ziloc: energy efficient wifi fingerprint-based localization ...guxxx035/paper/banghuiwcnc13.pdf ·...
TRANSCRIPT
ZiLoc: Energy Efficient WiFi Fingerprint-based
Localization with Low-Power Radio
Jianwei Niu∗, Banghui Lu∗†, Long Cheng∗†, Yu Gu†, Lei Shu‡
∗State Key Lab of Software Development Environment, Beihang University, Beijing 100191, China†Singapore University of Technology and Design, Singapore‡Guangdong University of Petrochemical Technology, China
{niujianwei, lu, chenglong}@buaa.edu.cn, [email protected], [email protected]
Abstract—Indoor localization is essential to enable location-based services in wireless pervasive computing environment. Inrecent years, WiFi fingerprint-based localization has receivedconsiderable attention due to its deployment practicability. Inorder to achieve on-the-fly localization, WiFi receivers (e.g.,mobile phones or laptops) being located need to scan WiFisignals continuously. Since they are normally battery driven,energy efficiency is a very important consideration in WiFifingerprinting localization systems. Motivated by the fact thatIEEE 802.11 (WiFi) and 802.15.4 (ZigBee) channels overlapin the 2.4GHz ISM band, in this work, we develop a WiFifingerprint-based localization system using ZigBee radio, calledZiLoc. We first present a novel RSS-location fingerprint modelto identify the features of surrounding APs. We then proposea simple yet effective method to compute the similarity of twoRSS fingerprints. Experimental results demonstrate that ZiLoccan achieve an average of 85% room-level localization accuracyand reduce more than 60% energy consumption compared withthe method using WiFi interfaces to collect RSS fingerprints.
I. INTRODUCTION
Location-based service (LBS) has a huge market potential,
such as navigation, tracking, and mobile advertising. One of
the key issues for LBSs is the accurate and efficient localiza-
tion. Since people tend to stay most of the time indoors (either
at work or at home), there is a growing demand for indoor
localization systems. While GPS (Global Positioning System)
is the most popular positioning system, it is not suitable for
indoor area due to the poor coverage. For this reason, various
indoor localization techniques have been proposed [1]–[6].
We are witnessing a ubiquitous deployment of 802.11 WiFi
access points (APs), in universities, offices, city parks, restau-
rants, and homes. Meanwhile, WiFi enabled portable devices,
such as laptops, mobile phones and tablets, have become
very common with the extensive coverage of 802.11 WLAN.
Therefore, WiFi fingerprint-based localization is considered
as a practical solution since it allows the design of an easily
deployable infrastructure-less low-cost localization system. A
WiFi-fingerprinting system works in two phases: an offline
training phase and an online localization phase. The first phase
collects fingerprints (RSS measurements of WiFi APs) from
different pre-known locations and stores them to a database
as the training set. The second phase infers the location
based on the observed RSS measurements, through finding the
closest match in the database. However, to achieve realtime
localization, WiFi receivers (e.g., mobile phones or laptops)
being located are required to scan WiFi signals frequently,
which is extremely power-hungry for battery driven mobile
devices.
1 6 11
WiFi Channels
ZigBeeChannels
2 3 4 5 7 8 9 10 12 13 14
11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
2405 2410 2415 2420 2425 2430 2435 2440 2445 2450 2455 2460 2465 2475 24802470 (MHz)
Fig. 1. 802.11 and 802.15.4 channels overlap in 2.4GHz ISM band
0
20
40
60
80
100
120
140
0 10 20 30 40 50 60
Cu
rre
nt
dra
w (
mA
)
Time (s)
WiFi NIC
ZigBee radio
Scanning
IdleScanning
Idle
Fig. 2. Our measurement shows WiFi NIC (network interface card) in the idlestate could consume more energy than Zigbee radio when scanning channels
Motivated by the fact that 802.11 and 802.15.4 channels
overlap in the 2.4GHz ISM band (as shown in Fig. 1), and the
low-power feature of 802.15.4-compliant radio transceivers,
this paper is an attempt to explore the 802.15.4 radio to collect
WiFi signals, where the objective is to improve the energy
efficiency of WiFi-fingerprinting localization. Fig. 2 illustrates
the power consumption comparison between ZigBee radio and
WiFi radio. The low-power operation of 802.15.4 radio makes
it attractive for a wide range of instruments, e.g., a plug &
play support for Zigbee radio on mobile phones has been
proposed [7]. We develop a WiFi fingerprint-based localization
system using ZigBee radio, called ZiLoc. Two important issues
are addressed in ZiLoc, the RSS-location fingerprint modeling
and location estimation. We implement a testbed to verify
the effectiveness of ZiLoc design. Three different k-Nearest
Neighbor (KNN) [8] classification methods are compared to
evaluate the location estimation performance. Experimental
results show that ZiLoc can effectively achieve 85% room-
level localization accuracy on average and save around 66%
energy compared with the WiFi NIC-based method.
