=Paper=
{{Paper
|id=Vol-3297/paper3
|storemode=property
|title=Real-Time Safe Route Finding and Visualization Using Shared Edge Devices
|pdfUrl=https://ceur-ws.org/Vol-3297/paper3.pdf
|volume=Vol-3297
|authors=Takenori Hara,Hideo Saito
|dblpUrl=https://dblp.org/rec/conf/apmar/HaraS22
}}
==Real-Time Safe Route Finding and Visualization Using Shared Edge Devices==
https://ceur-ws.org/Vol-3297/paper3.pdf
Real-Time Safe Route Finding and Visualization Using Shared
Edge Devices
Takenori Hara1* , Hideo Saito1
1
Keio University, 3-14-1 Hiyoshi, Kohoku-ku, Yokohama, Kanagawa, 223-8522, Japan
Abstract
Conventional safe route-finding systems display routes on a map to avoid dangerous areas
based on offline data such as statistical data on crime rates. However, such dangerous level
changes in a dynamic manner, which can be detected by people walking around. This paper
presents a system that shares the detected data measured by edge devices owned by walking
people for dynamic safe route finding in real-time. This system evaluates the safety of routes
in real time using and displays the route in AR on HMDs and smartphones. This paper reports
findings from a simulation and edge device (smartphone) implementation of spatial safety
assessment, safe route finding, and route visualization.
Keywords 1
Safe Routing, Edge Device, Visualization, AR
1. Introduction outdoors and 30% of cases occur in a house
(Figure 2). A higher percentage of assaults occur
1.1. Research Background on public transportation or in stores than assaults,
and we often hear news stories of station staff,
Violent crimes such as assault and injury shopkeepers, passersby, and passengers suddenly
(Assault is defined in Japan as a crime in which being violently attacked. Although transportation
the victim was not injured, while an injury is a and shopping are essential to daily life, the
crime in which the victim was injured) are crimes possibility of sudden violence and crime prevents
that occur close to people, have a significant people from leading their daily lives with peace of
impact on the victim's mind and body, threaten mind.
their peaceful daily life, and may increase social
unrest. According to a study on violent crime
conducted by the Ministry of Justice's Legal
Research Institute in Japan[1], the number of
assault cases has remained high at around 30,000
since 2006, and the number of injury cases has
also remained in the 20,000s since 2008 (Figure
1). In addition, these are the number of crimes
reported to the police; the actual number of
assaults and injuries is expected to be much higher.
The largest number of assault cases was from
people who did not know each other, while the Figure 1: Number of assaults and injuries in
number of injury cases was from people who Japan
knew each other, as typified by domestic violence.
About 30% of both assault and injury cases occur
APMAR’22: The 14th Asia-PacificWorkshop on Mixed and
Augmented Reality, Dec. 02-03, 2022, Yokohama, Japan
*Corresponding author.
EMAIL: gouki@hvrl.ics.keio.ac.jp (T.Hara); hs@keio.jp (H.
Saito);
ORCID: 0000-0002-7699-0376 (T.Hara); 0000-0002-2421-9862
(H.Saito);
© 2022 Copyright for this paper by its authors. Use permitted under Creative
Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Wor
Pr
ks
hop
oceedi
ngs
ht
I
tp:
//
ceur
-
SSN1613-
ws
.or
0073
g
CEUR Workshop Proceedings (CEUR-WS.org)
Sukru et al. propose a mobile app for the social
distancing of people in a COVID-19 environment.
It allows each user to manually register their
health status and then share their current location
on the network in real-time, predicting the
Figure 2: Location of injury and assault in Japan walking paths of other users and alerting them
with sound and vibration to the possibility of
approaching users with poor health status [4].
