Cooperative Evacuation Guidance Methods in Large-Scale

Disaster Situations Based on Wi-Fi Sensing Data

Atsuo Ozaki

Department of Information Science and Technology, Osaka Institute of Technology, Osaka, Japan

Keywords: Multi-Agent, Disaster, Cooperation, Evacuation, Guidance, Wi-Fi, Sensing Data.

Abstract: This study introduces a Wi-Fi packet sensor developed to acquire headcount distribution data, which are

obtained by deploying several of these sensors in a large-scale event, as well as describes the results of

evaluating the proposed distributed and coordinated evacuation guidance method in a disaster using multi-

agent simulation. The results confirm that by balancing the guide loads, it is possible to evacuate all evacuees

from the venue in a shorter time than before the addition of the load balance. Furthermore, it was confirmed

that if the evacuation route indicated by the guide did not significantly change the congestion situation, it was

important for evacuees to choose another exit route at their discretion.

1 INTRODUCTION

In recent years, crisis management measures have

become increasingly necessary because of the

intensification of natural disasters and the spread of

epidemics. Additionally, it is necessary to actively

promote business improvement and development to

overcome the economic slump caused by global

unrest and other factors.

We aim to address these problems by identifying

the number of visitors and human flow in real time in

large commercial facilities and large-scale event

venues where a large number of people gather to

efficiently guide evacuations in the event of a

disaster, prevent infectious diseases by avoiding

congestion, and promote economic activity to

optimize store layouts in regular times (Yamada et al.,

2022; Asano et al., 2024; Watanabe et al., 2024).

In this paper, a Wi-Fi packet sensor developed for

low-cost, real-time detection of the number of people

is described and a distributed cooperative evacuation

guidance method is proposed as its application for a

potential disaster in a large-scale event held monthly

in Hirakata City, Osaka Prefecture, Japan. The

proposed method was implemented using multi-agent

simulation technology, and its evaluation the results

are discussed.

https://www.oit.ac.jp/english/research/ist/computer/

index.html#lab3202

Figure 1: Developed Wi-Fi packet sensor.

2 WI-FI SENSING

TECHNOLOGY FOR

DETECTING THE NUMBER OF

PEOPLE

2.1 Issues and Initiatives

The use of cameras can be a highly accurate method

for detecting the number of people in a place;

however, it is often avoided because of limitations

regarding camera location and blind spots, as well as

the fact that the cameras capture the visitor's face.

Another method is to use the location information

service (e.g., GPS) of a smartphone held by a visitor

to analyse human dynamics; however, this method

can only be used when a specific application is

running, and its accuracy is reduced indoors and

underground.

Ozaki, A.

Cooperative Evacuation Guidance Methods in Large-Scale Disaster Situations Based on Wi-Fi Sensing Data.

DOI: 10.5220/0013110600003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 1, pages 273-278

ISBN: 978-989-758-737-5; ISSN: 2184-433X

273

One method that can address these problems is to

estimate the number of people by detecting Wi-Fi

radio waves that penetrate materials such as wood and

glass and can detect people behind obstacles, thus

eliminating the aforementioned problems of blind

spots and personal information disclosure. Therefore,

it can be used indoors and underground. This Wi-Fi

sensing method uses a Wi-Fi packet sensor

(hereinafter referred to as "PS") to collect radio waves

periodically transmitted from Wi-Fi terminals, such

as smartphones, and then extracts and analyses data

such as the terminal's MAC address, detection time,

and signal strength to estimate the number of people.

According to a survey by Japan's Ministry of Internal

Affairs and Communications, the smartphone

ownership rate as of 2022 was 77.3% (MIC, 2023).

However, excluding younger age groups, almost

100% of people own smartphones, and most are

considered to have their smartphones and Wi-Fi

functions turned on. Therefore, we promote this

research and development to consider effectively

estimating the number of people using this method

(Toyomi et al., 2021; Toyomi et al., 2022).

2.2 Developed Wi-Fi Packet Sensor

Figure 1 shows the Wi-Fi PS. The detected Wi-Fi

signals are processed by a Raspberry Pi computer.

