Computer Vision Based Smart Security System for Explainable Edge

Computing

N. Nisbet and J. I. Olszewska

School of Computing and Engineering, University of the West of Scotland, U.K.

Keywords:

Explainable Edge Computing, Security and Privacy, Computer Vision Applications, Explaining Object

and Obstacle Detection, Trustworthy Intelligent Vision Systems, Transparent Decision-Support Systems,

Explainable Artiﬁcial Intelligence.

Abstract:

Smart cities aim to reach a high quality of life for their citizens using digital infrastructure which in turn

need to be sustainable, secure, and explainable. For this purpose, this work is about the development of

an explainable-by-design smart security application which involves an intelligent vision system capable of

running on a small, low-powered edge computing device. Hence, the device provides facial recognition and

motion detection functionality to any electronic motion picture input such as CCTV camera feed or video. It

highlights the practicality and usability of the device through a support application. Our resulting explainable

edge computing system has been successfully applied in context of smart security in smart cities.

1 INTRODUCTION

Modern smart cities deploy a growing number of edge

computing technologies (Khan et al., 2020), which

include among others internet of things (IoT) de-

vices and artiﬁcial intelligence (AI) algorithms to pro-

cess such collected data (Mishra et al., 2024), for

diverse smart applications ranging from air quality

monitoring (Zhu et al., 2023) to building operation

(Slee et al., 2021), maintenance (Xiao et al., 2024)

or surveillance (Sharma and Kanwal, 2024b). How-

ever, newest regulations and standards (Houghtaling

et al., 2024), require these intelligent systems to be

trustworthy. Therefore, using explainable artiﬁcial in-

telligence (XAI) (Li et al., 2024) to develop transpar-

ent and reliable edge computing systems is not an op-

tional requirement but a cornerstone aspect of Society

5.0 (Chamola et al., 2023).

In particular, smart cities need trustworthy solu-

tions for smart security (Sharma and Kanwal, 2024a)

which integrates emerging technologies (such as IoT

devices, cloud computing, edge computing) (Sodiya

et al., 2024) and intelligent systems (like computer-

vision systems) into traditional security measures to

improve the safety and security of people, property,

and data (Utimaco, 2024).

Indeed, smart cities tend to an increase in security,

safety, and better living experiences for citizens, and

these smart cities are made possible due to smart cam-

eras and sensors. For example, the city of Barcelona

uses c. one million different sensors around the city,

monitoring everything from trafﬁc to temperature and

even waste management (Lea, 2020). Devices such as

sensors and cameras vary in the amount of bandwidth

required which is where edge computing is making

a big impact, because, quicker and more cost effec-

tive than ever, the sheer amount of data produced by

IoT devices is still time consuming to transport and to

store. Moreover, the ever-growing popularity of IoT

devices is creating more advanced versions producing

much more data (Naveen and Kounte, 2019).

While the majority of smart devices across our

cities are sensors that do not need to constantly send

data every few seconds or even milliseconds, some of

these sensors need quick decisions such as a smart

road detecting an accident and wanting to close a

lane. A single edge device could then control mul-

tiple sensors and enable these quick decisions to be

made without having to send the data to a main hub.

So cutting out this transfer time between multiple sys-

tems sitting in a queue waiting to be delt with can

save valuable seconds. With millions of sensors dot-

ted around a large area, there is also the possibility

data could arrive at different times with the decision

making delayed due to all the information not being

processed in time. Edge devices are thus an efﬁcient

solution to build networks within networks in order to

create fully automated systems with little or no impact

on the wider grid.

As smart cities grow, expand, and develop over

time, edge devices can work with cloud computing in

easing the strain on sudden expansion. Since smart

Nisbet, N. and Olszewska, J.

Computer Vision Based Smart Security System for Explainable Edge Computing.

