Driving Violation Prediction Based on an Emotional Style Transfer Network

Abstract

Emotions are closely related to driving behavior, and drivers with different emotions have different degrees of bad driving behavior. In order to explore the relationship between emotions and driving violations, a prediction model based on an emotional style transfer network is proposed. First, inspired by the idea of generative adversarial networks (GAN), the eigenvalues of emotions are extracted. Secondly, the one-way propagation method of the GAN network is improved to cyclic generation, which avoids the problems of non-convergence and long periods in the data training process, improving the utilization of training data. Thirdly, a driving violation prediction model is designed. In this model, the emotion factors are designed as time-related sequences, and by improving the Long Short-Term Memory (LSTM) model, the encoding and decoding processes of the time-related sequences are added to form the context, which improves the accuracy of prediction. Finally, the experimental and simulation data show that the proposed model has significant advantages in loss value, accuracy rate, macro-average score, and other indicators. At the same time, an emotion-induction scheme is given to reduce the possibility of driving violations. Furthermore, the proposed model can provide a theoretical basis for the impact of emotions on driving safety.

1. Introduction

The number of vehicles is increasing year by year, causing a series of problems such as road congestion, air pollution, and frequent traffic accidents. Among them, frequent traffic accidents have brought great harm to the safety of lives and property. According to data released by the Statistics Bureau of the People’s Republic of China, in 2020, 244674 traffic accidents occurred in China, with direct property losses of CNY 1313.606 million. The number of fatalities reached 61,703 and the number of injuries reached 250,723 [1]. One of the main causes of traffic accidents is the driver factor, that is, bad driving skills and emotions make it easier to break the law while driving, which leads to the occurrence of traffic accidents. Driving skills [2] reflect the driver’s ability to control the vehicle and how well he or she anticipates hazards. Generally, driving skills will improve as driving experience increases, but negative fluctuations in the driver’s emotions during driving often have a serious impact. So, driving factors, especially inappropriate driving behavior caused by emotions, need to be considered when developing vehicle assistance systems. At the same time, with the rise of intelligent vehicles, people prefer that the vehicle can provide assistance driving functions that detect the current state of the driver and automatically correct the driving behavior [3,4]. Therefore, it is very important to establish a prediction model based on emotions to improve the safety of vehicle driving and meet people’s needs for intelligent vehicle assistance functions.

Emotions are closely related to driving behavior [5], and drivers with different emotions have different degrees of bad driving behavior. For example, angry emotions often lead to behaviors such as high-speed driving and frequent lane changes, which have a higher risk of breaking the law. Therefore, building a driving behavior prediction model based on emotions can classify the driving violations and predict driving violations within a given period of time. This can not only assist intelligent vehicle decision-making and planning systems to correct the vehicle behavior trajectory and avoid accidents but also provide new ideas for improving the efficiency and safety of urban traffic.

At present, the methods for building driving violation prediction models can be divided into two categories: fuzzy rule-based driving behavior modeling and neural network-based driving behavior modeling [6]. The problem with these two methods is that the models trained on other drivers’ driving behavior data cannot accurately express the driving behavior of a specific driver. To solve this problem, an improved model is proposed that can adjust the model parameters to adapt to different drivers, but it is difficult to reflect the impact of the driver’s emotions, judgment, execution, and other psychological and physiological activities on the driving trajectory to a certain extent. So, the effect it achieves is still limited and cannot meet the needs of realistic intelligent assistance systems. In recent years, research on identifying driving style [7,8,9] has gradually become mainstream, but the definition of driving style is not unified due to various factors. Existing research methods mainly focus on driving style analysis based on questionnaires, which have the characteristics of high economy, short periodicity, easy quantification of results, and easy statistical analysis, and this allows for better classification and prediction of driving violations while being applied at scale.

