Automatic Detection of the Driver Distractions Based on the Analysis of

Face Videos

Artur Urz˛edowski and Kazimierz Choro

Department of Applied Informatics, Wrocław University of Science and Technology, Wyb. Wyspia´nskiego 27, 50-386

Wrocław, Poland

Keywords:

Driver Fatigue State, Driver Drowsiness Detection, Driver Distraction, Driver Conversing and Eating,

Daylight and Nightlight Conditions, Face Analysis.

Abstract:

The objective of the paper is to propose a new driver fatigue detection method using Percentage of Mouth

Openness (POM) and Percentage of Eye Closure (PERCLOS) as well as to show its robustness across vari-

ous real-world conditions. The openness of the eyes and mouth was determined by calculating Aspect Ratios

(AR) and checking if AR exceeded a given threshold. Six videos simulated different driving scenarios were

recorded to test detection performance under diverse lighting, with and without corrective glasses, moreover

with additional complexities such as blinking lights. Furthermore, the method ensures avoidance of misclas-

siﬁcation during such driver’s activities as conversing and eating. The method effectively detects fatigue in all

test scenarios in which the fatigue state occurred.

1 INTRODUCTION

According to the European General Safety Reg-

ulation (European Parliament and Council, 2019),

vehicle manufacturers are obliged to implement

ADDW (Advanced Driver Distraction Warning) sys-

tems in all new vehicles from July 2024 (Regulation

32019R2144, 2019). Its goal is to drastically reduce

the number of accidents related to distraction, mainly

fatal accidents and serious injuries. ADDW systems

are able to warn drivers when they are distracted and

to stimulate drivers to pay more attention to the traf-

ﬁc situation and to warn about dangerous incidents

on the road. These systems include warnings such as

(Laxton et al., 2022) (Fu et al., 2024):

• driver drowsiness and attention warning – analy-

sis of vehicle systems and driver warnings when

necessary;

• lane departure warning system – warning when

the vehicle is drifting out of its travel lane;

• advanced emergency braking system – automatic

detection of a potential collision and activation the

vehicle braking system to decelerate the vehicle

with the purpose of avoiding or mitigating a colli-

sion;

https://orcid.org/0000-0001-6969-976X

• emergency lane-keeping system – assistance of

the driver in keeping a safe position of the vehi-

cle with respect to the lane or road boundary to

avoid a collision;

• driver availability monitoring system – assessing

whether the driver is in a position to take over the

driving function from an automated vehicle in par-

ticular situations.

The ADDW systems focus on detecting visual dis-

tractedness in order to provide timely warnings to

the driver. They should trigger an alert by display-

ing a light on the car instrument panel or produce an

adequate warning sound. Driver monitoring can be

achieved through a driver-facing camera embedded in

the steering wheel or instrument cluster. The most

popular solutions of driver monitoring are to track

driver eye movements and gaze directions, as well as

to observe the head and its strange rotations. There

is a strong correlation between head and eye move-

ments. By combining eye movement tracking and

head tracking, it is possible to detect drowsiness, for

example, when the driver is tired or intoxicated (e.g.

after drinking alcohol), the head drops down or eyes

are closed or barely opened. The cameras installed

inside the vehicles play a key role in detecting and

analyzing potential dangers to the driver. The meth-

ods generally perform well in normal daylight condi-

tions, accurately detecting the onset of driver fatigue,

542

Urze¸dowski, A. and Choro

s, K.

Automatic Detection of the Driver Distractions Based on the Analysis of Face Videos.

DOI: 10.5220/0013316300003912

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

In Proceedings of the 20th Inter national Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2025) - Volume 2: VISAPP, pages

542-549

ISBN: 978-989-758-728-3; ISSN: 2184-4321

but some obstructions occur with additional complex-

ities such as blinking lights. In addition, yawning

should be distinguished from openness and closure of

the mouth while talking or eating.

The goal of the paper is to verify the efﬁciency

of the algorithms of distractedness detection by an-

alyzing the videos recorded by a camera under var-

ious real-world conditions: i.e. in day and night

conditions, driver both with and without corrective

glasses, and during activities such as conversing and

eating. These evaluations aim to highlight the relia-

bility of the algorithms in practical scenarios, provid-

ing insights into their potential deployment in modern

driver-assistance systems.