The outline of the paper is as follows. Section II surveys
the related work. Section III elaborates the ZiLoc design in
details. Section IV provides the experimental results. Finally,
conclusions are drawn in Section V.
II. RELATED WORK
Indoor localization is a critical enabler for LBS application-
s. Traditional indoor localization solutions normally require
dedicated infrastructures based on infrared [9], RFID [10],
or acoustic sensors [11]. Recently, there has been growing
interest in utilizing in-building communication infrastructures
(e.g., WiFi, WiMAX or FM) for indoor localization, especially
the WiFi fingerprint-based indoor localization techniques. The
idea of fingerprint-based localization is to utilize the received
signal strength (RSS) to estimate the location of a receiver. The
main challenging issue for accurate fingerprint-based indoor
localization is the design of robust and discriminative RSS
signatures. In [12], the authors report a comparative survey of
WLAN location fingerprinting methods.
With the proliferation of wireless technologies, several ap-
proaches have been proposed to promote the WiFi and ZigBee
coexistence [13], [14]. In [15], the authors present an analysis
of the IEEE 802.15.4 and 802.11 interference patterns at 2.4
GHz ISM band, and then design a MAC layer solution that
enables 802.15.4 nodes to coexist with WiFi networks. Li et al.
present WiBee, a method to build a WiFi radio map with Zig-
Bee sensor nodes [16]. Through empirical study, the authors
reveal that the difference of the ZigBee RSS and the WiFi RSS
is almost a constant. Our work differs from WiBee in that, we
present a complete design and implementation of a localization
system with ZigBee radio, including the training set collection,
RSS-location fingerprint modeling, RSS fingerprint similarity
measurement, and location determination. While WiBee only
focuses on how to match the RSS detected by ZigBee radio
with RSS detected by WiFi NIC, in which a linear model
between them is proposed.
III. ZILOC DESIGN
A. Motivation
−100
−80
−60
−40
0 500 1000 1500 2000 2500 3000
RS
S V
alu
e (
db
m)
Sequence Number of RSS Samples
Fig. 3. WiFi signals detected by a ZigBee radio
There exists frequency overlap across 802.11 based WiFi
and 802.15.4 based ZigBee on the 2.4GHz ISM band. How-
ever, the transmission power of WiFi senders typically are
10 to 100 times higher than that of ZigBee senders [15],
leading to much higher energy consumption when collecting
WiFi signals. This motivates us to design a WiFi-fingerprinting
localization system using the 802.15.4 radio to collect WiFi
signals for the energy saving. Since the channel frequencies
of off-the-shelf 802.15.4 radios (e.g., CC2420 [17]) can be
programmed, it’s practical for ZigBee radios to sense the
spectrum usage of WiFi network when WiFi and ZigBee radios
operating at the same or adjacent frequencies. Besides, beacon
frames are transmitted periodically to announce the presence
of an AP in 802.11 infrastructure networks. Therefore, we can
use ZigBee radio to detect the beacon signal strength of WiFi
channels and find periodic features of beacon frames through
folding of the RSS samples [14]. Fig. 3 shows an example
that WiFi signals can be easily detected by a ZigBee radio.
