Professional Documents
Culture Documents
Building A Practical Wifibased Indoor Navigation System 2014
Building A Practical Wifibased Indoor Navigation System 2014
Building a Practical
Wi-Fi-Based Indoor
Navigation System
This article presents the seven-step process involved in building a
practical Wi-Fi-based indoor navigation system, which was implemented
at the COEX complex in Seoul, Korea, in 2010. More than 200,000 users
downloaded the system in its first year of use.
A
s car navigation systems have The COEX Complex
become more ubiqu itous , Figure 1 provides a schematic view of the
researchers have recognized the COEX complex. The buildings, towers, and
potential for indoor navigation stores in the COEX area are mostly intercon-
systems to help pedestrians find nected via the COEX underground shopping
their way in complicated indoor areas such as mall area in the B1 level. The total length of
airport terminals, subways, shopping malls, the paths is more than 10 km, and the total size
and exhibition and convention centers. In addi- of the COEX underground shopping mall area
tion, the recent proliferation of Wi-Fi-equipped amounts to more than 450,000 m2 (comparable
smartphones and the rapid to 70 football stadiums). Large parking areas,
expansion of Wi-Fi zones have located on the B2, B3, and B4 levels, can accom-
moved this idea from wishful modate approximately 3,000 cars. More than
Dongsoo Han, Sukhoon Jung,
thinking to near reality. 200 e xhibitions and 2,000 conferences are held
Minkyu Lee, and Giwan Yoon
In October 2010, a Wi-Fi- at the COEX every year.
Korea Advanced Institute of
based indoor navigation system Although previous work confirmed the fea-
Science and Technology for the COEX complex (a land- sibility and potential of Wi-Fi-based localiza-
mark area in Seoul, Korea) was tion techniques,1 customers and vendors have
released by our team at the Ko- been reluctant to apply these techniques when
rea Advance Institute of Science and Technology developing location-based service (LBS) appli-
to smartphone users via the myCoex application. cations that require high localization accuracy,
The COEX area comprises Asia’s largest under- particularly in large indoor spaces. Most work
ground shopping mall area, three five-star hotels, failed to provide convincing evidence to custom-
one 55-story and one 41-story premier office ers and vendors, because the experiments were
tower, a large department store, a subway station, conducted in relatively small and confined in-
a city airport terminal, and more. door areas—not larger spaces, such as COEX.
Although we proved that it’s possible to build
a practical Wi-Fi-based indoor localization sys- Seven-Step Development Process
tem, developing such a system for a large indoor To address this issue, we decided to test exist-
area wasn’t an easy task. Here, we review the ing Wi-Fi-based indoor localization techniques
long and tedious steps required. and integrate them to build the COEX indoor
1 3 4 5 6
AP analysis/ Testing/
design goal set up Indoor map drawing WRM construction System build-up
2 7 service launching
AP analysis & installation Indoor map drawing Wi-Fi fingerprint Ul build-up Testing & evaluation
collection (dbm)
Items Results –10
Setting up design goals Marking of way and WRM construction System build-up System launching &
measure points user feedback
Items Design goals Indoor navigation app
AP1 AP2 AP3 AP4
average 5–8m ... ... ... ... ...
errors A7 –88 –20 –93 –82
Localization
engine
response A8 –90 –19 –95 –82
MST + 1 POI & WRM DB
B1 –22 –86 –62 –99
time seconds B2 –26 –76 –65 –94
Android
size of ~3 Mb B3 –33 –64 –74 –87 iPhone devices Nokia
B4 –47 –57 –77 –81
WRM B5 –54 –48 –83 –62
Figure 2. The seven steps in building an indoor navigation system for a large indoor area.
