A New Approach to Detect Driver Distraction to Ensure Traffic Safety and Prevent Traffic Accidents: Image Processing and MCDM

Abstract

One of the factors that threaten traffic safety and cause various traffic problems is distracted drivers. Various studies have been carried out to ensure traffic safety and, accordingly, to reduce traffic accidents. This study aims to determine driver-distraction classes and detect driver violations with deep learning algorithms and decision-making methods. Different driver characteristics are included in the study by using a dataset created from five different countries. Weight classification in the range of 0–1 is used to determine the most important classes using the AHP method, and the most important 9 out of 23 classes are determined. The YOLOv8 algorithm is used to detect driver behaviors and distraction action classes. The YOLOv8 algorithm is examined according to performance-measurement criteria. According to mAP 0.5:0.95, an accuracy rate of 91.17% is obtained. In large datasets, it is seen that a successful result is obtained by using the AHP method, which is used to reduce transaction complexity, and the YOLOv8 algorithm, which is used to detect driver distraction. By detecting driver distraction, it is possible to partially avoid traffic accidents and the negative situations they create. While detecting and preventing driver distraction makes a significant contribution to traffic safety, it also provides a significant improvement in traffic accidents and traffic congestion, increasing transportation efficiency and the sustainability of cities. It also serves sustainable development goals such as energy efficiency and reducing carbon emissions.

1. Introduction

As a result of the widespread use of individual vehicles, traffic accidents are increasing day by day. According to the data announced by the World Health Organization (WHO), approximately 1.3 million people die in road traffic accidents every year and between 20 and 50 million people are injured [1]. In 2021, traffic accidents occurring worldwide were ninth place among the causes of death. It is thought that if the current situation continues, it will rise to fifth in 2030. In response to this situation, the United Nations General Assembly aims to halve deaths and injuries in traffic accidents by 2030 [1,2].

One or more factors affect the occurrence of a traffic accident. In the WHO publications, a very wide array of possibilities have been discussed for the causes of traffic accidents and the risk factors associated with them. Distraction is one of the most important human factors affecting road safety [3,4]. According to the study conducted by the AAA, it was determined that 6 out of 10 traffic accidents were caused by distraction [5]. In another study, more than 80% of traffic accidents that occur are caused by using mobile phones, texting, talking to passengers, smoking, or drinking [6]. According to the National Highway Traffic Safety Administration (NHTSA) data, driver distraction was the cause of 8% of fatal traffic accidents and 14% of injury traffic accidents in 2020 [7]. According to WHO statistics, 45% of traffic accidents that occur in a year are caused by distracted drivers [8].

There are four physiologically different types of driver distraction. These are visual, cognitive, physical, and auditory [9]. One or more of the above-mentioned situations may occur simultaneously while driving and as a result, the driver becomes distracted. The activity that causes the most distraction, both while driving and in daily life, is mobile phones. Thanks to the extraordinary and rapid developments in technology, our lives are becoming more digital every day. When mobile phone usage is examined as a smart mobile phone, it is seen that there were 6.64 billion smartphone users in the world as of 2022 [10].

Mobile phones are frequently used for many functions, regardless of location. Undoubtedly, mobile phones are frequently used while driving. Around 660,000 drivers worldwide attempt to use their phones while driving at any time of the day. The National Safety Council reports that cell phone use while driving causes 1.6 million accidents each year [11]. Approximately 390,000 injuries occur annually in traffic accidents caused by texting or using a phone while driving [12]. About 14% of all fatal traffic accidents involve some form of (pedestrian and driver) mobile phone use [13]. Texting while driving is 6 times more risky than drunk driving in terms of the possibility of causing an accident. In addition, texting increases the risk of traffic accidents between 4 and 23 times compared to the normal driving situation [14]. Although using a mobile phone is one of the biggest causes of driver distraction, there are many reasons. Actions such as smoking or drinking, talking to passengers, hair and make-up, operating multimedia systems, reaching behind, etc., cause distraction. Detecting driver distraction is crucial, as studies have shown how perceptual blindness and distraction affect driving performance [15,16]. These works highlight the risks associated with insufficient attention to activities crucial for safe driving.