The paper is structured as follows. The next

section describes related work on methods for au-

tomatic detection of driver distractions. The third

section presents the method for detecting mouth and

eye openness. The next section discusses the prob-

lem of threshold estimation. The proposed approach

is described in the ﬁfth section. The next section

presents the results of the tests performed on real

videos recorded in various real-world conditions. The

ﬁnal conclusions are presented in the last section.

2 RELATED WORK

2.1 Methods Based on Artiﬁcial Neural

Networks

(Dwivedi et al., 2014) developed a system using a

Convolutional Neural Network (CNN) to capture la-

tent facial features and implemented a SoftMax layer

to classify driver drowsiness. They created a dataset

with 30 subjects of varying skin tones and eye sizes,

under diverse lighting conditions to simulate real-

world scenarios. Subjects played an open-source

avoidance game while their faces were recorded,

achieving accuracy of 92.33% on the validation set.

The algorithm was also tested on new subjects, yield-

ing an average accuracy of 78%.

(Zhao et al., 2020) developed a Multi-Task Cas-

caded Convolutional Neural Network (MTCCN) sys-

tem to detect facial features like the face, nose, mouth,

and eyes. This system uses CNNs to determine

whether eyes and mouth are open or closed, assess-

ing driver fatigue through eye closure (PERCLOS –

Percentage of Eye Closure) and mouth closure (POM

– Percentage of Mouth Openness) metrics. Eye clo-

sure above 0.25 and mouth closure above 0.5 indicate

potential fatigue. They used a dataset from Biteda,

featuring 4,000 real-world driving images categorized

by eye and mouth states. The system achieved accu-

racy of 93.62%, sensitivity of 93.64%, and speciﬁcity

of 60.89%.

(Chirra et al., 2019) developed a drowsiness detec-

tion system based on eye analysis. Using the Viola-

Jones algorithm, they identify the face and eye regions

from video frames, which are then analyzed by a 4-

layer CNN that classiﬁes images as drowsy or non-

drowsy. Their dataset of 2,850 images (1,450 drowsy,

1,400 non-drowsy) was split into training, validation,

and testing sets. The method achieved accuracy of

96.42% on the test data.

(Rajkar et al., 2022) developed a deep learning

system for driver drowsiness detection using OpenCV

for video streaming. Haar cascade classiﬁer was used

to detect the region of the face and eyes. A CNN

classiﬁed whether the eyes and mouth were open or

closed over time. If eye closure or mouth openness

exceeded a threshold, the system alerted the driver.

The authors used the YawDD and Closed Eyes In The

Wild datasets. They have used 80% of the data for

training and 20% for testing, the system achieved ac-

curacy of 97.9% for eye classiﬁcation and 95.76% for

yawning classiﬁcation.

(Vijaypriya and Uma, 2023) developed a Multi-

Scale CNN system for detecting drowsiness, using

Dlib to detect facial points from video frames. Each

frame was denoised with Cross Guided Bilateral ﬁl-

tering and wavelet transformation before feature ex-

traction. Multi-scale CNNs was used to classify pat-

terns for drowsiness detection. The YAWDD and

NTHUDDD datasets were split 80% for training and

20% for testing. The system achieved accuracy of

98.38%, precision 97.67%, and recall of 97.85% on

YAWDD, and accuracy of 98.26%, precision 99.45%,

and recall of 98.11% on NTHUDDD, outperforming

other methods.

(Yang et al., 2024) created a driver drowsiness

detection method called VBFLLFA (Video-Based

Driver Drowsiness Detection using Facial Landmarks

and Local Facial Areas). This approach was de-

signed to detect drowsiness by analyzing key facial

features. By analyzing regions around the eyes and

mouth, the model captured movement patterns indica-

tive of drowsiness. They used the Common Spatial

Pattern (CSP) algorithm to enhance the differentia-

tion between sample classes and to lower the model’s

computational complexity when analyzing each sam-

ple. They have designed two-branch multi-head at-

tention (TB-MHA) to extract features. The Multi-

layer Perceptron was used to make the ﬁnal classiﬁca-

tion of drowsiness. The authors have used three dif-

ferent datasets, the Video-Based Driver Drowsiness

Detection (VBDDD) dataset speciﬁcally created for

Automatic Detection of the Driver Distractions Based on the Analysis of Face Videos

543

this study, the Yawning Detection Dataset (YawDD),

and the National Taiwan University Driver Drowsi-

ness Detection (NTHU-DDD). The proposed method-

ology achieved average accuracy of 88.37%, preci-

sion of 0.90, recall of 0.92, and F1-score of 0.91 .