−10.00
−5.00
0.00
5.00
10.00
0 6 12 18 24
Devia
tion (
ms)
Time (hour)
(a) The relative deviation between themeasurements and the standard bea-con interval 102.4ms
0.70
0.75
0.80
0.85
0.90
0.95
1.00
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4
CD
F
Deviation (ms)
First 8 HoursMiddle 8 Hours
Last 8 Hours24 Hours
(b) CDF of relative deviation at threedifferent periods in a day
Fig. 4. Stability of 802.11 beacon frame period
Fig. 4 illustrates the relatively stable beacon frame period
of an AP. In this experiment, we use the sniffier tool wireshark
[18] and WiFi equipped laptop to measure the timing accuracy
of the beacon transmissions of a specific AP. During a period
of 24 hours, the laptop records all overheard beacon frames
with reception time stamps. To evaluate the impact of WiFi
traffic load on the beacon period, we divide the beacon
frame flows into three sections and each section is analyzed
separately.
We measure the interval between the reception timestamps
of two consecutive beacons. The relative deviation between the
measurements and the standard beacon interval (102.4 ms) are
reported in Fig. 4. We observe the period of beacon frame sent
by an AP is fairly stable. Thus the relative phases of APs can
be regarded as a feature of fingerprint. In other words, any two
different APs have the same relative phase when they appear
in different fingerprints. ZiLoc utilizes the stability feature
of relative phases for the similarity measurement of different
fingerprints.
B. System Architecture
RSS measurements with
unknown location
RS
S s
ampl
ing
RS
S fo
ldin
g
Fingerprintcollector
RS
S s
hapi
ng
Zig
Bee
rad
io
Locationdetermination
KNN based classification
Marked location
fingerprintTraining set
Offline
Online
Fin
gerp
rint m
odel
ing
Fig. 5. ZiLoc Architecture Overview
The ZiLoc architecture overview is shown in Fig. 5. As
a conventional fingerprint-based localization system, ZiLoc is
based on a client-server architecture. It is composed of offline
and online phrases. The client-side, e.g., a mobile device with
built-in ZigBee interface, serves as the fingerprint collector in
the offline phase. The fingerprint collector first samples the
RSS of a specific WiFi channel, leveraging the overlapped
frequency between ZigBee and WiFi networks. Then, the
client-side adopts the same approach in [14], called Common
Multiple Folding (CMF), to identify an AP and capture its
signal strength. Finally, a fingerprint is extracted from fuzzy
RSS samples obtained by ZigBee radio.
For each individual interested indoor location, a certain
number of location fingerprints marked with position labels
are collected and reported to the server. After collecting
enough fingerprints, the RSS fingerprint training set (also
called radio map or fingerprint database) is built. During the
online phase, the observed RSS measurements associated with
an unknown position is transmitted to the backend server-side.
In the location determination module, ZiLoc utilizes KNN-
based classification method [8] to assign an estimated position
to the unmarked fingerprint. Then, the server-side sends back
the determined location to the ZiLoc client to achieve real-time
indoor localization and tracking.
A key challenge in ZiLoc design is the similarity mea-
surement of raw fingerprints. Although we can identify the
existence and signal strength of an AP beacon with ZigBee
radio, the identification, e.g., BSSID, of each AP cannot be
obtained. We address this issue by leveraging the relative phase
between two AP beacons, which will be introduced in detail
in Section III-D.
C. RSS Sampling, Shaping and Folding
We use TelosB [19] mote equipped with ZigBee-compliant
CC2420 radio to sample WiFi signals. From [14], the sampling
period is set as 122µs. Based on our testbed experiment,
we find that the default WiFi beacon transmission period is
102.4ms1, during which around 839 RSS samples can be
collected.
We briefly introduce the shaping and folding methods of
CMF, full details can be found in [14]. Since each WiFi AP
broadcasts beacon frames periodically, a ZigBee radio may
scan multiple APs’ beacon signals in a certain period (e.g.,
3s). Unlike conventional approach to collect the AP signals
using WiFi NICs, ZigBee radios can neither decode any WiFi
beacon, nor determine if a signal is associated with a specific
AP. Consequently, it is non-trivial to distinguish multiple AP’s
signals. In our design, to obtain RSS-location fingerprint, we
keep scanning 802.15.4 channel 17 for N (e.g., 30) beacon
periods2. It will take about N ∗ 102.4ms to complete this
process, and around N ∗K samples can be collected, where
K is the total number of RSS samples collected in a beacon
period. We partition all the samples into N groups, and denote
each group as an RSS vector Ri[k] (i ∈ [1, N ], k ∈ [1,K]),where i is the group number and k is the sample index in a
group.
