Analysis of Wi-Fi-based and Perceptual Congestion
Masaki Igarashi, Atsushi Shimada, Kaito Oka and Rin-ichiro Taniguchi
Kyushu University, 744, Motooka, Nishi-ku, Fukuoka, Japan
{igarashi, atsushi, kaito}@limu.ait.kyushu-u.ac.jp, rin@kyudai.jp
Keywords:
Congestion Estimate, Perceptual Congestion, Probe Request.
Abstract:
Conventional works for congestion estimates focus on estimating quantitative congestion (e.g., actual number
of people, mobile devices, and crowd density). Meanwhile, we focus on perceptual congestion rather than
quantitative congestion toward providing perceptual congestion information. We analyze the relationship
between quantitative and perceptual congestion. For this analysis, we construct a system for estimating and
visualizing congestion and collecting user reports about congestion. We use the number of mobile devices as
quantitative congestion measurements obtained from Wi-Fi packet sensors, and user-report-based congestion
as a perceptual congestion measurement collected via our Web service. Base on the obtained quantitative and
perceptual congestion, we investigate the relationship between these values.
1 INTRODUCTION
Congestion measurements and estimates are useful
and important for various applications. Congestion
information can assist congestion avoidance and mit-
igation. It is also important that we grasp the num-
ber of visitors to retail stores for customer analysis
and marketing strategies. Additionally, evacuation
planning for disasters requires congestion information
(Choi et al., 2011).
The extent of congestion is measured or estimated
manually or using sensors such as cameras and Wi-
Fi packet sensors. Vision-based congestion estimates
using cameras have recently been developed, and the
accuracy of these methods improves with each year.
Cameras must be carefully installed in the target area
considering occlusion and blind areas, and the initial
costs tend to be high. Wi-Fi packet sensors estimate
the number of mobile devices (e.g., smartphones and
laptop computers). The number of mobile devices
tends to be proportional to the number of people, so
we can use them to roughly estimate congestion. Wi-
Fi packet sensors cover distances between dozens to
a hundred meters. A Wi-Fi radio wave has a higher
transmittance than visible light, therefore we can in-
stall packet sensors in typical situations without con-
sidering blind areas.
The above-mentioned techniques are aimed at es-
timating quantitative congestion measurements such
as people count, crowd density, and the number of
mobile devices. For customer analysis, the actual
(a) Dining hall (b) Bus stop
Figure 1: Two spots with similar crowd density.
people count and density are useful and important fac-
tors.
Meanwhile, in terms of providing congestion in-
formation to people, qualitative congestion measure-
ments such as a person’s perception are also impor-
tant. Figure 1 shows two spots with almost the same
crowd density. There are few vacant seats in the
dining hall, so we would feel that the dining hall is
crowded. The crowd density at the bus stop is similar
as the dining hall, but the bus stop cannot be consid-
ered crowded. Human perception about congestion
depends on the people count and density and also the
locations characteristics, such as area and seating ca-
pacity.
In this paper, we focus on the relationship between
quantitative and perceptual congestion toward provid-
ing perceptual congestion information. We use the
number of mobile devices as quantitative congestion
measurements obtained from Wi-Fi packet sensors,
and user-report-based congestion as a perceptual con-
gestion measurement collected via our Web service.
We investigate the relationship between these values.
Igarashi, M., Shimada, A., Oka, K. and Taniguchi, R-i.
Analysis of Wi-Fi-based and Perceptual Congestion.
DOI: 10.5220/0006206102250232
In Proceedings of the 6th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2017), pages 225-232
ISBN: 978-989-758-222-6
Copyright
c
2017 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
225
Probe request
Wi-Fi packet sensor
Server
Congestion estimate and forecast
Congestion visualization
Accuracy improvement of
congestion estimate by user reports
Visualization
User reports
Figure 2: Overview of our system.
2 RELATED WORK
2.1 Preliminaries
2.1.1 Probe Request and Counting Mobile
Devices
Probe request is a packet (or more precisely, a frame)
defined in IEEE 802.11, which is broadcast by a mo-
bile device with a Wi-Fi function when the device
searches for access points before establishing a con-
nection. The transmitting interval is between several
to several hundred seconds and depends on the device.
The probe request frame includes a MAC address for
the sender, so the receiver can identify the sender and
count the peripheral mobile devices. Additionally, the
receiver can obtain a received signal strength (RSS)
value. The RSS value is lower for larger distances be-
tween the sender and receiver. Therefore, the receiver
can roughly estimate the distance to the sender using
the RSS value.