(Cichocka and Ruminski, 2024) created a method

for the detection of drowsiness based on analysis of

Mouth Aspect Ratio (MAR) and Eye Aspect Ratios

(EAR). Fristly Haar’s cascade classiﬁers were used to

detect face and region of eyes. Convolutonal Neural

Network has been used to perform the ﬁnal drowsi-

ness classiﬁcation. They have trained the model using

MRL Eye Dataset which contains images of closed

and opened eyes. Then the model was tested on

the Drowsiness dataset (Perumandla, 2020) achieving

precision, recall, accuracy, and F1-score of 94%.

2.2 Other Machine Learning Methods

(Mandal et al., 2016) developed a bus driver monitor-

ing system with seven modules: head-shoulder detec-

tion, face detection, eye detection, eye openness esti-

mation, fusion, drowsiness measure (PERCLOS), and

fatigue level classiﬁcation. The system uses existing

in-vehicle cameras, making extra hardware unneces-

sary. It ﬁrst detects the driver’s head and shoulders,

and then uses two models to detect the face. Eye

detection was followed by estimation of eye open-

ness and ﬁnally PERCLOS was calculated to measure

drowsiness. The algorithm was trained on real-world

bus driver videos, with simulations involving 23 par-

ticipants, including those wearing glasses, under var-

ious lighting conditions.

(Saradadevi and Bajaj, 2008) developed a fatigue

detection system based on analysis of mouth and

yawning. Using a Viola-Jones classiﬁer, they track

the driver’s mouth and then apply SVM to classify

yawning. From over 1,000 collected images, they

trained on yawning and normal images, achieving ac-

curacy of 86% for normal images and 81% for yawn-

ing.

(Bakheet and Al-Hamadi, 2021) developed a

drowsiness detection system that uses adaptive his-

togram equalization to enhance image contrast, a

Haar AdaBoost classiﬁer for face detection, and

an active shape model (ASM) to locate the eyes.

HOG features are extracted from the eyes, and a

Naïve Bayes classiﬁer predicts eye status. Tested

on the NTHU Driver Drowsiness Detection dataset,

which includes various lighting conditions and sub-

jects with bare faces, glasses, and sunglasses, the

system achieved accuracy of 85.62%, F1-score of

87.84% for drowsiness, and F1-score of 81.09% for

non-drowsiness.

(Tang and Guo, 2024) proposed a method of de-

tecting driver fatigue by using infrared cameras and

classiﬁcation using Yolov8n + transfer learning. Due

to the lack of a publicly available dataset of the face

made by infrared cameras, the new dataset was specif-

ically developed for this study. In each frame, the

model classiﬁes whether the mouth and eyes were

closed or opened. The authors implemented the

method of fatigue detection based on Percentage of

Mouth Openness (POM) and Percentage of Eye Clo-

sure (PERCLOS). They were checking the 60 seconds

of the video and checking if the POM and PERCLOS

exceeded their thresholds. The method achieved ac-

curacy of 98% of fatigue detection.

3 MAR, LEAR, REAR FACTORS

Figure 1: Mouth landmarks.

Based on the preliminary research, it was decided

in our further experiments to use Mouth Aspect

Ratio (MAR) and Left/Right Eye Aspect Ratios

(LEAR/REAR) to classify the mouth and eyes open-

ness. MAR is calculated by summing the distance

between two points of the inner lips and the distance

between two points of the outer lips. Then the sum is

divided by the distance between mouth corners multi-

plied by 2 (Figures 1 and 2).

MAR =

∥m

− m

∥ + ∥m

− m

∥

2 · ∥m

− m

∥

(1)

where:

• MAR - mouth aspect ratio,

• m

...m

- mouth landmarks position.

The LEAR (analogously REAR) factors are cal-

VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications

544

culated in similar way as MAR.

LEAR =

∥e

− e

∥ + ∥e

− e

∥

2 · ∥e

− e

∥

(2)

where:

• LEAR - Left eye aspect ratio,

• e

...e

- Eye landmarks position.

