CN112215487A - A method of vehicle driving risk prediction based on neural network model - Google Patents

Abstract

The invention discloses a vehicle running risk prediction method based on a neural network model, which comprises the following steps: collecting vehicle running data to form vehicle historical data; step two: extracting the characteristics of the historical data of the vehicle by adopting a context time window to form statistical characteristics; step three: extracting the statistical characteristics, and step four: dividing the extraction result data in the third step, and the fifth step: constructing a neural network; construction of LSTM encoder-1 DCNN-LSTAn M decoder network architecture; step six: unlabeled dataset is denoted as { X_U}; the tagged data set is divided into a training set and a testing set, wherein the training set is marked as S_L＝{X_L，Y_L}; using tagged data sets S_LPre-training a neural network; then entering a self-learning stage to carry out a self-learning process on the set Y_PLPredicting, and directly taking the predicted value as a real label; after the completion, all the labels of the label-free data and the trained network model are obtained.

Description

A method of vehicle driving risk prediction based on neural network model

technical field

The present invention relates to the field of machine learning, and more particularly, to a method for predicting vehicle driving risk based on a neural network model.

Background technique

According to the World Health Organization's statistics in 2018, 1.35 million people are killed in traffic accidents every year, and the damage caused by road traffic collisions can reach 3% of the gross domestic product of most countries. In addition, about 94% of traffic-related fatalities are caused by drivers, and driver's improper driving behavior has become a major factor in traffic accidents. These driving behaviors are often caused by the driver's poor perception of the surrounding environment, rash or aggressive judgments and decisions, and inappropriate vehicle driving maneuvers. Driving risk assessment is based on the analysis of various driving characteristics (including driver, vehicle and surrounding environment) at the current and past moments, and gives the possibility of current vehicle collision or other traffic accidents. Evaluate and predict vehicle driving safety, and provide timely feedback to drivers to improve vehicle driving safety and reduce traffic accidents. Therefore, it is essential to evaluate and predict vehicle driving risks.

However, in the driving risk assessment task, labeling the driving data with risk is a rather challenging task. If the unsupervised learning method is used to classify the data, the obtained results may not be classified strictly according to the level of risk, and it is difficult to achieve satisfactory results in terms of accuracy. In addition, the task of driving risk assessment needs to deal with massive high-dimensional, time-series and unbalanced vehicle driving data. Finally, driving risk assessment has high requirements on accuracy, and it is difficult to accept the situation that high risk is judged as no risk. To sum up, it is a huge challenge to conduct a comprehensive, accurate and efficient assessment of driving risk.

After retrieval, Chinese invention patent CN201711234967.6 discloses a driving risk warning method and device, which uses a pre-established BP network to classify corresponding road sections into high-risk sections or low-risk sections; The above-mentioned vehicle issues a warning when driving to a high-risk road section; CN201910574565.3 discloses a method for analyzing the risk of illegal driving of vehicles based on the Beidou positioning system. The driving risk score generates a risk analysis report of illegal driving, reminds and urges the driver to improve driving behavior, and plays an early warning and assessment role for the driver's own driving behavior. However, the above methods do not fully consider the massive, high-dimensional, time-series, and unbalanced vehicle driving data of the vehicle driving process, and it is difficult to achieve refined driving risk prediction, which is poor in accuracy.

SUMMARY OF THE INVENTION

The present invention designs and develops a method for predicting vehicle driving risk based on a neural network model, which can only manually label a small part of data, and automatically learn potential features, establish a neural network model, and predict a certain period of time in the future. value at risk.

Another object of the present invention is a neural network model training method for vehicle driving risk prediction, which can manually label only a small part of data, and automatically learn potential features to build a neural network model.

A vehicle driving risk prediction method based on a neural network model,

Step 1: Collect vehicle driving data to form vehicle historical data;

Step 2: using the context time window to extract features from the vehicle historical data to form statistical features;

Step 3: extracting the statistical features, including: the type of the vehicle, the length and width of the vehicle; the steering entropy value; the parameters reverse collision time TTC ^-1 , reverse headway THW ^-1 and reverse head distance DHW ^-1 ; The duration of the vehicle pressing the dotted line when there is no intention to change lanes, the duration of the vehicle pressing the line, and the duration of the vehicle driving out of the solid line; the local traffic flow density, the local speed difference and the local acceleration difference.