To identify the presence of periodic AP beacon frames, RSS
shaping for each vector is performed. An RSS sample is set
to 0 if it is below −90dBm, otherwise, its value is set to 1
1The measurement is consistent with the results reported in [14]. Actually,the WiFi beacon transmission period can also be configured.
2The corresponding WiFi channel number is 6 as shown in Fig. 1, andchannels 1, 6, and 11 are the most common channels in WiFi networks.
1 0 1 ... 0 1 0
0 1 1 ... 0 0 1
1 1 0 ... 1 1 0
1 0 1 ... 0 1 1
0.3 0.8 0.2 ... 0.7 0.4 0.6
+
...
R'1
R'2
R'N-1
R'N
R*
1 2 3 ... K-2 K-1 K
+Folding
Shaping
Fig. 6. Example of the RSS shaping and folding
indicating the busy channel. Let R′
i[k] denote the shaped RSS
vector. Then, we fold R′
i[k], calculated as Eq. 1, and get the
normalized folded RSS vector R∗[k].
R∗[k] =1
N
N∑
i=1
R′
i[k], k ∈ [1,K] (1)
As shown in Fig. 6, R∗[k] can be considered as the probability
that an AP beacon frame is captured in the RSS series. For
obtaining the signal strength of AP beacon frames, we fold
the raw RSS vectors Ri[k] and get R[k], calculated as Eq. 2.
Each element of R[k] denotes the average signal strength of
a sampling point.
R[k] =1
N
N∑
i=1
Ri[k], k ∈ [1,K] (2)
D. RSS-location Fingerprint Modeling
0
0.2
0.4
α
0.6
0.8
1
0 100 200 300 400 500 600 700 800
Avg
. N
orm
aliz
ed
Fo
lde
d R
SS
Phase
Phase:100 Scan 1Scan 2
(a) We can easily find the relative phase of each AP withthe RSS folding results
−80
−70
−60
−50
0 100 200 300 400 500 600 700 800
Avg
. R
aw
RS
S (
dB
m)
Phase
[100,-75dBm]
[150,-65dBm]
AP1
AP2
AP3
AP4
AP5
AP6
AP1
AP2
AP3
AP4
AP5
AP6
Scan 1Scan 2
(b) RSS-location fingerprints are extracted
Fig. 7. Example of fingerprint modeling
R∗ are the normalized (also magnified) WiFi RSS series
and each periodic AP beacon signal is expected to generate
a peak in R∗. We consider that a peak in R∗ represents the
presence of an AP if the strength is larger than a threshold
α (we set α an empirical value as 0.5). Fig. 7(a) illustrates
the folded RSS series of two scans (each scan lasts 3072msin our experiment). As shown in the figure, after filtering out
the noises, 6 different AP beacons can be identified in each
scan. We observe that the RSS series in different scans show
similar signal envelopes, and the time difference between two
Sa
Sb
Match
Sa
Sb
Match
(b)(a)
APa1 APa2 APa3 APa4 APa5
APa1 APa2 APa3 APa4 APa6APa5
APa1 APa2 APa3 APa4 APa5
APa1 APa2 APa3 APa4 APa5APa6
Fig. 8. Example of matching two fingerprints
AP beacons within one scan is relatively stable. Let us define
the relative offset of a peak against the sampling start point as
the phase (f ) of an AP beacon. f is calculated by averaging
those adjacent samples whose strength are larger than α around
a peak. After obtaining the phase of an AP beacon, we can
easily get its signal strength h from R[k].
h = R[f ′] (3)
where f ′ is the closest sampling point to f in the RSS series.
Fig. 7(b) shows the corresponding signal strength of different
AP beacons in Fig. 7(a).
Since the AP beacon period and the time difference between
two AP beacons are stable, so the relative phase, i.e., the
relative distance of two peaks, also holds the stable feature.
We find that any AP can be uniquely identified by using a two-
tuple (phase, strength) in an RSS series. With this observation,
we use phase fi and strength hi to identify APi in one scan.