This paper focuses on a driving violation classification and prediction model with emotions. The aim of the model is to predict future driving violations by sensing changes in drivers’ negative emotions, thereby providing a basis for intelligent vehicle intervention and vehicle behavior correction. Particularly, the main contributions of the paper can be summarized as follows: (1) Designing methods for calculating the eigenvalue of emotions and the eigenvalue of driving violations. The calculation methods are used to form the training data. (2) The one-way propagation mode of network data in the generative adversarial networks (GAN) is improved to a cyclic propagation mode. This mode can avoid the defects of non-convergence and long periods of training data. (3) Designing a time-related sequence to reflect the eigenvalue of emotions and improving the Long Short-Term Memory (LSTM) model to support the time-related sequence. Based on this, an encoding and decoding scheme is proposed to reduce the loss value.

The remaining part of the paper is structured as follows: Section 2 describes the related work. Section 3 gives the system architecture, emotion-style transfer network, and prediction model of driving violations. Section 4 shows the simulation experiment and performance evaluation. Finally, Section 5 concludes this paper.

2. Related Work

With the rapid development of intelligent vehicles, the research on driving behavior analysis and intelligent assistance system design has gradually deepened. Existing research methods on driving skills and styles are mainly about neural networks [2], deep learning [10], data visualization [11], graph analysis [12], and so on. For driving behavior, the existing research mainly focuses on fuzzy control-based driving behavior modeling, neural network-based driving behavior modeling, and driving style-based behavior modeling.

The fuzzy control-based driving behavior modeling can directly reflect the self-learning and fuzzy reasoning abilities in the driving process, showing the fuzziness and self-learning characteristics of vehicle control. The authors in [13] constructed a driving model with fuzzy characteristics that has better tracking performance for target paths through experimental verification. This type of model is characterized by a large amount of computation and requires a large number of samples for training. The neural network-based driving behavior modeling can extract the driving behavior features from the pre-training model and then train the driving behavior model. The authors in [14] designed a three-layer feedforward neural network to extract driver behavior features and used real driving behavior data to achieve autonomous driving of vehicles. The authors in [15] improved the three-layer into a multi-layer, designed a multi-layer feedforward neural network, and built a driving behavior neural network model with the relevant nonlinear function approximation, which improved the convergence speed; however, the prediction results were not satisfactory when the size of the training data was small.

In recent years, research on driving style has become mainstream, mainly including two categories: style feature extraction and style feature classification. For the former, it can be divided into questionnaire-based methods, statistical-based methods, transformation-based methods, and deep learning methods. The questionnaire-based method [16,17] asks related people to fill out a form, then researchers collect and analyze the data to find the hidden relationships between results and factors. The statistical-based method is simple and effective, and it can extract the mean and variance of the driver behavior characteristics, thereby obtaining the driving style in a short time [18,19]. The disadvantage is that it can only focus on the macroscopic features of driving behavior and cannot extract the microscopic features. The authors in [20] proposed a principal component analysis method based on linear transformation dimensionality reduction, which can better cluster and classify by reducing the dimensions of the mean and variance. Extracting driving style features using deep learning methods is the latest research achievement. The authors in [21] used a deep embedding method to extract driving style features, and the authors in [22] adopted a convolutional autoencoder and t-stochastic neighbor embedding (t-SNE). The above methods have achieved good results, but they are all simulated experiments on fully connected autoencoders and convolutional autoencoders, which are only suitable for sequential autoencoder scenarios.

The main methods of style feature classification are supervised learning, unsupervised learning, and semi-supervised learning. The main idea of supervised learning is to classify driving styles in advance and train and predict the model based on the existing labels. For example, support vector machines (SVM) and Bayesian algorithms. For this type of method, a large amount of data needs to be collected and labeled. Unsupervised learning methods do not require labels, and typical methods include K-means and hierarchical clustering. For example, the authors in [23] and [24] used the K-means method to cluster the behavioral characteristics of different drivers. The authors in [25] used hierarchical clustering to evaluate the safety of driving behavior based on changes in vehicle coordinate position. The use of such methods has certain limitations. Semi-supervised learning requires partial label learning. The authors in [26] proposed an ensemble learning-online semi-supervised approach based on the k-nearest neighbor algorithm, which enables a small sample to have a better average recognition rate. This type of method tends to accumulate wrong labels during execution, resulting in larger errors.