Maria et al. proposed a method that uses BLE
1.2. Research Objectives beacons to detect and share a user's location and
provide a route to reach a destination without
Conventional navigation systems are designed passing other people indoors [5]. To successfully
to arrive at a destination smoothly and quickly. To avoid people, robot pathfinding methods can be
achieve a safe and secure society, we research, helpful: Kapil et al. proposed a pathfinding
develop, and implement navigation systems that method that predicts the movement path of a
can avoid crimes and unpleasant events by group of people [6], and Lucia et al. proposed a
detecting them in advance. The navigation system method to move around humans and obstacles in
uses edge devices such as surveillance cameras a social compatible and safe manner [7].
and smartphones to measure the level of danger in As a mechanism for sharing information
real-time and search for safe routes. The system sensed by the user, the Joint Tactical Information
also visualizes the detected dangerous elements Distribution System (JTIDS) and the Inter-
and the safe routes. vehicular Information System (IVIS), which are
We are convinced that this system can used in military applications, can be used as
contribute to solving various social problems. For references. These systems share the position,
example, the system can be used for urban heading, altitude, and speed of detected aircraft,
planning to prevent crimes, support services to ships, and vehicles with all platforms participating
maintain social distance in infectious disease in the network.
environments such as COVID-19, detect people in For safe route finding, it is necessary to
trouble and automatically ask for help from evaluate the safety of a location. According to
people around them, support services to protect Stacy et al., a person's gender, facial expression,
wild animals, and services to find lost cats. gaze, appearance, clothing, attitude, and physique
In this paper, we report on the findings and are known to help determine whether a person is
problems encountered in simulating the level of suspicious [8]. For example, if a person is silent
danger assessment, safe route search, and safe and stares at another person, has an aggressive or
route visualization of a location using edge predatory facial expression or body language, or
devices. We also report our findings and problems is carrying a weapon such as a gun or a knife, he
in a prototype system using a smartphone. or she is highly dangerous. Fan [9] and Fang [10]
et al. proposed a method for detecting abnormal
1.3. Related Works human behavior in real-time from video images,
and it is also provided as a service, such as the
Many methods have been proposed for safe Incident Detection Solution [11]. Salamon et al.
route finding that minimizes the level of danger proposed a method for sound classification [12],
and distance on the path and displays the path on we think this method evaluates the safety of a
a map. location by analyzing sounds such as
Galbrun et al. proposed a safe route search conversations or detecting gunshots or explosions.
method based on Total-Paths (total level of danger Kim et al. proposed a method to estimate the level
on a path) and Max-Paths (maximum level of of danger of a location based on the content of
danger on a path) [2]. Yizhou et al. conducted a geotagged messages on social media [13].
safe route search simulation experiment on an RedZone Map [14], a service that guides users to
actual congestion dataset for each Paris train safe routes for cars based on government-released
station based on another research result that crime map data; My Safety pin[15], a navigation
"higher congestion leads to higher crime rate" and application that uses the results of interviews with
confirmed the effectiveness of the method [3]. users about the safety of each location;
CrimeReports[16], which works with police
information systems to visualize crimes that have
occurred in; My Safe Map[17], a map service that 2. Study of a navigation system that
uses existing external data to search for safe
streets and blocks, have also been put to practical can detect and avoid criminal
use. Many methods have been proposed to damage and unpleasant events in
evaluate the safety of a location, such as those in advance
these previous studies. However, no method has
been established that runs on an edge device and
evaluates the safety of a location with sufficient
accuracy.
The freshness, spatial resolution, and accuracy 2.1. System Overview
of the original data are important for assessing the
level of danger. Ideally, the data should be We have conducted a system study to achieve
updated in real-time and evaluated by an impartial a navigation system that can detect and avoid
third party with a spatial resolution of about 3.5 m criminal damage and unpleasant events in
[18], which is defined by Edward et al. as a social advance. This system consists of six subsystems
distance that is hard to reach but easy to talk to. as shown in Figure 3.