An RSS-location fingerprint is modeled as a set of two-tuples,
denoted by S : {(fi, hi)|i ∈ [1, n]}, where n is the number of
identified peaks in one scan.
E. Location Estimation
Given two fingerprints Sa and Sb, we now describe the
similarity measurement method between them. The distance,
which is the output of similarity measurement, will be used in
k-Nearest Neighbor (KNN) classification algorithm in ZiLoc.
The similarity measurement is composed of two steps. First,
we perform vector conversion of Sa and Sb. Then, we compute
the distance (similarity) of Sa and Sb.
1) Vector Conversion: Let |Sa| = na and |Sb| = nb denote
the number of identified APs in Sa and Sb, respectively.
|Sa
⋂
Sb| = nab represents the number of common APs that
both Sa and Sb contain. Since the relative phases among
APs within a fingerprint are stable, if nab ≥ 2, we can find
the best match between Sa and Sb in terms of the phase
(not considering the signal strength) through cyclic shift. As
showed in Fig.8, we cyclically shift the phase sequence of Sb
and align it with Sa. Then, we can get the number of matched
APs. Fig.8(a) only has 2 matched APs, while Fig.8(b) shows
the best matching case. After na∗nb attempts, we can find the
best match with the largest number of matched APs. If there
exist more than one best matchings with the same number of
matched APs, we randomly select one as the best match.
Since some APs in Sa may not have corresponding APs in
Sb, as shown in Fig. 9, we add virtual APs with fixed signal
strength mR in order to let Sa and Sb have the exactly same
AP sequences. In our experiment, we set mR as −110dBm.
Given the best match between Sa and Sb, we build the signal
Sa
Sb
APa1
(fa1,ha1)APa2
(fa2,ha2)APa3
(fa3,ha3)APa4
(fa4,ha4)APa5
(fa5,ha5)
APb1
(fb1,hb1)APb2
(fb2,hb2)APb3
(fb3,hb3)APb4
(fb4,hb4)APb5
(fb5,hb5)
Va ha1 ha2 mR ha3 ha4 ha5 mR
hb1 hb3 hb4Vb mR hb2 mR hb5
Virtual AP
Fig. 9. Example of fingerprint vector conversion
strength vectors Va and Vb after inserting virtual APs into the
fingerprints. For example, as shown in Fig. 9, both Sa and Sb
contain 5 APs, but their relative phases are different. In this
case, after getting the best match and performing the vector
conversion, we have Va = (ha1, ha2,mR, ha3, ha4, ha5,mR),and Vb = (mR,hb1, hb2, hb3,mR, hb4, hb5).
2) Similarity Computing: From the first step, we get the
signal strength sequences Va and Vb ordered by the phase,
respectively. Now, different methods can be applied to measure
their distance, which is actually the similarity of Sa and Sb. We
investigate two methods in ZiLoc, one is the Euclidian distance
as Eq. 4, and the other one is the Manhattan distance as Eq.
5. We mark the similarity of any two arbitrary fingerprints Sa
and Sb as Distance(Sa, Sb) = d(Va, Vb).
d(Va, Vb) =
√
n∑
i=1
(ha,i − hb,i)2
(4)
d(Va, Vb) =n∑
i=1
(|ha,i − hb,i|) (5)
Manhattan distance is also called city-block distance be-
cause it is the distance in blocks between any two points in
a city. An influential hypothesis is that Euclidian distance is
valid when the dimensions are perceptually integral, whereas
city-block distance is appropriate when dimensions are percep-
tually separable [20]. In this work, the dimensions are actually
the signal strength of surrounding APs, hence they are not
integral dimensions to some extent.
IV. EXPERIMENTATION
In this section, we evaluate ZiLoc performance in room level
and compare different KNN classification algorithms. KNN
classification [8] is a method for classifying objects based on
majority vote, with the object being assigned to the class most
common amongst its k nearest neighbors. Three different KNN
algorithms that are used in this work are shown as following.
• Euclidian-KNN: The conventional Euclidian distance is
used to measure the similarity of fingerprints.