DOI: 10.5220/0013016300003886

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

In Proceedings of the 1st International Conference on Explainable AI for Neural and Symbolic Methods (EXPLAINS 2024), pages 147-154

ISBN: 978-989-758-720-7

147

devices can come in all shapes and sizes, these small,

low-cost edge devices can be fully customised to ac-

commodate any device and slip into the smart city jig-

saw with ease. As the technology develops, the up-

grading process is much easier and cheaper with only

the localised devices needing attention and not the

full network (Khan et al., 2020). Having this amount

of processing power available in small devices is en-

abling data processing closer to the source. Being

able to make sense of the data locally allows quicker

decisions as well as reducing transfer trafﬁc and stor-

age (Shi et al., 2016).

Furthermore, explainable edge computing can

contribute to provide transparent applications closer

to data sources such as IoT devices or local edge

servers. This proximity to data at its source can de-

liver beneﬁts, including not only faster insights but

also explainable decisions along with improved re-

sponse times and better bandwidth availability (IBM,

2024). Therefore, smart security often uses edge com-

puting in conjunction with intelligent vision applica-

tions (Sharma and Kanwal, 2024b) in new-generation,

real-time video surveillance systems (Visionplatform,

2024), in order to process and analyze visual data at

the network’s edge, close to where data is generated,

whilst reducing latency and bandwidth usage.

However, most existing edge computing systems

for smart surveillance embed deep learning algo-

rithms, (Perwaiz et al., 2024), (Aminiyeganeh et al.,

2024), (Sharma and Kanwal, 2024a) which are very

popular but still not considered as explainable or

transparent (Chamola et al., 2023).

So, in this study, we developed an explainable

edge computing system, including two complemen-

tary applications - one (i.e. the support application)

running on the main user’s device and the other one

(i.e. the edge application) integrated on the edge de-

vice, in order to detect, capture, and identify individ-

uals from live videos in context of smart cities.

It is worth noting that our intelligent vision system

is explainable by design and consists in the edge ap-

plication embedded on the edge device with face and

motion detection capabilities. Thus, the edge device

is able to make trustworthy decisions on facial recog-

nition by making the difference between known and

unknown faces which it categorises appropriately.

On the other hand, a user can receive as much

or as little information as they like with the support

application designed to split data from the edge ap-

plication to allow the user to receive all new infor-

mation, known/unknown faces, or motion detection.

This gives full ﬂexibility on the users’ end to select

what they require, how, and when.

Thence, the main contributions of this work are

the explainable design and user-friendly development

as well as the real-world deployment of the explain-

able edge computing system for smart security.

The paper is structured as follows. In Section 2,

we describe the developed explainable edge comput-

ing system. Then, we present some experiments using

our system in real-world environment, as reported in

Section 3. Conclusions are drawn up in Section 4.

2 DEVELOPED SYSTEM

2.1 System Design

2.1.1 Communication

The main aim for the smart security application is to

run remotely on a small edge device, and this im-

plies the method of communication needs to be sim-

ple but robust. To make the most of the connectiv-

ity features available in current small edge suitable

devices such as single board devices, the most com-

mon of these communication technologies had to be

explored. Hence, Bluetooth is a very cheap add-on

for any device. However, it was not considered in

this work due to its poor connection range. On the

other hand, Wi-Fi was considered and used for some

testing, but this could not be the only means of con-

nection due to not all devices having this capability.

Therefore, to cater for the most widely available op-

tion, Ethernet has been chosen as the main source

for connectivity, which allows for the best transfer

speeds, while Wi-Fi, which is almost as quick as Eth-

ernet, has been used when the range and needs for

extra equipment were considered as acceptable.

Besides, to provide the best ﬂexibility, one of

the most common web services which is called a

Representational State Transfer Application Program-

ming Interface (REST API) was selected (Au-Yeung,

2020). Indeed, a REST API enables applications to

communicate with a server across a network. Data

and commands are transferred as JavaScript Object

Notation (Json), meaning they are not language spe-

ciﬁc. A REST API works by handling GET, POST,

PUT and DELETE requests through URLs, control-

ling the direction and type of data being sent or re-

ceived (RedHat, 2020). The REST API is handled by

the micro web framework Flask enabling the applica-

tion to send and receive Hypertext Transfer Protocol

(HTTP) requests or responses.