However, most of the aforementioned studies and
services use statistical data such as crime maps,
which are not real-time, have a low spatial
resolution, and have low accuracy because they
Safty
are self-reported. And most of the existing
services that detect abnormal behavior in real-
time from video images are not available because (1) Edge Devices (2) Location danger level (3) Wide-area safe route
they do not disclose the detection results. evaluation system search system
Therefore, conventional navigation services are
limited to route guidance to avoid areas that are
statistically known to be dangerous. No safe
route-finding method has been proposed to
successfully pass by dangerous people in a narrow (4) Narrow-area safe route (5) Safe route (6) Data sharing server
space while maintaining social distance from search system visualization system
them. As for visualization, the system is limited to Figure 3: System Overview
displaying routes on a 2D map and alerting drivers
with sound and vibration. (1)Edge devices
Therefore, we are conducting research, A device equipped with various sensors such
development, and social implementation of an as a camera, microphone, thermo, GPS, etc. We
information-sharing system and navigation assume surveillance cameras, smartphones, and
system that can detect and avoid criminal damage wearable devices.
and unpleasant events in advance, featuring the
following three elements. (2) Location danger level evaluation system
The system evaluates and shares the safety of
(1) Edge devices accurately measure and share the a location based on information acquired by
level of danger of a location in real-time. various sensors in edge devices. We can use other
information such as sound (conversations,
(2) A wide-area safe route-finding method based gunshots, explosions) or temperature (human
on the level of danger measured by edge devices, body, road surface, fire, freezing) to evaluate the
and a narrow-area safe route-finding method for safety of the location. This information can be
successfully passing dangerous people in a used to set various conditions in addition to safety,
narrow space while maintaining social distance. such as easy walking routes, cool routes, and
routes with many pedestrians.
(3) Social Distance and safe route visualization
and human guidance methods (3) Wide-area safe route search system
This system performs a wide-area route search
to maximize safety and minimize distance based
on the safety level and distance of the location
evaluated by the edge device.
The location and attributes of PFs, PDs, and PSs
(4) Narrow-area safe route search system detected by the device and their measured level of
This system searches for a safe route in a danger are sent to the data-sharing server as a
narrow space (e.g., a street or a station) to pass by Danger Point (DP). Although the DP is quantified
a dangerous person while maintaining a social by the location safety evaluation system by
distance from him or her. analyzing the facial expressions and attitudes of
each person, we set the DP uniformly for each
(5) Safe route visualization system attribute as shown in Table 1 for simplicity in this
This system visualizes shared information like simulation.
dangerous persons detected by other edge devices.
This system also visualizes the routes and social Table 1
distance calculated by the safe route search Attributes and DP of people
system. Wide-area safe routes are displayed on a Attributes Color DP Number
map, and narrow-area safe routes for passing each
PE
other are displayed on the HMD. person with an - 100
edge device
(6) Data-sharing server
A server that accumulates real-time spatial PF
person with a 1 10
safety information. It stores information uploaded fever
from each edge terminal and distributes it to each
terminal in real-time. PD
10 10
dangerous person
2.2. Simulator
PS
police officer or -10 10
We will evaluate the usefulness of the security guard
designed system. However, our system includes
subsystems that are difficult to implement at this Figure 4 shows the results of mapping the
time. In particular, as mentioned in section 1.3, current positions of PF, PD, and PS detected by
there is no established method to evaluate the 100 edge devices on a 2D map. We drew bubbles
safety of a location with sufficient accuracy on an (spheres) colored according to the attributes at the
edge device. In Addition, there are also no small, detected person's location. This bubble is drawn at
lightweight, safe, optical see-through HMDs that the location of the PF, PD, and PS at the time the
can be easily worn. Because of these difficulties PE detected them, but they could move. Also, the
in implementing our system in the real world, we map would be filled with bubbles of newly
first conducted our system evaluation in a detected PFs, PDs, and PSs. Therefore, we set the
simulator. transparency of the bubbles to indicate how new
We constructed a virtual Akihabara city [19] the information is, and set them to gradually
area of approximately 625 m x 625 m. All disappear in one minute so that the map would not
subsystems operate ideally in a virtual Akihabara. be filled with bubbles. Users can see that the
For example, virtual Akihabara has an ideal edge disappearing bubble is old information.
