2.
distributed in the environment. This is impractical for “object” are interchangeable terms in this paper.
most of the outdoor scenarios like road vehicular Reference tags are the active RFID tags used for
tracking, disaster site, and battlefield. calibrating environmental parameters. Every reference
In this paper, we present Scout, an easy to deploy tag also has a unique ID. Once being activated, the
and cost-effective outdoor localization system based on active tags will periodically emit signals with their IDs.
off-the-shelf active RFID system and a probabilistic
localization algorithm. Considering the uncertainty
caused by the varying environments, Scout
incorporates a probabilistic scheme based on on-site
calibration and Bayesian inference to improve the
localization accuracy. Bayesian inference was also used
for cellular and WLAN-based localization systems [6,
7]. To the best of our knowledge, Scout is the first
outdoor localization system integrating active RFID
technology and probabilistic localization algorithm. Figure 1. Scout system structure
The rest of this paper is organized as follows:
Section 2 describes the Scout system, including the RFID readers: The whole detection area is covered
basic system assumptions, system overview, and the with a RFID readers network such that any object tag
probabilistic localization algorithm. In Section 3, the located in the area should be within the read range of at
system performance metric is introduced, followed by least three RFID readers. The problem of reader
the simulation results and related discussions. Section 4 deployment to satisfy this requirement has been
concludes the paper. comprehensively studied in the literature, such as in
[9]. The reader location is fixed and known. A reader is
2. Scout System responsible for 1) Collecting and decoding the signals
emitted by the active tags within its coverage, 2)
The Scout system uses off-the-shelf long range Measuring the RSSI for each tag within its range, 3)
active RFID systems, similar to the Mantis system [1, Reporting tag ID, corresponding RSSI, and its own ID
2]. According to [2], this system works at 433MHz number to the server. To realize the above
frequency range, with a minimal range of 100 meters functionalities, each RFID reader is equipped with two
and maximum range of 500 meters. The MANTIS interfaces: a RF interface that detects RFID tags within
reader can provide Received Signal Strength Indicator its coverage; and a communication interface (Ethernet
(RSSI) readings for each tag within its range. In the or 802.11b wireless interface) for communicating with
next section, we discuss the network layout of the the servers.
Scout system. Servers: Each RFID reader should be within the
reach of at least one server, so that the readers can
2.1. Network Layout communicate the measured location information of the
tags to the servers. The number of servers deployed in
Scout system consists of an RFID sensing network a detection area depends on the size of the detection
that measures the RSSI information from the objects area and the system design. A server is responsible for
(or active tags), and a communication network that 1) Collecting location information sent from the
enables the transmission and processing of the location readers within its coverage area, and 2) Calculating the
information. The Scout system follows a hierarchical location of the object tags according to the localization
architecture comprising of the following three tiers: algorithm.
Servers, RFID readers, and RFID tags. For simplicity, The Scout system works as follows: After the
in Fig. 1, we just show the hierarchical architecture system is setup, the RFID readers will begin to detect
within the coverage of one server, although in practice the signal sent by RFID tags. If any tag is located
the whole detection area may be covered by several within its read range, the reader will collect and report
servers. the identification and the RSSI of the tag along with its
RFID tags: Two types of active tags are included in own ID to the corresponding server. All the
the system, object tags and reference tags. Every information reported by the RFID readers will be
tracked object will be attached with a unique active processed at the server(s). The server(s) will estimate
RFID tag, called “object tag”, used for identifying and the location of the objects (to the readers) according to
tracking the object. The ID of an object tag is a unique the Scout localization algorithm using the reported RSS
identifier for a tracked object. So, “object tag” and information and the position of the readers.