2.1.2 Execution

To enable any camera to be smart implies external

processing and the possible use of a third-party li-

EXPLAINS 2024 - 1st International Conference on Explainable AI for Neural and Symbolic Methods

148

brary. Facilitating this for a low-powered, limited pro-

cessing edge device requires careful considerations.

Indeed, the process of taking an image and transform-

ing it into usable data involves advanced computer

vision techniques (Sharma and Kanwal, 2024a), and

there are different algorithms for face recognition to

choose from (Rostum and Vasarhelyi, 2024), (Chen

and Krzyzak, 2024), (Wood and Olszewska, 2012).

In particular, this system needs to integrate im-

age/video processing methods for detecting, captur-

ing and identifying an individual from live videos.

Detecting an individual involves locating a human

face in a live video or a still image. Capturing is

the process of taking that information and storing it

in a way that can be used, and identifying it is know-

ing who the detected person is (Thales, 2020). More-

over, this system aims to be an explainable edge com-

puting system. Therefore, the Histogram of Oriented

Gradients (HOG) algorithm which is based on an ex-

plainable machine learning approach (Chamola et al.,

2023), has been selected. It uses a black and white

or grey scale version of an image, where each pixel

is compared to the ones round about. From this, the

change in gradient is noted, creating a detailed pattern

of the image that can be matched with a face database

for face recognition and identiﬁcation purposes. An-

other beneﬁt of these patterns is that it is also very

powerful at detecting shapes or moving objects which

is why it is the best option for this work.

2.1.3 Storage

One of the main advantages of a small edge device

is the ability to make quick decisions, reducing thus

the amount of data transfer, but limiting the scope for

storage types that can be considered - with cloud or re-

mote varieties not suitable. Additionally, not all edge

devices are capable of running an operating system

(OS) such as MS Windows. Therefore, cross compat-

ibility is required (Wheeler and Olszewska, 2022).

Despite MySQL being a highly popular manage-

ment system and available for all OS’s, its recom-

mended memory requirements of 4-8GB make it not

suitable for this type of device. Another popular

database called PostgreSQL is also unsuitable due to

its high storage requirements of 512 MB.

Indeed, for a device with limited storage, process-

ing power and possible transfer rates, the database

needs to be lightweight. With these requirements,

the Python-based TinyDB ﬁts the bill. Storing data

in Json format ﬁts also well with the REST API side

of the application. In addition, the database takes up

minimum storage and is user friendly. A TinyDB can

be setup like any other with a single database contain-

ing different tables. Due to the miniscule space each

one takes up, it is easier, and more manageable, to

have a database for each section, as shown in Fig. 1.

Figure 1: Edge application storage diagram.

Figure 2 illustrates how the data ﬂows in our sys-

tem and how the different databases are managed.

Hence, the users make requests to the edge device

from the application on their computer to access the

data through the API call. The API can then control

the retrieval or insertion of the data to the correct lo-

cation. From the edge software, the detected data is

programmatically directed to the correct location to

be stored until required.

Figure 2: System database management diagram.

2.2 System Applications

Our system consists of two different applications,

namely, the main edge application and a support ap-

plication, as illustrated in Fig. 3. The edge applica-

tion is performing the main data and video process-

ing, while the support application demonstrates how

the edge device can be integrated into other systems.

2.2.1 Edge Application

The edge application has been written mainly us-

ing Python 3 version and has been structured in ﬁve

classes (see Fig. 4), with the main class api server

handling the communication between the application

and external network. The video processing has been

split into a class each for facial recognition and mo-

tion detection with another for processing both. Keep-

ing these separate enable them to be run individually

or together with these processes being managed by

the start process class.

Computer Vision Based Smart Security System for Explainable Edge Computing

149

Figure 3: Our system architecture.

Figure 4: Class diagram for the Edge Application.

To allow for multiple camera feeds to be

processed simultaneously, the edge application

runs the concurrent.futures library. Available

since Python 3.2, it has two subclasses, namely,

ThreadPoolExecutor and ProcessPoolExecutor,

using multi-threading and multi-processing, respec-

tively. What makes this better than other multi-

processors is that tasks are added to a pool which

allocates them to available resources. This pooling

increases speed and reduces computing power by cre-

ating a repository of all the active tasks and only re-

leases them when the resources are available (Forbes,

2017).