Step 4: Divide the extraction result data of Step 3, and randomly extract no more than 5% of the data for labeling to form a labeled data set; the remaining data is an unlabeled data set, which is used for unlabeled training and semi-supervised learning. test;

Step 5: Build a neural network; build an LSTM encoder-1DCNN-LSTM decoder network architecture;

Step 6: The unlabeled data set is recorded as {X _U }; the labeled data set is divided into training set and test set, and the training set is recorded as _SL ={X _L , Y _L }; the labeled data set _SL is used Pre-training the neural network;

Then enter the self-learning stage, use the unlabeled data set {X _U } to use the pre-trained network to generate pseudo-labels {Y _P }; for each generated pseudo-label, there is a certain degree of confidence ε, the confidence degree Compared with the threshold ε _th , the set greater than the threshold is denoted as SP ^t ₌ {X _Uh ^t ,Y _Ph ^t }, the set smaller than the threshold is denoted as {Y _PL ^t }, and t is the number of iterations; for the set {Y _Ph t ^t }, according to the manifold hypothesis, the pseudo-label is considered to be the real label; the set _SL and _SP ^t are combined to form a new set _SL ^t , which is then used in the training of the network; for {X _U ^t } , use the retrained network to regenerate pseudo-labels; predict the unlabeled data set {X _U ^mst } in the final stage of self-learning, and use the predicted value directly as its true label; after completion, all unlabeled data sets are obtained. The labels of the data and the trained network model.

As a preference, the penalty function of the neural network is:

where the probability mass function f of the distribution P can be defined as

AOBC(0,k)=AOBC(k)

为预测值。Among them, N(t,k)=| _SL,k ^t | in AOBC(k), N(0,k)=| _SL,k ⁰ | in OBC(k); N is in the mini-batch The number of data, m is the number of categories, y _ik is the real value,

is the predicted value.

As an option, the limit value penalty function is also included, as shown in the following formula:

Among them, ev is the limit value.

As a preference, the steering entropy value SRE:

As a preference, the inverse collision time TTC ^-1

As a preference, the specific calculation formula of the local traffic flow density is as follows:

其中σ_x与σ_y通过下式定义:Among them, X _j =(x _j ,y _j ) ^T is the vehicle of interest, μ=(x, y) ^T is the center coordinate of the target vehicle,

where σ _x and σ _y are defined by:

σ _x =|v _x |+k ₁ L

σ _y =|v _y |+k ₂ W

Among them, v _x , v _y are the lateral and longitudinal speeds of the vehicle, and k ₁ and k ₂ are compensation factors.

As a preference, the local velocity difference is calculated as follows:

As a preference, the local acceleration difference is calculated as follows:

The beneficial effects of the present invention:

The method first extracts the characteristics of target vehicle driving, vehicle-road interaction, and local traffic conditions. Feature extraction is performed on the massive vehicle driving data, and a small part is manually labeled, so as to obtain a small data set with labels and a large data set without labels. A convolutional neural network combining a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory network (LSTM) is built, and an adaptive over-balanced cross-entropy is used as the loss function of the neural network. The above neural network is embedded into the semi-supervised learning framework, and the final label results and trained network are obtained after pre-training, self-learning and fine-tuning on the two datasets. Finally, a limit penalty block is added to refine the model.

A cost-sensitive semi-supervised deep learning method is used to analyze the vehicle driving data to obtain the current and future driving risk scores, and the score is a continuous value between 0 and 3, which makes the risk assessment more refined. This method can be applied to the driving risk warning system in ADAS to give timely feedback to the driver.

Good results can be obtained by using only a small part of labeled data, which greatly solves the labeling problem of a large amount of unlabeled data. The adaptive over-balanced cross-entropy loss function is adopted, which greatly improves the training performance of semi-supervised deep learning with unbalanced categories. The loss function can make the network in an over-balanced state during the entire training process, which can effectively improve the detection accuracy of high-risk data. This method can be applied to other related similar scenarios.