Figure 2: Left eye landmarks.

Multiple tests have been performed to analyze

the Mouth Aspect Ratio and Eye Aspect Ratio while

performing different actions. The tests consisted of

capturing the frames from the desktop camera and

simulating different scenarios such as talking, eat-

ing, yawning, blinking, and normal state. Perform-

ing these tests allowed us to estimate MAR and EAR

thresholds. It was discovered that a person had closed

their eyes when the EAR was smaller or equal to 0.19,

therefore this value was selected as the EAR thresh-

old. A more challenging threshold estimation was

for the mouth aspect ratio, as the mouth aspect ratio

must be considered when the person is talking, eat-

ing, and yawning. All of these actions involve mouth

openness, making the MAR value higher. It has been

found that when the person is talking or eating the

MAR does not exceed 0.4, therefore this value was

set as the MAR threshold.

4 PERCLOS AND POM FACTORS

Fatigue detection checks whether the eye and mouth

are closed or open. Since constant blinking and fre-

quent yawning may suggest a fatigued state, the sys-

tem checks the 120 seconds and analyzes the frequen-

cies of eye closure and mouth openness. Each frame

comes with some time interval. This time interval

is then used to check if the PERCLOS or the POM

thresholds are exceeded or not. The Percentage of

Eye Closure (PERCLOS) checks the eyes state on

each frame and assigns the check result to it. After

that the system checks if the PERCLOS threshold has

been exceeded by summing all time intervals of the

frames where both eyes were closed. If the sum ex-

ceeds the PERCLOS threshold, the system will alert

about the fatigue state.

A similar logic was assigned to calculate the

Percentage of Mouth Openness (POM). The system

checks all frames in a given time period (120 seconds)

and sums all time intervals of the frames where the

mouth was opened (which may suggest yawning). If

the sum exceeds the POM threshold the system will

alert about the fatigue state.

5 PROPOSED APPROACH

Figure 3: Diagram of the drowsiness detection algorithm.

The diagram (Figure 3) shows how the drowsiness de-

Automatic Detection of the Driver Distractions Based on the Analysis of Face Videos

545

tection algorithm works. Two additional image pro-

cessing steps are performed after capturing the frame

from the camera. Firstly, the brightness of the image

is adjusted by using gamma correction:

out

= V

(3)

The γ is set to

2.5

(Setiawan et al., 2022). The

Gaussian blur is then used to smooth the image and

reduce the noise. Gaussian blur is based on the Gaus-

sian function that determines the weights of the blur

effect (Desai et al., 2020):

G(x, y) =

2πσ

−

2σ

(4)

where:

• x, y - coordinates of a pixel relative to the center

of the blur area,

• σ - standard deviation, which controls the spread

of the blur.

Those two additional image processing steps help

the frontal face detector to detect faces in low-light

images. Then, the Dlib package is used on each frame

to detect the face and key landmarks such as mouth

points and eye points. The authors have used the

shape-predictor-68-face-landmarks model which was

trained on iBUG 300-W face landmark dataset. The

dataset consists of several thousand labeled images

covering a wide range of ethnicities, ages, and gen-

ders (Kazemi and Sullivan, 2014).

Next the mouth aspect ratio (MAR), left eye as-

pect ratio (LEAR) and right eye aspect ratio (REAR)

are calculated. The openness or closure was deter-

mined by the threshold, if AR exceeded the given

threshold the mouth/eye was classiﬁed as open, oth-

erwise, it was classiﬁed as closed. Then we calculate

two factors PERCLOS and POM. Having these values

we can evaluate if the person is fatigued or not.

6 TESTING

To thoroughly examine the system’s performance un-

der various conditions, several tests have been de-

signed. A series of six videos have been created, each

lasting seven minutes, recorded under different light-

ing conditions, with and without corrective glasses.

These videos are created to simulate real-world driv-

ing scenarios and to test the system’s ability to detect

the driver’s fatigue state. The ﬁrst video is recorded

in normal daylight conditions, providing a baseline

scenario where the lighting is consistent and natural.