3. The Proposed Driving Violation Prediction Model

Traditional models, such as fuzzy rule-based and learning-based methods, have problems such as poor adaptive effects and low accuracy of personalized driving behavior expression. In order to solve the above problems, transfer learning theory in our model is introduced. The idea of the proposed model is to transfer prior knowledge (including driving violation eigenvalues and parameters) from multiple source domains (a dataset on driving violation behavior with emotions) and apply the extracted prior knowledge in the target domain. Our dataset used for the proposed model was derived from questionnaires filled out by illegal persons while they were being educated and punished by the vehicle management department. So, the authenticity of the data from these questionnaires is guaranteed. When we aggregated and analyzed the data from the questionnaires, the dataset was divided into two parts, one as training data, forming the source domain, and the other as test data, forming the target domain. So, the existing driving violation data is transferred to the drivers in the target domain, and a personalized model is established to predict the probability of the driving violation in the future.

3.1. Architecture

The proposed driving violation prediction model contains the driving violation data of the source domain and the target domain. It is necessary to construct the personalized model through the source domain data, as shown in Figure 1.

In the proposed architecture, driving violation data with emotions in the source domain needs to be reused to build a personalized behavior model. If the emotion data in the source and target domains are similar, the two parts of the data can be merged and transformed into a simple machine learning problem. But in fact, due to the complexity of emotions, it is impossible to directly use the combined data to train a driving violation prediction model. That is, it is necessary to filter out samples in the source domain with a similar distribution of emotion data in the target domain. In our paper, maximum mean discrepancy is used to measure the distance between emotion data in the source and target domains. Specifically, we first obtain the mathematical expectation of a continuous function of samples with different distributions in the two spaces and then construct a function such that the mean of the difference between the above two mathematical expectations has a maximum value.

The proposed prediction architecture draws on the idea of GAN, but the way GAN generates training data cannot achieve the goal of classifying and predicting driving violations. So, we introduce the concept of transfer learning on the basis of GAN to construct an emotional style transfer network, which transfers the knowledge learned from the driving violation dataset with emotions to the new driving emotion data. In our model, D is defined as the emotional driving violation data domain, which is decomposed into two parts, the feature space $\Gamma$ and the marginal probability distribution P(S), $X=\left(s_1,s_2,\cdots ,s_n\right)$ is defined as the input sample of each driver. In a particular domain $D=\left(\Gamma ,P\left(S\right)\right)$ task T is defined by $T=\left(t,f\left(·\right)\right)$, y is the label space, $f\left(·\right)$ is the prediction function, and $f\left(s\right)=P\left(t\left|s\right.\right)$. According to the above definitions, when given a source domain $D_S=\left(\left(s_1,t_1\right),\left(s_2,t_2\right),\cdots ,\left(s_n,t_n\right)\right)$ and its task target $T_S$, and target domain $D_T=\left(\left(s_1^{\prime },t_1^{\prime }\right),\left(s_2^{\prime },t_2^{\prime }\right),\cdots ,\left(s_n^{\prime },t_n^{\prime }\right)\right)$ and its target $T_T$, we improve the learning ability of prediction function $f_T\left(·\right)$ in the target domain by acquiring knowledge in $D_S$ and $T_S$, thereby constructing a personalized emotional driving violation model.

