ITM Web of Conferences 54, 01011 (2023) https://doi.org/10.

1051/itmconf/20235401011
I3CS-2023

Enhanced Driver Drowsiness Detection using Deep

Learning
Dipender Singh1* and Avtar Singh2
1Department of Computer Science and Engineering, Dr B R Ambedkar National Institute of Technology, Jalandhar, 144011,
Punjab, India
2
Department of Computer Science and Engineering, Dr B R Ambedkar National Institute of Technology, Jalandhar, 144011,
Punjab, India

Abstract- The primary reason for road accidents is drowsiness reported by National Highway
Traffic Safety Administration (NHTSA). To overcome this issue, researchers have proposed and
implemented various methods based on driver behaviour and vehicle movements. Vehicle-based
methods often rely on a set of predetermined parameters to detect drowsiness, such as changes in
steering wheel angle or lane deviation. However, these parameters may not always accurately reflect
a driver's level of alertness. Therefore, it is essential to develop an effective approach for driver
drowsiness detection. Deep learning techniques such as convolutional neural networks (CNN) are
structured solutions to detect drowsiness based on drivers' facial features. The proposed approach
based on CNN focuses on the eyes and mouth region using the nose as a central point. CNN is
operated with rectified linear activation function (ReLU) which gives 94.95% accuracy as compared
to existing methods even in different situations namely low light, different angles, and transparent
glasses.

Keywords- Convolutional Neural Network (CNN), Electrooculography (EOG), and Rectified

linear activation function (ReLU).

1 INTRODUCTION
One of the primary causes of road accidents is driver drowsiness [1]. Regularly driving for very long period,
drivers are certainly exhausted, which leads to drivers being tired and sleepy. After many researches, it was found
that the main reason for road accidents is drowsiness. Different countries have distinct reports. From all reports,
it is found that driver fatigue is the main cause of accidents. Developing a method that determine driver fatigue is
a challenge. According to the report of "Ministry of Road Transport and Highways" there were 16231 accidents
in India in 2022. Due to the increasing number of vehicles, the number of accidents is increasing day by day.
Numerous vehicles are driving at night, including heavy equipment trucks. Drive these kinds of vehicles
continuously for such a long period of time become more prone to these drastic conditions.
During last two decades various methods including GPS systems that detects when a driver begins to drift out
of their lane due to inattention. Some other techniques used electrocardiographic equipment, which measure the
driver's heart rate to determine the driver's sleepiness. One of the main goals of intelligent transportation systems
(ITS) is to increase people's well-being and minimize the number of accidents. One of the main factors in
accidents, especially in gridlock roads is the drowsiness and monotony of drivers. The ability to make decisions
and the awareness to drive a vehicle is reduced due to drowsiness. According to some research, drivers feel drowsy
after short period of driving, most particularly in the early afternoon, after and late at night, when the driver's
tiredness is high and the state of mind is far from normal. Moreover, hypnotic substances and alcohol use cause
the driver to lose awareness. In other Nations, completely different types of accident stats were reported, which
were caused by driver’s distraction. The most common cause of two hundred accidents is lack of concern and the
transitory condition of the driver. Studies show that the use of a drowsiness detection system significantly reduces
the number of accidents by at least 20% [1].In order to detect the driver's drowsiness, a series of facial images are
captured in which eye blink and head movements are monitored. Facial pattern exploration can be a fashionable
area of analysis that has various applications such as facial recognition, security systems for people authentication,
and virtual tools.
Proposed approach focuses on the localization of the driver’s eyes and mouth which is achieved by capturing
a complete representation of the face from the data set that includes 1400 images taken from Kaggle and then
detecting the position of the eyes and mouth using a Convolution neural network. After detecting the desired
regions, the method detects whether the driver is drowsy or not based on eye and mouth movements. Develop a

* Corresponding author : [email protected]

© The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons
Attribution License 4.0 (https://creativecommons.org/licenses/by/4.0/).
ITM Web of Conferences 54, 01011 (2023) https://doi.org/10.1051/itmconf/20235401011
I3CS-2023

CNN-based system that can accurately detect driver drowsiness in the data set which includes 1400 images of
different people with different conditions. The system should be able to identify common signs of drowsiness,
such as heavy eyelids, yawning, and nodding off, and alert the driver to take a break. The need is to design a
system that works under various lighting conditions that must be robust enough to handle variations in head
movements and facial expressions. The paper is divided into subparts which include the previous works,
limitations of the present system, methodology and experimental results.