• Euclidian-DW-KNN: This scheme also uses Euclidian
distance for similarity measurement. However, we as-
sign different fingerprint distances with different weights,
which is to mitigate the noisy fingerprints. Given the
distance Distance(Sa, Sb) between two fingerprints Sa
and Sb, its weight is calculated as 1/Distance(Sa, Sb)2.
• Manhattan-DW-KNN: It differs from Euclidian-DW-KNN
in that the Manhattan distance is used for fingerprint
similarity measurement.
We also report the performance of fingerprinting-based
localization using WiFi NIC to scan the MAC address and AP
RSS. In this test, we collect the same number of fingerprints
as ZiLoc experiment, and use the Manhattan-DW-KNN as the
classification algorithm.
A. Experimental Setup
1048
1049
1028
1030
1038
1039
1045 1044 1043 1042
Fig. 10. Floor plan of the experimental environment
Our experiment was conducted on the 10th floor of the New
Main Building at Beihang University, Beijing, China. The floor
plan is showed in Fig. 10 and the area of each room is about 30
square meters. There are 10 rooms involved in our experiment.
In our experiment, we didn’t deploy any extra AP and only
scanned the existing WiFi infrastructures. The devices used
to collect location fingerprints include a ThinkPad laptop and
a TelosB mote connected to the USB port of laptop. Based
on our field survey, a considerable number of APs work on
channel 6, so we set the ZigBee radio to scan channel 17,
which has a full overlap with WiFi channel 6. There are at
least 6 APs and 11 APs on average can be detected within
a room. For each room we collect 18000 fingerprints and we
randomly select 3000 as test samples, while the rest 15000 are
used to build training set. Each fingerprint is extracted from
the RSS samples of 30 beacon periods, so it costs about 3s to
collect one fingerprint.
B. ZiLoc vs. WiFi NIC-based method
In this comparison, the k value is fixed to 10 in KNN
classification. We report the ZiLoc performance in terms of
localization accuracy and energy consumption in Fig. 11 and
Fig. 12, respectively.
0
0.2
0.4
0.6
0.8
1
1028 1030 1038 1039 1042 1043 1044 1045 1048 1049
Acc
ura
cy
Room NumbersEuclidian-KNN
Euclidian-DW-KNN
Manhattan-DW-KNN
Manhattan-DW-KNN (WiFi NIC)
Fig. 11. Localization accuracy of room level evaluation
1) Localization Accuracy: Fig. 11 shows the localization
accuracy of different methods. Overall, ZiLoc under three
classification algorithms can provide comparable performances
to the WiFi NIC-based method. DW-KNN algorithm outper-
forms the one without distance-weighted KNN. It is interesting
that Manhattan-DW-KNN shows a higher accuracy compared
to the other two classification algorithms. WiFi NIC-based
method shows around 88% accuracy on average, while ZiLoc
with DW-KNN algorithm can achieve an average of 85%
accuracy, which is very close to the accuracy of WiFi NIC-
based method.
2) Energy Consumption: To measure the practical respec-
tive energy consumption of ZigBee radio and WiFi radio, we
use an external WiFi NIC. Fig. 12(a) shows the measurement
equipments in our experiment. Fig. 12(b) plots the energy
consumption as the time elapsed. As seen in the figure, ZiLoc
costs only about 34% energy of the WiFi NIC-based method.
(a) Measurement equipments
0
5
10
15
20
25
30
0 10 20 30 40 50 60
En
erg
y C
on
su
mp
tio
n (
J)
Time Elapsed (s)
WiFi NIC
ZigBee Radio
(b) Energy consumption
Fig. 12. Comparison of energy consumption
C. Impact of k in KNN
Fig. 13 shows the localization accuracy of each individual
room with the change of k values. As shown in Fig. 13(b),
room 1030 has the highest and most stable classification
accuracy. This is because, this room is relatively isolated
and far from the other interested rooms as seen in Fig. 10.
Intuitively, the fingerprints in room 1030 would show more
differences from those fingerprints in other rooms, hence it
suffers fewer interference when performing the classification.
Surprisingly, we find that those rooms, whose two neighboring
rooms are also involved into our localization experiment, such
as 1043, 1044 and 1049, can achieve a relatively higher
accuracy. In contrast, the other rooms, that have only one
neighboring room involved, such as 1028, 1038, 1039 and
1042, show a comparatively lower accuracy.