3.2. Emotional Style Transfer Network

The proposed model realizes knowledge transfer between driving violation data in different source domains, which provides data support for the subsequent training of the personalized driving model in the target domain. That is, a part of the existing data is selected as the input, and the emotional style transfer network is used to construct a data transfer learning model that conforms to the driving behavior in the target domain so as to extract the essential characteristics of driving violations. Inspired by the idea of GAN, two combined network models of generator and discriminator are established in the proposed model for confrontation and game. The generator generates emotional violation data according to the input, and the discriminator calculates the probability that the data sequence is a real data sequence until it is close to the emotional driving violation data in the target domain, which provides data for further prediction and assessment of driving violations, as shown in Figure 2.

In view of the fact that the GAN network is one-way in the confrontation process, there are problems such as difficult convergence and a long training period when training driving violation data sets with emotions. At this time, it is necessary to manually intervene with the generator and the discriminator to achieve a dynamic balance and prevent the gradient from disappearing. So, we adopt a recurrent adversarial network as shown in Figure 3. $S^{\prime }$ is a copy of S, $T^{\prime }$ is a copy of T, and the cyclic mapping of the data domain from $S^{\prime }$ to S and from $T^{\prime }$ to T is done using the generator and discriminator, respectively. At the same time, we add consistency constraints and consistency loss, and minimize the error to ensure that the generator network can produce an efficient mapping from the source domain to the target domain, thereby generating emotional driving violation data that are close to the data distribution in the target domain.

The recurrent adversarial network consists of generator F and generator G, discriminator D and discriminator $D^{\prime }$, which can realize the mutual mapping of source domain S and target domain T. In which emotion data are $s_i\in S$ and $t_i\in T$, and F and G are $S\to T$ and $T\to S$, respectively. Specifically, use F to convert $s_i$ to $t_i^{\prime }$, and D to discriminate the results, then adopt G to obtain $s_i^{\prime }$, and then use a consistency loss to measure the distance between $s_i$ and $s_i^{\prime }$. In the same way, use G to convert $t_i$ to $s_i^{\prime }$, and D to discriminate the results, then use F to obtain $t_i^{\prime }$, and then use the consistency loss to measure the distance between $t_i$ and $t_i^{\prime }$. An adaptive estimation optimizer is adopted to adjust the learning rate autonomously to gradually check the above two distance values, and finally the mapping function between S and T can be obtained.

F is a map of S→T, and its adversarial loss is:

$$L_{GAN}\left(F,D,S,T\right)=E_{t\sim Pdata\left(t\right)}\left[log\left(D\left(t\right)\right)\right]+E_{s\sim Pdata\left(s\right)}\left[log\left(1-D\left(F\left(s\right)\right)\right)\right]$$

(1)

Its optimization objective is:

$$\underset{F}{min}\underset{D}{max}\left(L_{GAN}\left(F,D,S,T\right)\right)$$

(2)

Similarly, G is a mapping of $T\to S$, and its adversarial loss is:

$$L_{GAN}\left(G,D^{\prime },T,S\right)=E_{s\sim Pdata\left(x\right)}\left[log\left(D^{\prime }\left(s\right)\right)\right]+E_{t\sim Pdata\left(t\right)}\left[log\left(1-D^{\prime }\left(G\left(t\right)\right)\right)\right]$$

(3)

Its optimization objective is:

$$\underset{G}{min}\underset{D^{\prime }}{max}\left(L_{GAN}\left(G,D^{\prime },T,S\right)\right)$$

(4)

The consistency loss is:

$$L_{cyc}\left(F,G\right)=E_{s\sim Pdata\left(s\right)}\left[{‖G\left(F\left(s\right)-s\right)‖}_1\right]+E_{t\sim Pdata\left(t\right)}\left[{‖F\left(G\left(t\right)\right)-t‖}_1\right]$$

(5)

According to the GAN principle, the optimization goal of setting the emotional style transfer network model is:

$$L\left(D,D^{\prime },S,T\right)=L_{GAN}\left(F,D,S,T\right)+L_{GAN}\left(G,D^{\prime },T,S\right)+\mu L_{cyc}\left(F,G\right)$$

(6)

The maximum value $F^{\prime }$ and minimum value $G^{\prime }$ serve as the optimal solution to Equation (6).