3.
environment, such as shadowing. Thus the path loss
2.2. Probabilistic Localization Scheme can only be considered as an average value. To take
these factors into consideration, it has been shown that
The proposed algorithm estimates objects’ locations the received signal strength usually demonstrates a
based on RSSI of the object tags and the reference tags Gaussian normal distribution [8]. Hence, we take a
provided by the readers. However, signal strength probabilistic approach and model the path loss at
based RF localization systems suffer from the inherent distance d as a random variable PL(d) by using a
errors associated with the RF propagation due to Gaussian random variable N [0, 2 ] with zero-mean,
factors such as fading, shadowing, interference, etc. and standard deviation- .
These factors cause random variations of the received d
PL ( d ) PL ( d ) N [0, ] PL ( d ) 10n log( ) N [0, 2 ] (2)
signal strength (RSS), and thus make it impossible to 0
d0
come up with a deterministic relationship between the With a given transmitting antenna power Pt ,
RSSI and the propagation distance.
The proposed Scout algorithm is robust to transmitting antenna gain Gt , and receiving antenna
randomness caused by propagation variations. It adapts gain G r , the received signal strength Pr ( d ) (in dB) at
to varying outdoor environmental conditions by
following the subsequent steps: 1) Calibrates the distance d is given by:
propagation parameters using on-site reference tags, 2) Pr (d ) Pt Gt Gr PL(d ) (3)
2
Estimates the distance between the object and the Pt Gt Gr PL(d0 ) 10n log(d ) N [0, ]
readers based on a probabilistic model deduced from Because the first four terms in (3) are all constant,
general RF path loss model, and 3) Determines the we can use one constant to represent them. So the
target location by applying Bayesian inference. above equation can be rewritten as:
The Scout localization algorithm is based on the Pr (d ) Pr (d ) N (0, ) 10n log(d ) N (0, 2 ) (4)
following assumptions: 1) Any object located in the
area can be detected by at least three RFID readers, 2) where is a constant. It equals the median signal
The readers are fixed, and their locations are known, 3) received at d 0 1m .
There is one reference tag per reader deployed in the From (4), we can see that given the actual distance
reader’s vicinity, and the location of the reference tags d PHY between the tag and the reader, the received
is known, 4) The transmission power of all tags is
signal strength Pr (d PHY ) is a Gaussian random variable
identical.
2
In the following subsections we introduce the with mean 10n log(d PHY ) and variance . So the
probabilistic model, the calibration of the propagation
probability density function (PDF) of observing a
parameters, and the Bayesian inference based
certain value of Pr (d PHY ) (i.e, RSSI reported by the
localization algorithm.
reader) is given by:
2.2.1. Probabilistic Model 1 (RSSI 10n log(dPHY ))2 (5)
p(P (dPHY ) RSSI )
r exp
In order to estimate the distance between a 2 2 2
transmitter (i.e., tagged object or reference tag) and a Or given the actual distance d PHY between the tag
receiver (i.e., reader), we use the log-distance path loss
model that has been used extensively in the literature and the reader, the PDF of the estimated
[8]: distance, d estimate , is given by following the usual
PL(d ) PL(d0 ) 10n log(d / d0 ) (1) transformation of variable approach:
where d represents the distance between the transmitter (RSSI ) 1 (RSSI 10n log(dPHY ))2
p(destimate ) exp
and receiver, PL(d ) is the median propagation loss (in (destimate ) 2 2 2 (6)
dB) measured at distance d; n is the path loss exponent 10n (10n log(destimate / dPHY ))2
exp
which indicates the decreasing rate of signal strength in destimate ln10 2 2 2
an environment; d0 is a reference distance normally (
RSSI
)
where d estimate 10 10 n (7)
chosen close to the transmitter. In general, the exponent
n is environment-dependent. In free space, n is equal to The PDF of d estimate is plotted in Fig. 2. From Fig. 2,
2. In more complicated environments, n will generally we observe that as d PHY (actual distance in the figure)
be larger.