A local traffic condition descriptor is proposed to describe the traffic conditions, speed and acceleration differences around the target vehicle, and a descriptor is given to simplify the description of the surrounding scene. The descriptor can also be used in other similar fields.

Description of drawings

FIG. 1 is a schematic diagram of the context window of the present invention.

FIG. 2 is a schematic diagram of a vehicle of interest based on a target vehicle according to the present invention.

FIG. 3 is a deep learning network model of the present invention.

Figure 4 is the semi-supervised learning framework of the present invention.

Fig. 5 is the general framework of the present invention.

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings, so that those skilled in the art can implement it with reference to the description.

The invention proposes a driving risk assessment and prediction method based on adaptive cost-sensitive semi-supervised deep learning. The method first extracts the characteristics of target vehicle driving, vehicle-road interaction, and local traffic conditions. The feature extraction is carried out on the massive vehicle driving data, and a small part is manually labeled, so as to obtain a small data set with labels and a large data set without labels. A convolutional neural network combining a one-dimensional convolutional neural network (1D-CNN) and a long short-term memory network (LSTM) is built, and an adaptive over-balanced cross-entropy is used as the loss function of the neural network. The above neural network is embedded into the semi-supervised learning framework, and the final label results and trained network are obtained after pre-training, self-learning and fine-tuning on the two datasets. As a preference, finally, a limit value penalty module is added to refine the model. This method can manually label only a small part of the data, and automatically learn the potential features, and can predict the risk value in a period of time in the future, so as to provide timely feedback to the driver.

Step 1: Convert and count the vehicle driving history data. Directly using the original vehicle driving data as input will result in too sparse features and too low feature order, thus increasing the difficulty of training. The context time window (as shown in Figure 1) is used to extract features from a fixed window. The selected window width is W _w =3 seconds, and the sampling frequency per second is not less than 10 Hz. A total of 5 seconds of historical data is selected, and the overlap rate of each time window is Ov=66.67%, that is, 2 seconds. Therefore, a total of 3 time windows are obtained, and all data in the window are counted.

Step 2: Extract statistical features. A total of 47-dimensional features are extracted, including basic vehicle information, vehicle driving and interaction information, and vehicle surrounding environment information. Each statistical data has mean, maximum, minimum and variance statistics, so each data has 4 statistics.

A) Basic information of the vehicle:

Type of vehicle, length L and width W of the vehicle.

B) The driving characteristics of the target vehicle:

Define the longitudinal travel direction of the vehicle as the x-axis and the lateral travel direction as the y-axis. Select the horizontal and vertical speed and acceleration data of the vehicle, and extract the mean, maximum, minimum, and variance of the vehicle's horizontal and vertical speed and acceleration data in each time window as input features, with a total of 16-dimensional features. When the driver is driving steadily and normally, the steering behavior will be relatively smooth, but the steering behavior may be chaotic when the driver is fatigued or distracted. Steering entropy is used to quantitatively indicate the driver's directional handling characteristics to reflect the degree of steering stability and driving safety.

The heading angle in the time window is recorded as θ=(θ ₁ , θ ₂ , θ ₃ , L, θ _m ), and m is the number of data in the time window. Use Taylor's second-order expansion for a given time to predict the next corner as follows:

which is:

Error function definition:

为第n时刻θ_n的预测值。设置一个α值，作为一种优选α取值0.035；使90％数据的预测误差落在-α与α之间。将这个预测误差区间划分为9段，即共9个区间，然后求取各段分布频率，即p_i，i＝1,2,..9。运用下式计算转向熵值SRE：In the formula,

is the predicted value of θ _n at the nth time. Set an α value, as a preferred α value of 0.035; make the prediction error of 90% of the data fall between -α and α. This prediction error interval is divided into 9 sections, that is, a total of 9 sections, and then the distribution frequency of each section is obtained, that is, p _i , i=1,2,..9. Use the following formula to calculate the steering entropy value SRE:

C) Interaction characteristics between vehicles:

The speed of the target vehicle is denoted as _{ve , the longitudinal coordinate x e ,} the speed of the preceding vehicle is denoted as v _p , and the longitudinal coordinate x _p . THW head-to-head distance refers to the time interval between the end of two consecutive vehicles passing through a certain section in the vehicle queue driving on the same lane; DHW head-to-head distance refers to the vehicle queue driving on the same lane. , the distance interval between the front ends of two consecutive vehicles passing through a certain section; TTC collision time, that is, the collision time of the two vehicles when both the rear vehicle and the front vehicle are traveling at the current speed. The calculation formulas of THW, DHW, and TTC are as follows.

