Participants

The number of subjects was 30, there were 15 professional drivers and 15 staffs and students from Wuhan University of Technology (20 male subjects and 10 female subjects). They were aged from 22 to 55 years old (Mean age = 32.2, SD = 8.2). They had driving experience from 2 to 18 years and 6.9 years’ driving experience on average, the driving distance was up to 6500 km in the experiment. It should be noted that there are many people who have driving license without actual driving experience in China. In order to avoid interfering with the credibility of the experiment, driving mileage was used instead of driving license age as an indicator of drivers’ proficiency in this research. In terms of driving mileage, they drove an average of 110,000 kilometers, ranging from 400 to 400,000 kilometers.

Vehicle instrumentation

The road traffic experimental car was developed and integrated based on the GAC Trumpchi car in Intelligent Transport System Research Center, Wuhan University of Technology, which provided platform support for a real vehicle road test and data collection.The field test was carried out on the third ring expressway in Wuhan from September 14^th to 30^th, 2016 and from March 12^th to 28^th, 2017. The experimental car was equipped with a variety of data collection equipment. A road environment image acquisition system was used to record data, such as driving operation and road scene, and was also used for the calibration of emergency braking behavior(S1 Fig [18]). Driving behavior detection systems and an on-board data acquisition system are used to obtain the vehicle running information and record vehicle traffic conditions. Information such as the driving operation and vehicle movement is extracted from the collecting platform by configuring the CANoe analyzer.

Test route

Considering the influence of different types of roads, the mixed-route in and around Wuhan (city road, city expressway, highway) were selected for experiments. The city roads section consisted of two or three lanes in each direction and the speed limits varied from 40 to 60 km/h. The total distance of urban roads was about 12 km. The urban expressways section consisted of three or four lanes in each direction and the speed limit was 80 km/h. The total distance of urban expressways was about 34 km. The freeways section consisted of three or four lanes in each direction and the speed limits varied from 100 to 120 km/h. The total distance of freeways was about 45 km. As shown in S2 Fig. The first segment consisted of a short adaption drive through an expressway which speed limits was 70 km/h indicated by yellow in the figure, in order to let drivers familiar with the vehicle condition. The second segment was the freeways which indicated by blue in the figure. The next two segments were urban expressways and urban roads indicated by red and green in the figure, respectively. Participants were given route guidance instructions provide by an experiment assistant present in the vehicle all the time in the experiment. All participants drove the same instrumented vehicle throughout the study.

Subjective style evaluation

Real driving style was obtained by the subjective global assessment. The experimental driver reported their driving style after the experiment, their driving style were shown in S1 Table. In the thirty drivers’ driving style samples list, ten drivers were of aggressive driving style samples, ten were of moderate driving style samples and ten were of mild driving style samples.

Feature extraction and coding

In order to analyze the braking characteristics, the typical braking data collected in the experiment is analyzed in this paper. After the driver received the braking signal, the relationship between the brake pedal force, braking deceleration and braking time is as shown in S3 Fig. The CAN bus was used to obtain vehicle data such as brake pedal information, and so on. All data was synchronized with the master time that was transmitted by the monitoring software at every 0.1s.

Feature extraction is one of the key points in pattern recognition. As shown in the study, the driver’s acting force of brake pedal presents as obvious priorities in the braking process. The driver exerts a greater force on the brake pedal in a short time to avoid a collision. The braking pressure changing rate is shown as an important feature, and the driver’s braking behavior is influenced by driving style [19]. This feature is considered as the focus. Considering the braking performance of vehicle, the critical value of conflict severity was obtained by the ratio of braking distance to instantaneous speed. The quantization value of conflict severity was established taking conflict time and critical value as parameters, and the severity of traffic conflict was quantified and classified [20, 21].So the typical braking data collected in the experiment was analyzed in this paper.