Note that the model introduced in (1) does not or (v in the figure) increases, or as n decreases, the
consider the variable factors in the surrounding spread of the PDF become flatter, and the uncertainty
4.
associated with the propagation increases (in the figure, the following equations:
the unit for the estimated distance is meter). n j (i )
nj ;
2
i 1,2,3
i j (9)
RSSI j ( j )- RSSI j (i )
n j (i ) ;(i j)
dist (i, j )
10log( )
dist(j , j )
j (i )
j ; (10)
i 1,2,3 2
i j
j (i ) RSSI j (i ) 10n j (i ) log(dist (i, j )); (i j)
The standard deviation of detection area can be
calculated by:
Nr No
Figure 2. PDF of the estimated distance 1 (11)
[ std ( RSSI j (i ))]2
Nr No j 1 i 1
2.2.2. Calibration of the Propagation Parameters where N r and N o denotes the number of the readers and
In order to calculate the probability using (5), or (6)
the tags respectively. For the above example
and (7), we need to know the value of the propagation
scenario, N r 3, N o 4 ; std ( RSSI j (i )) is the standard
parameters: , n , . In Scout localization scheme, we
use on-site reference tags and object tags to calibrate deviation for vector [ RSSI 1j (i), RSSI 2 (i),....RSSI w (i) ].
j j
the propagation parameters during the process of
localization. As assumed before, every reader has a 2.2.3. Bayesian Inference based Localization
corresponding reference tag deployed in its vicinity. Algorithm
The reference tags are static and their location is Notations and definitions:
known. After detection, the readers report the RSSIs of 1) Basic detection area: The area overlapped by the
both the reference tags and the object tags. This section range of three adjacent readers (Fig. 3).
describes how the Scout algorithm calibrates the
propagation parameters.
Let us consider the scenario depicted in Fig. 3. There
are 3 readers, reader_1, reader_2, and reader_3 for a
specific area. Tag_1, tag_2, tag_3 are the reference tags
corresponding to the three readers and tag_4 is the
tracked object. Suppose during a period of time, time
slot 1 to time slot w, the three readers detect w times
(one detection during each time slot) and report the
tags’ signal strength as {RSSI tj (i)} , (j=1,2,3; i=1,2,3,4;
Figure 3. Grid and basic detection area
t=1,2,3,4….w). Here, RSSI tj (i ) denotes the signal
strength of tag_i reported by reader_j at detection time
t. If w is large enough, then from (4):
w
RSSI (i)
j RSSI t (i)) / w P (dist (i, j))
j r j 10n log(dist (i, j)) (8)
j
t 1
where RSSI j (i ) is the average received signal strength of
tag_i reported by reader_j calculated from the
measurements obtained during w slots; dist (i, j )
denotes the Euclidean distance between reader_j
(j=1,2,3) and tag_i (i=1,2,3 for reference tags, 4 for the
tracked object); j , n j denote the propagation constant Figure 4. Observations of an object
and the path loss exponent of reader_j, respectively.
2) Grid cell: The detection area covered by a server
According to (8), the propagation constant and the
is divided into a uniform grid (Fig. 3). Each square
path loss exponent of reader_j can be calculated using
patch is called a grid cell, the center of a grid cell is
5.
