CN112215487A - Vehicle driving risk prediction method 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

Vehicle driving risk prediction method based on neural network model

Technical Field

The invention relates to the field of machine learning, in particular to a vehicle driving risk prediction method based on a neural network model.

Background

According to the statistics of the world health organization in 2018, 135 thousands of people are lost due to traffic accidents every year, and the loss caused by road traffic collision can reach 3% of the total domestic production value of most countries. In addition, about 94% of traffic-related death accidents are caused by drivers, and improper driving behavior of drivers becomes a major factor in traffic accidents. These driving behaviors are often caused by poor driver perception of the surrounding environment, misapprovals or aggressive decisions and decisions, inappropriate vehicle driving maneuvers. The driving risk assessment is to analyze various driving characteristics (including drivers, vehicles and surrounding environments) at the current time and the past time to give the possibility of collision of the current vehicle or other traffic accidents. The vehicle driving safety is evaluated and predicted, and the feedback is timely carried out on the driver, so that the vehicle driving safety is improved, and further, traffic accidents are reduced. Therefore, it is essential to evaluate and predict the vehicle running risk.

However, tagging risky travel data is a challenging task in a travel risk assessment task. If the data is classified by using the unsupervised learning method, the obtained result may not be strictly classified according to the risk level, and it is difficult to achieve a satisfactory result in terms of accuracy. Furthermore, the driving risk assessment task needs to face high-volume, time-series, category-imbalanced vehicle driving data. Finally, the driving risk assessment has high requirements on accuracy, and it is difficult to accept the case where a high risk is determined to be risk-free. In summary, there is a great challenge to comprehensively, accurately and efficiently evaluate the driving risk.

Through retrieval, the Chinese invention patent CN201711234967.6 discloses a driving risk warning method and device, which classifies the corresponding road sections into high risk road sections or low risk road sections by using a pre-established BP network; sending warning information to the vehicle to control the vehicle to send a warning when the vehicle runs to a high-risk road section; CN201910574565.3 discloses a vehicle illegal driving risk analysis method based on a Beidou positioning system, which is characterized in that a risk score of illegal driving of a vehicle is accurately calculated, a risk analysis report of illegal driving is generated by a user according to the risk score of illegal driving of the vehicle, a driver is reminded and urged to improve driving behaviors, and the functions of early warning and checking the driving behaviors of the driver are achieved. However, the above method is not thorough in consideration of vehicle driving data with high dimensionality, time sequence and unbalanced category in the vehicle driving process, and is difficult to achieve refined driving risk prediction, and poor in accuracy.

Disclosure of Invention

The invention designs and develops a vehicle running risk prediction method based on a neural network model, which can manually label only a small part of data, automatically learn potential characteristics, establish the neural network model and predict a risk value in a future period of time.

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 running risk prediction method based on a neural network model,

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 the completion, all the labels of the label-free data and the trained network model are obtained.

Preferably, 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.

Preferably, a limit penalty function is also included, as shown in the following formula:

wherein ev is a limit value.

As a preference, the steering entropy value SRE:

preferably, the time to collision TTC^-1

Preferably, the specific calculation formula of the local traffic flow density is as follows:

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.

Preferably, the local velocity difference is calculated as follows:

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

the invention has the following beneficial effects:

the method firstly extracts the characteristics of target vehicle running, vehicle-road interaction, local traffic conditions and the like. And (3) carrying out feature extraction on massive vehicle driving data, and manually labeling a small part of the vehicle driving data to obtain a small labeled data set and a large unlabeled data set. A convolutional neural network combining a one-dimensional convolutional neural network (1D-CNN) and a long-term memory network (LSTM) is built, and self-adaptive over-balanced cross entropy is adopted as a loss function of the neural network. The neural network is embedded into a semi-supervised learning framework, and a final label result and a trained network are obtained after pre-training, self-learning and fine-tuning are carried out on the two data sets. Finally, a limit penalty module is added to refine the model.

A cost-sensitive semi-supervised deep learning method is adopted to analyze vehicle driving data to obtain current and future driving risk scores, and the scores are continuous values between 0 and 3, so that risk assessment is more precise. The method can be applied to a driving risk warning system in the ADAS, so that timely feedback is given to a driver.