DHW=(x _p -x _e )

However, taking TTC as an example, the following two situations may occur when using TTC directly: when the speed of the following vehicle is lower than the speed of the preceding vehicle, that is, the following vehicle can never catch up with the preceding vehicle at the current relative speed, and the TTC is negative; If the speed of the car behind is only a little faster than the speed of the car in front, that is, it takes a long time for the car behind to catch up with the car in front at the current relative speed, and the TTC is a large positive value. Therefore, the value range of TTC is theoretically (-∞, +∞), and the real high-risk TTC interval is very small. Therefore, directly using TTC as an input may lead to a decrease in the accuracy of the model. In addition, the high-risk TTC interval may be compressed again through subsequent feature standardization. Therefore, the inverse collision time TTC ^-1 is used, as shown in the following equation.

In addition, for negative TTCs, in order to simplify the features, all negative TTCs are uniformly assigned a sufficiently large positive value TTC ^-1 _max (usually 50 seconds), that is, the target vehicle will collide with the preceding vehicle after a long time. This method of eliminating and replacing irrelevant values can reduce the negative confounding effect on the model and improve the training accuracy. The formula is expressed as follows:

其中，THW^-1 _max＝10；

Wherein, THW ^-1 _max =10;

其中，DHW^-1 _max＝200；

Wherein, DHW ^-1 _max =200;

After processing, the safe TTC value is compressed to between 0 and a small value, while the high-risk TTC value is enlarged, and the value range is also enlarged, thereby improving the accuracy of the model. In the same way, THW and DHW are transformed into reverse headway time distance THW ^-1 and reverse headway distance DHW ^-1 to enlarge the value range of high-risk THW and DHW. Finally, the maximum value, mean and variance are selected as output features, with a total of 9 dimensions.

When there is an intention to change lanes, the driver's perception of changing lanes is very important. If you do not have a good perception of the current lane and the lane to be changed, and directly change the lane, it may bring a higher driving risk. Therefore, it is selected whether there are vehicles running side by side in the lane to be changed when the lane change intention appears, and the 3-dimensional features of the maximum TTC ^-1 and THW ^-1 of the vehicle in front of and behind the lane to be changed are used as input.

D) Interaction between vehicle and road

The length of time that the vehicle presses the dotted line when there is no intention to change lanes, the length of time that the vehicle presses the line, and the time that the vehicle drives out of the solid line.

E) Local Traffic Situation Descriptor

In order to better describe other running vehicles, roads, obstacles, etc. during the driving process of the vehicle, corresponding descriptors are required for description. First, consider factors such as vehicles and obstacles driving around the target vehicle. A Gaussian weight-based local traffic density descriptor (LTD) is proposed. Taking the target vehicle as the benchmark, consider the front vehicle, the left front vehicle, the right front vehicle, the left vehicle, the right vehicle, the left rear vehicle, the right rear vehicle, and the rear vehicle as the vehicle of interest (as shown in Figure 2, if there is no such vehicle) If the vehicle is a vehicle, it is recorded as 0), and the contribution of the vehicle of interest to the traffic density of the target vehicle is calculated.

The specific calculation formula is as follows:

其中σ_x与σ_y通过下式定义:Among them, X _j =(x _j , y _j ) ^T is the vehicle of interest, x _j , y _j are the abscissa and ordinate of the vehicle of interest respectively; μ = (x, y) ^T is the center coordinate of the target vehicle,

where σ _x and σ _y are defined by:

σ _x =|v _x |+k ₁ L

σ _y =|v _y |+k ₂ W

Wherein, v _x , v _y are the lateral and longitudinal speeds of the vehicle, and k ₁ and k ₂ are compensation factors. As an example, k ₁ =0.625 and k ₂ =1.25.