On the other hand, the edge application uses the

Open Source Computer Vision Library (OpenCV)

with the Dlib library to capture, store and retrieve data

from images and video. The Dlib library is a toolkit

originally written in C++ containing machine learn-

ing algorithms. A number of these tools have been

made available for use with Python by means of an

API. OpenCV handles a number of these pretrained

algorithms which for this application is mainly the

facial landmark detector for HOG encoding. There-

fore, using OpenCV alongside the facial recognition

library (Geitgey, 2020) enables the edge application

to inspect live video footage for face identiﬁcation.

In particular, as OpenCV stores colour in the BGR

colour space, this needs to be converted in the more

usable RGB. Using the video capture class to gain the

footage as an array of images, the application can then

scan each frame using the facial recognition face loca-

tions class. This uses the pretrained Dlib HOG SVM

to ﬁnd any facial patterns and the popular library

NumPy to return the patterns as array. These pat-

terns can then be compared to the pre-saved images

of known faces through the face recognition compare

faces class and the similarity can be conﬁrmed using

the face distance class. The found faces are shown us-

ing the OpenCV rectangle class which enables a box

to be displayed using the outer most co-ordinates of

the detected pattern. If the face is known, i.e. matches

a stored image, the image ﬁlename, date, and time are

saved to the tiny database table detected face us-

ing the insert method from the table class (Siemens,

2023). If the detected face is not known, i.e. does not

match any of the stored images, the image with the

newly detected face is saved to a local folder using

the OpenCV function imwrite. The details are then

saved to the unknown image table in the TinyDB, all

the details are also saved to the unread data table.

The motion detection also utilises the video cap-

ture class to gain the live video footage as a sequence

of images. This allows the ﬁrst and second frame to

be captured using the function read which returns de-

coded images that can then be compared to each other

using the function absdiff which calculates the ab-

solute difference for each element in the two arrays.

When motion or a difference in between the

frames has been detected, the new image containing

the difference is simpliﬁed to a gray-scale, Gaussian-

EXPLAINS 2024 - 1st International Conference on Explainable AI for Neural and Symbolic Methods

150

(a) (b) (c) (d)

Figure 5: Support Application screens: (a) splash screen; (b) login/register screen, (c) home screen; (d) add-device screen.

blurred image by applying OpenCV functions This

image is then ﬁltered to reduce noise by removing

pixels with too high and low values. These values are

then dilated to over-emphasise the edges and enable

the contours to be used to box off the different area or

detected motion. As with the face detection, the date

and time of the detected motion is inserted into the

TinyDb along with the details of the device IP from

which the detection was found.

2.2.2 Support Application

To highlight the versatility of the edge application and

RESTful API, the support application (see Fig. 5 (a))

has been built on the .Net Core platform or Universal

Windows Platform desktop application. The layout

for the page views are created using the Extensible

Application Markup Language (XAML) using grids

to control the position of visual tools, e.g. TextBlock.

The main and most important screen is the lo-

gin/register page, as displays in Fig. 5 (b), as this

is how the user will connect to the actual edge device.

From here, there is a home screen (see Fig. 5 (c))

which provides the user with an overview of the avail-

able statistics together with the functionality to start

and add devices. Using the add-device screen (see

Fig. 5 (d)), all the details can be added for a camera

with options for the detection setting with face and/or

motion selectable.

The data provided on the data screen (see Fig. 5

(c)) gives a breakdown of all the data that has been

detected since last viewing. When the API from this

page has been called, the data is wiped meaning no

new repeat data will be sent providing the user with a

data stream that can be stored locally.

