-
corridor 1without human presence 1.0 · 10−6 2.45 0.85 dent of each other, and so are target node locations. Conse-
foyerwithout human presence 2.0 · 10−6 2.29 0.81 quently the joint conditional PDF can be written as
foyerwith human presence 1.0 · 10−6 2.41 0.77
corridor 2without human presence 2.0 · 10−6 1.75 0.94 
O 
M 
N

corridor 2with human presence 2.5 · 10−6 1.89 0.91 p(P|d, α, β) = p(Pijk |dij , α, β). (7)
i=1 j=1 k=1
The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’07)

The conditional PDF p(P|d, α, β) is similar to (4). However, model parameters αn and βn
the two model parameters α and β are included in the condition 
as these parameters are estimated together (jointly) with target ∂L2 (d̂n , αn , βn )
nodes locations. This yields a conditional PDF defined as ∂αn
αn =α̂n
    

 O  M
1
N k
1 Pijk ∂ k Pij /N
p(Pijk |dij , α, β) = exp − = N ln − =0
α · d−β α · d−β
. (8) ∂αn i j αn · dˆ−β
ijn
n
αn · dˆ−β
ijn
n
ij ij αn =α̂n

∂L2 (d̂n , αn , βn )
This conditional PDF expresses the probability of receiving a ∂βn
signal of power Pijk given the distance dij and the path loss 
βn =β̂n
 

∂ 
O M N k
model parameters α and β. The log-likelihood function is 1 k Pij /N
L(d, α, β) = ln p(P|d, α, β). = N ln − = 0.
∂βn i j αn · dˆ−βn
ijn αn · dˆ−βn
ijn βn =β̂n
The proposed conditional PDF does not assume any prior
(10)
knowledge of the channel model parameters α and β, although
it assumes (1) and (4) as a valid channel model. Estimating Maximizing the log-likelihood function L2 (d̂n , αn , βn ) yields
the values of the parameters separately for each target node lo- the channel model parameter estimates α̂n and β̂n common to
cation yields different α and β parameters. This is not cor- the desired location estimation area. These two estimated chan-
rect since these parameters should be common to the given nel model parameters are used in stage one of the next iteration
location estimation area. Therefore, the maximization of the cycle n = n + 1 to re-estimate all O target locations Xn and
log-likelihood function L(d, α, β) can be solved by an iterative Yn . Next iteration cycle n = n + 1 starts after the stage two.
algorithm divided into two stages, namely two separate maxi- The iteration process, and therefore the estimation of both
mization problems. locations and parameters, is terminated when
In stage one, the locations Xn = [X1 , . . . , XO ] and Yn =
[Y1 , . . . , YO ] of O target nodes are estimated, where n is the |α̂n − α̂n+1 | < tolerance α
(11)
number of iteration cycle. The ML estimates X̂n and Ŷn |β̂n − β̂n+1 | < tolerance β
are obtained when the log-likelihood function of stage one
L1 (dn , αn−1 , βn−1 ) is maximized with respect to the target that is, when the estimated channel model parameters esti-
node locations Xi and Yi mated at iteration cycles n and n + 1 vary less than tolerance
  α and tolerance β. Figure 2 shows the flowchart of the algo-
∂L1 (dn , αn−1 , βn−1 ) rithm. Figures 3 and 4 in Section V show the average location
=
∂Xi Xi =X̂i estimation accuracy in meters against each iteration cycle n.
 O  
N k
 ∂  M
1 k Pij /N
N ln −β
− −β IV. E XPERIMENTAL S ETUP AND M EASUREMENT
i
∂Xi j αn−1 ·d n−1 αn−1 ·d n−1
ijn ijn Xi =X̂i C AMPAIGN
=0
The conventional and proposed methods were compared by ex-
  periments conducted in real environments. The environments
∂L1 (dn , αn−1 , βn−1 ) were a corridor of a shopping center and an office room. In
=
∂Yi Yi =Yˆi addition, the experimental measurement campaigns were con-
 O  
N k ducted with human activity in the environments which corre-
 ∂  M
1 k Pij /N
N ln −β
− −β sponds to a typical condition.
i
∂Yi j αn−1 ·d n−1 αn−1 ·d n−1
ijn ijn Yi =Yˆi The corridor was 6.0 meters wide and the experimental area
= 0. was 26.4 meters long. The height of the ceiling was 3.0 me-
(9) ters. The flooring material was granite. Shop windows and
entrances to shops were on both sides of the corridor. The ceil-
Equation (9) can be solved by maximizing the O pairs of joint ing of the corridor was covered by metal plates. We used 17
log-likelihood functions separately for each target node loca- wireless sensor nodes in total for the measurement campaign.
tion. The result yields the estimated target nodes locations X̂n Fifteen nodes represented anchor nodes fixed to the ceiling at
and Ŷn , and therefore the distances d̂n . Note that in the first the height of 2.70 meters in a 3 by 5 grid. Another node, rep-
iteration cycle n = 1, an arbitrary α0 and β0 initialization val- resenting a target node, was fixed to a tripod at the height of
ues are used. In the following iteration cycle n = n + 1, α̂n−1 0.8 meter. The last node was used as a sink node. This node
and β̂n−1 are used estimated in following stage two. was connected to a computer which stored the measured data
In stage two, the two channel model parameters αn and βn for future processing. We measured and collected data from 26
are estimated by maximizing the log-likelihood function with different target node locations.
respect to the parameters using the estimated target nodes lo- The office room was around 7.0 meters wide and 7.0 meters
cations X̂n and Ŷn from stage one. The ML estimates α̂n and long. The height of the ceiling was 2.61 meters. The office
β̂n are obtained when the log-likelihood function of stage two room contained wooden desks, plastic chairs and a computer
L2 (d̂n , αn , βn ) is maximized with respect to the two channel on every desk. The room walls were made of concrete and
The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’07)