called grid point, and the distance between two nearest RSSI for reader_j as:
t
grid points is called grid_step. Assume N grid cells 1 (13)
RSSI tj RSSI ij
exists in the detection area, the grid cells are denoted wi t w 1
by C( gx , gy ) , C( gx , gy ) ,…., C(gx ,gy ) , using the coordinates of We denote these three readers by reader_1,
1 1 2 2 N N
their grid points ( gxk , gyk )(k=1,2,…N). reader_2, reader_3. Then, a new observation matrix
which includes only the observations from these three
3) State of an object at time t, x t : The state is the
readers will be presented as:
tracked object’s location. In our grid-based localization RSSI1t w 1 RSSI1t w 2 ... RSSI1t (14)
system, we use grid cells to denote the location, so we [RSSI 't w 1
, RSSI 't w 2
,...RSSI 't ] t
RSSI 2 w 1 t
RSSI2 w 2 t
... RSSI2
t w 1 t w 2 t
represent the state of the object at time t by discrete RSSI3 RSSI3 ... RSSI3
random variable x t , xt {C( gx , gy ) , C( gx , gy ) ,...C( gx , gy ) } . 3) Calibration of the Propagation parameters: We
1 1 2 2 N N
4) Observation RSSI : RSSI is the RSSI of the t t use the method described in section 2.2.2 to calibrate
j j
the propagation parameters of the three readers in the
object (or reference tag) reported by reader_j at time t. basic detection area. Given location of the three readers
Fig. 4 shows the sequence of RSSI tj (j=1,2…R, and the corresponding reference tags and the
t=1,2,…T) from R readers located within the server’s observations falling in the current window, the
coverage during time interval [1, T]. propagation parameters can be calculated by (9) and
5) Belief Bel ( xt ) : At time t, a probability (10); Given [ RSSI 't w 1 , RSSI 't w 2 ,...RSSI 't ] of the
distribution over possible value of x t , called Belief, reference tags and the tracked objects, the propagation
standard deviation can be calibrated by (11).
represents the uncertainty. Bel ( xt C( gx , gy ) ) represents
k k 4) According to the observation vector at time
the probability that the object is located within grid cell t, RSSI 't , we can estimate the object’s location using
C( gx , gy ) at time t. This probability will be computed by
k k Bayesian inference.
using a series of observations. The Bayesian inference is used in the following
6) Window_size, w: The number of observations way: Whenever the sensors (i.e., the readers) provide a
per reader that will be used for calculating Bel ( x t ) in new observation at time t, RSSI 't , the server first
one recursion (Fig. 4). predicts state at time t according to the state at time t-1:
7) Recursion_time, r: The number of times we Bel ( xt ) p ( x t | xt 1 ) Bel ( xt 1 )dxt 1 (15)
execute the recursive Bayesian inference to estimate the
location of the tracked object. where x t 1 represents state of the object at time t-1,
8) Estimated Area Threshold, P: The estimation Bel ( xt ) is called predictive belief, p( xt | xt 1 ) is the
area is given by the grid cells with belief higher or system dynamic model. Then the server corrects the
equal to P. predicted belief using the observation at time t by:
Input: The tracked object ID, location of the Bel ( x t ) t
p ( RSSI 't | x t ) Bel ( x t ) (16)
readers ( x j , y j ) , reference tags (rxi , ryi ) , and grid points where p( RSSI 't | xt ) is the perceptual model which
( gxk , gyk ), threshold P, observations for the object and describes the likelihood of having observations
reference tags collected in interval [t-w+1, t] given by: RSSI 't given that the object is at location x t , t is a
RSSI1t w 1
RSSI1t w 2
.... RSSI1t
(12) constant used to normalize the resulting Bel ( xt ) .
RSSI 2 w
t 1
RSSI 2 w
t 2
... t
RSSI 2
. . . . Implementing Bayesian filters requires specifying
[ RSSI t w 1
, RSSI t w 2
,...RSSI t ]
. . . . the perceptual model p( RSSI 't | xt ) , system dynamics
. . . . model p( xt | xt 1 ) , and the representation of the
RSSI R w
t 1
RSSI R w
t 2
... t
RSSI R
belief Bel ( xt ) .
Output: estimated area of the object.
Procedure: 4a) System Dynamics Model p( xt | xt 1 ) :
1) Initialization: The initial belief is given by: For moving objects, especially fast moving object,
1 1 1 this model depends on the object motion parameters,
Bel ( x 0 ) { , ,... } .
N N N such as direction, speed, etc.
t t 1 t t 1
2) Determine the basic detection area: Use the For the static objects, p( x | x ) 1 if x x .
observations of the object as defined in (12). The basic p( xt | xt 1 ) 0 xt xt 1
detection area is determined by the three readers with
So, according to (15), we have:
the strongest average RSSI. We define the average
Bel ( x t ) p ( x t | x t 1 )Bel ( x t 1 ) dx t 1
Bel ( x t 1 ) (17)
6.