3 EXPERIMENTS

3.1 Test Setup

To fully test both applications, the developed edge

software was installed on a Raspberry Pi device

(Raspberry Pi 4 Model B, Broadcom BCM2711,

Quad core Cortex-A72 (ARM v8) 64-bit SoC

1.5GHz, 4 GB RAM, Ubuntu 20.10 OS). The support

application was installed on a laptop (Dell Latitude

E5530, Intel Core i7-3610QM CPU 2.30GHz, 8GB

RAM, Windows 10 Professional (20H2) OS) which

was communicating with the edge device, as shown

in Fig. 6. The support application was then used to

receive, collect and display data from the edge device.

Figure 6: Overview of the system hardware setup deployed

in real-world environment.

3.2 Face Detection

The initial testing of our software using the laptop’s

built-in webcam device involved ﬁve different indi-

viduals and was performed to test the functionality for

the face detection where the face was detected as ‘un-

known’ after comparing the image against the saved

images in the database, as exempliﬁed in Fig. 7(a).

This unknown detected image was then saved through

the support application as a ‘known’ person who was

then detected when carrying out the same test again,

as illustrated in Fig. 7(b).

Both devices detected the same number of faces

in each of the 35 carried out tests in this series, and

the results on both the laptop and the Raspberry Pi

devices gave an accuracy of 97.14%.

A further series of 200 tests was carried out to

test the full system using a CCTV test video from

Panasonic and containing nine different persons who

were evolving in that video over time. Both devices

were set to 24fps with the same test video and sub-

ject images. Each test was repeated twice - at ﬁrst,

Computer Vision Based Smart Security System for Explainable Edge Computing

151

(a) (b)

Figure 7: Example of face detection: (a) a detected ‘un-

known face’; (b) added to the database and then updated to

‘known face’.

all the faces were ‘unknown’ and secondly, all the

faces were ‘known’. For these latter series of tests on

the CCTV video, the cumulative computational time

were measured on both devices. For the series of tests

which were implemented with no known faces, the

laptop completed the full batch of tests in 12 hours

8 minutes, while the Raspberry Pi took only 6 hours

17 minutes. The other series of tests using all the

known faces was completed much quicker, with the

laptop taking 9 hours 20 minutes, and the Raspberry

Pi only 4 hours 44 minutes. These results were repre-

sentational of the real-world situation, where a much

higher processing time is needed for detecting an un-

known face rather than a known face when perform-

ing a one-to-many search in the database, leading to

the computational processing time of the overall sys-

tem being not only dependent of the image processing

approaches but also of the one-to-many search algo-

rithm complexity. In terms of performance, the ﬁrst

series of tests showed a precision of over 90%, while

the second series of tests reached a precision of 100%

where no person was missed in any of the tests, and

everyone was detected correctly. Furthermore, even

with the lighting reﬂections and background clutter,

there were no false positives detected. Therefore,

these tests highlight that the system performance is re-

lated to the training dataset which quality needs to be

carefully considered when deploying such edge com-

puting system.

Besides, during testing on the laptop, it was noted

through its Task Manager that the application used

20-25% CPU and peaked at almost 2GB memory but

steadied out at 250-300MB with the power usage very

high. Having the application running in the back-

ground did not appear to cause any performance is-

sues to other applications being used. In comparison,

the Raspberry Pi monitored via its activity manager

has used more processing power, between 25-35% but

the memory usage was 250-400MB with a peak of

almost 600MB. When using video for testing on the

Raspberry Pi, it was noted that if the device ran hot,

with the temperature symbol ﬂashing in the top cor-

ner, its CPU usage rose to 75%. So, additional cau-

tion should be exercised with handling the Raspberry

Pi device as it should be kept in a cool place out of the

sun, with an optional fan ﬁtted.

3.3 Motion Detection

For the motion detection, the system sensitivity can be

adjusted, depending on the device used and subjects

needed. This is achieved by increasing or decreasing

the contourArea which is the area used to compute

the object motion.

Hence, tests were carried out for two different sce-

narios, as shown in Fig. 8, with a close webcam and

distant CCTV camera which required very different

values.

On one hand, Fig. 8 (a) illustrates a typical envi-

ronment this type of device is used for and performs

well - only detecting the motion of people and cars.

It should be noted this test was carried out on quite a

still day. However before settling on, the ﬁnal value

motion was detected on the trees.