Start, n = 0
V. R ESULTS AND D ISCUSSION
Figure 3 shows the average location estimation accuracy in me-
Initialize 0 and 0 values ters against the iteration cycle n for the experiment conducted
in the corridor of a shopping center. In addition, each curve
n=n+1 in the figure represents different α0 and β0 initialization val-
ues according to the legend. The first and the second values
stage ,
n-1 n-1
represent α0 and β0 , respectively. We can observe that the pro-
one posed method quickly converges to the same steady state av-
Estimate target node
locations Xn, Yn
erage location estimation accuracy regardless of the α0 and β0
initialization values.
Figure 4 shows the same characteristic for the experiment
stage Xn, Yn
two
Estimate channel model
conducted in the office room. The results are very similar to
parameters n, n the previous ones. The proposed method quickly converges to
the same average location estimation accuracy regardless of the
α0 and β0 initialization values.
In both of the environments, the steady state was reached
within five iteration cycles. This would satisfy (11) setting tol-
erance α to 1 · 10−8 and tolerance β to 0.001.
| n-1 - n| < tolerance No
| n-1 - n| < tolerance
In addition, we measured the true values of the channel
model parameters in the two environments to compare the ac-
Yes
curacy of conventional and proposed methods. Table 3 summa-
End rizes the results. A good agreement can be observed between
the values of measured parameters and estimated parameters
Figure 2: Flowchart of the algorithm. for the corridor, whereas there is a disagreement for the office
room. However, the purpose of the joint estimation is to ob-
tain the target nodes locations without a prior knowledge of
windows were at one side of the room. We used 11 wireless the channel model parameter values, therefore only the achiev-
nodes for the experiment. Nine nodes served as anchor nodes able location accuracy is of interest. We can observe from Tab.
fixed to the ceiling at the height of 2.61 meters in a 3 by 3 grid. 3 that the accuracy of the proposed method is around 10 cm
One node was placed on a box at the height of 0.73 meter which higher than the accuracy of the conventional method for both
represented the target node. The last node was again used as of the environment.
a sink node. The estimation was performed for 24 randomly
chosen target node locations.
VI. ACKNOWLEDGEMENT
The measurement campaign was conducted as follows. Tar-
get node was temporally placed to randomly selected but The authors would like to thank Daisuke Anzai for his valuable
known locations. During the measurement process the target comments and suggestions.
node location was not changed. The target node was transmit-
ting a packet at power level of 0 dBm every second for about
VII. C ONCLUSIONS
2 minutes. The packets were received by the anchor nodes
which measured the received signal powers and sent the mea- The proposed method is to our knowledge the first method ad-
sured data, represented as RSSIs, to a sink node. The sink node dressing the issue of target node location estimation using RSSI
was connected to a computer which stored the measured data. without a prior knowledge of channel model parameters. The
In our experiments, we used 10 RSSI measurements. advantage of the proposed method is that a common two-layer
The wireless nodes used in the experiments were based on channel model can be applied to any environment for the IEEE
the IEEE 802.15.4 standard. The nodes operated in 2.4 GHz 802.15.4 signal. In addition, the parameter values of the chan-
ISM band and were equipped with 1/4 wavelength monopole nel model do not have to be obtained in advance, thus avoiding
antenna. Table 2 summarizes the main node parameters. pre-measurement campaign to determine them.