2.1.2 Wi-Fi Packet Sensor
A Wi-Fi packet sensor is a sensor for collecting probe
request frames. We can make a prototype Wi-Fi
packet sensor using a single-board computer such as
a Raspberry Pi
1
. Commercial Wi-Fi packet sensor
products have recently become available
2
. A Wi-
Fi packet can be sent from over one hundred meters
away, so they can cost effectively cover a large area.
2.2 Estimation and Analysis of Crowd
Density and Pedestrian Flow
Fukuzaki et al. analyzed pedestrian flow using Wi-
Fi packet sensors (Fukuzaki et al., 2014). Yaik et
1
https://www.raspberrypi.org/
2
http://aibeacon.jp/
al. compared the number of Wi-Fi frames and crowd
counting manually (Yaik et al., 2016). They evalu-
ate the correlation between manual and Wi-Fi frames
counting. Above-mentioned methods used Wi-Fi
probe requests. Xi et al. use Channel State Infor-
mation (CSI) of Wi-Fi for counting crowd (Xi et al.,
2014). Their method outperforms the state-of-the-art
approaches in terms of accuracy and scalability.
Weppner et al. estimated crowd density by Blue-
tooth scan data (Weppner and Lukowicz, 2013).
Schauer et al. proposed a hybrid method for esti-
mating crowd density and pedestrian flow using Wi-Fi
and Bluetooth (Schauer et al., 2014). They estimated
the crowd density (defined as the number of people
per unit area) for each spot using Wi-Fi and Blue-
tooth, and compared the estimated number of people
with the ground truth.
These works considered the number of people or
the number of devices, but not the peoples perception
of the congestion.
3 SYSTEM FOR CONGESTION
ESTIMATE AND COLLECTING
USER REPORTS
In this section, we describe our system for estimating
and visualizing the congestion, and collecting user re-
ports.
3.1 System Overview
Figure 2 gives an overview of this system. We capture
probe request frames and store tuples of the received
time, location ID, and MAC address to the database.
Then, we calculate the extent of the congestion us-
ing the data stored in the database. Our Web service
plots a time series of the congestion for each location.
This service has a function for receiving user reports
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
226
Big Sand 1F Big Sand B1F
QASIS
Learning
Space
Bus stop
(Nishitetsu)
Bus stop
(Showa)
Dining hall: Big Sand 1F
Past Forecasted
Congestion
Figure 3: Web service for visualizing congestion.
about the congestion. The details of the system are
described in following subsection.
3.2 Probe Request Capturing and
Filtering
We used Wi-Fi packet sensors located in various loca-
tions to capture probe request frames. Wi-Fi packets
can be received when the receiver is several hundred
meters away from the sender. In this study, we esti-
mated the congestion at dining halls and bus stops in a
university campus. We filtered out packets with weak
signal strengths (under -80dB) so that we only col-
lected data from close devices. The received time t of
the packet, place ID (sensor ID) p, and MAC address
m are stored to the database D
3
.
D D {(t, p, m)} (1)
3.3 Congestion Degree based on Probe
Requests
We define the congestion degree c(t, p) for each
time and place using probe request data without prior
3
Actually, we store hash values to the database instead
of MAC addresses because of privacy issues.
Please report about congestion.
Crowded
Medium
Vacant
Submit
Figure 4: User report form.
knowledge of the location as
c
0
(t, p) =
|{m |(t
0
, m, p
0
) D, t 180 t
0
t, p
0
= p}| (2)
c(t, p) =
c
0
(t, p)
α max
t
(c
0
(t, p))
, (3)
where c
0
(t, p) is the number of unique MAC address
observed during the three minutes. We obtain c(t, p)
by normalizing c
0
(t, p). The value α is determined
empirically (in this paper, α = 0.75).
3.4 Visualizing Congestion
We developed a Web service to visualize the extent of
the congestion. Figure 3 plots the congestion degree
calculated using Eq. 3. The red line represents the
congestion during the last 30 minutes and the green
line represents the forecasted congestion. Readers can
browse other locations using the upper tabs.
3.5 User Report
We collected user reports as perceptual congestion
measurements. Figure 4 shows the form for report-
ing the congestion degree on our website. There are
five radio (option) buttons. Users can only select one
radio button. After a user pushes the submit button
in the form, the selected item, time, and location are
submitted to the server.