device that measures the safety of the location
with 100% accuracy. Then, We randomly placed
autonomously roaming people with various
attributes PE, PF, PD, and PS shown in Table 1 in
a virtual Akihabara. PE is a person with an edge
device (100 people), PF is a person with a fever
(10 people), PD is a person who can be judged
dangerous for some reason such as a facial
expression or attitude (10 people), and PS is a
person who can reduce the level of danger in that
location such as a police officer or security guard
(10 people). The PE has an ideal edge device that
detects PF, PD, and PS within a 30-meter radius
without error and measures the level of danger.
Figure 5: Wide-area safe route search
Figure 4: Mapping results of detected PF(Red),
PD(Yellow), and PS(Green)
When a PE with an edge device sets a
destination, a wide-area safe route system
performs the route search. We have extended the
Dijkstra method, which is used for route finding
in navigation systems, to perform wide-area safe
route finding considering distance and level of
danger in space. Dijkstra's method is an algorithm
in graph theory that can find the shortest distance
from a start node to a goal node and its path. We
generated a graph structure from map data in
which the intersections are represented as nodes I
and the roads connecting the intersections as
edges E. We defined the Edge Danger Level (EDL)
as the level of danger between intersection In and
the adjacent intersection In+1, where EDL is the sum
of the DP of PF, PD, and PS that exist between the Figure 6: Narrow-area safety route search
intersections. We set a danger threshold value
DLth, and if EDL exceeds the DLth, the path is Figure 5 shows an example of wide-area safe
considered impassable and the node connection is routing. When a PE with an edge device starts
disconnected. In this simulator, we set the from point S and goes to G, Figure 5(a) is usually
threshold DLth = 5. The system evaluates the EDL the shortest path. If one PF is detected at E(I1, I2),
of all edges, updates the graph structure, and finds one PF and one PD at E(I2, I3), and one PF, PD,
the shortest and safest path using the Dijkstra and PS at E(I3, I4), the level of danger EDL for each
method. edge is as follows
EDL(I1, I2)＝ 1 ：passable
EDL(I2, I3)＝ 11 ：impassable
EDL(I3, I4)＝ 1 ：passable
Considering the threshold DLth = 5, edge E(I2,
I3) is cut off because it is impassable, and the
shortest path search is performed on the remaining
edges, resulting in the path shown in Figure 5(b).
E(I2, I3) and E(I3, I4) are passable, but there are PF
and PD, so a narrow-area safe path search is
performed to find a path that allows them to pass or PD on the route, the system performs a narrow-
each other while maintaining the social distance. area safe route search and shows PE0 a route to
Figure 6 shows an example of a narrow-area safe pass them while maintaining social distance.
path search. The system generates a graph Figure 8 shows the view from PE0. In this
structure by dividing the space to be traversed into experiment, PE0 is assumed to be wearing a see-
meshes. The size of the mesh was set to 3.5 m, a through HMD. System display PF, PD, and PS
social distance defined by Edward et al[18]. as detected and shared by other edge devices. The
hard to reach but easily conversable. In the system also displays a narrow-area safe route
narrow-area safe path search, the Social Distance search result. The system superimposed bubbles
according to the attributes of PF, PD, and PS was with sizes corresponding to the social distance for
considered for the EDL evaluation. For PF with each attribute of PF, PD, and PS, and the safe
fever, the Social Distance in the COVID-19 passing route (green).