Various measures should be taken to reduce the losses in traffic accidents. Considering that the distraction of drivers is one of the most effective factors in traffic accidents, this issue should be carefully considered. The aim is to partially prevent the losses in traffic accidents by reducing the distraction of drivers. In order to determine the measures that can be taken for driver distraction, the violation must first be determined. Within the scope of this study, the actions that cause distraction of drivers were obtained as a result of various survey results and research. The main 23 actions that cause distraction are obtained. The aim is to detect driver distraction by creating a new dataset in addition to the general datasets. In order to prevent errors that may occur in the creation of the dataset of 23 driving activities and to increase accuracy in deep learning algorithms, the first 9 actions that cause distraction are taken into account in the study. Information obtained from the literature, experts, and drivers was used to determine 23 factors that cause driver distraction. The analytical Hierarchy Process (AHP) method was used to obtain the importance level of these factors. The AHP method was chosen because it provides a systematic and structured approach to multi-criteria decision-making problems. This method transforms complex decision-making processes into a simple and understandable hierarchical structure, allowing us to determine the relative importance of each criterion. Therefore, the AHP method is ideal for objectively and scientifically determining the importance levels of driver-distraction factors. The AHP, one of the multi-criteria decision-making (MCDM) methods, is used in the determination of nine actions. To detect distraction, the YOLOv8 approach, one of the deep learning methods, is preferred in this study.

Another aim of the study is to contribute to sustainable transportation and development goals in terms of traffic accidents, traffic congestion, fuel consumption, carbon emissions, etc., by partially preventing driver distraction through instant warning of drivers as a result of driver-distraction detection and various penal sanctions. By reducing driver distraction, improvements in traffic flow, energy efficiency, and other environmental concerns can be achieved. Thus, a vital step is taken to create a safer and environmentally friendly traffic flow in line with sustainable transportation goals.

The main purpose of this study is to improve traffic safety by accurately detecting driver distraction and to contribute to sustainable transportation goals by reducing traffic accidents. Key achievements include the successful implementation of the AHP method to prioritize deceptive classes and the effective use of the YOLOv8 algorithm to detect these classes with high accuracy. By detecting driver distraction, traffic accidents and the negative situations they create can be partially prevented. This not only significantly contributes to traffic safety, but also improves traffic flow and reduces congestion, thereby improving transport efficiency and the sustainability of cities. In addition, these efforts are also in line with sustainable development goals such as energy efficiency and reducing carbon emissions.

This article aims to detect driver distraction using a scientific and strategic approach. To do so, a four-stage object-detection approach is applied. First, actions that cause driver distraction are determined using various surveys and research results. AHP is used to obtain the order of importance of the determined actions. Secondly, a large dataset is obtained by adding new data in addition to the public dataset. Third, a new annotated file is created for each image in the dataset, the data are annotated, and the annotated data are divided into two groups: train and validation. Finally, to train the model and detect driver distraction, the YOLOv8 algorithm is used and trained with data. As a result of the training, autonomous detection of driver distraction is carried out.

2. Literature Review

Since traffic accidents and driver distraction are a general problem, many studies have been conducted by researchers. Researchers have used various algorithms, mathematical models, and methods including verbal explanations. In this section, studies on detecting driver distraction using deep learning algorithms are presented. In addition, for the reader’s ease of reading, Table 1 contains additional information on literature reviews.

There is only one study in which the AHP method and the YOLO algorithm are used together. Asif et al. proposed disaster classifications for emergency response. Deep learning and object-recognition algorithms are used together to automate emergency response. VGG-16 and YOLO algorithms were preferred in order to analyze the images related to a disaster and to classify disaster types. The ranking process was done using the AHP method to determine the appropriate emergency response [17].

Xing et al. designed a driver activity-recognition system based on deep convolutional neural networks in their study to understand driver behaviors. Specifically, seven driving activities were determined: normal driving, right mirror control, rearview mirror control, left mirror control, using an in-car multimedia device, messaging, and answering mobile phones. Transfer learning algorithms AlexNet, GoogLeNet, and ResNet50 models were applied to reduce training costs and fine-tune pre-trained Convolutional Neural Network (CNN) models. According to the results, AlexNet 81.6%, GoogLeNet 78.6%, and ResNet50 74.9% accuracy rates were obtained. Then, the binary classification deep learning method was used to detect distraction. Accordingly, distraction was detected with an accuracy of 91.4% [18].