The Wi-Fi antenna mounted on the device can detect

radio waves within a radius of approximately 70 m

around the PS. The device is also equipped with a

battery that can run for approximately 12 h. A Long-

Term Evolution (LTE) antenna and Subscriber

Identity Module (SIM) card are also installed, and the

detected data can be transmitted to and stored in the

cloud so that the detection results can be checked in

real time from a remote site. The PS is lightweight

and compact, making it highly portable, and the

measurements can be started immediately after the

power is turned on. Therefore, the PS can be used for

last-minute events and for a variety of venues, both

large and small, as the number of PSs can be

increased according to the size of the venue.

At the main venue of "Hirakata-shuku Kurawanka

Goroku-ichi" (commonly called "Goroku-ichi")

, a

large-scale event held monthly in Hirakata City,

Osaka Prefecture, Japan, the number of visitors was

detected using the developed PS for 2 h, from 11:00

to 13:00, on November 8, 2020 (Sunday). The results

confirmed that the PS could detect the number of

visitors with reasonable accuracy, with a correlation

value of approximately 80% with the manually

https://www.gorokuichi.net/ (Japanese)

measured value (true value) (Toyomi et al., 2022).

The system is under continuous development and

improvement, and it currently exhibits a performance

of approximately 90% (Toyomi et al., 2024).

Figure 2: Goroku-ich image map (Japanese version).

3 APPLICATION TO

EVACUATION GUIDANCE IN

DISASTER SITUATIONS

3.1 Background

Goroku-ichi is a large-scale event held on the second

Sunday of every month in Hirakata City, which was

once an inn town located on a historical highway that

developed during the Edo period (Figure 2). As

Goroku-ichi attracts an average of 8,000 people and

more than 10,000 people at times, it is necessary to

take crisis management measures to deal with natural

disasters such as earthquakes, typhoons, and floods,

as well as the spread of diseases, such as the

coronavirus disease. In response to requests from the

Goroku-ichi office, we developed a PS to determine

the distribution of people in the Goroku-ichi venue,

which is important for crisis management measures,

and conducted a variety of field experiments.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

274

Because cars, motorcycles, bicycles, and other

vehicles travel in the Goroku-ichi venue, there are

four locations where traffic control is required, and

approximately ten security guards are deployed to

take charge of this task at each location. In this

section, we propose a decentralized cooperative

evacuation guidance method that assumes the case

where multiple security guards cooperate to guide

visitors to evacuate in the event of an earthquake in

the Goroku-ichi venue, and discuss the results of

implementing and evaluating the effectiveness of the

proposed method using multi-agent simulation

technology. Hereafter, the security guards are

referred to as "guides" who guide visitors during the

evacuation, and the Goroku-ichi visitors are referred

to as "evacuees."

3.2 Proposed Behavior Models for

Evacuation Guide

We examined two methods to efficiently guide a large

number of evacuees to evacuate the Goroku-ichi

venue. In both methods, it was assumed that each

guide had his own means of communication and

shared information with the others. It was also

assumed that the information on the number of

evacuees in the venue from multiple PSs was shared.

 Method 1: Dynamically determine the zone that

each guide is in charge of Goroku-ichi venues are

divided into several areas for management. In this

method, each guide travels to one of the nearest

areas from the current location immediately after a

disaster occurs. However, to avoid overlapping the

areas they are in charge of, they pass through the

areas where there are already guides for the

evacuees, go to the area where no one is in charge,

and start evacuation guidance for evacuees.

However, if the information from the PS indicates

that the evacuation route (exit) is more crowded

than other routes, the guide informs the evacuees of

the route to the next nearest exit. After the

evacuation guidance is completed in the assigned

area, if there is still an area that has not been

evacuated, the guide will go to that area to begin the

evacuation guidance.

 Method 2: Load equalization added to Method1.

This method divides the venue area equally into

numbers proportional to the number of guides and

assigns these divided areas to the guides. In this

case, the divided areas are assigned such that each

guide is in charge of an equal number of areas.