corners, and at intersections. Usually, be drawn was determined by a simple of collecting Wi-Fi fingerprints by ar-
two waypoints are enough to repre- function, [w/i], where w is the width tificially generating Wi-Fi fingerprints
sent a straight line, but when a path is of a corridor and i is a predefined in- using the relatively smaller number of
rather long or bent, multiple waypoints terval between two adjacent measure manually collected Wi-Fi fingerprints.5
might need to be used so that the path points—3 meters in our case. However, the results of applying the
can be represented by connecting the interpolation technique could hardly
waypoints. Wi-Fi Radio Map Construction outperform the accuracy achieved on
Once the network is prepared, the After marking the measure points on the manually collected Wi-Fi finger-
measure points are marked to collect the map, the next step requires col- prints. With COEX, the interpolation
the Wi-Fi fingerprints on or alongside lecting precise Wi-Fi fingerprints at technique was applied only for the
the links of the network. The tool in each marked measure point, which is parking areas.
Figure 3a shows the task of marking critical for accurate localization.6 The For data collection, the collectors
the measure points. In the tool, a circle Wi-Fi fingerprints should be densely download the maps marked with the
represents a waypoint, and an arrow collected for an accurate localization, reference points to their smartphones
connecting the circles represents a link. and at least 20–30 Wi-Fi fingerprints (Figure 3b), choose and locate the
A line can be drawn by marking each should be collected at each point for marked reference points, and then
end of the line with blue cross marks reliability. Consequently, the Wi-Fi fin- start collecting the Wi-Fi fingerprints
as reference points. The planners must gerprint collecting activity in a large- along the lines connecting the reference
decide on and mark the reference points scale indoor space can be difficult and points. They use their smartphones’
in advance. Once both ends of a line are time-consuming. In COEX, dozens Wi-Fi fingerprint-collecting tool for
decided, the measure points on the line of collectors performed the collecting this purpose (Figure 3c). The tool sup-
are automatically generated in the same activity in parallel, after dividing the ports the collectors to check if the lo-
interval by the tool. entire area into several subareas—each cation where the collection has started
When a corridor was too wide to of which was assigned to a particular is the correct location by providing
be covered by a single measurement collection team. information on every reference point.
line, multiple lines were drawn along We developed a radio map interpo- The interval between two a djacent col-
the way link. The number of lines to lation technique to reduce the efforts lection points is set to a predetermined
Figure 3. Waypoint marking: (a) the tool to mark the measure points and waypoints on a PC, (b) a map downloaded to a
smartphone marked with the reference points, and (c) the tool to collect Wi-Fi fingerprints on a smartphone.
interval, 3 meters in our case, to map the phone, so we instructed collectors localization engine,7 because it was
them to their corresponding mea- not to hold the bottom of the phone and slightly more accurate than other meth-
sure points afterwards. In the case of not to move or change the height of the ods at COEX,8 and it was flexible for
COEX, the collector could ensure the phone while collecting 20 fingerprints. fine-tuning and further enhancing the
interval relatively easily because the Although the collector’s body could engine’s accuracy. The kWNN simply
floors were tiled with 40 × 40 cm square cause the attenuation of signal strengths, outputs the weighed centroid of the
tiles. Without the tiles, we might need we asked the collectors to hold the top k locations, with the weights de-
to use a distance measurement tool to phone in front of their chests facing one termined by the distance between the
ensure the interval. direction. We confirmed that the accu- captured fingerprint and each nearest
The Wi-Fi fingerprint collecting ac- racy was not, or was only slightly, influ- neighbor. However, the accuracy level
tivity was performed by 15 collectors enced by the body and direction in a very required for the practical use of the sys-
for two weeks. We organized seven crowded area such as COEX. tem couldn’t be achieved solely with
teams: six teams with two members in the localization engine.