Alotaibi and Alotaibi have developed a set of models consisting of CNN and hierarchical recurrent neural networks for detecting distracted drivers. “StateFarm” and “AUC” datasets published on the Kaggle platform were used as the dataset. ResNet, HRNN, and Inception models were used as a single group model. The model has a ResNet block and two HRNN layers integrated with the Inception module, followed by two dense layers and a softmax classifier. It is observed that the proposed model has an accuracy rate of 96.23% and 92.36% for the “StateFarm” and “AUC” datasets, respectively [19].

Xiong et al. conducted a study to determine the distraction of drivers due to mobile phones. They proposed a two-step approach for detecting distraction. In the first step, Progressive Calibration Networks architecture, which detects and monitors the facial region, is used. In the second step, CNN architecture was preferred in order to detect the mobile phone in the determined area. The aim was to detect the presence of a mobile phone by considering the pairwise comparison situations. For this, 22,216 pictures of drivers dealing with mobile phones and 25,464 normal drivers were included in the study. The proposed approach is stated to have an accuracy rate of 96.56% [20].

Li et al. conducted a study to detect driver distraction using advanced deep-learning techniques. In the study, AlexNet, VGG-16, and ResNet-18 algorithms were applied to the open-source StateFarm, Origin, and 3MDAD datasets, and the results were compared. Accordingly, success rates of 99.92%, 100%, and 99.99% were achieved, respectively. It was observed that the ResNet-18 algorithm achieved a lower success rate in the two datasets [21].

Ezzouhri et al. performed a study on a current public dataset to determine the distraction of drivers. The key feature of the proposed solution is the extraction of the body parts of the driver by applying deep learning-based segmentation before detecting and classifying the distraction. The mask R-CNN model was applied as the segmentation method. In addition, CNN algorithms VGG-19 and Inception-V3 were preferred as classification methods. In the proposed approach, 38,327 driver images were obtained and the “AUC” dataset was used as a comparison. The results of the study show that the segmentation module significantly improves the performance in classification. An accuracy rate of 96% was obtained for the “AUC” dataset [22].

Various studies are presented in Table 1. Table 1 summarizes the applied method, number of classes, used dataset, and mAP values.

Table 1. Studies on the detection of driver distraction in the literature.

Author	Method/Algorithm	Number of Classes	Dataset	mAP
[23]	CDCNN and VGG-16	10	StateFarm	99.73%
[24]	ResNeXt-34, ResNeXt-50, VGG-16, and VGG-19	2	Author’s	92.88%
[25]	AlexNet, Inception V3, VGG-16, and 3D-CNN	10	StateFarm	98.05%
[6]	E2DR/ResNet50 and VGG-16	10	StateFarm	92%
[26]	AlexNet, VGG16, EfficientNet B0, Vanilla CNN, Modified DenseNet, and InceptionV3 + BiLSTM	10	StateFarm and AUC	99.75%
[27]	MobileNet, ResNet, Inception-v3, and DenseNet	10	StateFarm	94.3%
[28]	SqueezeNet	10	StateFarm	99.93%
[29]	VGG-16 and VGG-19	10	StateFarm	99.39%
[30]	MobileNet-SSD, YOLOv3, and Faster R-CNN	1	Author’s	99%

Table 1. Studies on the detection of driver distraction in the literature.

Author	Method/Algorithm	Number of Classes	Dataset	mAP
[23]	CDCNN and VGG-16	10	StateFarm	99.73%
[24]	ResNeXt-34, ResNeXt-50, VGG-16, and VGG-19	2	Author’s	92.88%
[25]	AlexNet, Inception V3, VGG-16, and 3D-CNN	10	StateFarm	98.05%
[6]	E2DR/ResNet50 and VGG-16	10	StateFarm	92%
[26]	AlexNet, VGG16, EfficientNet B0, Vanilla CNN, Modified DenseNet, and InceptionV3 + BiLSTM	10	StateFarm and AUC	99.75%
[27]	MobileNet, ResNet, Inception-v3, and DenseNet	10	StateFarm	94.3%
[28]	SqueezeNet	10	StateFarm	99.93%
[29]	VGG-16 and VGG-19	10	StateFarm	99.39%
[30]	MobileNet-SSD, YOLOv3, and Faster R-CNN	1	Author’s	99%

The differences between this study and the studies in the literature, and its contribution to the literature, are listed below.