Immediately after a disaster occurs, the guide first

moves to the nearest area in the group of areas that

he is in charge of and begins to guide the evacuees.

The evacuee guiding method is the same as that

used in Method 1. After the guidance of the first

area is completed, the guide moves to the area

adjacent to the current area in his/her/their area

group that has the largest number of evacuees and

has not yet been evacuated, and guides the

evacuees. When all the guidance in the area group

for which he/she/they are in charge of is completed,

the guide moves to the area in charge of another

guide who has not yet completed the guidance, and

guides the evacuation.

3.3 Criteria for Judging Congestion for

Determining Evacuation Routes

The proposed methods assume that PSs are placed in

appropriate areas and that information on the number

of evacuees is shared by each guide to improve the

evacuation guidance efficiency. In the evaluation of

these methods in Goroku-ichi, we introduced α,

which calculates congestion of evacuees from a PS

placed at the intersection in a venue exit, where heavy

congestion is expected, and uses this as a decision-

making tool to select the appropriate evacuation route

(see Figure 3). α is expressed as follows:

α: = n/A (1)

where n is the number of evacuees present within the

PS measurement range and A is the road area within

the measurement range. Specifically, each guide must

sum all α on the evacuation route and communicate

the route with the smallest total value to the evacuees

as an optimal evacuation route.

Figure 3: Sensor placement near exits to calculate the

congestion degree α.

3.4 Evacuee Models

Evacuees remain in place until the guide tells them to

evacuate; when the guide provides them with an

evacuation route, they start evacuating and go to the

venue’s exit. If some evacuees started to move in the

vicinity, they would look at the surrounding situation,

start moving in the direction of most evacuees in

progress, and follow other evacuees, even if the route

was not communicated to them by the guide.

Cooperative Evacuation Guidance Methods in Large-Scale Disaster Situations Based on Wi-Fi Sensing Data

275

However, if the guide communicates a different route

in this state, the evacuees follow that route. During

the process of moving along the route, the evacuator

avoids a place by moving to the side if there is a

person in the place where he/she/them wants to go. If

there are no places to avoid, the evacuee waits for a

predetermined period. If the waiting state continues

due to congestion, the evacuation route and exits are

changed based on the route selection decision

criterion β, as follows:

β: = Tp-Tw-Td,

if β > 0, then: Maintain the current situation;

else: Select another route and exi

(2)

where Tp is the allowable time to maintain the current

situation, Tw is the time spent waiting due to the

inability to move from the current position, and Td is

the time to move from the current position to the

current target exit when uncrowded. In other words,

only Tp can be changed at the will of the evacuee,

indicating that the farther away from the exit, the

more likely the evacuee is to change his/her/their

intended route to the exit.

We also set the probability γ of selecting another

exit when β < 0. γ indicates the probability (%) of

selecting another nearest exit or any other exit, both

excluding the current intended exit. For example, if γ

= 0%, the exit is toward another nearest exit;

conversely, if γ = 100%, the exit is toward the other

exit.

4 EVALUATION

To evaluate the proposed method, we developed a

simulator for evacuation guidance after an earthquake

in Goroku-ichi and estimated the time required for all

evacuees to complete the evacuation from Goroku-

ichi venues. As shown in Figure 4, Goroku-ichi is

held within 1.1 km from Hirakata City Station to

Hirakata Park Station, and is a large-scale event with

approximately 200 stores located along a road parallel

to the Yodo River on the north side. Goroku-ichi

consists of ten management areas (Figure 4: a–j).

Additionally, points A–D in the figure are locations

where traffic control is required, such as when cars

cross venue and where security guards are placed at

all times.

4.1 Conditions and Settings

For the evaluation, the number of evacuees and their

assignments were based on the total number of

visitors (860) measured using the PSs in each of the

aforementioned areas on October 13, 2019 (Table 1).

Regarding the number of guides, as the labour cost of

security guards is an operational issue, two cases were

evaluated: one was the same as the current situation,

with a total of ten guides (A: 2, B: 3, C: 2, and D: 3),

and the other was a total of four guides, one assigned

to each of A–D.