So far, each piece of data in each time window forms a corresponding local traffic density descriptor. In addition, using the local traffic density as the weight, the local speed difference is solved, and the local speed difference descriptor is obtained, which is used to describe the speed difference (LVD) of the surrounding vehicles with the target vehicle as the benchmark. Similarly, solve for the local acceleration difference (LAD). As shown in the following formula:

v _e is the speed of the target vehicle, as mentioned above, v _j is the speed of the vehicle of interest, a _e is the acceleration of the target vehicle, a _j is the acceleration of the vehicle of interest, _Ni is the number of vehicles of interest, the maximum value is 8, such as As shown in Figure 2, if there is no corresponding vehicle, it will be reduced, at least 0, that is, there is no vehicle of interest, that is, there is no vehicle around.

As for the obstacle, think of it as a vehicle traveling at zero speed. Finally, the statistical indicators such as the mean, maximum, minimum, and standard deviation of the above-mentioned three descriptors are obtained respectively, with a total of 12-dimensional features.

The following table is the characteristic statistics:

Table 1 Statistical input features

Step 3: Divide all the statistical data, randomly extract no more than 5% of the data for labeling, and use the remaining data for unlabeled training and testing of semi-supervised learning. The risk score at the current moment is evaluated according to the upper quantile value of 2 to 5% (that is, the data exceeds 95 to 98% of the value of all data), and the marked values are 0 (good), 1 (fair), 2 ( poor), 3 (very poor). The labeling method is to label the data, and the final result is averaged and rounded to an integer.

Step 4: Build the neural network. Building the LSTM encoder-1DCNN-LSTM decoder network architecture is shown in the figure. Among them, Convolution is 1D-CNN, MaxPooling is the maximum pooling layer, Dropout is the discarding layer, FC is the fully connected layer, and Softmax is the activation layer. Except for the last layer, all other convolutional layers and fully connected layers have activation functions of ReLU.

The details are as follows: Build a neural network. Build the LSTM encoder-1DCNN-LSTM decoder network architecture as shown. First, the statistics are embedded, and the data is mapped to the LSTM encoder through an embedding layer. The number of input nodes in the embedding layer is the feature dimension, and the output is 128. The LSTM encoder has a total of 128 hidden units. Finally, the tensor of the last hidden unit is taken for activation and deformation, and a one-dimensional tensor is obtained. The tensor undergoes three 1D convolutions - activation - 1D pooling - random discarding to obtain a convolved tensor. The convolution process is mainly to extract deep patterns between hidden features under different time series, which can better reflect risk information. The number of convolutional channels of the three convolutional layers is 64, 128, and 256, respectively. After the output of the last layer is expanded, the first branch connects two fully connected layers and Softmax layer to obtain the current risk score, and the number of hidden units is 128, 64, and 4, respectively. The other branch is connected to an LSTM decoder, decodes the current hidden features, and obtains the prediction of future risk value. The environment used by the computer is Win10, the software name is Python3.7, the deep learning framework is Keras2.2.4, and the backend is Tensorflow1.14

The LSTM encoder encodes the input and adopts the Dropout method, that is, in the process of deep learning network training, for the neural network unit, it is randomly discarded from the network according to a certain probability to reduce overfitting. A certain probability is 0.2; next is a 1DCNN (one-dimensional convolutional neural network) to reduce the error caused by the convolution between the features of the 2DCNN. With the maximum pooling Maxpooling, the features that contribute greatly to the risk value can be selected. After three convolution-pooling-Dropout, the first branch connects two fully connected layers and Softmax layers to get the current risk score. The other branch taps into an LSTM decoder to decode the current high-level features to obtain predictions of future risk values.

Step 5: Embed the above network into a semi-supervised learning architecture, as shown in Figure 4. The unlabeled dataset is denoted as {X _U }. The labeled data set is divided into a training set and a test set, wherein the training set is denoted as S _L ={ _XL , Y _L }. First, the network is pre-trained with the labeled dataset _SL . The Early-Stopping method is used in the training process, that is, when a certain monitoring value does not improve the network performance after patience iterations, the network automatically terminates, and the network weight with the best performance before the stop is returned. The monitoring value selects loss, that is, the loss will not decrease after patience times. The choice of patience is relatively large here, the reason is to allow the network to learn enough from a small amount of data.