In the iterative stage of a discriminative network, the maximization optimization objective along the positive gradient direction is:

$$g_{dis}={\nabla }_{{\theta }_{dis}}\left[L_{GAN}\left(F,D,S,T\right)+L_{GAN}\left(G,D^{\prime },T,S\right)+\mu L_{cyc}\left(F,G\right)\right]$$

(7)

The minimization optimization objective along the negative gradient direction is:

$$g_{gan}={\nabla }_{{\theta }_{gan}}\left[L_{GAN}\left(F,D,S,T\right)+L_{GAN}\left(G,D^{\prime },T,S\right)+\mu L_{cyc}\left(F,G\right)\right]$$

(8)

The resulting generative network $F^{\prime }$ is an efficient mapping function from S to T.

The influencing factors of driving violations mainly come from two parts: driving experience and emotions. Driving experience refers to the basic information that can affect vehicle driving, such as driving age, occupation, education, drinking habits, smoking habits, insomnia, etc. Emotions refer to sudden factors that can affect vehicle driving, such as social sensitivity, degree of fatigue, anxiety, aggressive driving, being overly nervous, paranoia, etc. These factors are the input data of the style transfer network model, and different inputs make the driving violation data obey different probability distributions. S_i in the source domain is decomposed into driving experience S_i1 and emotions S_i2, and T_i in the target domain is decomposed into driving experience T_i1 and emotions T_i2, as shown in Figure 4. Each factor in S_i and T_i corresponds to a different input sample, and the parameters of the emotional style network are trained through the driving experience, emotion sequence, and driving violation sequence in the source domain, thereby extracting the characteristics of the driving violations. The emotional style network model is then obtained, which is validated with the remaining data from the source domains.

3.3. Driving Violation Prediction

Time-related sequences can be formed due to the persistence of emotions during driving. On this basis, a driving violation prediction model based on an improved long short-term memory (LSTM) network is designed. In this model, an encoder and a decoder are constructed. The encoder is responsible for encoding the source domain data of the emotion transfer network into a fixed-length context feature vector, and the decoder is responsible for extracting the required information from the obtained feature vector to predict the violations in the next period, as shown in Figure 5.

The neural network model has the disadvantage that it can only deal with fixed-length sequences; that is, it needs to ensure that the length of the input sequences is equal to the length of the output sequences. However, the persistent sequence of the violation prediction model is unstable; that is, the length of the collected valid historical sequence is inconsistent with the length of the predicted sequence in the next period. Therefore, a structure including an encoder, context feature vector, and decoder is designed to solve the above problem of inconsistent sequence length.

The unit size of our proposed model is the same as that of LSTM. Assuming that the size of the encoder step is fixed and the history dataset of driving violations is t_i. In each step of LSTM, the driving skill data and emotion data of t_i are input, and the hidden layer state h_t of LSTM at time t is calculated:

$$h_t=f\left(\alpha w_t+\beta h_{t-1}+b\right)$$

(9)

Where $w_t$ is the driving violation sequence at time t, $h_{t-1}$ is the hidden layer state of LSTM in the previous step, $f\left(·\right)$ is the activation function, $\alpha$ and $\beta$ are the weights of the input layer, and b is the bias of the input layer. According to formula (9), the hidden layer state of the next step is calculated sequentially until the preset step is reached and the final hidden layer state h_n is set as the context feature vector.

The decoder takes the context feature vector generated by the encoder as the initial hidden layer state, and it receives the output of the hidden layer state of the recurrent unit in the previous step and calculates the output of the hidden layer state of the current step:

$$w_{t-1}=f^{\prime }\left(\delta h_{t-2}+b^{\prime }\right)$$

(10)

$$h_t=f\left(\alpha w_{t-1}+\beta h_{t-1}+b\right)$$

(11)

where $w_{t-1}$ is the output of the previous step, $h_{t-2}$ is the input of the hidden layer state of the previous step, $h_{t-1}$ is the output of the hidden layer state of the previous step, $f\prime \left(·\right)$ is the activation function tanh of the LSTM output layer, $\delta$ is the weight of output layer of LSTM, and $\delta$ is bias of the network.