Table 1: Initial number of evacuees in each area.

cdef g h i j

45 46 119 145 73 95 64 116 114 43

(Total:860)

Therefore, in Method 1, because the number of

areas is ten, if the number of guides is set to four, the

assignment of guides to each area is dynamic;

however, if the number of guides is ten, there is no

dynamic assignment because the number of guides is

equal to the number of areas. Meanwhile, for Method

2, the area of the Goroku-shi venue is divided equally

into 20 areas, and each guide is assigned to the same

number of areas. In other words, if there were four

guides, each was in charge of five areas, and if there

were ten guides, each was in charge of two areas.

Additionally, it was assumed that the guide would

spend a certain amount of time in each area guiding

the evacuees and that they would remain in an area

for a certain amount of time, which was set at 30 s for

this evaluation based on empirical evidence. The

communication of the evacuation route by the guide

to the evacuees was to be done by voice, and it was

assumed that it could be communicated to evacuees

within an 18-m radius around the guide's position,

this was also based on empirical evidence. Regarding

the evacuees' movement after obtaining information

on the evacuation route from other evacuees around

them, because this movement is propagated, it is

assumed that the information can be obtained from

people within 1 m of them. If no one was in front

Figure 4: Goroku-ich's ten management areas (a–j), guard

locations (A–D), and exits.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

276

of the evacuee, the moving speed of the guide and

evacuee was set to 1 m/s, following (Liberto et al.,

2020), considering the congestion situation. There

were six exits from the venue to the south that the

evacuees targeted, and these six exits were also set as

destinations for the evacuees in this evaluation (see

Figure 4). Furthermore, as the Goroku-shi venue is a

single road, when selecting an exit other than the

currently intended and next nearest exits, the opposite

of the next nearest exit is selected based on γ.

4.2 Results

Table 2 lists the evacuation time simulation results

(average of 100 trials) until all evacuees leave the

venue, assuming that evacuees do not change the

route (exit) communicated to them by the guides.

These results are based on a model that reflects the

mentality of evacuees who trust the routes

communicated to them through guides who know the

situation.

Table 2: Evacuation time (s) according to each method

when the number of guides is 10 and 4.

＃Guides＝10 ＃Guides＝4

Method 1 716 908

Method 2 639 692

Figure 5: Evacuation time based on γ according to Tp.

Figure 5 also shows the results of changing γ in

20% increments applied to changing to other routes

(exits) and then changing Tp from 0 s to1000 s for

Method 2, which could evacuate quickly in the case

of ten realistic guides. In each case, the average value

was obtained from 50 trials. The results show that

when γ = 0% and Tp = 100 s, the minimum

evacuation time is 441 s; when γ = 100% and Tp = 0

s, the maximum evacuation time is 838 s. When Tp

was greater than 400 s, the difference in evacuation

time was almost the same, approximately 20 s, for

different γ values.

5 DISCUSSION

As can be seen in Table 2, Method 2 with four guides

is able to guide the evacuation in a shorter time than

Method 1 with ten guides. This is believed to be due

to the fact that in Method 2, after completing the

evacuation of their own area group, each guide went

to help other guides in unguided areas, thus balancing

the workload among the guides relatively well.

Conversely, in Method 1, there was a large imbalance

in the load between guides who were in charge of

high-load areas and those who were not.

Figure 6: Minimum evacuation time based on γ and Tp.

In Figure 5, for any γ, there is a minimum

evacuation time when Tp is less than 300 s. Figure 6

shows the minimum evacuation time and the γ and Tp

combination at that time for each of the cases shown

in Figure 5. The case with no route (exit) change

(Table 1: Method 2 with ten inductors) is also

included in this figure for reference. It was confirmed

that by changing other routes (exits) according to the

situation, the evacuation time could be reduced by

approximately 31% (Figure 6). Based on this result,

it can be concluded that the most efficient evacuation

method is to leave immediately and go to the next-

nearest exit if you are stopped by an obstacle during

evacuation and the congestion is not resolved after

waiting for some time. Although the route

information provided by the guide is important, the

situation changes from moment to moment.