This video is intended to test the functionality of the

system under optimal driving conditions. The second

video is recorded under similar daylight conditions

but with the driver wearing corrective glasses. This

setup aims to evaluate whether the glasses have any

impact on the system’s performance and its ability to

detect fatigue. The third video shifts to a more chal-

lenging environment by being recorded under night-

light conditions. This scenario tests the system’s per-

formance in low-light situations, which are common

during nighttime driving and can signiﬁcantly affect

the detection algorithms. The fourth video is recorded

under nightlight conditions with additional complex-

ity: a blinking bright screen is used to imitate the ef-

fect of passing car lights. This setup is designed to

test the system’s robustness in handling sudden and

intermittent light changes, which can be distracting

and may affect the drowsiness detection. The ﬁfth

video is recorded in normal daylight conditions, but

the driver is eating. This scenario tests whether the

proposed methodology is capable of distinguishing

yawning from chewing food. The sixth video is also

recorded in normal daylight conditions, but the driver

is talking. This setup is designed to check the ability

to differentiate between conversation and yawning.

Each of these videos is divided into three distinct parts

to simulate different states of the driver. In the ﬁrst

part, the driver simulated a non-fatigue state, main-

taining a normal state and vigilance. This serves as

the control segment, where the system should ideally

detect no signs of fatigue. In the second part, lasting

for two minutes, the driver began to blink more fre-

quently and yawn, simulating a state of fatigue. This

is the critical segment where the system’s ability to

detect fatigue and issue alerts is tested. Finally, in the

third part, the driver returned to a normal state.

The purpose of this structured simulation is to as-

sess whether the system can accurately detect the on-

set of fatigue, alert the driver in a timely manner, and

then stop the alerts once the fatigue state is no longer

present. By doing so, the system’s efﬁciency in main-

taining driver alertness and safety can be thoroughly

evaluated. These tests are crucial to ensure that the

system can adapt to various real-world conditions and

reliably support the driver in maintaining a high level

of alertness, thus enhancing overall road safety. In

each of the charts, the red line represents the simu-

lated state, and the blue line represents the state de-

tected by the system.

6.1 Daylight Conditions Without

Corrective Glasses

The ﬁrst video provides a baseline scenario for

the testing phase, serving a controlled environment

to evaluate the system’s performance. This video

VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications

546

Figure 4: Chart showing detected and simulated state for

video with daylight conditions without corrective glasses.

was recorded in natural daylight without corrective

glasses. This scenario ensures minimal interference

from external variables that could affect the system’s

accuracy. Figure 4 shows the simulation of the fa-

tigue state which started at 180 seconds of the video

and stopped at 300 seconds. The system detected the

fatigue state at 236 seconds and stopped detecting it

at 352 seconds of the video.

6.2 Daylight Conditions with Corrective

Glasses

Figure 5: Chart showing detected and simulated state for

video with daylight conditions with corrective glasses

The second video introduces a new variable to the

baseline scenario by adding corrective glasses. This

is designed to evaluate the system’s ability to detect

fatigue when visual obstructions or distortions caused

by the glasses are present. Figure 5 shows a graphical

representation of the result, similarly, as before the fa-

tigue state has been simulated after third minute and

lasts for the next two minutes. The system detected

a fatigue state at 244 seconds and stopped detecting

it at 415 seconds of the video. Even though the sys-

tem detected the fatigue and returned to normal, the

duration was extended. Comparing these results with

the baseline scenario suggests that corrective glasses

may increase the system’s detection latency and false-

positive duration.

6.3 Daylight Conversing

Figure 6: Chart showing detected and simulated state for

video with daylight conditions while conversing.

The third video was also taken in normal daylight

conditions but the subject was talking with another

person. Figure 6 shows a graphical result of this sce-

nario. The video is three minutes long, the fatigue

state was not simulated, and the system has not de-

tected it either. However, despite the physical move-

ment of the mouth during the conversation, the system

did not determine that such activity met its threshold

to recognize a fatigue state. This indicates the ability

to differentiate between ordinary facial activities and

fatigue-related indicators.

6.4 Daylight Eating

Figure 7: Chart showing detected and simulated state for

video with daylight conditions while eating.