environment is defined as 6 feet by the Centers for
Disease Control and Prevention and 1 m by the
WHO [20][21]. We defined Social Distance as 5
m, five times the WHO standard. According to
Matsunaga et al., when the distance to a person is
less than 6 m, the likelihood of assault
increases[22]. Therefore, we set the social
distance to PD (person with dangerous facial
expressions/attitudes) as 10 m. We also set the
social distance as 20 m assuming that the PS
(police officer or security guard) has an eye range
of 20 m. Based on these social distances, the
system calculates EDL for all edges in the same
way as in a wide-area safe route search, updates
the graph structure, and finds the shortest path
using the Dijkstra method. In this way, we can Figure 7: Results of wide-area safe route search
avoid dangerous people by computing travel paths
that can maintain safe social distances. In addition,
these social distances can be changed dynamically
according to the user's attributes, season, time of
use, and level of danger of the person to make it
safer and more convenient. For example, the
social distance can be adjusted to take more time
when the user is a child or a woman, or when it is
late at night.
2.3. Simulation Results
Figure 8: Visualization detected and shared PF,
Figure 7 shows the mapping results of the PD, and PS, and Narrow-area safety route search
current positions of PF, PD, and PS detected and results
shared by 100 edge devices and the wide-area safe
route on a 2D map. We deployed PE0, a person 2.4. Smartphone prototype system
tasked with moving from the start point S to the
goal point G. PE0 is equipped with an edge device Simulator experiments have shown that the
and moves along the route indicated by the edge system works effectively within the virtual
device's wide-area safety route search. PE0 is Akihabara. Therefore, we developed a prototype
equipped with an edge device and moves along system using a smartphone as shown in Figure 9.
the route indicated by the edge device's wide-area The edge device on the simulator we implemented
safe path search. Although the shortest route is to was an ideal device that could detect people and
go straight from S to G, the blue color indicates their attributes within a radius of 30 meters
that the route search is performed in consideration without error. However, since such an ideal
of the level of danger. When PE0 needs to pass PF device does not exist in reality, our prototype
system simply detects a person from a 2.5. Smartphone Prototype Result
smartphone's RGBD camera image (without a
level of danger assessment) and shares the
As shown in Figure 11 we placed PF, a person
person's location coordinates. We developed a
assumed to have a fever, in the park. Next,
smartphone application in Unity and installed it
smartphones were placed at PE1, PE2, and PE3.
on an iPhone 13 pro max (Figure 10). The app
The location of PE1 is in the same park as PF and
obtains the position coordinates (latitude,
has no obstacles, so PF can be detected directly,
longitude, and altitude) of the smartphone from
but from the location of PE2 and PE3, PF cannot
the Global Navigation Satellite System (GNSS),
be detected directly due to the obstacles (height
the azimuth and orientation of the device from the
difference and trees).
magnetic compass and gyro, and detects the
person using the RGB camera and LiDAR, and
also gets the distance to the person. The distance
to the person is also obtained. The application
then calculated the position coordinates of the PE1
detected person and displayed a translucent PF PE3
bubble (red sphere) with a diameter of 5 meters.
The location information of the detected person is
shared by Unity's photon plug-in and is
superimposed on the same location from another
smartphone. The location coordinates of the
detected person are uploaded to the data server PE2
(Nifty Cloud mobile backend), and the
Figure 11: PF and edge devices position
information of the detected person can be viewed
from a web browser (Google Map). Figure 12 shows the screen when looking at PF
from PE1, where PF is detected and a red bubble
Data Server Web Server with a diameter of 5 m is superimposed.
Edge Device Edge Device Web Browser
Figure 9: Smartphone prototype system
position azimuth Person Detection
coordinates orientation RGB-Cam/LiDAR
Figure 12: Detected PF from PE1
Figure 13 and 14 show the view of PF from PE2
and PE3. From PE2 and PE3, the PF cannot be
seen directly due to the height difference and the
trees, respectively. However, a red bubble is
displayed at the location of the PF based on the
information from PE1, so that the user can
iphone13 recognize the presence of the PF.
pro max
Figure 10: Prototype Application
Figure 13: View of PF from PE2 Figure 15: Floated Bubble from PE2
Figure 14: View of PFs from PE3 Figure 16: Floated Bubble from PE3
A red bubble with a diameter of 5m Figure 17 shows the results of mapping the PFs
representing the PF is shown in Figures 12, 13, detected by PE1 on Google Map (aerial photo)
and 14, but these bubbles are not occluded by with images. This allows users in remote locations
obstacles such as the ground or trees in real space, to determine the location of PFs.