The test network for driving violations takes a multilayer perceptron network as the core. In this network, by inputting multiple pieces of driving violation data, the multi-layer perceptron classifies the data to be tested and outputs the probability that this data is a sequence of driving violations. In order to improve the training effect and prediction accuracy, the loss function C of the test network is designed:

$$C=\kappa C_p+\left(1-\kappa \right)C_f$$

(12)

where C_p is the cycle loss function, C_f is the emotional factor loss, and $\kappa$ is the coefficient of proportionality.

C_p consists of two parts, the encoder-decoder loss and the MLP loss, and its calculation method is the same as that of reference [27]. The definition of C_f is:

$$C_f=min{‖S_{i2}-S_{j2}‖}_{\mathcal{l}1}$$

(13)

S_i2 and S_j2 are the time-related sequences of emotions for predicting violations, and the sequence of emotions for actual violations, ${‖·‖}_{\mathsf{\ell }1}$ is the $\mathsf{\ell }1$ -norm on ${\Re }^n$.

During the operation of the network, the prediction network generates a new prediction sequence and calculates the loss value. The loss value decreases with the increase in the number of iterations until the output of the prediction network conforms to the driving violation sequence of the target domain for a period of time in the future.

4. Performance Evaluation

4.1. Evaluation

To evaluate the performance of the proposed model, we established a driver simulation experiment platform, designed a questionnaire to collect driver emotions, and developed a data collection module and a data processing module for the platform. The data collection module collects road condition information (such as traffic light changes, road separation line changes, sidewalks, etc.), and the data processing module implements functions of the prediction model. The hardware environment of the vehicle running on the platform is mainly composed of a driving recorder, a vehicle camera, a GPS (Global Positioning System) locator, a vehicle radar, etc. The platform runs on a server cluster, where the CPU (Central Processing Unit) is configured as: Intel [email protected], memory 128G, hard disk 2T. Moreover, we used Python to develop the platform. In addition, we have a driving simulator, which looks like a game console and consists of a screen, a steering wheel, a gas pedal, gears, etc. It displays the simulated traffic conditions on the screen, and the drivers judge and operate according to the road conditions to simulate the real driving scene. The dataset was from the vehicle management department. On this basis, we collected 1500 questionnaires filled out by illegal persons, in which the questionnaires contain basic information, emotional information, and driving violation information, which is shown in Figure 6. In the process of data preprocessing, we used the interpolation method to supplement some missing data. Then, for the 1500 samples obtained, some samples were selected as the driver data in the target domain, and the remaining samples were used as driver data in the source domain. In addition, emotions were evaluated using the Self-Assessment Manikin (SAM) as the standard, and the driver’s current emotions were determined through process measurements during the initial stage of the simulation. After a period of time, the driver’s motion was changed and evaluated by playing different types of music. This evaluation process takes three attempts to finally determine the driver’s current motion. The Delphi method was used to establish an emotion induction music library, and samples were selected using the music effect scale based on the Thayer model as the standard [28].

Considering the scale of the data and the actual step size, the data sample frequency is set to 2 Hz, and finally contains a time sequence of 120 steps within 60 s. In experiments, a 2-layer LSTM is adopted as the encoder and decoder, and the length of the eigenvalues of driving experience and emotions, which are output by the encoding layer, is 20. The learning rate of LSTM is set to 0.0001, the batch size is 128, the relu function is used as the activation function, and the Adam optimization algorithm is used to update the weight parameter matrix. First, the loss function is calculated, as shown in Figure 7. With the maximum training times set to 100, the error fluctuates continuously during training, but the trend tends to decrease and the model stabilizes until the final result is obtained.