The superiority of the YOLOv8 algorithm used in the study compared to other deep learning algorithms has been proven by various scientific studies [31,32]. However, the use of YOLO algorithms in large datasets is not preferred due to the high workload in annotating the dataset. This has led to the use of less reliable algorithms to escape the workload. In this study, despite the size of the dataset, each item in the dataset is annotated manually to use the correct and reliable YOLO algorithm. Thus, the difference in the study is revealed. In addition, for the first time in the field of “distraction of driver” by using the YOLO algorithm, a study is conducted with such a large dataset, thus contributing to the literature.

When the literature review is examined, generally public datasets are used in studies. In this study, a new dataset study is carried out in addition to the public datasets. Thus, the dataset is further expanded and updated. Unlike the studies in the literature, a total of 103 drivers from 5 different countries and 19 different cities are included in the study in order to reflect a global scale of the dataset. Approximately 45,000 pictures are used in the study. In this way, it is ensured that the study included various cultural driving characteristics. In addition, for the first time in the literature, a dataset that offers such a wide variety has been used.

The number of studies in which MCDM methods and deep learning algorithms are used together is very limited. In one study, a scientific approach is obtained with the use of MCDM methods. In this study, MCDM methods and deep learning algorithms are used together. For the first time in the literature, the AHP method is used for classes considered in deep learning algorithms. The use of MCDM methods and deep learning algorithms together as a package program is presented as an innovation and contribution to the literature.

3. Materials and Methods

Brief information about the dataset that formed the main structure of the study is given in the section below. After obtaining the dataset, the reasons for driver distraction are also given in this section.

3.1. Data Collection and Preparation

Public access and previously prepared StateFarm, AUC, and Amal Ezzouhri datasets are used in the study [22,33,34]. Using public access datasets, a large dataset from Egypt, India, America, and Morocco is included in the study. In addition to the public access dataset, a new dataset is created using drivers in four different cities in Turkey. The aim is to reflect different cultural driver characteristics in the study by using a large dataset. In order for the dataset to be reliable, pre-processing work is carried out for the entire dataset, and data that are not suitable for the model are removed from the dataset.

The study flow chart for detecting driver distraction is given in Figure 1. As can be seen from the figure, a six-step roadmap was taken into consideration.

As a result of various studies, survey results, and field research, 23 causes of driver distraction are determined. While some of the identified causes of distraction are very important, some of them are considered rarely seen causes. The reasons for driver distraction are given in Figure 2.

In order to determine the ranking of the causes, the AHP method, one of the MCDM methods, is preferred in this study. By using the AHP method, distraction classes to be considered in deep learning will be determined.

3.2. Analytic Hierarchy Process

AHP was introduced as a basic approach developed by Thomas L. Saaty in the late 1970s [35]. It was developed to determine the best and most consistent of multiple options compared to each other on different criteria bases. The AHP method aims to solve difficult decision-making problems by using pairwise comparison matrices. The AHP approach is a method frequently used in decision-making problems. It reaches the weight of the option/criterion by taking into account the individual’s judgments hierarchically and comparing them according to similar and common features [36].

The AHP approach has a systemic and strategic infrastructure. By applying this infrastructure to different problems, the solution is easily reached. The steps of the AHP method are listed below [37].

• Step 1. Defining the decision problem and determining the criteria/options;
• Step 2. Creating the hierarchical structure;
• Step 3. Establishment of pairwise comparison matrices and comparison of criteria.

Pairwise comparison matrices are used to compare the criteria or alternatives considered for the method. Considering that there are n elements in the AHP hierarchy, n(n − 1)/2 comparisons are made [38]. The [1,2,3,4,5,6,7,8,9] scale developed by Saaty is used in pairwise comparison matrices [39].