4b) Perceptual Model p( RSSI 't | xt ) : Euclidean distance between the object and the center of
For a single reader, for example reader_1, we use the cell (i.e., the grid point corresponding to the cell):
the model described in Section 2.2.1. The PDF given E (C( gxk , gyk ) ) dist (( gxk , gyk ), ( x0 , y0 ))
(20)
by (5) is used to calculate p1 ( RSSI1t | xt ) , reader_1’s ( gxk x0 )2 ( gyk y0 ) 2
perceptual model at time t. Combining the observations where ( gxk , gyk ) are the coordinates of the grid point,
from m readers to estimate the object’s location, we use
the following equation: ( x0 , y0 ) are the coordinates of the object, and dist ( )
m
(18) is the Euclidean distance between two points.
p( RSSI 't | xt ) p j ( RSSI tj | xt )
j 1
As described in Section 2, the Scout localization
where p j ( RSSI tj | xt ) represents the perceptual model of algorithm is a probabilistic localization algorithm,
which uses beliefs to represent the probability that the
a single reader, reader_j. For our algorithm, we use object is located within a given grid cell. Therefore, the
m=3.
error distance is a discrete random variable. Given the
4c) Calculating belief Bel ( xt ) :
estimated area A, the AED is given by:
We calculate the probability belief Bel ( xt ) using E (C( gxk , gyk ) ) Bel ' ( x t C( gxk , gyk ) )
(16). Bel ( xt ) is obtained by (17) and p( RSSI 't | xt ) is AED k
Bel ' ( x t C( gxk , gyk ) ) (21)
obtain by (18), t is a constant for time-t which is used k
to normalize the resulting Bel ( xt ) . In our algorithm, we ( gxk x0 ) 2 ( gyk y0 ) 2 Bel ' ( x t C( gxk , gyk ) )
k
use grid point ( gxk , gyk ) to calculate the Bel ' ( x t C( gxk , gyk ) )
belief Bel ( xt C( gxk , gyk ))
. k
where C( gx , gy ) A ; and Bel ' ( xt C( gxk , gyk ) ) is the
5) Determination of the resulting estimated area: k k
After r times of Bayesian inference, the resulting relative belief of grid cell C( gx .
k , gyk )
estimated area is decided by the relative belief, Bel ' ( xt ) : As defined above, AED is the average error
Bel ( xt ) (19) distance between the object and the grid points within
Bel ' ( xt )=
| max (Bel ( xt C( gxk ,gyk ) )) | the estimated area. AED reflects the localization
k 1,2.. N
accuracy of the Scout algorithm. Large AED values
where N is the number of the grid cells in the detection mean poor accuracy while small AED values reflect
area. Given threshold P, the resulting estimated area high localization accuracy. The minimum value of
comprises of the states with relative probability belief AED is zero, which will happen when there is only one
larger or equal to P. cell in the estimated area and its grid point is exactly
where the object is. The maximum value of AED
3. Simulation Results increases with the distance between two adjacent
readers.
In this section, we evaluate the performance of the
proposed localization scheme. First, we will develop a 3.2. Visual Results
performance metric for Scout localization accuracy.
Then, we illustrate visual results that depict the We use MATLAB to simulate a basic detection
resulting estimated areas for multiple objects. At last, area covered by reader_1, reader_2, and reader_3. Five
we investigate the scheme performance dependence on tracked objects are distributed within the area, as
a number of system, algorithm, and environmental shown in Fig. 5(a). All the tracked objects are static. In
parameters. the simulation, we consider the environmental factors
such as shadow and fading, which cause received
3.1. Performance Metrics signal strength measurement errors (refer (4)). The
Gaussian random variable in (4) will be generated
To characterize the localization accuracy of the randomly using MATLAB randn function. The
proposed Scout algorithm, we define a performance simulation parameters are: Grid_step=1, window_size
metric denoted as the Average Error Distance (AED). w=20, recursion_time r=10, path loss constant 0,
The output of the localization algorithm is an path loss exponent n=4, distance between two adjacent
estimated area that is composed of several grid cells as readers D=80m, RSSI standard deviation 3 dB
defined in Section 2. We define the Error (according to [1], 3 dB is a typical standard variance
Distance, E (C( gx , gy ) ) , for grid cell C ( g x , g y ) , as the for the long range active RFID systems). Fig. 5(b)-(f)
k k k k
7.