2 Previous Works
This section followed the various recent works to achieve the better results in drowsiness detection.

Sukrit et al. in [2] have used the most commonly used technique which is the Eye Aspect Ratio (EAR) and Eye
Closure Ratio (ECR). Same accuracy for people wearing spectacles. Useful in conditions when drivers drive for
longer distances. Adaptive thresholding varies greatly from person to person. So, In [2] Deng and Wu in [3] have
used video Images that were used to detect Blinking, yawning and duration of eye closure. They drafted an
identification technique for facial regions which is based on the Golden Ratio of 68 main points. There were not
any publicly available image-based driver sleepiness recognition datasets. algorithm works with an accuracy
reaching 95% when the Euclidean distance is within 20 pixels. Average accuracy regardless of environmental
conditions is approximately 92%. The limitation is that it requires a high-performance machine with heavy
processors and good RAM.
An artificial Neural Network is used in [4]. Focusing more on EEG (Electroencephalography) and EOG signals
measurement and image classification. High complexity using devices Signals. Works effectively in bad
conditions. The problem is Collecting data from signaling devices is difficult to handle these complex devices. In
[5] an important ensemble machine learning is used based on hybrid sensing. Making observations based on
driving performance using counterfeit and driver observation systems. It has a high accuracy of 82.4% using
driving. The number of candidates who participated was not sufficient and all were males in their 20s.Anchan and
Saritha have used Image Processing Algorithms to detect the face, position of the head, and Blinking of the eye[6].
This system is available as an android application. So, there is no extra cost. It has a low accuracy of 81% and
thus is unreliable.
Yan et al. [7] have used face detection systems to detect dullness and lethargy using the Percentage of eye
closure, eyelid separation and eye shutdown rate. The eyelid separation and percentage of eye closure are used to
find how much a person is already exhausted and the eye separation is used for the detection of diversions. This
method focuses on three different parts of the face, thus providing more options for detecting fatigue and
distraction. The limitations of this method are unable to measure fatigue and dizziness. Therefore, object
evaluation is not possible using this directly. PERCLOS (Percentage of Eye Closure) and Grayscale Image
Processing are used for drowsiness detection by kuo and Fan-Lin. It uses grey-scale images which greatly reduces
the memory space required to store images compared to RGB images. Accuracy is high, more than 90% when the
driver’s eyes are open, and the proportion is greater than 50%. The system requires additional calculation steps.
The publication[8] suggests using blink patterns and horizontally symmetrical eye features to detect sleepiness.
Detects eye blinks with over 90% accuracy and 1% false positives. It has an advantage over systems that use
statistical information from past frames, because this system is not dependent on changes in head position, it works
on the same time frame. It only provides results for blink detection as there is no common database to compare its
result for sleepiness. Its performance is affected by the presence of glasses. Electrooculography is proposed by
Parisa et al. in [9]. It is based on detecting eye blinks, and horizontal symmetry features of eyes. For saccade
detection, the proposed algorithm has false drowsy rate less than 5% and true positive greater than 80%. The eye
blink discernment using this suggested algorithm will perform better than the normal median filter by at least 20%.
In this system, 3 means clustering is required to prevent confusion of driver’s reduced eyelash drops including any
eye movements which suspect drowsiness.
In [10] has used various sensors including heart rate, breathing and visual signal for the act. It has an accuracy
of 93% and it focuses on four different types of distractions. It was an analysis for detecting different types of
distractions but the driver was not alerted when he/she got distracted. Schwarz et al [11] have used NADS
(National Advanced Driving simulator). The models are useful in detecting low levels also. Eyelash Drops are
predictive microsleeps.[12] Discusses drowsiness of the operator using Image Processing which includes sensing
using driver operation, physiological factors and vehicle response. This system can handle various conditions such
as light changes, shadows, reflections, etc. The system also works regardless of the texture and color of the face.
During intensive use, water vapor can get on the used sensors, which leads to a reduction in the utilization capacity.
In [13] Proposed an Eye Aspect Ratio to detect features of the user from various angles to drowsiness detection.
It shows good results in Rainy Weather, Wear glasses (transparent) with bright light and dim light. The method
fails to detect drowsiness under some conditions like Wear sunglasses and Night drive.