• Step 4. The formulation in Equation (1) is used to normalize the pairwise comparison matrices [40].

$$a_{ij}^{\prime }={\displaystyle \frac{a_{ij}}{{\sum }_{i=1}^n{a}_{ij}}}$$

(1)

• Step 5. The formulation in Equation (2) is used to calculate the priority vector [41].

$${\mathcal{w}}_i=\left({\displaystyle \frac{1}{n}}\right){\sum }_{j=1}^n{a}_{ij}^{\prime }\hspace{1em}i,j=\mathrm{1,2},\dots ,n$$

(2)

• Step 6. To test the AHP consistency, the Consistency Ratio (CR) is calculated using Equations (3) and (4) [38].

$$CR={\displaystyle \frac{CI}{RI}}$$

(3)

$$CI={\lambda }_{max}-{\displaystyle \frac{n}{n-1}}$$

(4)

$a_{ij}$: decision alternatives; ${\lambda }_{max}$: eigenvalue

The Random Index value can be calculated according to the table given in Saaty’s study for cases where the number of criteria in the hierarchical structure is up to 15 [42]. For cases with more than 15 evaluation criteria, the study of Alonso and Lamata can be referred to [43].

3.3. You Only Look Once (YOLO)

YOLO, which detects objects using convolutional neural networks, was first developed in 2015 [44]. The YOLO algorithm is constantly being developed and the YOLOv8 model was developed by ‘Ultralytics’ in the first half of 2023 [45]. With the YOLO algorithm, class probabilities and bounding boxes can be estimated by making a single scan [46]. Unlike CNN architectures that provide regional recommendations, the YOLO algorithm passes the input image once through the fully convolutional neural network. Thanks to this approach, the YOLO algorithm is faster than other models and can run in real time [47]. The original YOLO architecture consists of 24 convolutional layers, followed by two fully connected layers and a final detection layer [44]. The YOLO algorithm performs a linear regression using two fully connected layers. It uses a final detection layer to protect bounding boxes with high confidence scores. In the YOLO algorithm, the input image is divided into N*N grid cells. The grid cell containing the center of the image performs the estimation task. Each of the grid cells predicts only one object. Also, each grid cell stipulates a certain number of bounding boxes [48].

The output network of the YOLO algorithm is the vector N*N*(5B + C). When the output network is examined, it is seen that parameter B is the grid cell that estimates the bounding boxes. Each bounding box has a certain confidence score. The C parameter is defined as a set of class scores for the bounding boxes. Each of the bounding boxes consists of five items: $x,y,\mathsf{\delta },\mathrm{w},\mathrm{h}$. The previously mentioned confidence score reflects how an object is found in each of the bounding boxes and its accuracy. In addition to the classification of objects, the YOLO algorithm also determines the location of the detected object. Since it is an algorithm that is constantly being developed, it is observed that better results are obtained with each new model [49].

Performance comparisons of YOLOv8 and previous versions are made according to the COCO dataset. YOLOv8 is seen as faster in terms of some parameters and speed. In this study, the YOLOv8 algorithm is used. In addition, the model was solved with the YOLOv8 algorithm and stable results could not be obtained.

4. Results and Discussion

First, the dataset is obtained, cleaned, and ranked. Then, the AHP method is applied to determine which class will be used. After the determination of the classes, the data are annotated. Finally, training is performed with the appropriate dataset and classes. In this section, the applications of the methods used in the study are explained.

4.1. Analysis of AHP

The 23-person decision-maker team is formed to prioritize the classification of actions that affect driver distraction. The decision-maker team includes experts and academics in the field of traffic safety. The judgments of the decision-making team and the action classes in the literature are examined and included in the study. In addition, drivers are included in the team in order to be suitable for real life. Thus, The aim is to ensure the objectivity of the classes used in the study by adopting a broad perspective. In the application of the AHP method, firstly, pairwise comparison matrices are created using the Saaty scale. Since the number of classes considered is more than 15, the formulation given in Equation (5) is used while calculating the consistency ratio [43]. In addition, the consistency ratio must be less than 0.1 in order to prove the consistency of the study. The consistency ratios calculated in 23 pairwise comparison matrices created by the decision-maker team are averaged. In this study, the consistency ratio is calculated as 0.087. After obtaining the consistency ratio, the priority order of the classes is determined by taking the average of the prioritization weights obtained from the pairwise comparison matrices in order to prioritize the classes.

$$RI\left(n\right)=0.00149n^3-0.05121n^2+0.59150n-0.79124$$

(5)

The sum of the weights obtained in the AHP method should be 1. In

Table 2. Weight values of activity classes.