depict the resulting estimated areas for each object for distance between the object location and the estimated
different values of threshold P, P=1, 0.75, 0.5. Each location corresponding to P=1 are less than 2 meters
estimated area includes all the grid cells which have for all the five objects. Furthermore, we observe that
relative belief probability equal to or larger than the the shapes of the estimated areas are decided by the
threshold P. The solid line and the dash line depict the objects’ distances from the three readers. The closer
corresponding estimated area of P=0.75 and P=0.5 readers have more influence on the results. For
respectively. example, (see Fig. 5(f)) object_5 is located very close
to reader_2 while very far away from the other two
readers. So, the nearest reader, reader_2 dominates its
estimation result, and the estimated areas are ring
shaped around reader_2. Object_3 is in a similar
situation with respect to reader_1.
3.3. Numerical Results
We conduct a series of experiments to evaluate the
performance of the Scout localization algorithm. We
(a) Topology (b) Object_1 (0, 23) use MATLAB to simulate the impact of the following
parameters on the localization accuracy represented by
Average Error Distance (AED): window_size (w),
recursion_time (r), distance between two adjacent
readers (D), RSSI standard deviation ( ), and Path
Loss Exponent (n). The default topology is shown in
Fig, 5(a). The received signal strength at a certain
distance from the reader is simulated as a Normal
random variable (refer to (4)).
(c) Object_2 (0,0) (d) Object_4 (10, 40)
3.3.1. Impact of window_size and recursion_time
Experiment 1: The simulation is based on the
following system configuration: grid_step=1, threshold
P=0.5, D=80m, n=4; =0; =3dB. We run the
simulation for the following three scenarios: 1) ideal
calibration, i.e. the propagation parameters (n, , )
obtained from the on-site calibration procedure are
identical to the real environmental parameters; 2)
window_size w=20;and 3) window_size w=2.
(e) Object_3 (0, 65) Fig. 6(a) depicts AED as a function of the
recursion_time for the three scenarios. We observe that
AED decreases as we increase the recursion_time.
Moreover, we notice that the window_size has limited
influence on the accuracy of the results. This is due to
the fact that the use of the recursive Bayesian inference
will eliminate the uncertainty caused by the limited
sampling.
Fig. 6(b) illustrates AED as a function of the
recursion_time for different objects in the above first
(f) Object_5 (-35,10) scenario. We observe that AED convergence speed is a
Figure 5. Visual results function of the object locations within the basic
detection area. For object_1 and object_2, the
The results demonstrate that the proposed algorithm localization accuracy increases much faster than that of
works well for all the five objects. It can effectively object_3, with increasing of the recursion_time. From
estimate the locations of the tracked objects. As shown (18) we can see that Bayesian inference makes use of
in the figures, all the five objects are within the the observations from the three readers to recursively
corresponding estimated area of P=0.75. And the diminish the uncertainty caused by the RF propagation.
8.
Furthermore, from (6) and its plot, we learn that for the readers. We notice that as the distance between the
distances far away from reader, the PDF curve becomes readers increases the localization accuracy, expressed
flatter and the uncertainty increases. Since object_3 is in terms of AED, decreases.
very far from reader_2 and reader_3, the observations From the probabilistic model presented in (6), we
from reader_2 and reader_3 contain huge uncertainty observe that increase in d PHY (the physical distance
and provide little information. This is equivalent to the between the object and the reader) increases the spread
situation that Bayesian filter only make use of of the PDF, increasing the localization uncertainty
observation from reader_1 for localizing object_3. which decreases the estimation accuracy.