The average accuracy of the three KNN algorithms with
different k values are showed in Fig. 14. We can see that all
the three algorithms are descending along with the increase of
k values. It is interesting that when k is small, e.g., less than
5, there is a small increase of the localization accuracy. This
indicates our methods can get the best performance when k is
small. While when k is larger than 6, the accuracy of all the
three algorithms are descending with the increase of k. Thisis because the sample spaces of different rooms are highly
correlated with each other. Therefore, when more neighbors
are involved into the classification, it also brings noises.
0
0.2
0.4
0.6
0.8
1
1 200 400 600 800 1000
Accura
cy
k
Euclidian-KNNEuclidian-DW-KNN
Manhattan-DW-KNN
(a) 1028
0.8
0.85
0.9
0.95
1
1 200 400 600 800 1000
Accura
cy
k
Euclidian-KNNEuclidian-DW-KNN
Manhattan-DW-KNN
(b) 1030
0
0.2
0.4
0.6
0.8
1
1 200 400 600 800 1000
Accura
cy
k
Euclidian-KNNEuclidian-DW-KNN
Manhattan-DW-KNN
(c) 1038
0
0.2
0.4
0.6
0.8
1
1 200 400 600 800 1000
Accura
cy
k
Euclidian-KNNEuclidian-DW-KNN
Manhattan-DW-KNN
(d) 1039
0
0.2
0.4
0.6
0.8
1
1 200 400 600 800 1000
Accura
cy
k
Euclidian-KNNEuclidian-DW-KNN
Manhattan-DW-KNN
(e) 1042
0
0.2
0.4
0.6
0.8
1
1 200 400 600 800 1000
Accura
cy
k
Euclidian-KNNEuclidian-DW-KNN
Manhattan-DW-KNN
(f) 1043
0
0.2
0.4
0.6
0.8
1
1 200 400 600 800 1000
Accura
cy
k
Euclidian-KNNEuclidian-DW-KNN
Manhattan-DW-KNN
(g) 1044
0
0.2
0.4
0.6
0.8
1
1 200 400 600 800 1000
Accura
cy
k
Euclidian-KNNEuclidian-DW-KNN
Manhattan-DW-KNN
(h) 1045
0
0.2
0.4
0.6
0.8
1
1 200 400 600 800 1000
Accura
cy
k
Euclidian-KNNEuclidian-DW-KNN
Manhattan-DW-KNN
(i) 1048
0
0.2
0.4
0.6
0.8
1
1 200 400 600 800 1000
Accura
cy
k
Euclidian-KNNEuclidian-DW-KNN
Manhattan-DW-KNN
(j) 1049
Fig. 13. Accuracy of each room with different k values
0
0.2
0.4
0.6
0.8
1
1 200 400 600 800 1000
Accura
cy
k
Euclidian-KNN
Euclidian-DW-KNN
Manhattan-DW-KNN
0.75
0.8
0.85
0.9
2 4 6 8 10 12 14 16 18 20
Fig. 14. Average accuracy with different k values
Remarkably, for any k values, Manhattan-DW-KNN algorithm
always shows the highest accuracy. With these observations,
we conclude that Manhattan-DW-KNN algorithm is a simple
yet effective algorithm in ZiLoc.
V. CONCLUSION
In this work, we have developed ZiLoc system, an energy
efficient WiFi fingerprint-based localization with ZigBee radio.
We present a novel RSS-location fingerprint model to uniquely
identify APs’ beacon signal features. We then propose a simple
yet effective method to compute the distances between the
testing and training RSS fingerprints. We have implemented
ZiLoc, and conducted extensive testbed experiments to study
the performances. Interestingly, through experimental results,
we find that Manhattan distance is more suitable for RSS
fingerprint similarity measurement than the Euclidian distance.
Experimental results also demonstrate that ZiLoc can achieve
an average of about 85% room-level localization accuracy and
reduce around 66% energy consumption compared with WiFi
NIC-based method when collecting RSS fingerprints.