2
ITM Web of Conferences 54, 01011 (2023) https://doi.org/10.1051/itmconf/20235401011
I3CS-2023

Table 1. Comparative analysis of various techniques

Ref Problem Need Technique Used Parameters Solution Environment Future

Direction
[3] Purposed a To detect Optimizes the Detect the eyes Developed a Data Set is Data set
DriCare the driver algorithm KCF with and mouth facial region generated with should be
System that drowsines the help of purposed movements with detection the help of the based on
detects the s with Multi convolutional the help of the technique different team different
yawning, eye more Neural Network. Dlib Library based on 68 members to public in
closure tracking essential collect the different
duration and accuracy. components dataset from the conditions.
blinking with with92%accu vehicle camera
the help of the racy. in same
video images. conditions.
[2] Develop a To Eye Closure Ratio Detection of the System works Method based Adaptive
system(applica provide and Eye Aspect face with the with almost on the real-time thresholdin
tion) install on the more Ratio is used detect help of linear same detection of the g varies
the driver car accuracy the driver SVM and accuracy with driver’s face greatly
to track the in drowsiness. histogram spectacles with the help of from
eyes movement different and useful in the application person to
for driver conditions conditions installed on the person.
drowsiness. . when drivers system. Thresholdin
drive for long g based on
distances. the other
criteria
rather than
sleep count

[5] Detection Classified For the Combination of System works Driver is System is
system based the Physiological the three on every field covered with expensive
on the factors: driver’s signals: techniques: using sensors on body using these
vehicle, alert state electroencephalogra ECG and EEG combination and methods.
physiological mainly m and for the of the surrounding. System
indicators and slightly electrocardiogram, physiological technologies must be
behavioural drowsy driving simulators signals, with accuracy adaptable
using hybrid state for for behavioural simulators for of 82.4%. for
sensing. early indices. driver behaviour everyone.
drowsy and vehicle
detection. classification
using sensors.
[4] Three types of Detection Sensors of EEG and Detection of Work Sensors and Collecting
methods of driver EOG fixed for the brain, eyes efficiency of camera used for data and
introduced for drowsines brain and eyes activity done system is physiological handling
driver s using activity. ANN is using sensors good in bad signals and EEG and
drowsiness: different used for analysis of fixed. Images conditions. driver’s images. EOG
EEG, EOG and methods the images. analyse using signaling
image for better ANN. devices is
analyzation. results. difficult.
Artificial
Neutral
networks used
to analyse the
eyes state.
[14] Driver’s eyes Detection Two methods are Detect eyes and AdaBoost Image is Accuracy
and mouth of drowsy introduced: mouth location. algorithm captured using of method
location find using state Percentage and used for face camera and is less in
using their of eyes Time That Mouth is detection carrying out for bad
geometric and mouth Open (PTMIO), including further process. conditions
features which for better Percentage and edge and line and
helps in accuracy. Time that Eyelids features different
drowsiness cover the Pupils which leads people.
detection. (PATECP). to 84%
accuracy.
[11] Detection of For NADS (National Detection of Models Sensors fixed in Blinks have
driver analysing Advanced vehicle-based performed different been shown
drowsiness level of Driving simulator) activities like well in locations of to be
using sensors drowsines and signals from the handle identifying vehicle. predictive
for monitoring s from vehicle's sensors movement, mild tiredness of
driver low level integrated with the phone use and from microsleeps
movements. to high driver monitoring distractions. moderate to .
level. system. severe
drowsiness.

3
ITM Web of Conferences 54, 01011 (2023) https://doi.org/10.1051/itmconf/20235401011
I3CS-2023

Shen et al. in [14] cites a user's sleepiness based on multiple facial features using the percentage and time of
the eyelids covering the pupil, the percentage of time the mouth is visible to the camera. Detects sleepiness using
eyes and mouth. The disadvantage is that the extraction of targets is difficult in some cases and sometimes the
behavior of the target is not detectable. Khosro et al. [15] suggested computing the drowsiness of operators using
Haar-like features and AdaBoost classifiers. Local Binary Pattern is employed to fetch characteristics of eye
features. Mohit Dua et al.[16] suggested combination of different deep learning models like Rest Net and Alex net
using RGB videos as input and detect drowsiness. Model uses GPU for better accuracy to detect facial expressions
and hand gestures. Model is divided into subparts and every deep learning model has their own functioning
respectively.