Therefore, it is important for evacuees to make their

own decisions.

6 CONCLUSIONS

This paper describes a Wi-Fi packet sensor developed

to obtain data on the distribution of the number of

people, which is important for crisis management

measures against natural disasters and epidemics as

400

450

500

550

600

650

700

750

800

850

900

0 100 200 300 400 500 600 700 800 900 1000

Evacuation Time (s)

Tp (s)

γ = 0

γ = 20

γ = 40

γ = 60

γ = 80

γ = 100

Cooperative Evacuation Guidance Methods in Large-Scale Disaster Situations Based on Wi-Fi Sensing Data

277

well as for improving operations and developing

businesses. Using the data measured by this sensor on

the distribution of the number of people in a large-

scale event, the proposed evacuation guidance

method was evaluated using a multi-agent simulation.

As a result, it was confirmed that all evacuees could

be evacuated from the venue in less than one minute

by adding a way to balance the load assigned to each

guide. Furthermore, it was confirmed that even if the

evacuation route (exit) indicated by the guides did not

change the congestion situation, it was important for

evacuees to choose another route (exit) at their own

discretion. The simulation results showed that it is

important to take action based on one's own judgment

while referring to information rather than relying on

others for one's own life. In this study, the evaluation

was based on the distribution of the number of people

during daytime hours from 11:00 to 13:00. It is also

necessary to evaluate the situation during opening

and closing of the event (e.g. when visitors come

from Hirakata City and Hirakata Park Stations and

when they leave), when the distribution of people is

uneven around the two ends of the venue, which will

be investigated in future work. Additionally, methods

to further shorten the evacuation time must also be

investigated.

REFERENCES

Liberto, C., Nigro, M., Carrese, S., Mannini, L., Valenti, G.,

Zarelli, C. (2020). Simulation framework for pedestrian

dynamics: modelling and calibration. IET Intelligent

Transport Systems, Vol. 14 Iss. pages 1048-1057.

Yamada, C., Ozaki, A. (2022). Cooperative Evacuation

Guidance Method for Large-Scale Event. 37th

International Technical Conference on

Circuits/Systems, Computers, and Communications

(ITC-CSCC 2022), pages 922-925.

Watanabe, K., Toyomi, R., Ozaki, A. (2024). Study of

Distributed Cooperative Evacuation Guidance Model

Using Multiple Mobile Robots. 39th International

Technical Conference on Circuits/Systems, Computers,

and Communications (ITC-CSCC 2024), pages 1-6.

Asano, M., Toyomi, R., Ozaki, A. (2024). Distributed

Cooperative Evacuation Guidance Model Based on

Self-Driven Autonomous Agents. 29th International

Symposium on Artificial Life and Robotics (AROB-

ISBC-SWARM 2024), pages 24-28.

Ministry of Internal Affairs and Communications (MIC).

(2023). Telecommunications Usage Trends Survey. 2.

Mobile handset ownership. (Japanese)

https://www.soumu.go.jp/johotsusintokei/whitepaper/ja/r0

5/html/datashu.html#f00279

Toyomi, R., Fujino, Y., Ozaki, A. (2021). A Study on

Improved Method for Obtaining the Distribution of

People Using Wi-Fi Probe Request. The 36th

International Technical Conference on

Circuits/Systems, Computers and Communications

(ITC-CSCC 2021), pages 208-211.

Toyomi, R., Fujino, Y., Ozaki, A. (2022). Wi-Fi packet

sensing method for obtaining the number of people. The

Twenty-Seventh International Symposium on Artificial

Life and Robotics 2022 (AROB 27th 2022), pages 857-

860.

Toyomi, R., Ozaki, A. (2024). Development and evaluation

of a portable crowd estimation system using Wi-Fi.

International Journal of Artificial Life and Robotics,

Springer. https://doi.org/10.1007/s10015-024-00977-0

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

278