Analysis of Wi-Fi-based and Perceptual Congestion
227
Table 1: Location of Wi-Fi packet sensors.
ID Location Floor Type Purpose
1 Dining hall A GF Indoor Breakfast, lunch and dinner
2 Dining hall B GF Indoor Breakfast, lunch and dinner
3 Learning space 3F Indoor Learning
4 Dining hall C B1 Indoor Lunch
17 Bus stop A N/A Outdoor Returning home
18 Bus stop B N/A Outdoor Returning home
2. Dining hall B (GF)
1. Dining hall A (GF)
4. Dining hall C (B1)
3. Learning space (3F)
18. Bus stop B
17. Bus stop A
50m
Figure 6: Wi-Fi packet sensors on a campus.
4 ANALYSIS OF CONGESTION
AND USER REPORTS
In this section, we analyze the congestion and user
reports obtained using our system.
4.1 Operation of Our System
We installed six Wi-Fi packet sensors on an university
campus (Figures 5 and 6). Table 1 shows the installed
spot, whether or not the spot is located indoors. We
have been operating these packet sensors since Jan-
uary 2016.
Figure 5: Wi-Fi packet sensor.
We have been operating our Web service for vi-
sualizing congestion and recording user reports since
July 2016. We received over three hundred user re-
ports about congestion via our Web service during the
first four weeks.
4.2 Time Series of Congestion
Figure 7 shows the congestion calculated using Eq. 3
for a typical week. We can see that the dining halls
(IDs 1, 2, and 4) have a steep peak around 12:00 be-
cause of lunchtime. The congestion at the bus stops
(IDs 17 and 18) tends to fluctuate intensely. This is
because busses arrive every 5 to 15 minutes to col-
lect passengers. Both bus stops are mainly used to go
home, so peak congestion occurs around the evening.
4.3 Correlation Analysis of User
Reports
We analyzed the correlation between user reports and
the congestion recorded using Wi-Fi packet sensors.
Figure 8 shows the scatter diagrams and correlation
coefficients for the user reports and Wi-Fi-based con-
gestion for each location. The correlation coefficients
for locations 1, 3, and 4 are over 0.6, so we can
say that the quantitative and perceptual congestion of
those spots have moderate correlations. Meanwhile,
the correlation coefficient for location 2 is less than
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
228
00:00 06:00
12:00
Jul. 25 (Mon)
18:00 00:00 06:00
12:00
Jul. 26 (Tue)
18:00 00:00 06:00
12:00
Jul. 27 (Wed)
18:00 00:00 06:00
12:00
Jul. 28 (Thu)
18:00 00:00 06:00
12:00
Jul. 29 (Fri)
18:00 00:00 06:00
12:00
Jul. 30 (Sat)
18:00 00:00 06:00
12:00
Jul. 31 (Sun)
18:00
0
25
50
75
100
Congestion degree
1
(a) Location 1
00:00 06:00
12:00
Jul. 25 (Mon)
18:00 00:00 06:00
12:00
Jul. 26 (Tue)
18:00 00:00 06:00
12:00
Jul. 27 (Wed)
18:00 00:00 06:00
12:00
Jul. 28 (Thu)
18:00 00:00 06:00
12:00
Jul. 29 (Fri)
18:00 00:00 06:00
12:00
Jul. 30 (Sat)
18:00 00:00 06:00
12:00
Jul. 31 (Sun)
18:00
0
25
50
75
100
Congestion degree
2
(b) Location 2
00:00 06:00
12:00
Jul. 25 (Mon)
18:00 00:00 06:00
12:00
Jul. 26 (Tue)
18:00 00:00 06:00
12:00
Jul. 27 (Wed)
18:00 00:00 06:00
12:00
Jul. 28 (Thu)
18:00 00:00 06:00
12:00
Jul. 29 (Fri)
18:00 00:00 06:00
12:00
Jul. 30 (Sat)
18:00 00:00 06:00
12:00
Jul. 31 (Sun)
18:00
0
25
50
75
100
Congestion degree
3
(c) Location 3
00:00 06:00
12:00
Jul. 25 (Mon)
18:00 00:00 06:00
12:00
Jul. 26 (Tue)
18:00 00:00 06:00
12:00
Jul. 27 (Wed)
18:00 00:00 06:00
12:00
Jul. 28 (Thu)
18:00 00:00 06:00
12:00
Jul. 29 (Fri)