Some disadvantages of driver drowsiness detection systems at present include high false alarm rates, difficulty
in detecting microsleeps, and the need for calibration or individual customization. Additionally, some systems can
be intrusive or uncomfortable for the driver, and their effectiveness can be affected by environmental factors.
Present drowsiness detection systems use complex approaches and costly equipment, such as an Electrooculogram
(EOG) which monitored the movements of muscles acting on the eyes. Some methods work on face detection
using 68 facials points which store coordinates in dynamic storage. But these methods are not able to find the
accurate desired regions if the full face is not in the frame as all coordinates are correlated to each other. In recent
years most of the methods follow the Convolution neural networks (CNN)which give more accuracy but fail in
different conditions like different angles, low light, and transparent glasses. The main reason for failure is the
Working of CNN as it required different angles (0, +200, -200) to analyze the object. Angle problem in all driver-
based techniques remain same because every technique just focuses on increasing the number of hidden layers
which causes the loss of the lot of data.

3 Methodology
Main motive of this paper to provide a method which gives more accuracy in different conditions. For this, method
is divided into subparts that work efficiently. The proposed method implement on dataset includes 1400 images
taken from Kaggle which includes different conditions. The following steps are used to solve the issues of driver
drowsiness as shown in figure 1.

Fig. 1. Steps of Methodology.

Step1:
First step of methodology after loading dataset in machine is testing and training. Most common split ratio of
training and testing is 80% and 20% respectively for dataset of images more than 1000. But chances of overfitting
are very high to achieve more accuracy. To overcome this situation, k-fold cross validation method is used which
gives 1078 images as training and 3022 images for testing which eliminates overfitting chances.

Step 2:
Next Step of methodology to detect eyes and mouth region for further process. As earlier mention in paper, driver
looked at different angle so concerned regions sometimes lost. To overcome this problem, detection of eyes and
mouth region done based on nose coordinates. Some of sample figures are mentioned below for reference how
interested region is detected.

4
ITM Web of Conferences 54, 01011 (2023) https://doi.org/10.1051/itmconf/20235401011
I3CS-2023

Fig. 2. Detection of Eyes Region

The region is segmented with Convolutional Neural Network (CNN) model based on pre-trained model of human
eye, mouth and face. The trained CNN model uses the Images pre-processed by TensorFlow library. Method uses
CNN python libraries to detect eyes and mouth region without any manual image processing. The trained model
can be applied to different images with different features and values. Class feature of TensorFlow to differentiate
between closed, open and other facial expressions. We use an open-source library named OpenCV for training
and analysing images.

Step 3:

Model takes a trained Convolutional Neural Network (CNN) and converts the weights into simplex and
maxpooling layers. The Sequential model is a higher-level neural network architecture that implements sequence
learning. The model takes a sequence of zero or more random inputs and predicts a single output, which can be
used as an input to subsequent layers in the network. Model fit over a training set of 30 epochs with validation
data that are generated based on the output of that model. The validation steps are chosen to yield the best training
accuracy, while minimizing time and memory usage. The final model is then flattened, dropped outed, and trained
on the given dataset.

Fig. 3. Working of CNN with ReLU

This model has 3 convolutions, 4 max pooling and 1 flatten on top of the previous layers. The final output of this
model is a dense vector with 64 dimensions per class. In a CNN, ReLU is usually applied after the convolutional
and pooling layers. The output of the convolutional layer is passed through the ReLU activation function, which
sets all negative values to zero and leaves positive values unchanged. This helps to eliminate any negative values
that might arise in the output of the convolutional layer, which can hinder the performance of the network. The
use of ReLU with CNNs has been shown to improve the accuracy of the final results, especially in image
recognition tasks.

5
ITM Web of Conferences 54, 01011 (2023) https://doi.org/10.1051/itmconf/20235401011
I3CS-2023

4 Experimental Results

In this study, the proposed approach aimed to develope a new CNN-based driver drowsiness detection
system using a dataset of frontal face images. The dataset consisted of 1400 images with a balanced ratio of
normal images and drowsy images. The images were preprocessed to resize them to 145x145 pixels and
normalize the pixel values.
We trained and evaluated the proposed CNN model using Keras with the TensorFlow backend. The model
consisted of three convolutional layers, each followed by a max pooling layer, and two fully connected (dense)
layers. The proposed approach uses ReLU activation function for the convolutional layers and the SoftMax
activation function for the output layer. The model was trained with a batch size of 39 andfor 30 epochs,
using the Adam optimizer and categorical cross-entropy loss function. CNN model was able to achievean
accuracy of 94.95%, which is better than other existing techniques for driver drowsiness detection.