On the other hand, Fig. 8 (b) demonstrates the

other end of the motion scale, with a webcam used

to test close quarters’ motion. For this test, the

contourArea setting was much higher, to allow for

the closer and therefore larger objects’ motion detec-

tion.

Besides, caution should be taken when using the

motion detection feature of the edge computing sys-

tem that the device is ﬁxed and so it does not move,

in order to avoid false positive detections due to the

camera shake, as one can observed in Fig. 8 (c).

(a) (b) (c)

Figure 8: Examples of motion detection: (a) on high sensi-

tivity; (b) on low sensitivity; (c) caused by camera shake.

3.4 Discussion

The developed applications, namely, edge applica-

tion and support application, were both successfully

tested using standard and advanced software testing

approaches (Black et al., 2022), with some series of

tests described in Sections 4.2-4.3, in order to deliver

a reliable edge computing system.

Furthermore, this system’s development process

also included user tests involving the participation

of ﬁve individuals with different technical knowledge

and backgrounds, to provide feedback helping the de-

veloper to achieve a user-friendly system.

EXPLAINS 2024 - 1st International Conference on Explainable AI for Neural and Symbolic Methods

152

Besides, during the development of this explain-

able edge system, ethical considerations were anal-

ysed as well, e.g. to ensure that data and sensitive in-

formation were stored and handled in an ethical way.

It is worth noting that the use of CCTV with smart

technology such as face recognition is a relatively new

concept with the laws and e guidelines still being cre-

ated at national and international levels (e.g. the Gen-

eral Data Protection Regulation (GDPR), EU AI Act

2024, etc.). Although this application does not ven-

ture into this area of detail or intrusiveness, it high-

lights many possible issues the public had and still

has with this area. As CCTV cameras become more

powerful and smaller with high quality lenses and im-

age processing, it is important they are used appropri-

ately (Van Noorden, 2020). Albeit for this applica-

tion, consent would not be available due to the device

being aimed at the security sector, full consideration

into the handling, storing, and removal of data has

been considered. Indeed, full control over the nam-

ing and storing has been programmed in to make it

easy to add or remove images and details. Moreover,

all companies using this edge computing system are

expected to adhere to government guidelines.

It is worth noting that many of the issues relat-

ing to ethics can be improved by the security of the

device (Quincozes et al., 2024) and careful consid-

eration over storage (Ahmadi, 2024). Thence, this

application is running on an Ubuntu device which

is password protected. The communication is also

secured with a password, while the main source of

storage of the Raspberry Pi, which is the mem-

ory card, can be encrypted should it need to be.

For a user to access the edge device, they need

to register their details by sending them using the

HttpClient which encodes them and sends them

through a Basic AuthenticationHeaderValue (Mi-

crosoft, 2024). The API server on the edge side stores

the submitted details in the TinyDb users db table

and encrypts the password using the Python library

Werkzeug function generate password hash. The

encryption is created using SHA256 hash algorithm

with salt, which is random extra data for added secu-

rity in the format method$salt$hash. The added

salt encryption helps disguise the same password

used by more than one user. SHA265 is a vari-

ant of the highly popular SHA-2 which sets the size

to 256 bits, meaning all hashed codes are the same

size. To check the credentials on login, the function

check password hash compares the received pass-

word against any stored password and returns true or

false. For added security, a random twelve-digit pub-

lic ID is generated during registration which will be

used for creating a token on login. When a regis-

tered user logs into the edge device, their details are

checked against the contents of user db for a match.

If all the details are conﬁrmed, a token is generated

using the library PyJWT or Json web token which

uses the HS256 algorithm. HS265 is a Hash-based

Message Authentication Code (HMAC) combining,

in this case, a predeﬁned secret key to encode the gen-

erated user public ID using H265 (Otemuyiwa, 2017).

For extra security, an expiry date can be added to the

token to force users to have to login and to stop old to-

kens being active longer than they need to be. On the

Windows support application, the token is returned on

locally in memory. Any future requests made to the

edge device for data are thus sent using this token as

Digest authorisation. These received requests are then

decoded using the secret key to gain the user public

key which is check against the results in the user db.