While for the object_1 and object_2, the observations D and ADE
Average Distance Error (meters)
4
from all the three readers provide useful localization
Object 1
information. As a result, the localization of object_3 3.5
Object 2
converges very slow comparing to object_1 and 3
object_2. 2.5
2
1.5
1
50 60 70 80 90 100 110 120
Distance Between Readers, D (meters)
Figure 7. AED vs distance between readers, D
3.3.3. Impact of Propagation parameters, , n
Experiment 3: In this experiment we plan to
evaluate the impact of the propagation standard
deviation on the localization accuracy. We perform
the simulation for object_1, object_2, and object_3
depicted in Fig. 5(a) using the following system
configuration: Grid_step=1, D=80m, w=20, r=10,
n=4, 0 , P=0.5. Fig. 8(a) plots AED as a function
of the standard deviation .
Experiment 4: In this experiment we evaluate the
impact of the path loss exponent n on the localization
accuracy. We perform the simulation for object_1,
object_2, and object_3 depicted in Fig. 5(a) using the
following system configuration using the following
system configuration: Grid_step=1, D=80m, w=20,
r=10, 0, =3dB, P=0.5. Fig. 8(b) plots AED as
a function of the path loss exponent.
Figure 6. Impact of window_size and recursion_time From Fig. 8(a) we observe that AED increases (i.e.
on the system performance the localization accuracy decreases) as the standard
deviation increases. From Fig. 8(b) we conclude
3.3.2. Impact of the distance between two adjacent that AED decreases (i.e. the localization accuracy
readers increases) as the path loss exponent, n, increases. From
Experiment 2: Assume the distances between the the probability model defined in (6) and (7), increase
readers are the same. We perform the simulation for the in or decrease in n, increases the spread of the PDF,
object_1 and object_2 with the same geometry but increasing the uncertainty associated with the
different scale. As depicted in Fig. 5(a), Object_1 is propagation which results in decrease of the
located at the gravity center of the basic estimated area localization accuracy.
and the object_2 is located at the midpoint of reader_2 Moreover, from Fig. 8(a) and (b) we also observe
and reader_3. To eliminate the impact of inaccuracy that the AED increase rate depends on the relative
caused during propagation parameter calibration, the position of the object to the readers. The change of the
simulation is run for the ideal calibration scenario. The propagation parameters has the smallest impact on
system configuration is set as following: Grid_step=1, object_3 and the largest impact on object_1. Since
w=20, r=10, n=4, 0 , =3dB, P=0.5. Fig. 7 object_3 is dominated by reader_1, only change of the
illustrates AED as a function of the distance between
9.
reader_1’s measurements (caused by the change of the In this section, we test the system localization
propagation parameters) influences object_3 performance using the following experiment.
localization. Since object_1 is located equidistant from Experiment 5: This experiment is designed to
the three readers, object_1 localization is influenced by probabilistically evaluate the system performance
changes of all the three readers’ measurements. under the following settings: Using the same reader
network topology as shown in Fig. 5(a), we track 1000
objects which are uniformly distributed within the
basic detection area. The other simulation parameters
are: Grid_step=1, D=80m, n=4, 0, 3 dB, P=1,
0.8, 0.6. We test the system performance for three
scenarios: 1) w=20, r=10; 2) w=20, r=20; 3) ideal
calibration, r=10.
Fig.9 shows the cumulative Average Error Distance
distribution obtained from scenario 1 and 2. We notice
that over 90% percent of the object the localization
AED falls within 7 meters and 5 meters for the scenario
of r=10 and 20 respectively. As expected, increasing
recursion_time r will leads to higher localization
accuracy. Fig.10 shows the cumulative Average Error
Distance distribution obtained from scenario 1 and 3.