18:00 00:00 06:00
12:00
Jul. 30 (Sat)
18:00 00:00 06:00
12:00
Jul. 31 (Sun)
18:00
0
25
50
75
100
Congestion degree
4
(d) Location 4
00:00 06:00
12:00
Jul. 25 (Mon)
18:00 00:00 06:00
12:00
Jul. 26 (Tue)
18:00 00:00 06:00
12:00
Jul. 27 (Wed)
18:00 00:00 06:00
12:00
Jul. 28 (Thu)
18:00 00:00 06:00
12:00
Jul. 29 (Fri)
18:00 00:00 06:00
12:00
Jul. 30 (Sat)
18:00 00:00 06:00
12:00
Jul. 31 (Sun)
18:00
0
25
50
75
100
Congestion degree
17
(e) Location 17
00:00 06:00
12:00
Jul. 25 (Mon)
18:00 00:00 06:00
12:00
Jul. 26 (Tue)
18:00 00:00 06:00
12:00
Jul. 27 (Wed)
18:00 00:00 06:00
12:00
Jul. 28 (Thu)
18:00 00:00 06:00
12:00
Jul. 29 (Fri)
18:00 00:00 06:00
12:00
Jul. 30 (Sat)
18:00 00:00 06:00
12:00
Jul. 31 (Sun)
18:00
0
25
50
75
100
Congestion degree
18
(f) Location 18
Figure 7: Congestion for a typical week.
Analysis of Wi-Fi-based and Perceptual Congestion
229
0
Vacant
25 50
Medium
75 100
Crowded
User reports
0
25
50
75
100
Congestion
Correlation coefficient: 0.64
(a) Location 1
0
Vacant
25 50
Medium
75 100
Crowded
User reports
0
25
50
75
100
Congestion
Correlation coefficient: 0.66
(b) Location 2
0
Vacant
25 50
Medium
75 100
Crowded
User reports
0
25
50
75
100
Congestion
Correlation coefficient: 0.68
(c) Location 3
0
Vacant
25 50
Medium
75 100
Crowded
User reports
0
25
50
75
100
Congestion
Correlation coefficient: 0.58
(d) Location 4
0
Vacant
25 50
Medium
75 100
Crowded
User reports
0
25
50
75
100
Congestion
Correlation coefficient: 0.49
(e) Location 17
0
Vacant
25 50
Medium
75 100
Crowded
User reports
0
25
50
75
100
Congestion
Correlation coefficient: 0.48
(f) Location 18
Figure 8: Correlations between the congestion and user reports.
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
230
Table 2: Time table of bus stop (Location 17).
8:11 8:25 8:41 9:11 9:41 10:21 10:41 11:16 11:41 12:11 12:46 13:01
13:22 13:47 14:17 14:37 14:47 14:57 15:12 15:42 16:12 16:27 16:32 16:46
16:50 16:57 17:17 17:42 18:17 18:22 18:42 18:57 19:27 19:42 20:02 20:31
21:01 21:31 22:01
Table 3: Time table of bus stop (Location 18).
6:57 6:59 7:11 7:21 7:37 7:39 7:46 7:56 8:13 8:21 8:34 8:39
8:44 8:49 8:54 8:59 9:04 9:14 9:30 9:36 9:46 9:57 10:12 10:17
10:26 10:36 10:44 10:57 11:01 11:12 11:27 11:41 11:51 11:56 12:04 12:11
12:16 12:21 12:26 12:41 12:57 13:06 13:12 13:26 13:36 13:46 13:57 14:11
14:21 14:26 14:37 14:41 14:44 14:47 14:51 14:54 14:58 15:01 15:04 15:11
15:16 15:21 15:26 15:37 15:41 15:46 15:56 16:01 16:04 16:09 16:14 16:19
16:24 16:27 16:31 16:34 16:38 16:43 16:46 16:51 16:56 17:04 17:06 17:12
17:14 17:19 17:22 17:24 17:26 17:29 17:35 17:39 17:42 17:44 17:48 17:52
17:55 18:01 18:04 18:06 18:11 18:17 18:21 18:26 18:31 18:37 18:41 18:44
18:53 18:56 18:59 19:06 19:11 19:14 19:17 19:26 19:31 19:36 19:46 19:51
19:54 19:57 20:06 20:11 20:14 20:26 20:29 20:41 20:49 21:01 21:06 21:09
21:14 21:24 21:26 21:41 21:59 22:01 22:06 22:21 22:48 22:53 22:59 23:13
00:00 01:00 02:00 03:00 04:00 05:00 06:00 07:00 08:00 09:00 10:00 11:00
12:00
Jul. 27 (Wed)
13:00 14:00 15:00 16:00 17:00 18:00 19:00 20:00 21:00 22:00 23:00
0
25
50
75
100
Congestion degree
17
(a) Location 17
00:00 01:00 02:00 03:00 04:00 05:00 06:00 07:00 08:00 09:00 10:00 11:00
12:00
Jul. 27 (Wed)
13:00 14:00 15:00 16:00 17:00 18:00 19:00 20:00 21:00 22:00 23:00
0
25
50
75
100
Congestion degree
18
(b) Location 18
Figure 9: Congestion of bus stops for a typical day.