Then enter the self-learning stage, and use the unlabeled dataset {X _U } to use the pre-trained network to generate pseudo-labels {Y _P }. For each generated pseudo-label, there is a certain confidence ε (the pre-trained network is directly obtained after passing through the Softmax layer). Comparing the confidence with the threshold ε _th , the set greater than the threshold is denoted as S _P ^t ={X _Uh ^t ,Y _Ph ^t }, and the set (t is the number of iterations) less than the threshold is denoted as {Y _PL ^t }, t is the number of iterations. For the set {Y _Ph ^t }, according to the manifold hypothesis, the pseudo-label is considered to be the real label. The set _SL and _SP ^t are merged to form a new set _SL ^t , which is then used in the training of the network. For {X _U ^t }, (unlabeled dataset not accepted by an iterative process) pseudo-label regeneration using the retrained network. The process is shown in the following formula:

S represents the training data set, L represents the label, t represents the number of iterations, X represents the input feature, Y represents the label, U represents the unlabeled, P represents the pseudo-label, h represents greater than the threshold ε _th , l represents less than the threshold ε _th , m _st is the total number of iterations. Correspondingly, _SL ^t is the labeled training data set at the t-th iteration. The rest are the same.

When the value of i is smaller, the number of iterations of the threshold decreases successively, because the lower the confidence, the easier it is to introduce noise. Similarly, the number of patient selections will also decrease as the confidence decreases.

As an option, fine-tuning of the model can also be performed. The fine-tuning process sets all CNN and LSTM layers as non-trainable, and only fine-tunes the fully connected layers. The reason is that the feature extraction layer has already been trained and the latent features of the unlabeled dataset have been learned, and it is only necessary to adjust the weights of the fully connected layer. Predict the unlabeled data set {X _U ^mst } in the final stage of self-learning, and use the predicted value directly as its true label. After completion, all unlabeled data labels are obtained, and the trained network model is obtained.

Step 6: Set up the neural network loss function. The performance of the network is degraded due to class imbalance, i.e., high-risk data is always far less than risk-free data. Therefore, the loss function is set to compensate the category imbalance. The multi-class cross-entropy loss function (CE) is selected as the basic loss function as the multi-class loss function, as shown in the following formula:

分布P下的数学期望；N为mini-batch中数据的数量，m为类别数，y_ik为真实值，

为预测值。对上述损失函数添加过权重(OBC)，即：where E _yi:P is a random variable _yi with distribution P, where _yi,k

Mathematical expectation under distribution P; N is the number of data in the mini-batch, m is the number of categories, y _ik is the true value,

is the predicted value. Overweight (OBC) is added to the above loss function, namely:

in,

However, during the iterative process of semi-supervised learning, the amount of labeled data is constantly changing, but each category varies unevenly. Therefore, it is necessary to make adaptive conditions for this point, as shown in the following formula:

N(t,k)=|S _L ^t |

AOBC(0,k)=OBC(k)

Among them, N=|S _L ⁰ | in OBC(k).

In the above formula || represents the number of elements in the set; AOBC(t,k) is the weight of the k-th type of loss with the change of the number of iterations t.

As a preference, step 7: add a limit value penalty module to the above network. In order to make up for the very special serious classification errors of the neural network, a limit value penalty module needs to be introduced. The module uses fuzzy logic, and when certain values are closer to the parameters at the moment of collision, the higher the penalty is based on. As shown in the following formula:

Among them, ev is the limit value; generally, the data exceeding 99.99% of all values is taken as the limit value.

As an option, step 8: pre-training, self-learning and fine-tuning of the network. The optimizer of the network is selected as the Adam optimizer, the learning rate is 10 ^-3 , the decay is 10 ^-6 , and the ε _th is selected as 0.999999, 0.99999, 0.9999, 0.999, 0.99, 0.95, 0.9. After fine-tuning, all the data labels and the trained network are obtained. Each output label corresponds to a confidence level, and the current risk score is obtained by multiplying the values of all labels by the confidence level. The trained network can then be used directly (ie, to evaluate new data without semi-supervised methods). The overall framework is shown in Figure 5.