As mentioned before, increasing w will make the
system performance incline to that of ideal calibration.
So, Fig. 10 shows the upper bound of the system
performance for r=10. Furthermore, from Fig. 9 and
Fig. 10, we can get the conclusion that in most of the
situation, localization accuracy with larger P (P<=1) is
better.
Figure 8. Impact of the propagation standard
deviation and the path loss exponent
3.3.4. System Performance
CDF of AED
1
0.9
r=10,P=1
0.8 r=10,P=0.8
r=10,P=0.6
0.7 r=20,P=1
Cumulative Percentage
r=20,P=0.8
0.6
r=20,P=0.6
0.5
0.4
0.3
0.2
0.1
0
0 1 2 3 4 5 6 7 8 9 10 11 12
Average Error Distance (meters)
Figure 9. AED distribution for r=10 and r=20
10.
CDF of AED
1
0.9
0.8
w=20,P=1
0.7
w=20,P=0.8
Cumulative Percentage
w=20,P=0.6
0.6
ideal, P=1
ideal, P=0.8
0.5
ideal, P=0.6
0.4
0.3
0.2
0.1
0
0 1 2 3 4 5 6 7 8 9 10
Average Error Distance (meters)
Figure 10. AED distribution for ideal calibration and w=20
4. Conclusion [2] Mantis: http://www.rfcode.com/data_sheets/mantis.pdf
[3] J. Hightower, C. Vakili, G. Borriello, and R. Want,
“Design and calibration of the SpotON Ad-Hoc Location
In this paper, we presented Scout, a cost-effective Sensing System,” UW CSE 01-08- University of
outdoor localization system based on active RFID Washington, Seattle, WA, August 2001
systems as well as a probabilistic approach to the [4] A. Sayed, A. Tarighat, and N. Khajehnouri, “Network-
localization problem using only RSSI readings. Scout based wireless location,” IEEE Signal Processing Mag.,
system relies on a hierarchical architecture to cover a vol. 22, no. 4, pp. 24–40, July 2005.
large outdoor environment. The proposed approach [5] L. M. Ni, Y. Liu, Y. Cho Lau, A. P. Patil,
addresses the localization uncertainty due to the “LANDMARC: indoor location sensing using active
varying environments. Using MATLAB based RFID”, In Proceedings of the First IEEE International
simulations we have evaluated the performance of the Conference on Pervasive Computing and
Communications (PerCom 2003), 23-26 March 2003
proposed approach. For the example system settings,
[6] T. Roos, P. Myllymaki, and H. Tirri, “A Statistical
the simulation results show that in 90% percent of the Modeling Approach to Location Estimation,” IEEE
localization estimation, the system provides objects Trans. on Moible Computing, vol.1, Issue 1, pp 59-69,
location with a AED (Average Error Distance) less Jan. 2002
than 7 meters. Also, the simulation shows that the [7] T. Roos, P. Myllymaki, H. Tirri, P. Misikangas, and J.
system performance improves with the higher values Sievanen, “A probabilistic Approach to WLAN User
of recursion_time, window_size, reader density, path Location Estimation,” International Journal of Wireless
loss exponent, and with lower values of the Information Networks, 9 (3), July 2002
propagation standard deviation. The simulation results [8] T. S. Rappaport and T. Rappaport, Wireless
Communications: Principles and Practice, 2nd Edition.
prove that the proposed system is an accurate and cost-
Pretice Hall, Dec.2001
effective candidate for outdoor localization. [9] C.-F. Huang and Y.-C. Tseng, “The coverage problem in
a wireless sensor network,” ACM Int’l Workshop on
5. References Wireless Sensor Networks and Applications (WSNA),
pages 115-121, 2003
[1] BSR INCITS 371.2, Real Time Location Systems
(RTLS) – Part 2: 433-MHz Air Interface Protocol”.
Views
Actions
Embeds 0
Report content