ACKNOWLEDGEMENTS
This work was supported in part by the Research Fund of the StateKey Laboratory of Software Development Environment under GrantNo. BUAA SKLSDE-2012ZX-17, National Natural Science Foun-dation of China under Grant No. 61170296, 61190120, Program forNew Century Excellent Talents in University under Grant No. NECT-09-0028, Singapore-MIT IDC IDD61000102a, IDG31100106a, andNRF2012EWT-EIRP002-045.
REFERENCES
[1] R. Want, A. Hopper, V. Falcao, and J. Gibbons, “The active badgelocation system,” ACM Trans. Inf. Syst., vol. 10, no. 1, pp. 91–102,Jan. 1992.
[2] T. He, C. Huang, B. M. Blum, J. A. Stankovic, and T. Abdelzaher,“Range-free localization schemes for large scale sensor networks,” inProc. MobiCom ’03, 2003.
[3] K. Whitehouse and D. Culler, “A robustness analysis of multi-hopranging-based localization approximations,” in Proc. IPSN ’06, 2006.
[4] M. Youssef, M. Mah, and A. Agrawala, “Challenges: device-free passivelocalization for wireless environments,” in Proc. ACM MobiCom ’07,2007, pp. 222–229.
[5] D. H. Kim, Y. Kim, D. Estrin, and M. B. Srivastava, “Sensloc: sensingeveryday places and paths using less energy,” in Proc. ACM SenSys ’10,2010, pp. 43–56.
[6] S. P. Tarzia, P. A. Dinda, R. P. Dick, and G. Memik, “Indoor localizationwithout infrastructure using the acoustic background spectrum,” in Proc.
MobiSys ’11, 2011, pp. 155–168.[7] C. Chen, X. Zhang, J. Zhang, and Y. Tang, “usd card: a plug play
solution for mobile device to access wireless sensor networks,” in Proc.
WASA’11, 2011, pp. 354–365.[8] V. Nitin Bhatia, “Survey of nearest neighbor techniques,” Journal of
Computer Science and Information Security, vol. 8, no. 2, 2010.[9] D. Hauschildt and N. Kirchhof, “Advances in thermal infrared localiza-
tion: Challenges and solutions,” in Proc. IPIN ’10, 2010, pp. 1–8.[10] A. Jimenez Ruiz, F. Seco Granja, J. Prieto Honorato, and J. Gue-
vara Rosas, “Accurate pedestrian indoor navigation by tightly couplingfoot-mounted imu and rfid measurements,” IEEE Transactions on In-
strumentation and Measurement, vol. 61, no. 1, pp. 178–189, jan. 2012.[11] E. Mangas and A. Bilas, “Flash: Fine-grained localization in wireless
sensor networks using acoustic sound transmissions and high precisionclock synchronization,” in Proc. ICDCS ’09, 2009.
[12] V. Honkavirta, T. Perala, S. Ali-Loytty, and R. Piche, “A comparativesurvey of wlan location fingerprinting methods,” in Proc. WPNC ’09.
[13] J. Huang, G. Xing, G. Zhou, and R. Zhou, “Beyond co-existence:Exploiting wifi white space for zigbee performance assurance,” in Proc.
ICNP ’10, 2010, pp. 305–314.[14] R. Zhou, Y. Xiong, G. Xing, L. Sun, and J. Ma, “Zifi: wireless lan
discovery via zigbee interference signatures,” in Proc. MobiCom ’10.[15] C.-J. M. Liang, N. B. Priyantha, J. Liu, and A. Terzis, “Surviving wi-fi
interference in low power zigbee networks,” in Proc. SenSys ’10, 2010.[16] W. Li, Y. Zhu, and T. He, “Wibee: Building wifi radio map with zigbee
sensor networks,” in Proc. INFOCOM ’12, 2012.[17] “Cc2420: 2.4 ghz ieee 802.15.4 zigbee-ready rf transceiver.” [Online].
Available: http://www.ti.com.cn/cn/lit/ds/symlink/cc2420.pdf[18] “wireshark.” [Online]. Available: www.wireshark.org[19] “Telosb datasheet.crossbow technology inc.” [Online]. Available:
http://www.xbow.com.cn/wsn/pdf/TelosB Datasheet.pdf[20] D. M. Ennis, F. Gregory Ashby, “Similarity measures,” Scholarpedia,
vol. 2, no. 12, 2007.