making it difficult to get a sense of distance. In the
simulator, the 3D geometry of the entire
townscape is known, so the bubbles can be
occluded, but the prototype system cannot
measure the 3D geometry of the real world, so
occlusion is not possible. The LiDAR in the
iPhone13 can measure real-world 3D geometry at
close range (about 15m). The AR Foundation tool
provides an Automatic Environment Occlusion
function using LiDAR, but it does not work on
objects at long distances as in this experiment.
Therefore, We generally assume that there are
fewer obstacles in the sky, we floated the bubble Figure 17: Detected persons on Google Map
at 30 m above the detected PF and displayed a thin
cylinder up to the bubble. The results are shown
in Figures 15 and 16.
3. Discussion, and Future Works 3.2. Smart Phone and Future Works
We have developed a prototype system that As mentioned in section 3.5, the bubble
can detect and avoid criminal damage and representing the location of the detected PF is not
unpleasant events in advance, using a simulator occluded by objects in the real world, making it
and a smartphone. Through experiments, we difficult to get a sense of the distance. We floated
confirmed that the system works and that the the bubble at 30 m above the PF, but the same
information about the detected persons can be problem occurs when there is an object in the sky.
shared. In the future, we would like to estimate the 3D
geometry of the object in front of the PF from the
3.1. Simulator and future works video and perform bubble occlusion. Another
issue is the accuracy of GNSS. We experimented
in a park with no obstacles in the sky, but the
Once the simulator was run and the locations
GNSS accuracy was at best about 5 m horizontally
of PFs, PDs, and PSs were identified on the 2D and 3 m vertically, with errors of 30 m or more in
map, we felt that "we should avoid that dangerous
some situations. Therefore, a method to estimate
area where PFs and PDs are densely located," and self-position by measuring the 3D geometry of an
did not feel that much need for wide-area safe object is possible. For example, position
route search. It may be sufficient to have
estimation can be performed by matching point
information that "that area is dangerous". In cloud data obtained by LiDAR sensors installed in
addition, when we had to pass by a PF or PD, the recent smartphones with real space, and
AR display of social distance in the form of a something similar has already been implemented
bubble made us feel secure, but when we had to in automatic driving. For this purpose, it is
pass through their bubble, we felt great stress
necessary to convert the entire real space into
despite it being a simulation. On the other hand,
point cloud data. However, point cloud data in
when there was a PS that was supposed to be a cities and buildings are available through the
police officer, etc., I was able to move with ease PLATEAU project [23] promoted by Japan's
even if there was a PD. However, the bubble was Ministry of Land, Infrastructure, Transport, and
often invisible behind a building, and when we
Tourism, and we would like to consider
turned a corner, we were sometimes surprised
improving the accuracy of self-position
when the bubble suddenly appeared, and we felt
estimation using this method in the future.
that the AR display needed to be improved in this Through these research activities, we will
regard. The edge device on the simulator is an
promote the social implementation of
ideal device to detect PF, PD, and PS within a information-sharing and navigation systems that
radius of 30m without error, but in reality, such an
can detect and avoid criminal damage and
ideal device is difficult to implement. Therefore, unpleasant events in advance.
in the future, we would first like to narrow down
the detection target to people with fever and
conduct experiments in the real world to 4. References
determine the social distance from people with a
fever to verify the effectiveness of the system. We [1] Study on Violent Offenders (Ministry of
would also like to study a safer route-finding Justice, Japan), URL:
method by incorporating PF, PD, and PS path http://www.moj.go.jp/housouken/housouken
forecasts into the wide-area and narrow-area safe 03_00104.html, Sep 2021
route-finding method proposed in this study. [2] E. Galbrun, K. Pelechrinis, E. Terzi, "Urban
In this paper, we have evaluated the navigation beyond shortest route: The case of
effectiveness of the system subjectively, but in the safe paths", Information Systems, 57, 160-
future, we would like to evaluate the effectiveness 171.