S3 Fig was applied to analyze the typical braking progress and show the change of braking characteristics with time directly, which contained braking impulse, maximum value of braking force and the time window. The abscissa, ordinate and corresponding symbol of this figure are shown as below. In the braking process, suppose that there was not slip phenomenon caused by lock of the wheels, when the brake pedal force F_p or brake fluid pressure p rises to a certain value (brake fluid pressure p_a), brake force F_μ reaches to the ground adhesion F_φ and stop rising, with the invariability of brake structure and external conditions, the vacuum booster pressure p_μ is in direct proportion to brake force F_φ and brake pedal force F_p [22]. p_μ (the unit is Mpa) is up to the peak and down slightly. When there is pressure to the brakes, slippage is usually 20% 50% of the peak [23], and then a slight upward trend appears. The reason for p_μ is at early stage of the brake, p_μ is increasing with the increases of tire-road adhesion coefficient φ; p_μ is declining when there is a big difference between the adhesion coefficient and sliding adhesion coefficient. At a later stage of the brake, is increasing with the lower runtime, which leads to a slight increase of p_φ [22]. The braking curve is divided into two parts; braking process within the brake work time τ₁ is called as transient process, and it is called stationary process in the continuous braking time τ₂. The essence of the transient process is the constantly increasing tire deformation. On the other hand, tire deformation of the smooth process remains unchanged basically. τ₁ is decided by the structure of the brake, and it is also an important influence on braking distance; τ₂ is the real reason for slowing the car.

All data is output force (the unit is Mpa) of vacuum booster in S3 Fig, the original data has been uploaded already. It is found that the degree of the emergency brake is related to braking impulse per unit area p_i, and the maximum value of braking p_max in a certain time T based on a lots of experiments.

(1)

(2)

As shown from the experiments, T is the time window, it is 1.6s, f is the sampling frequency, the value is 10, so the Eqs 1 and 2 can be simplified as: (3) (4)

As shown in real vehicle tests for many times on the third ring expressway in Wuhan, p* = 80Mpa⋅Hz, p_max = 9Mpa is the critical value of general braking and emergency braking. Braking behavior data was coded, with 1 representing the condition of general braking and 2 representing the condition of emergency braking. General braking requires that braking impulse was more than 80Mpa⋅Hz, and the maximum value of braking force was less than 9Mpa in the time window (1.6s, after start of braking behavior); otherwise, braking behavior that not fitted conditions of general braking was called as emergency braking. Coding rules are as follows: (5)

As shown in studies, the fit of data is the best when each observation sequence in eight data. Considering model simplicity, each observation sequence has eight braking data [24]. The data preprocessing is completed by MATLAB. The brake pedal pressure for starting point of sampling is bigger than 1Mpa, each duration is 1.6 s, the recorded 30 subjects (each subject have 60 braking data) with a total of 1800 times car brake data is the sample of study, such as braking time, brake information of pressure booster. The braking feature coding C is calculated by Eq 5.

Driving style modeling and solution

The theoretical foundation of HMM is established by Baumet al., which is promoted by Rabiner and others, which has strong modeling ability to the time domain signal so that it has become a research hotspot [16, 25]. It has been successfully used in speech recognition, behavior recognition, and character recognition and fault diagnosis. HMM is a powerful statistical tool on describing discrete time data samples, it is a double random process, one of which is the Markov chain, which is a basic stochastic process to describe the transfer of state. Another stochastic process describes the statistical correspondence between States and observations. Because the state cannot be seen directly, HMM is a random process to perceive the existence of the state and its characteristics. A hidden Markov model can be defined by:

1. {S} –a set of state including an initial observation state S_M and a hidden state S_N;
2. A –the transition probability matrix, A = a_ij,where a_ij is the transition probability of taking the transition from state i to j;
3. B –the output probability matrix, B = b_j(O_k) for discrete hidden Markov model or B = b_j(x) a continuous hidden Markov model, where O_k stands for a discrete observation symbol, and x stands for continuous observations of k–dimensional random vectors.

If the initial state distribution π = {π_i}, the complete parameter set of the hidden Markov model can be expressed compactly as λ = {N, M, A, B, π}.

Recognition result by hidden Markov model

There is a total of 1800 times car braking data, each subject has 60 times car braking data, 48 times braking data of each subject is chosen to train, and the 48 times braking data is divided into 6 samples, a sample means to take 8 constant breaking data as a sequence, the number of samples that used for training is 80% (1440/1800) of the total samples. That is to say each observation sequence consists of 8 braking feature values, which means each sequence includes 8 brake, in the whole experiment, 30 subjects had an average of 60 braking for each one, that is 30 × 60 = 1800 brake, then the 60 braking for each person is divided into 6 non repetitive sequences with each observation sequence as a the sample, there are 30 × 6 = 180 samples. The Baum-Welch algorithm was used to train and model the vector data 30 × 6 among them. Training log-likelihood changes were shown, according to the research situation. The log-likelihood of three driving style could reach a steady state with at most 40 times iterative training; they were stable about -31, -37, and -50, respectively. The corresponding driving style type of the maximum logarithmic likelihood values were homologous to three driving style such as aggressive (S5 Fig), moderate (S6 Fig) and mild (S7 Fig).