each team for collection, and one team System Build-Up Thus, we developed a signal filter, an
with three staff members for supporting After preparing the WRM, indoor adaptive Kalman filter, and a delay filter,
the collection activities of the six teams. map, and way network, we installed a and integrated them with the localiza-
The Wi-Fi fingerprints were collected at localization engine on the WRM and tion engine to further improve accu-
about 10,000 measure points, and 20 the path-finding routine on the indoor racy. The localization engine deals with
Wi-Fi fingerprints were collected at map. Then, we installed the navigation only a single Wi-Fi fingerprint at a time,
each point, resulting in a data collection applications on top of the localization whereas the filters use the sequences of
of around 200,000 total Wi-Fi finger- engine and path-finding routine. the consecutively scanned Wi-Fi finger-
prints. Also, the 20 Wi-Fi fingerprints The localization engine installed on prints or locations (that is, historic data).
at each point were merged into an aver- the WRM is the core component of the Given that Wi-Fi fingerprints can be
age Wi-Fi fingerprint, so approximately indoor navigation system. Along with the scanned every two or three seconds for
10,000 average Wi-Fi fingerprints were WRM, the performance of the engine is indoor navigation, the approach taken
collected into the Wi-Fi radio map critical for a more accurate localization. by the filters is reasonable.
(WRM—available for download at Accordingly, various techniques and Strong Wi-Fi signals are sometimes
https://kailos.io/wrm/seoul/coex). filters should be integrated together with missed in a captured fingerprint be-
We used a Samsung Galaxy S1 for the localization engine. cause of limited scanning time. The
the collection. In the Galaxy S1, the We adopted the k-weighted near- signal filter imputes the Wi-Fi finger-
Wi-Fi-antenna is on the bottom left of est neighbors (kWNN, k = 3) for our print captured for localization using
–20 556
557
–25
–30
–35
–40
–45
–50
–55
–60
–65
–70
–75
–80
–85
–90 400
–95
–100
1 2 3 4 5 6 7 8
40 Location 350
30 300
20 250
78
56
8
0
67
205
34
–5
10 –10 204
5
207
12
–20 206
4
201
–25
3
–30
2
0 –35
–40
1
0 –45
–50
100 –55
–60
–65
200 –70
–75
300 –80
–85
400 –90
–95
500 –100
1 2 3 4 5 6 7 8
600
Location
700 0
800
Figure 5. Measured accuracy and Wi-Fi signal distributions excerpted from the box areas in the B1 level of COEX. The accuracy of
the areas with only weak Wi-Fi signals was always worse than that of the areas with strong Wi-Fi signals. Black, blue, violet, and
red bars represent 3-, 6-, 9-, and more than 12-meter error distances, respectively.
of 6.0 meters at area A, 5.9 meters Figure 5 shows this situation for the c orrectly used, the effect of applying
at area B, and 7.1 meters at area C. B1 level of the COEX area. The length the filters would be great. Despite the
Considering that no special APs were of the bars at a location indicates the system’s inability to achieve 100 per-
installed to support localization, the average error distance there. The blue cent accuracy in the floor-level esti-
results aren’t so bad. The floor-level and black bars represent the locations mation, almost all of the inaccurate
accuracies obtained at areas A, B, and with good accuracy (less than 3 and floor-level estimations were removed
C were 99.6, 96.2, and 90.6 percent, less than 6 meters in error distances, or deactivated, mainly by the applica-
respectively. respectively), and the violet and red tion of the filters. The adaptive Kalman
The accuracy differences in each area bars represent the locations with bad filter had the most prominent effect in
indicate that the strength of AP signals accuracies (less than 9 and greater than this aspect. The delay filter was effec-
in Wi-Fi fingerprints is more impor- 12 meters in error distances). As the tive in controlling outliers because it
tant to accuracy than the number of two side graphs illustrate, the accuracy could deactivate the display of loca-
AP signals. The accuracy at area B was of localization depended on the distri- tions if the newly estimated location
slightly better than that at area A. It butions of the Wi-Fi signal strengths. deviated too far from the previous loca-
seemed that the gain by the strong AP The locations with strong signals (that tions. The signal filter imputed missed
signals in area A was offset by that ar- is, locations near the peak of a particu- AP signals for the captured Wi-Fi sig-
ea’s complex structure. Area A included lar AP signal) showed better accuracy nals, which helped improve the overall
a mixture of meandering paths and than the locations with only weak or localization accuracy. After activating
small- and medium-sized open spaces, medium signal strengths. all the filters integrated into the system,
whereas area B had a grid structure As the figure illustrates, the l ocations the displayed locations proceeded on
with relatively low ceilings. with low accuracies were usually the way network in a stable manner.