of the system quantitatively. For example, as [3] Z. Yizhou, X. Yuetian, A. Shohreh. "On
indicators of quantitative evaluation, we are Integration of Any Factor with Distance for
considering the frequency with which PDs can not Navigation : Walk Safely and Fast Enough",
be avoided, how many were able to take the safe Proc. 2019 IEEE 23rd International
route, the average of the shortest distance to a PD, Enterprise Distributed Object Computing
and the distance increased by the safe route Workshop (EDOCW).
proposed by the system.
[4] G. Sukru, C. Mustafa, A. Bilin, "Pedestrian [16] CrimeReports,
Path Modification Mobile Tool for COVID- URL:https://www.crimereports.com, Sep
19 Social Distancing for Use in Multi-Modal 2021.
Trip Navigation", arXiv.org, May 2021. [17] My Safe Map,
[5] F. Maria, B. Alina, G. Antonino, C. Antonio, URL:https://web.mysafemap.com/, Sep
V. Massimo, "Viale S., A proximity-based 2021.
indoor navigation system tackling the [18] Edward T. Hall, The Hidden Dimension,
COVID-19 social distancing measures", Doubleday, New York, 1990.
Proc. 2020 IEEE Symposium on Computers [19] ZENRIN City Asset Series,
and Communications (ISCC). URL:https://www.zenrin.co.jp/contents/pro
[6] K. Kapil, G. Yuxiang, M. Jared, W. I-Jeng, duct/service/3d/asset/index.html, Sep 2021.
H. Chien-Ming, "Group-Aware Robot [20] CDC "How to Protect Yourself & Others",
Navigation in Crowded Environments", URL:https://www.cdc.gov/coronavirus/2019
arXiv.org, Dec 2020. -ncov/prevent-getting-sick/prevention.html,
[7] L. Lucia, D. Daniel, C. Gianluca, S. Roland, Sep 2021.
D. Renaud, "Robot Navigation in Crowded [21] WHO " Coronavirus disease (COVID-19)
Environments Using Deep Reinforcement advice for the public",
Learning", Proc. IEEE/RSJ International URL:https://www.who.int/emergencies/dise
Conference on Intelligent Robots and ases/novel-coronavirus-2019/advice-for-
Systems (IROS) 2020. public, Sep 2021.
[8] B. Stacy, A. Ali, E. William, S. Morgan, R.
Erick, "Is Someone There? Do They Have a
Gun- How Visual Information about Others
Can Improve Personal Safety Management
for Blind Individuals", Proc. ACM
SIGACCESS Conference on Computers and
Accessibility 2017.
[9] Z Fan, J Yin, Y Song, Z Liu, Real-time and
accurate abnormal behavior detection in
videos - Machine Vision and Applications,
2020.
[10] M Fang, Z Chen, K Przystupa, T Li, M
Majka, O Kochan. Examination of abnormal
behavior detection based on improved
YOLOv3, Electronics, 2021
[11] Incident Detection Solutions,
URL:https://info.hitachi-
ics.co.jp/topics/news/20210309a.html, Mar
2021.
[12] J. Salamon and J. P. Bello, "Deep
Convolutional Neural Networks and Data
Augmentation for Environmental Sound
Classification," in IEEE Signal Processing
Letters, vol. 24, no. 3, pp. 279-283, March
2017.
[13] K. Jaewoo, C. Meeyoung, S. Thomas, "
SocRoutes: safe routes based on tweet
sentiments", Proc. 23rd International
Conference on World Wide Web, April 2014,
179-182
[14] RedZone Map, URL:
https://www.redzonemap.com/, Jul 2019.
[15] My Safetipin, URL:http://safetipin.com/,
Sep 2021.