The Forward-Backward algorithm was used to assess vector data 30 × 1 of eight braking operations for each group in the remaining 30 groups. The maximum log-likelihood was obtained by identify verification, which showed that the similarity of driving style and matching degree of judgment between characteristics and model reaches maximum. There were 8 times for wrong identification, which the aggressive driving style and mild driving style identified falsely for 4 times respectively; the remaining 26 were correctly identified.

In this paper, the method of identifying the correct rate is used to evaluate the driving style discrimination. The Eq 12 is used to calculated the accuracy of driving style recognition, the accuracy μ of driving style recognition shows the rate of times for correct recognition in the total times of recognition, c_k stands for driving style, in which k ∈ {aggressive, moderate, mild}.

(12)

In the Eq 12, T_aggressive is the times that samples are recognized as c_aggressive and actually are attributed to c_aggressive, T_moderate is the times that samples are recognized as c_moderate and actually are attributed to c_moderate, T_mild is the times that samples are recognized as c_mild and actually are attributed to c_mild, S is the whole times for samples recognition.

Recognition accuracy and model recognition time were used for the evaluation model. In the 30 times tests, there were 8 times identification errors in a total, and the recognition accuracy was 88.86%. The program runtime was recognized through the MATLAB command tic/toc recognition program. It was concluded that the hidden Markov model recognition runtime was an average of 0.02s after repeated experiments.

Recognition result by back propagation neural network

The data source of back propagation (BP) neural network is the same as the former two schemes. The structure of the BP network features is introduced as follows. Firstly, the data is modularization processed, eight input neurons is set by eight consecutive braking information which are also the characteristics of the eight dimensions; and then two output neurons are built and coded. The first kind of coding is “01”, the second kind of coding is “10” and the third kind of coding is “11”. Three output neuron represents three driving style. Due to a hidden layer which can fit various kinds of classification, the BP network is designed as a layer. The number of internal nodes in the hidden layer is namely 30 hidden layer neurons according to 2 layers the design of the number of input layer.

The vector data of 30 × 6 is taken as the training sample. The BP neural network is modeled with the help of the MATLAB neural network toolbox. A random number of weight threshold (−1, 1) generate and sigmoid function is used in the activation function.

Training was stopped when all the training samples of mean square error (MSE) and less than 0.01 or the number of iterations reached 1000 times. The gradient descent method is used to correct each sample weight, the update rate of the decision value is determined by learning factor and the learning factor is set as 0.8 after repeated experiments. The optimization method is to add momentum in this paper; the momentum factor is set as 0.9. The BP neural network is used to judge the rest of the 30 × 1 vector data. Driving style model identification accuracy is 77.27% and the recognition time is 1.25s.

Recognition result by support vector machine

In this paper, the problem to be solved is about one class classification. Support vector machine involves two classification methods. Thus, the three categories need to be divided into three binary classification problems, which can use a directed acyclic graph (DAG) to describe the relationship of these classifiers [31]. Driving style is represented by A, B, C. At the time of extraction of the training set, a corresponding vector is extracted as set respectively and C corresponding vector is extracted as a negative set; a corresponding vector is as positive set and B corresponding vector as negative set; when B corresponding vector is the positive set, the C corresponding vector is the negative set. The three training sets are trained respectively, and three training results files are obtained, such as A (aggressive) and B (moderate), C (mild).

At present, each support vector machine classifier training is under the help of the modeling forecast simulation platform (CMSVM) that more widely used [32], three support vector machine are needed: AvsC-SVM, AvsB-SVM, the BvsC-SVM.

In order to optimize the model, all data samples need to randomly be divided into three parts: the training sample subset used for modeling, the experimental sample subset used to optimize the model parameters and sample subset used to test model generalization of the test [32]. The three sample set are input into CMSVM in sequence, and the recognition accuracy of AvsC-SVM, AvsB-SVM and BvsC-SVM are 83.32%, 81.84% and 83.33%, respectively. Recognition runtime is 0.15 s, 0.18 s and 0.17 s, so the recognition accuracy based on CMSVM is a product of three support vector machine recognition accuracy is 82.83%, and the model runtime is the sum of three support vector machine: 0.15s+max(0.18s, 0.17s) = 0.32s.