The fourth video was taken in daylight conditions

and the subject was eating for three minutes without

simulating the fatigue state. This scenario checks if

the system detects such activity as fatigue. Figure 7

Automatic Detection of the Driver Distractions Based on the Analysis of Face Videos

547

shows a graphical representation of the detection re-

sult. Those ﬁndings demonstrate that the system oper-

ates effectively in this real-world scenario, as it accu-

rately retained from detecting any fatigue state during

this activity. Even though eating involves noticeable

mouth movements, the system did not classify such

activity as a fatigue state. This highlights the sys-

tem’s ability to distinguish between context-speciﬁc

facial movements and the physiological signs associ-

ated with fatigue such as yawning. The system proved

once more its robustness and reliability in a real-world

scenario.

6.5 Nightlight Conditions

Figure 8: Chart showing detected and simulated state for

video with nightlight condition.

This video was recorded in nightlight conditions and

the subject did not wear corrective glasses during

recording to ensure that the system was tested without

interference from reﬂective surfaces. Figure 8 shows

the detected and simulated state from the recording.

Similarly, as before when fatigue was simulated, the

subject started the simulation after third minute and

ﬁnished the simulation after two minutes. The system

detected a fatigue state at 236 seconds and stopped

detecting at 339 seconds of the video. This scenario

demonstrates the system’s capability to detect fatigue

states even in challenging lighting conditions. These

ﬁndings underscore the robustness of the system in

maintaining reliable performance in low-light envi-

ronments. This highlights its potential utility in night-

time scenarios without the usage of infrared cameras.

6.6 Nightlight Conditions with Blinking

Screen

This video was recorded in nightlight conditions and

the subject did not wear corrective glasses. This sce-

nario introduces a blinking screen to simulate the ef-

fect of passing car lights. The aim was to replicate

Figure 9: Chart showing detected and simulated state for

video with nightlight condition with blinking light.

a common-world scenario such as driving at night,

where intermittent light sources may inﬂuence visual

clarity and fatigue detection. Figure 9 shows a graph-

ical result of the detection. The fatigue simulation

started in the third minute and lasted for the next two

minutes. The system detected a fatigue state at 218

seconds of the video and stopped detecting it at 369

seconds. The accurate detection of fatigue under these

conditions demonstrates the robustness and reliabil-

ity of the system. Importantly, the blinking screen

did not cause any false positives in the accuracy of

the system, indicating its resilience to environmental

lighting ﬂuctuations. This result showcases the sys-

tem’s potential in scenarios where users are exposed

to different lighting conditions.

6.7 General Results

Table 1: Detection results for different test scenarios.

Table 1 shows results for different test scenarios with

simulation duration and fatigue duration. In tests 6.1,

6.2, 6.5, and 6.6 the blue line (fatigue detection) is

shifted to the right. This is expected behavior since

the system checks the 120 seconds and then examines

whether the threshold has been exceeded. The same

applies to the time after the simulation of the fatigue

state when the system still detects drowsiness even

when it was not simulated. This phenomenon results

from the inertia of the detection process.

VISAPP 2025 - 20th International Conference on Computer Vision Theory and Applications

548

7 CONCLUSIONS

The results indicate that the system performs effec-

tively in both daylight and nightlight scenarios, accu-

rately identifying fatigue states when they were simu-

lated. These results suggest that the method is robust

and reliable, even in the presence of corrective eye-

wear. Importantly, the system also demonstrated its

ability to ignore non-fatigue activities, such as con-

versing and eating, under daylight conditions, further

conﬁrming its accuracy and relevance in real-world

driving scenarios. Nightlight conditions presented a

more challenging environment, yet the system still

detected fatigue with similar accuracy, even with the

presence of ﬂuctuating light sources, such as a blink-

ing screen simulating passing car lights. These ﬁnd-

ings show that the method holds promise for practical

applications in driver monitoring systems, especially

in varying environmental conditions.

In conclusion, this fatigue detection method has

the potential to signiﬁcantly improve road safety by

providing a reliable real-time solution to identify fa-

tigued drivers. Further research and testing can be

conducted to reﬁne the response times and adapt-

ability of the system to other driving scenarios, but

the results indicate a strong foundation for future de-

velopment. In addition, the authors plan to conduct

broad research to compare the presented method with

other existing SOTA methods, ensuring a comprehen-

sive evaluation of its performance and potential ad-

vantages. Furthermore, the authors intend to test the

performance of the system under dynamic lighting

changes to assess its robustness and reliability in vary-

ing environmental conditions.