Table 2. Different Models Comparison

Model Technique Accuracy (%)

Proposed Approach Eyes and Mouth Region 94.95

CNN-based Eyes Tracking 92.3

SVM-based EEG 88.9

Random Forest Heart Rate Variability 91.2

LSTM-based Driving Behavior Analysis 87.5

Rule-Based Steering Wheel Movements 82.1

The results showed that the model achieved an accuracy of 94.95%, which is significantly better than other
techniques used for driver drowsiness detection. This high accuracy level demonstrates the effectiveness of the
improved CNN modelin detecting driver drowsiness.
The graph in figure 4 shows the accuracy of a CNN model over 30 epochs, with two labels: train accuracy and
test accuracy. The train accuracy represents the accuracy of the model on the training dataset, which is the data
that the model was trained on. Initially, the train accuracy is low, as the model is still learning and improving its
performance. As the model is trained on more data, the train accuracy improves, reaching a plateau where it no
longer improves significantly.
The test accuracy, on the other hand, represents the accuracy of the model on a separate dataset that it has not
seen before. This is an important metric, as it indicates how well the model will perform on new, unseen data.
Initially, the test accuracyis low, as the model has not yet learned to generalize well to new data. The train
accuracy and test accuracy constantly increase as the number of epochs increases.

Fig. 4. Model Accuracy

Ideally, the train accuracy and test accuracy should follow a similar trend, with both increasing as the model is
trained on more data.

6
ITM Web of Conferences 54, 01011 (2023) https://doi.org/10.1051/itmconf/20235401011
I3CS-2023

The graph in figure 5 shows the model’s loss over 30 epochs, with two labels: train loss and test loss. It indicates
how well the model is fitting the data, with a lower loss indicatingbetter performance. Overall, both graphs
provide valuable insight into the performance of the CNN model over time, allowing for adjustments to be made
to improve its accuracy and generalization capabilities.


Fig. 5. Model Loss
In figure 6, the Proposed Model experimental results showed that our CNN-based driver drowsiness detection
system achieved an accuracy of 94.95% in detecting driver drowsiness, with a precision of varies from 80%
to 99% and recall of 82% to 98% for different factors including in dataset. The system also achieved an F1
score of 88% to 97% andan AUC of 0.95. The chances of model overfitting increaseas the number of
epochs increases. The change in the model remains overall same after 30 epochs. Overall results of the model
are better in different conditions.

Fig. 6. Model Performance Summary

5 Conclusion
To overcome the problem of road accidents that significantlyhappened due to driver drowsiness. The proposed
model focuses on closed and open eyes with yawning cases whichare major factors in detection. The main aim
is to detect drowsiness in different conditions for better accuracy. Overall, the model experimental results
demonstrate the potential of using CNN for driver drowsiness detection, which can befurther improved by using
larger datasets and more advanced CNN architectures. The experimental results also highlight the importance of
developing robust and reliable driver drowsinessdetection systems to improve road safety and prevent accidents.

References

[1] K. Kumari, “Review on Drowsy Driving: Becoming Dangerous Problem,” International Journal of Science and Research,
(2014).
[2] S. Mehta, S. Dadhich, S. Gumber, and A. J. Bhatt, “Real-Time Driver Drowsiness Detection System Using Eye Aspect
Ratio and Eye Closure Ratio Fatigue Detection Non-Intrusive Methods Driver monitoring system,” (2019).
[3] W. Deng and R. Wu, “Real-Time Driver-Drowsiness Detection System Using Facial Features,” IEEE Access, vol. 7, pp.
118727–118738, (2019).
[4] T. Vesselenyi, S. Moca, A. Rus, T. Mitran, and B. Tǎtaru, “Driver drowsiness detection using ANN image processing,”
in IOP Conference Series: Materials Science and Engineering, vol. 252, no. 1, (2017).