To better understand accuracy fluc- detected in combination with the However, activating the filters
tuation, we depicted the accuracy levels locations with good or high a ccuracies, caused a significant delay in response
obtained at each location in a figure. implying that if the trace data were time. Most of this was caused by the
delay filter. When we deactivated this a known indoor map. Tagging the s ystems are integrated with the indoor
fi lter, we could reduce the response locations of the assembled fingerprint navigation system to direct users from a
time within around MST + 0.2 sec- sequences by referencing only indoor source to a destination, hopefully with-
onds, but it wasn’t worth it. The us- maps is a challenging problem. out any disconnection. Of course, sev-
ers were more tolerant of the delay The reconfiguration of spaces, eral technical issues, including seamless
than of the possible incorrect location which frequently happens (particu- integration of indoor and outdoor navi-
estimation. larly in exhibition areas), is another gation systems, should be addressed to
issue to be addressed. As a solution, realize this system.
Service Launching and the indoor map and its associated way Yet the COEX system has already
User Feedback network should be changed. Then, affected the marketplace. After expe-
After testing and evaluating a naviga- the WRM should be reconstructed. riencing the COEX navigation system,
tion system, the system should be in- However, if the change o ccurs too many customers and vendors recognize
tegrated with a more general-purpose frequently, this approach requires the potential of Wi-Fi-based indoor
smartphone app so it can be distributed considerable time and cost. A walk- localization and navigation systems for
to users. Then, the app can collect, ac- ing survey, in which fingerprints are their businesses. For example, Incheon
cumulate, and analyze user feedback to collected while collectors are walking International Airport decided to pro-
further improve the navigation system. in a space and the collected l ocations vide an indoor navigation service to
In the case of COEX, we integrated are later estimated for tagging the help passengers find their way in the
the COEX indoor navigation system locations of the fingerprints, is one airport terminal, and some hospitals
into the myCoex app. The myCoex approach used to cope with this situ- are applying localization technology
app provides general services such as ation. Crowdsourcing is another. We to monitor the locations of doctors,
exhibition and event notification, fa- developed a crowdsourcing WRM nurses, staff, and even patients.
cility finding, and providing informa- construction method to collect finger-
tion about stores and offices within the prints of POIs through user partici- Acknowledgments
complex as well as the indoor navi- pation. 13 Wi-Fi-SLAM 14 and Zee 15 This work was funded by the Korea Trade NET-
gation service. The myCoex app was incorporate the capability of smart- work (KTNET) and Korea International Trade
downloaded more than 200,000 times phone-integrated sensors such as Association (KITA). We thank the anonymous
referees for their insightful and constructive
its first year, with approximately 20 three-axis accelerometers, compasses, comments, which greatly improved the quality
percent of the downloads from abroad. gyroscopes, and pressure sensors to and readability of this article.
Meanwhile, the ever-changing AP build WRMs automatically.
environment is one of the nagging The more user feedback data we
issues of Wi-Fi-based indoor localiza- have, the further we can enrich the ser- References
tion and navigation systems. We ad- vices of the app. Identifying crowded or 1. P. Bahl and V. Padmanabhan, “RADAR:
dressed this problem to some extent by less crowded areas and analyzing visi- An In-building RF-based User Loca-
utilizing user feedback data. We con- tors’ moving patterns are possible for tion and Tracking System,” Proc. 19th
Ann. Joint Conf. IEEE Computer &
firmed that the addition of new APs the valuation of POIs. Comm. Societies (INFOCOM 00), 2000,
and the removal of existing APs can be pp. 775–784.