REFERENCES

Bakheet, S. and Al-Hamadi, A. (2021). A framework for in-

stantaneous driver drowsiness detection based on im-

proved hog features and naïve bayesian classiﬁcation.

Brain Sciences, 11(2):240.

Chirra, V. R. R., Uyyala, S. R., and Kolli, V. K. K.

(2019). Deep cnn: A machine learning approach for

driver drowsiness detection based on eye state. Rev.

d’Intelligence Artif., 33(6):461–466.

Cichocka, S. and Ruminski, J. (2024). Driver fatigue detec-

tion method based on facial image analysis. In 2024

16th International Conference on Human System In-

teraction (HSI), pages 1–6. IEEE.

Desai, B., Kushwaha, U., Jha, S., and NMIMS, M. (2020).

Image ﬁltering-techniques algorithms and applica-

tions. Applied GIS, 7(11):970–975.

Dwivedi, K., Kumar, B., and Sethi, A. (2014). Drowsy

driver detection using representation learning. In 2014

IEEE International Advance Computing Conference

(IACC), pages 995–999. IEEE.

European Parliament and Council (2019). Regulation (EU)

2019/2144 of the European Parliament and of the

Council of 27 November 2019. https://eur-lex.eur

opa.eu/eli/reg/2019/2144/oj. Accessed: 2024-11-06.

Fu, B., Boutros, F., Lin, C.-T., and Damer, N. (2024). A

survey on drowsiness detection – modern applications

and methods. IEEE Transactions on Intelligent Vehi-

cles, doi: 10.1109/TIV.2024.3395889.

Kazemi, V. and Sullivan, J. (2014). One millisecond face

alignment with an ensemble of regression trees. In

Proceedings of the IEEE conference on computer vi-

sion and pattern recognition, pages 1867–1874.

Laxton, V., Oliveira, E., Stuttard, N., Avis, A., and Hosten,

P. (2022). Technical support to assess the upgrades

necessary to the advanced driver distraction warning

systems. Final report. European Commission, Brus-

sels. https://trl.co.uk/uploads/trl/documents/XPR118

---Technical-support-to-assess-the-upgrades-neces

sary-to-the-advanced-driver-distraction-warning-sys

tems-(6).pdf.

Mandal, B., Li, L., Wang, G. S., and Lin, J. (2016). Towards

detection of bus driver fatigue based on robust visual

analysis of eye state. IEEE Transactions on Intelligent

Transportation Systems, 18(3):545–557.

Perumandla, D. (2020). Drowsiness dataset, Kaggle.

https://www.kaggle.com/datasets/dheerajperumandla/

drowsiness-dataset/data. Accessed: 2024-11-11.

Rajkar, A., Kulkarni, N., and Raut, A. (2022). Driver

drowsiness detection using deep learning. In Applied

Information Processing Systems: Proceedings of IC-

CET 2021, pages 73–82. Springer.

Saradadevi, M. and Bajaj, P. (2008). Driver fatigue de-

tection using mouth and yawning analysis. Interna-

tional journal of Computer science and network secu-

rity, 8(6):183–188.

Setiawan, W., Suhadi, M. M., Pramudita, Y. D., et al.

(2022). Histopathology of lung cancer classi-

ﬁcation using convolutional neural network with

gamma correction. Commun. Math. Biol. Neurosci.,

2022:Article–ID.

Tang, X.-X. and Guo, P.-Y. (2024). Fatigue driving de-

tection methods based on drivers wearing sunglasses.

IEEE Access.

Vijaypriya, V. and Uma, M. (2023). Facial feature-based

drowsiness detection with multi-scale convolutional

neural network. IEEE Access, 11:63417–63429.

Yang, L., Yang, H., Wei, H., Hu, Z., and Lv, C. (2024).

Video-based driver drowsiness detection with opti-

mised utilization of key facial features. IEEE Trans-

actions on Intelligent Transportation Systems.

Zhao, Z., Zhou, N., Zhang, L., Yan, H., Xu, Y., and Zhang,

Z. (2020). Driver fatigue detection based on convolu-

tional neural networks using em-cnn. Computational

intelligence and neuroscience, 2020(1):7251280.

Automatic Detection of the Driver Distractions Based on the Analysis of Face Videos

549