0.5 even though it is from the same category (dining
hall) as locations 1 and 4. After analyzing user re-
ports in more detail, we found this low correlation was
caused by a lot of submissions of ’Crowded’ in a short
time. During this period, the Wi-Fi packet sensors did
not estimate that the location was crowded. There-
fore, we believe that these submissions were mali-
cious. We will deal with such malicious submissions
in the future.
The correlation coefficients of two bus stops (lo-
cations 17 and 18) are not large (around 0.5). This is
because of the intense fluctuations in the congestion
calculated using Wi-Fi sensors. Figure 9 shows the
congestion of two bus stops. Tables 2 and 3 show the
timetables of two bus stops. Because busses run fre-
quently, congestion curve fluctuates intensely. Con-
sequently, the correlation coefficients of the bus stops
are not high, and forecasting congestion is not easy.
5 CONCLUSION AND FUTURE
WORK
In this paper, we described a system for estimating
and visualizing congestion using Wi-Fi packet sen-
sors. We also collected user reports about conges-
tion via our system. We analyzed the relationship
between quantitative congestion measurements using
Wi-Fi packet sensors and perceptual congestion mea-
Analysis of Wi-Fi-based and Perceptual Congestion
231
surements based on user reports. Based on our anal-
ysis, we found correlations between the quantitative
and perceptual congestion measurements for each lo-
cation.
We plan to install Wi-Fi packet sensors at more lo-
cations (e.g., lecture rooms, laboratories, and confer-
ence rooms) and then analyze the congestion in more
detail. Based on the relationship between quantitative
and perceptual congestion, we will improve the accu-
racy of congestion estimates and provide congestion
information via our system.
REFERENCES
Choi, J., Hwang, H., and Hong, W. (2011). Predicting the
Probability of Evacuation Congestion Occurrence Re-
lating to Elapsed Time and Vertical Section in a High-
rise Building, pages 37–46. Springer US, Boston,
MA.
Fukuzaki, Y., Mochizuki, M., Murao, K., and Nishio, N.
(2014). A pedestrian flow analysis system using wi-fi
packet sensors to a real environment. In Proceedings
of the 2014 ACM International Joint Conference on
Pervasive and Ubiquitous Computing: Adjunct Publi-
cation, UbiComp ’14 Adjunct, pages 721–730, New
York, NY, USA. ACM.
Schauer, L., Werner, M., and Marcus, P. (2014). Estimating
crowd densities and pedestrian flows using wi-fi and
bluetooth. In Proceedings of the 11th International
Conference on Mobile and Ubiquitous Systems: Com-
puting, Networking and Services, MOBIQUITOUS
’14, pages 171–177, ICST, Brussels, Belgium, Bel-
gium. ICST (Institute for Computer Sciences, Social-
Informatics and Telecommunications Engineering).
Weppner, J. and Lukowicz, P. (2013). Bluetooth based
collaborative crowd density estimation with mobile
phones. In Pervasive Computing and Communica-
tions (PerCom), 2013 IEEE International Conference
on, pages 193–200.
Xi, W., Zhao, J., Li, X. Y., Zhao, K., Tang, S., Liu, X., and
Jiang, Z. (2014). Electronic frog eye: Counting crowd
using wifi. In IEEE INFOCOM 2014 - IEEE Confer-
ence on Computer Communications, pages 361–369.
Yaik, O. B., Wai, K. Z., Tan, I. K., and Sheng, O. B. (2016).
Measuring the accuracy of crowd counting using wi-
fi probe-request-frame counting technique. Journal of
Telecommunication, Electronic and Computer Engi-
neering (JTEC), 8(2):79–81.
ICPRAM 2017 - 6th International Conference on Pattern Recognition Applications and Methods
232