W
reflected to the WRM by analyzing the
2. H. Liu et al., “Survey of Wireless Indoor
appearance patterns of APs in the feed- e can make the indoor Positioning Techniques and Systems,”
back Wi-Fi fingerprints.12 The feed- navigation system more IEEE Trans. Systems, Man, and Cyber-
back data was collected without the complete by incorpo- netic: Applications & Reviews, vol. 37,
no. 6, 2007, pp. 1067–1080.
intervention of users while they were rating pedestrian dead-
using the service. Meanwhile, the ac- reckoning techniques and using various 3. A. Kushki, K. Plataniotis, and A. Venet-
curacy wasn’t significantly influenced sensors embedded in smartphones. For sanopoulos, “Kernel-based Positioning in
by a slight change of the AP environ- example, floor-level accuracy can be Wireless Local Area Net-works,” IEEE
Trans. Mobile Computing, vol. 6, no. 6,
ment—for example, from 10 to 20 per- drastically improved with the help of a 2007, pp. 689–705.
cent. However, when the change was pressure sensor. In building a WRM,
made for more than half of the APs, the crowdsourcing approach will play a 4. D. Madigan et al., “Bayesian Indoor
Positioning Systems,” Proc. 24th Ann.
accuracy drastically degraded. critical role in reducing time and costs Joint Conf. IEEE Computer & Comm.
For a more precise WRM update, we in the future. Societies (I N FOCOM 05), 20 05,
need to assemble a sufficient number of We are currently developing a door- pp. 1217–1227.
fingerprint traces into long sequences to-door navigation system, in which 5. M. Lee and D. Han, “Voronoi Tessel-
and fit the assembled sequences into bus, subway, and train navigation lation Based Interpolation Method for
7. V. Moghtadaiee, A. Dempster, and S. Lim, Suk-Hoon Jung is a PhD student of information and communications engi-
“Indoor Localization Using FM Radio neering at Korea Advanced Institute of Science and Technology. His research
Signals: A Fingerprinting Approach,” Proc. interests include real-time locating systems and pervasive computing. Jung has
2011 IEEE Indoor Positioning Indoor an MS in information and communications engineering from KAIST. Contact
Navigation (IPIN 11), 2011, pp. 1–7. him at sh.jung@kaist.ac.kr.
Call Articles
12. J. Lim et al., “Radio Map Update Auto-
mation for WiFi Positioning Systems,”
IEEE Comm. Letters, vol. 7, no. 4, 2013,
pp. 693–696.
for
13. M. Lee et al., “Elekspot: A Platform for
Urban Place Recognition via Crowdsourc-
ing,” Proc. 2nd IEEE/IPSJ Int’1 Symp. IEEE Software seeks practical, readable
Applications and the Internet (SAINT articles that will appeal to experts and nonexperts
12), 2012, pp. 190–195.
alike. The magazine aims to deliver reliable
14. B. Ferris, D. Fox, and N. Lawrence. information to software developers and managers
“WiFi-SLAM Using Gaussian Process to help them stay on top of rapid technology
Latent Variable Models,” Proc. 20th Int’l
Joint Conf. Artificial Intelligence (IJCAI change. Submissions must be original and no
07), 2007, pp. 2480–2485. more than 4,700 words, including 200 words
for each table and figure.
15. A. Rai et al., “Zee: Zero-effort Crowd-
sourcing for Indoor Localization,” Proc.
18th Int’l Conf. Mobile Computing Author guidelines:
and Networking (MobiCom 12), 2012, www.computer.org/software/author.htm
pp. 293–304. Further details: software@computer.org
www.computer.org/software
Selected CS articles and columns
are also available for free at
http://ComputingNow.computer.org.