Drive Activity Class	1	2	3	4	5	6	7
Prioritization Weights	0.0182	0.0115	0.0096	0.0106	0.0086	0.0106	0.0010
Drive Activity Class	8	9	10	11	12	13	14
Prioritization Weights	0.0163	0.0585	0.0605	0.0681	0.0086	0.0576	0.0393
Drive Activity Class	15	16	17	18	19	20	21
Prioritization Weights	0.0873	0.1142	0.0979	0.0969	0.0701	0.0557	0.0096
Drive Activity Class	22	23
Prioritization Weights	0.0086	0.0806

, the priority order of the action classes is given according to the priority weighting obtained.

It has been seen that the class of action that causes distraction the most is texting/paying attention to a mobile phone. When Table 2 and its prioritization weights are examined, the activity classes causing the most distraction are 16, 17, and 18, in that order. Prioritization continues as 15, 23, 19, 11, 10, 9, 13, and 20, in that order. Although the prioritization weights of the first 11 classifications are relatively close to each other, the weight value of the No. 1 class in the 12th position is quite low. For this reason, the first 11 activity classes are considered in this study. Activity classes 16, 17, and 18 fall within the scope of dealing with mobile phones. Therefore, the three classes mentioned are considered as a single class of activity. In addition, since the activity classes 9, 10, and 11 are eating, drinking, and smoking, these three classes are evaluated as a single class.

A comprehensive literature review is conducted to evaluate the prioritization weights obtained as a result of the AHP analysis. Accordingly, one-to-one correspondences are observed with the activity classes stated to be effective in the literature [6,23,26,34]. Thus, it was decided to use the prioritization weights obtained in the AHP analysis in the study.

4.2. Analysis of YOLO Algorithm

The classes to be used are determined after the prioritization weights obtained by the AHP method. The activity classes used are given in Table 2. LabelImg program was used for the annotating processes of the data in the determined classes. Each class is annotated separately to avoid class conflicts. After the annotating process was completed, cleaning and file preparation processes were carried out so that the data could be suitable for the training model. The model is trained using Google Colab Pro and the epoch count for training is considered 200. In order to understand that the training is carried out in a healthy way and the model is suitable, it is necessary to examine various measurement parameters. Training results can be evaluated with output parameters such as F1-score, precision, recall, mAP, etc. The formulations for the output parameters are given in Equations (6)–(10) [50,51]. The model outputs obtained in this context are presented in Figure 3, Figure 4 and Figure 5.

$$Precison={\displaystyle \frac{True Positive}{True Positive+False Positive}}$$

(6)

$$Recall={\displaystyle \frac{True Positive}{True Positive+False Negative}}$$

(7)

$$AP={\sum }_{i=0}^{i=n-1}\left[\left(Recall\left(i\right)\right)-\left(Recall\left(i+1\right)\right)\right]\ast Precision\left(i\right)$$

(8)

$$mAP={\displaystyle \frac{1}{n}}{\sum }_{i=0}^{i=n}{AP}_i$$

(9)

$$F_1=2\ast {\displaystyle \frac{precision\ast recall}{precision+recall}}$$

(10)

Figure 3: This graph expresses the change in F1-score at different epoch levels during the training process of the model. The F1-score is a measure of a test’s accuracy and is calculated using the precision and recall of the test.

Figure 4: This graph shows the average mAP at the threshold of 0.5, which indicates the model’s ability to accurately detect and classify objects with a threshold of 0.5 IoU.

Figure 5: This graph represents the map on a series of IoU thresholds between 0.5 and 0.95 and provides a more comprehensive assessment of the model’s performance at different levels of rigidity in object detection.

As can be seen from the training results mAP 0.5, a success rate of 99.53% (mAP 0.5) has been achieved in terms of accuracy. In addition, when the mAP 0.5:0.95 parameter is examined, it is seen that a success rate of 91.17% is achieved. It is understood from the F1-score value that there is no wrong model selection in the datasets without ignoring the extreme cases. The F1-score value, in which all measurement metrics are examined throughout the model, was obtained as 0.859. The consistency of the results is also proven by the images given in Figure 6. Some detection images in the new dataset created by the model are presented in Figure 7.