Although the embodiment of the present invention has been disclosed as above, it is not limited to the application listed in the description and the embodiment, and it can be applied to various fields suitable for the present invention. For those skilled in the art, it can be easily Therefore, the invention is not limited to the specific details and illustrations shown and described herein without departing from the general concept defined by the appended claims and the scope of equivalents.

Claims (8)

1. A vehicle running risk prediction method based on a neural network model is characterized in that,

the method comprises the following steps: collecting vehicle running data to form vehicle historical data;

step two: extracting the characteristics of the historical data of the vehicle by adopting a context time window to form statistical characteristics;

step three:extracting the statistical features, including: the type of vehicle, the length and width of the vehicle; a steering entropy value; time to counter collision (TTC) of parameter^-1Headway time THW^-1And reverse headwear distance DHW^-1(ii) a When the vehicle has no lane changing intention, pressing the dotted line for a long time, compacting the line for a long time, and driving the vehicle out of the solid line for a long time; local traffic flow density, local speed differential, and local acceleration differential.

Step four: dividing the extraction result data in the step three, and randomly extracting no more than 5% of data from the extraction result data to carry out labeling to form a labeled data set; the rest data is a label-free data set and is used for label-free training and testing of semi-supervised learning;

step five: constructing a neural network; constructing an LSTM encoder-1 DCNN-LSTM decoder network architecture;

step six: unlabeled dataset is denoted as { X_U}; the tagged data set is divided into a training set and a testing set, wherein the training set is marked as S_L＝{X_L，Y_L}; using tagged data sets S_LPre-training a neural network;

then entering a self-learning stage, and setting the unlabeled data set { X_UApply the pretrained network to generate a pseudo label (Y)_P}; each generated pseudo label is provided with a certain confidence coefficient epsilon, and the confidence coefficient is compared with a threshold value epsilon_thComparing, the set greater than the threshold is marked as S_P ^t＝{X_Uh ^t,Y_Ph ^tThe set smaller than the threshold is denoted as { Y }_PL ^tT is iteration times; for the set { Y_Ph ^tAccording to the manifold assumption, the false label is considered as a real label; will gather S_LAnd S_P ^tMerge to form a new set S_L ^tThen the data is used for training the network; for { X_U ^tRe-generating the pseudo label by applying the retrained network; for data set { X) not labeled in the final stage of self-learning_U ^mstPredicting, and directly taking a predicted value as a real label of the predicted value; after completion, all the labels and trainings of the label-free data are obtainedAnd (5) training a network model.

2. The neural network model-based vehicle driving risk prediction method according to claim 1, wherein the penalty function of the neural network is:

wherein the probability mass function f of the distribution P can be defined as

AOBC(0,k)＝AOBC(k)

Wherein N (t, k) ═ S in aobc (k)_L,k ^tN (0, k) ═ S in obc (k)_L,k ⁰L, |; n is the number of data in the mini-batch, m is the number of categories, y_ikIn order to be the true value of the value,

is a predicted value.

3. The neural network model-based vehicle driving risk prediction method of claim 2, further comprising a limit penalty function as shown in the following formula:

wherein ev is a limit value.

4. The neural network model-based vehicle travel risk prediction method of claim 1, characterized in that the steering entropy value SRE:

5. the neural network model-based vehicle travel risk prediction method of claim 4, wherein the time to collision TTC^-1

6. The neural network model-based vehicle travel risk prediction method according to claim 1 or 5, characterized in that the local traffic flow density is specifically calculated according to the formula:

wherein, X_j＝(x_j,y_j)^TFor the vehicle of interest, μ ═ x, y^TIs the coordinates of the center of the target vehicle,

wherein sigma_xAnd σ_yIs defined by the formula:

σ_x＝|v_x|+k₁L

σ_y＝|v_y|+k₂W

wherein v is_x,v_yIs the transverse and longitudinal speed of the vehicle, k₁And k is₂Is a compensation factor.

7. The neural network model-based vehicle travel risk prediction method of claim 1 or 6, wherein the local speed difference is calculated as follows:

8. the neural network model-based vehicle travel risk prediction method according to claim 1 or 7, characterized in that the local acceleration difference is calculated as follows:

Images