The driving violation data in the source domain is used to predict driving violations in the target domain over time. In order to measure the matching degree of the above two, MSE and SDR are used as evaluation indicators, where MSE is defined as $\frac{1}{n}{{\displaystyle {\sum }_{i=1}^n\left({\phi }_i-\stackrel{\sim }{{\phi }_i}\right)}}^2$, SDR is defined as $log\frac{{\displaystyle {\sum }_{i=1}^n{\phi }_i^2}}{{\displaystyle {\sum }_{i=1}^n{\left({\phi }_i-\stackrel{\sim }{{\phi }_i}\right)}^2}}$. n is the sample size, ${\phi }_i$ and $\stackrel{\sim }{{\phi }_i}$ are the actual and predicted values, respectively.

Taking three types of drivers with different driving experiences and emotions as examples, the validity of the prediction model of driving violations based on an emotional style transfer network is verified. In the experiment, each driver is set as the source driver and the target driver, and two transfer learning experiments are performed, respectively. The data transfer between different types of drivers is bidirectional, and 100, 160, 220, 280, 340, and 400 data points are extracted from 1500 samples as target driver data. The experimental results are shown in

Table 1. Data transfer results of driving violations.

The Number of Target Driver Data Transfer Direction/Indicators		100	160	220	280	340	400
Driver Type 1 to Type 2	MSE	5.35	5.78	3.56	3.18	2.77	2.31
	SDR	16.66	17.46	18.75	20.17	21.98	22.94
Driver Type 1 to Type 3	MSE	6.49	6.38	3.38	4.16	2.11	2.21
	SDR	15.49	16.36	18.24	21.13	21.23	22.38
Driver Type 2 to Type 3	MSE	9.11	6.13	4.39	3.26	3.36	3.18
	SDR	17.21	16.23	19.21	21.18	23.65	23.17
Driver Type 2 to Type 1	MSE	6.27	5.16	4.29	4.33	3.18	2.88
	SDR	16.24	18.18	22.32	23.44	24.19	24.91
Driver Type 3 to Type 1	MSE	6.13	6.61	5.64	3.77	2.51	2.81
	SDR	16.88	19.34	21.51	23.52	25.11	25.85
Driver Type 3 to Type 2	MSE	7.15	6.32	4.26	4.28	3.13	3.02
	SDR	17.19	19.3	21.17	22.44	23.45	24.93

. It can be seen from the table that as the amount of target driver training data increases, the MSE value tends to decrease, and the SDR value tends to increase. This shows that when the training data of the target driver is large and the historical data of the source driver is sufficient, it is beneficial to improve the performance and effectiveness of the prediction model.

The accuracy of the prediction results is measured by the accuracy rate and the macro-average score. The accuracy rate is defined as the ratio of the sum to the number of samples, where the sum is the number of samples with positive predicted values and positive true values plus the number of samples with negative predicted values and negative true values. The macro-average score is defined as the ratio of the composite score of F1 to the number of samples. The results are compared with SVM, LightGBM (Light Gradient Boosting Machine), GRU (Gate Recurrent Unit), and CNNS (Composite Neural Networks) [29] as shown in

Table 2. Comparison results with existing methods.

Models	Accuracy Rate	Macro-Average Score
SVM	0.834	0.825
LightGBM	0.880	0.871
GRU	0.904	0.920
The proposed model	0.917	0.931

. As the number of iterations increases, the value of the loss function becomes smaller and smaller, resulting in the proposed model outperforming other algorithms in terms of accuracy and macro-average score. LightGBM outperforms SVM on low-dimensional dense data, but its disadvantage is that the input sequence cannot contain time information. GRU has superior performance on small-scale datasets but requires more parameters. CNNS is a hierarchical model combining CNN (Convolutional Neural Network) and LSTM, which can fully mine meaningful information but ignores the processing of time-sequence data. The proposed method improves the utilization of data by adopting data transfer and increases accuracy and the macro-average score.