7
ITM Web of Conferences 54, 01011 (2023) https://doi.org/10.1051/itmconf/20235401011
I3CS-2023

[5] J. Gwak, A. Hirao, and M. Shino, “An investigation of early detection of driver drowsiness using ensemble machine
learning based on hybrid sensing,” Applied Sciences, vol. 10, no. 8, (2020).
[6] S. L.R. and S. Anchan, “Human Drowsiness Detection System,” International Journal of Engineering and Advanced
Technology, vol. 8, no. , pp. 316–319, Apr. (2019).
[7] J. J. Yan, H. H. Kuo, Y. F. Lin, and T. L. Liao, “Real-time driver drowsiness detection system based on PERCLOS and
grayscale image processing,” IEEE International Symposium on Computer, Consumer and Control, IS3C ,pp. 243–
246,(2016).
[8] T. Danisman, I. M. Bilasco, C. Djeraba, and N. Ihaddadene, “Drowsy driver detection system using eye blink patterns,”
in 2010 International Conference on Machine and Web Intelligence, ICMWI 2010 - Proceedings, pp. 230–233,(2010).
[9] P. Ebrahim, W. Stolzmann, and B. Yang, “Eye movement detection for assessing driver drowsiness by
electrooculography,” in Proceedings - 2013 IEEE International Conference on Systems, Man, and Cybernetics, SMC
2013,pp. 4142–4148,(2013).
[10] M. Gjoreski, M. Z. Gams, M. Luštrek, P. Genc, J. U. Garbas, and T. Hassan, “Machine Learning and End-to-End Deep
Learning for Monitoring Driver Distractions from Physiological and Visual Signals,” IEEE Access, vol. 8, pp. 70590–
70603, (2020).
[11] C. Schwarz, J. Gaspar, T. Miller, and R. Yousefian, “The detection of drowsiness using a driver monitoring system,”
Traffic Inj Prev, vol. 20, no. sup1, pp. S157–S161, (2019).
[12] S. Himani Parmar, M. Jajal, and Y. Priyanka Brijbhan Electronics, “Drowsy Driver Warning System Using Image
Processing | Drowsy Driver Warning System Using Image Processing,” IJEDR, (2019).
[13] C. B. S. Maior, M. J. das C. Moura, J. M. M. Santana, and I. D. Lins, “Real-time classification for autonomous drowsiness
detection using eye aspect ratio,” Expert Syst Appl, vol. 158, (2020).
[14] P. Wang and L. Shen, “A method of detecting driver drowsiness state based on multi-features of face,” in 2012 5th
International Congress on Image and Signal Processing, CISP 2012, pp. 1171–1175,(2012).
[15] M. Sabet, R. A. Zoroofi, K. Sadeghniiat-Haghighi, and M. Sabbaghian, “A new system for driver drowsiness and
distraction detection,” in ICEE 2012 - 20th Iranian Conference on Electrical Engineering, pp. 1247–1251,(2012).
[16] M. Dua, Shakshi, R. Singla, S. Raj, and A. Jangra, “Deep CNN models-based ensemble approach to driver drowsiness
detection,” Neural Comput Appl, vol. 33, no. 8, pp. 3155–3168, (2021).
[17] S. Liu, Y. Li, C. Wen, and S. Wang, ”A Novel Driver Drowsiness Detection Method Based on Heart Rate and Its
Variability,” IEEE Access, vol. 9, pp. 114103-114114,(2021).
[18] F. Li, X. Zhao, M. Liu, and H. Han, ”Driver drowsiness detection based on multi-task learning and EEG signals,”
Computer Methods and Programs in Biomedicine, vol. 205, (2022).
[19] S. Anwar, K. Khan, A. Rehman, and M. Baig, ”A novel driver drowsi- ness detection system using non-contact vital signs
and machine learning techniques,” Biomedical Signal Processing and Control, vol. 68, (2021).
[20] C. Liu, J. Lu, Y. Zhang, and Y. Yang, ”Driver drowsiness detection using multi-task learning and a convolutional neural
network,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 4, pp. 1748-1758,(2022).
[21] H. He, Y. Zhou, X. Zhang, and J. Zhang, ”Driver drowsiness detection using multi-view convolutional neural networks,”
IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 3, pp. 1248-1258,(2022).
[22] H. Li, Z. Li, Y. Li, and Y. Li, ”A novel driver drowsiness detection method based on deep learning and physiological
signals,” Measurement, vol. 190,(2022).
[23] N. K. Ibraheem and S. S. Yahya, “Real-time driver drowsiness detection system using machine learning algorithms,” SN
Computer Science, vol. 2, no. 3,(2021).
[24] A. Khan, A. Qayyum, and H. Bokhari, “Driver drowsiness detection system using deep learning and handcrafted
features,” IEEE Access, vol. 9, pp. 64416–64427,(2021).
[25] A. Sharma and R. Pandey, “Efficient driver drowsiness detection using transfer learning with deep neural networks,”
Computers Electrical Engineering, vol. 91,(2021).
[26] S. Reddy, S. K. Arora, and S. K. Rautaray, ”Driver Drowsiness Detectionusing Deep Learning and Effective Fusion of Multi-modal Data,”
PatternRecognition Letters, vol. 154, pp. 11-20,(2022).