The AHP method and the YOLO algorithm are used together in order to automatically detect driver distraction. Thanks to the broad perspective data used in the study, the inclusion of various driver characteristics in the study ensured that the study area is in a global framework. It is seen that the action classes obtained in the AHP method exactly match the action classes in the literature. In addition, thanks to the detection of action classes using the AHP method, the workload is reduced by annotating and modeling the less important action classes, and a scientific and strategic approach is presented by using the AHP method. The authors reached a consensus on this issue by meeting with traffic safety experts. A more sensitive analysis process is carried out by re-detailing the action classes obtained as a result of the AHP method, taking into account the studies in the literature (left- and right-hand talking or texting). Obtaining the results of the study takes a long time due to the size of the dataset. Due to the speed of the algorithms, the dataset used in the study was made at a certain limit.

Reducing driver distraction is not only a topic of study with a focus on traffic safety but can also contribute significantly to strategies for sustainable transportation and development goals. The measures are expected to reduce traffic accidents, traffic congestion, delays, and the resulting economic and social losses while reducing the consumption of fossil fuels and greenhouse gas emissions. A healthier and more robust traffic flow reduces congestion, thereby minimizing drivers’ time loss and reducing environmental damage. In addition, the reduction in the number of traffic accidents reduces the risks and costs to human and public health.

By reducing distraction, drivers are encouraged to be more aware and attentive, which has a positive impact on traffic safety and public health. All these factors contribute in an important way to sustainable transportation and development goals. Therefore, studies on reducing driver distraction play a critical role in achieving sustainable transportation and development goals, offering important contributions from different perspectives. These measures are essential for a sustainable future and should be considered as an important step to improve the well-being of societies.

Detailed information about the study was given to all participating drivers in order to obtain experimental images. However, due to data-protection laws, it is not possible for us to directly record the daily activities of drivers. In order to find a solution to this problem, we have requested drivers to use their vehicles on a closed traffic route and exhibit normal driving habits in a controlled environment. This approach ensured that the collected data was both ethically compatible and reflected typical driving behavior, while at the same time respecting the privacy of participants.

5. Conclusions

The aim of this study is to determine the action classes that cause driver distraction with the MCDM-based method using the global dataset and to detect the forbidden action classes while driving with a deep learning algorithm. The proposed methodology approach consists of two steps: In the first step, data collection, arrangement, and determination of action classes by the AHP method are performed. In determining the action classes, traffic safety experts, drivers, transportation engineers in the local government, academics, and studies in the literature are used. After the action classes are determined, the data are annotated using the LabelImg package program. As a second step, the images in the dataset are divided into two groups, train and validation, in accordance with the deep learning algorithm. The validation dataset is approximately 20% of the dataset. Model trainings as a deep learning algorithm in terms of stability and speed are trained with the YOLOv8 algorithm. When the YOLO algorithm results are examined, mAP 0.5–99.53% and mAP 0.5:0.95–91.17% values are obtained. The F1-score value, which is an important indicator in datasets that are directly related to recall and precision and that are not homogeneously distributed, is 0.859. It is observed that, with the YOLO parameters, the model successfully detects driver distraction. Thanks to the study, it is predicted that traffic accidents will be prevented by detecting driver distraction automatically and with high accuracy. This study aimed to shed light on future studies by using MCDM and deep learning algorithms together. In addition, the aim is to prevent traffic accidents by using deep learning algorithms and MCDM methods in the field of traffic safety.

The results of our study have several important points in terms of road safety: The high detection accuracy shows that our system accurately and reliably identifies distracted driving behaviors. Thus, thanks to developments in future studies, drivers can be warned and possible accidents can be prevented. By prioritizing distraction in the classroom, the most dangerous distractions can be identified and interventions and educational campaigns aimed at reducing these elements can be carried out. Thanks to our detection and prioritization system, there is a significant reduction potential in traffic accidents. It is in line with the sustainable development goals by reducing the likelihood of traffic accidents and contributing to energy efficiency and lower carbon emissions through improved traffic management. Reducing distractions can ensure a smoother traffic flow and reduce congestion. This improvement in traffic efficiency not only improves road safety but also contributes to the sustainability of urban transport systems.

This article is a guide for those who want to work in the fields of globally generated datasets and driver-distraction detection. An infrastructure has been created for the datasets to be used in the studies to be carried out for the detection of driver distraction of different deep learning algorithms. A scientific framework for traffic safety issues is presented by using the AHP method and the YOLO algorithm together.