On the basis of existing datasets, new data are collected through a driving simulator to verify the effectiveness of the proposed model and find out the relationship between emotions and driving violations. During the experiment, drivers with driving experience were invited to induce different emotions through music, and the Self-Assessment Manikin (SAM) emotion report is used as the standard. Moreover, emotional valence and emotional arousal are used to measure emotion and its magnitude. The experimental steps are as follows: (1) The driver’s initial emotion is determined by SAM, and the target driver data is formed at the initial moment. (2) The driver’s emotion is induced by music, and the driver’s current emotion is determined by SAM, which forms the current moment data of the target driver. (3) The simulated driving experiment is carried out, and the driving violation data is recorded. (4) Play music while driving to ease or aggravate current emotions and record data on violations. For example, the driver was first tested with SAM to determine the initial motion, and then music was played to induce a motion change for 4 minutes at a volume of 70 dB. The driver was then tested again with SAM to determine the current motion to ensure a successful motion change. The obtained time-related sequences are input into the proposed model, and some results are shown in

Table 3. Results of the relationship between emotions and driving violations (average times).

Emotions	Speeding	Running a Red Light	Violating the Line	Not Giving Way to Pedestrians	Changing Lanes Illegally
Emotions with high valence and high arousal	1.84	0.94	3.45	1.40	2.31
Emotions with high valence and low arousal	0.55	0.30	0.90	0.74	0.70
Emotions with low valence and low arousal	0.35	0.18	0.79	0.49	0.50
Emotions with low valence and high arousal	1.26	1.15	2.87	1.20	1.86
Peace emotions	0.86	0.73	2.04	1.30	1.47

.

It can be seen from Table 3 that emotions have an impact on driving violations, but driving violations can be improved after adjusting emotions. Therefore, if the driver’s emotion recognition function and emotion regulation are added to the intelligent vehicles, the number of driving violations can be effectively reduced.

4.2. Discussion

The proposed model can provide a solution for predicting driving violations in current intelligent vehicles. This model, which takes the current emotional data of the driver sensed by the in-vehicle sensors as the target domain and the known data as the source domain, predicts driving violations that may occur over a period of time. Then, with the vehicle assisted driving module as the core, it can provide reminder functions to prompt the drivers to pay attention to or change the current abnormal driving behavior. For example, if speeding behavior is predicted, equipment such as in-vehicle voice or a multi-functional seat will remind the driver by using sound, light, vibration, etc., and even if necessary, the vehicle will actively decrease the speed to reduce the probability of driver violations, thereby improving traffic safety.

Limited by the factors describing the driver’s emotions, the questionnaire in our experiment only focuses on the classification of emotions and does not distinguish the degree of a certain emotion. For example, the emotion of sadness is not subdivided into grief or bitterness. So, it is necessary to further study the influence of different expression degrees of the same emotion on driving violations. In addition, this study only focuses on the relationship between negative emotions and driving violations and does not consider the relationship between positive emotions such as excitement and driving violations. According to the above limitations, we will explore the relationship between the degree of emotional factors and the types of driving violations and pursue the relationship between positive emotional factors and the types of driving violations in the future. On the basis of emotional factors, we will further consider other relevant factors, including driving habits, the driver’s education level, and third-party factors such as road conditions and pedestrian violations during driving, to predict driving violations.

5. Conclusions

In this paper, we propose an emotional style transfer network model with driving experience and emotions as the main factors, which aims to predict driving violations. Inspired by the idea of the GAN network, we design the eigenvalue of emotions, improve the data propagation method of the GAN network, propose time-related sequences with emotions, and introduce the encoding and decoding processes into the GAN network. From the perspective of experiments and simulated driving, the proposed model has achieved a significant improvement in the evaluation indicators. Moreover, an emotion induction scheme is designed, and it provides a feasible idea for reducing the probability of driving violations.