Table 2: Node specification. Table 3: Comparison of conventional and proposed method


Frequency band 2.4 GHz (acc. - accuracy).
Transmission rate 250 kbit/s Location Conventional method Proposed method
Modulation OQPSK estimation Measured acc. Estimated acc.
Spreading technique DSSS (2 MHz) area α β (m) α β (m)
Antenna 1/4 wavelength monopole Corridor 2.5·10−6 1.89 2.34 3.6·10−6 1.93 2.24
Transmission power 0 dBm (1 mW) Office 1.1·10−6 1.61 1.34 7.5·10−6 2.71 1.27
The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC’07)

3.4 3.0
Average Location Estimation Accuracy (m)

Average Location Estimation Accuracy (m)


2.5E-6, 1.89, true 1.1E-6, 1.61, true
2.8
3.2 1.0E-7, 1.50 1.0E-7, 1.50
1.0E-7, 2.00 2.6 1.0E-7, 2.00
1.0E-7, 2.50 1.0E-7, 2.50
3.0 2.4
1.0E-6, 1.50 1.0E-6, 1.50
1.0E-6, 2.00 2.2 1.0E-6, 2.00
2.8 1.0E-6, 2.50 1.0E-6, 2.50
1.0E-5, 1.50 2.0 1.0E-5, 1.50
2.6 1.0E-5, 2.00 1.0E-5, 2.00
1.8
1.0E-5, 2.50 1.0E-5, 2.50
2.4 1.6
1.4
2.2
1.2
2.0 1.0
0 5 10 0 5 10
Iteration n Iteration n

Figure 3: Average location estimation accuracy against itera- Figure 4: Average location estimation accuracy against itera-
tion cycle n and various α0 , β0 initialization values for experi- tion cycle n and various α0 , β0 initialization values for experi-
ments conducted in the corridor. ments conducted in the office room.

R EFERENCES [13] R. Zemek, M. Takashima, S. Hara, K. Yanagihara, K. Fukui, S. Fuku-


naga, and K. Kitayama, “An Effect of Anchor Nodes Placement on a Tar-
[1] K. Romer and F. Mattern, “The Design Space of Wireless Sensor Net-
get Location Estimation Performance,” in TENCON 2006, Hong Kong,
works,” IEEE Wireless Communications, vol. 11, no. 6, pp. 54 – 61, De-
November 2006.
cember 2004.
[2] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci, “A Survey [14] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, Nu-
on Sensor Networks,” IEEE Communications Magazine, vol. 40, no. 8, merical Recipes in C: The Art of Scientific Computing. Cambridge Uni-
pp. 102 – 114, August 2002. versity Press, 1992.
[3] N. Patwari, J. N. Ash, S. Kyperountas, A. O. Hero III, R. L. Moses,
and N. S. Correal, “Locationg the Nodes: Cooperative Localization in
Wireless Sensor Networks,” IEEE Signal Processing Magazine, vol. 22,
no. 4, pp. 54 – 69, July 2005.
[4] “Wireless medium access control (MAC) and physical layer (PHY) speci-
fications for high layer wireless personal area networks (WPANs),” 2003.
[5] C. Savarese, J. M. Rabaey, and J. Beutel, “Locationing in Distributed
Ad-Hoc Wireless Sensor Networks,” vol. 4. Salt Lake City, USA: IEEE
International Conference on Acoustics, Speech, and Signal Processing
(ICASSP’01), May 2001, pp. 2037 – 2040.
[6] I. J. Myung, “Tutorial on Maximum Likelihood Estimation,” Journal of
Mathematical Psychology, vol. 47, no. 1, pp. 90 – 100, February 2003.
[7] M. Takashima, D. Zhao, K. Yanagihara, K. Fukui, S. Fukunaga, S. Hara,
and K. Kitayama, “Location Estimation Using Received Signal Power
and Maximum Likelihood Estimation Method in Wireless Sensor Net-
work,” IEICE Tran. on Communication (Japanese edition), vol. J89-B,
no. 5, pp. 742 – 750, May 2006.
[8] S. Hara, D. Zhao, K. Yanagihara, J. Taketsugu, K. Fukui, S. Fukunaga,
and K. Kitayama, “Propagation Characteristics of IEEE 802.15.4 Radio
Signal and Their Application for Location Estimation,” in IEEE VTC
2005 Spring, vol. 1, Stockholm, Sweden, May 2005, pp. 97 – 101.
[9] J. Arias, A. Zuloaga, J. Lzaro, J. Andreu, and A. Astarloa, “Malguki: An
RSSI Based Ad Hoc Location Algorithm,” IEEE Microprocessors and
Microsystems, vol. 28, no. 8, pp. 403 – 409, October 2004.
[10] F. Mondinelli and Z. M. Kovacs-Vajna, “Self-Localizing Sensor Net-
work Architectures,” IEEE Transactions on Instrumentation and Mea-
surement, vol. 53, no. 2, pp. 277 – 283, April 2004.
[11] N. Patwari, A. O. Hero III, M. P. N. S. Correal, and R. J. O’Dea, “Relative
Location Estimation in Wireless Sensor Network,” IEEE Transactions on
Signal Processing, vol. 51, no. 8, pp. 2137 – 2148, August 2003.
[12] R. Mino, K. Iwamoto, , M. Takashima, R. Zemek, K. Yanagihara,
S. Hara, and K. Kitayama, “A Belief Propagation-based Iterative Loca-
tion Estimation Method for Wireless Sensor Networks,” in PIMRC 2006,
Helsinki, Finland, September 2006.

You might also like