4 CONCLUSIONS

In this work, we developed an explainable edge com-

puting prototype for the next-generation smart envi-

ronments which require privacy, security and trust.

Hence, we built a system running two applications,

i.e. a user-friendly application to support the main

edge application integrating an intrinsically explain-

able computer vision technologies together with a

small, low-powered edge computing device, in order

to provide an accurate whilst responsible AI-based so-

lution for real-world smart security.

REFERENCES

Ahmadi, S. (2024). Security Iimplications of edge com-

puting in cloud networks. Journal of Computer and

Communications, 12:26–46.

Aminiyeganeh, K., Coutinho, R. W. L., and Boukerche, A.

(2024). IoT video analytics for surveillance-based

systems in smart cities. Computer Communications,

224:95–105.

Au-Yeung, J. (2020). Best practices for

REST API design. Available online at:

https://stackoverﬂow.blog/2020/03/02/best-practices-

for-rest-api-design/.

Black, R., Davenport, J. H., Olszewska, J. I., Roessler, J.,

Smith, A. L., and Wright, J. (2022). Artiﬁcial Intelli-

gence and Software Testing. BCS Press.

Chamola, V., Hassija, V., Sulthana, A. R., Ghosh, D., Dhin-

gra, D., and Sikdar, B. (2023). A review of trustwor-

thy and explainable artiﬁcial intelligence (XAI). IEEE

Access, 11:78994–79015.

Chen, G. and Krzyzak, A. F. (2024). Face recognition

via selective denoising, ﬁlter faces and HOG fea-

Computer Vision Based Smart Security System for Explainable Edge Computing

153

tures. Signal, Image and Video Processing. Springer,

18:369–378.

Forbes, E. (2017). Executors and pools. In Learning Con-

currency in Python, pages 159–188. Packt.

Geitgey, A. (2020). face recognition pack-

age. Available online at: https://face-

recognition.readthedocs.io/en/latest/face recognition.html.

Houghtaling, M. A., Fiorini, S. R., Fabiano, N., Goncalves,

P. J. S., Ulgen, O., Haidegger, T., Carbonera, J. L.,

Olszewska, J. I., Page, B., Murahwi, Z., and Prestes,

E. (2024). Standardizing an ontology for ethically

aligned robotic and autonomous systems. IEEE Trans-

actions on Systems, Man, and Cybernetics: Systems,

54(3):1791–1804.

IBM (2024). What is edge computing? Available online at:

http://www.ibm.com/topics/edge-computing .

Khan, L. U., Yaqoob, I., Tran, N. H., Kazmi, S.

M. A., Dang, T. N., and Hong, C. S. (2020).

Edge-computing-enabled smart cities: A compre-

hensive survey. IEEE Internet of Things Journal,

7(10):10200–10232.

Lea, P. (2020). IoT and Edge Computing for Architects. 2nd

ed. Packt.

Li, Y., Xiao, Y., Gong, Y., Zhang, R., Huo, Y., and Wu, Y.

(2024). Explainable AI: A way to achieve trustworthy

AI. In Proceedings of the IEEE International Confer-

ence on Big Data Security on Cloud, pages 150–155.

Microsoft (2024). .NET documentation. Available online

at: https://docs.microsoft.com/en-gb/dotnet/.

Mishra, A. K., Reddy, R. R., Tyagi, A. K., and Arowolo,

M. O. (2024). Artiﬁcial intelligence-enabled edge

computing: Necessity of next generation future com-

puting system. In IoT Edge Intelligence, page 67–109.

Springer.

Naveen, S. and Kounte, M. R. (2019). Key technologies and

challenges in IoT edge computing. In Proceedings of

the IEEE International Conference on IoT in Social,

Mobile, Analytics and Cloud, pages 61–65.

Otemuyiwa, P. (2017). Brute Forcing HS256 is

Possible: The Importance of Using Strong

Keys in Signing JWTs. Available online at:

https://auth0.com/blog/brute-forcing-hs256-is-

possible-the-importance-of-using-strong-keys-to-

sign-jwts/.

Perwaiz, N., Fraz, M. M., and Shahzad, M. (2024). Smart

surveillance with simultaneous person detection and

re-identiﬁcation. Multimedia Tools and Applications.

Springer, 83:15461–15482.

Quincozes, V. E., Quincozes, S. E., Kazienko, J. F., Gama,

S., Cheikhrouhou, O., and Koubaa, A. (2024). A

survey on IoT application layer protocols, security

challenges, and the role of explainable AI in IoT

(XAIoT). International Journal of Information Secu-

rity, 23:1975–2002.

RedHat (2020). What is a REST API? Available on-

line at: https://www.redhat.com/en/topics/api/what-

is-a-rest-api.

Rostum, H. M. and Vasarhelyi, J. (2024). Comparing the

effectiveness and performance of image processing al-

gorithms in face recognition. In Proceedings of the

IEEE International Carpathian Control Conference,

pages 1–5.

Sharma, H. and Kanwal, N. (2024a). Smart surveillance

using IoT: A review. Radioelectronic and Computer

Systems, 109(1):116–126.

Sharma, H. and Kanwal, N. (2024b). Video surveillance

in smart cities: Current status, challenges and fu-

ture directions. Multimedia Tools and Applications.

Springer, pages 1–46.

Shi, W., Cao, J., Zhang, Q., Li, Y., and Xu, L. (2016). Edge

computing: Vision and challenges. IEEE Internet of

Things Journal, 3(5):637–646.

Siemens (2023). TinyDB. Available online at:

https://github.com/msiemens/tinydb.

Slee, D., Cain, S., Vichare, P., and Olszewska, J. I. (2021).

Smart lifts: An ontological perspective. In Proceed-

ings of the International Joint Conference on Knowl-

edge Discovery, Knowledge Engineering and Knowl-

edge Management (KEOD), pages 210–219.

Sodiya, E. O., Umoga, U. J., Obaigbena, A., Jacks, B. S.,

Ugwuanyi, E. D., Daraojimba, A. I., and Lottu, O. A.

(2024). Current state and prospects of edge comput-

ing within the Internet of Things (IoT) ecosystem. In-

ternational Journal of Science and Research Archive,

11(1):1863–1873.

Thales (2020). Facial recognition:

top 7 trends. Available online at:

https://www.thalesgroup.com/en/markets/digital-

identity-and-security/government/biometrics/facial-

recognition .

Utimaco (2024). What is Smart Security?. Available

online at: https://utimaco.com/service/knowledge-

base/emergency-communications-and-public-

warnings/what-smart-security.

Van Noorden, R. (2020). The ethical questions that haunt

facial-recognition research. Nature, 587:354–358.

Visionplatform (2024). What is edge computing for

computer vision and what are the beneﬁts?. Available

online at: https://visionplatform.ai/what-is-edge-

computing-for-computer-vision-and-what-are-the-

beneﬁts/ .

Wheeler, D. and Olszewska, J. I. (2022). Cross-platform

mobile application development for smart services. In

Proceedings of the IEEE International Symposium on

Computational Intelligence, pages 203–208.

Wood, R. and Olszewska, J. I. (2012). Lighting-variable

AdaBoost based-on system for robust face detection.

In Proceedings of the International Conference on

Bio-Inspired Systems and Signal Processing, pages

494–497.

Xiao, Y., Huo, Y., Cai, J., Gong, Y., Liang, W., and

Kolodziej, J. (2024). ERF-XGB: An Edge-IoT-based

explainable model for predictive maintenance. IEEE

Transactions on Consumer Electronics, 70(1):4016–

4025.

Zhu, Y., Al-Ahmed, S. A., Shakir, M. Z., and Olszewska,

J. I. (2023). LSTM-based IoT-enabled CO2 steady-

state forecasting for indoor air quality monitoring.

Electronics, 12(1):1–12.

EXPLAINS 2024 - 1st International Conference on Explainable AI for Neural and Symbolic Methods

154