Good results can be obtained only by using a small part of label data, and the problem of labeling of a large amount of label-free data is greatly solved. The self-adaptive over-balanced cross entropy loss function is adopted, and the training performance of the semi-supervised deep learning with unbalanced category is greatly improved. The loss function can enable the network to be in an over-balanced state in the whole training process, and the state can effectively improve the detection accuracy of high-risk data. The method can be applied to other related similar scenes.

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

Drawings

FIG. 1 is a diagram illustrating a contextual window of the present invention.

FIG. 2 is a schematic diagram of a target vehicle-based vehicle of interest in accordance with the present invention.

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

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

Fig. 5 is an overall framework of the invention.

Detailed Description

The present invention is further described in detail below with reference to the attached drawings so that those skilled in the art can implement the invention by referring to the description text.

The invention provides a driving risk assessment and prediction method based on self-adaptive cost-sensitive semi-supervised deep learning. The method firstly extracts the characteristics of target vehicle running, vehicle-road interaction, local traffic conditions and the like. And (3) carrying out feature extraction on massive vehicle driving data, and manually labeling a small part of the vehicle driving data to obtain a small labeled data set and a large unlabeled data set. A convolutional neural network combining a one-dimensional convolutional neural network (1D-CNN) and a long-term memory network (LSTM) is built, and self-adaptive over-balanced cross entropy is adopted as a loss function of the neural network. The neural network is embedded into a semi-supervised learning framework, and a final label result and a trained network are obtained after pre-training, self-learning and fine-tuning are carried out on the two data sets. As a preference, a limit penalty module is added to refine the model. The method can manually label only a small part of data, automatically learn potential characteristics, and predict risk values in a future period of time, so as to timely feed back to a driver.

Step 1: and converting and counting the vehicle running history data. Because the original vehicle driving data is directly used as input, the characteristics are too sparse, and the characteristic order is too low, so that the training difficulty is increased. The extraction of features is performed on a fixed window using a contextual time window (as shown in fig. 1). Width of the selection window is W_wThe sampling frequency per second is not less than 10Hz at 3 seconds. A total of 5 seconds of historical data were selected, and the overlap ratio Ov was 66.67%, i.e., 2 seconds, for each time window. A total of 3 time windows are obtained and all data within the windows are counted.

Step 2: and extracting the statistical characteristics. The 47-dimensional features are extracted in total and include vehicle basic information, vehicle traveling and interaction information, vehicle surrounding environment information, and the like. There are mean, maximum, minimum, variance statistics per statistical data, so there are 4 statistics per data.

A) Basic information of the vehicle:

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

B) Target vehicle travel characteristics:

the longitudinal driving direction of the vehicle is defined as an x axis, and the transverse driving direction is defined as a y axis. And selecting the speed and acceleration data of the vehicle in the transverse direction and the longitudinal direction, extracting the mean value, the maximum value, the minimum value and the variance of the speed and acceleration data of the vehicle in the transverse direction and the longitudinal direction in each time window as input features, and totaling 16-dimensional features. The steering behavior of the driver is smoother during steady and normal driving, but the steering behavior of the driver may be confused during fatigue, distraction, etc. The steering entropy is used for quantitatively indicating the direction control characteristics of the driver so as to reflect the steering stability and the driving safety.

Recording the course angle in the time window as theta ═ theta₁,θ₂,θ₃,L,θ_m) And m is the number of data in the time window. Predicting the next corner by using Taylor second-order expansion in a given time as shown in the following formula:

namely:

error function definition:

in the formula,

is the nth time theta_nThe predicted value of (2). Setting an alpha value as an optimal alpha value of 0.035; the prediction error for 90% of the data is made to fall between-alpha and alpha. Dividing the prediction error interval into 9 segments, i.e. 9 intervals in total, and then obtaining the distribution frequency of each segment, i.e. p_i，i＝1,2,..9。Calculating the steering entropy value SRE using the formula:

C) interaction characteristics between vehicles:

target vehicle speed is noted as v_eLongitudinal coordinate x_eFront vehicle speed is denoted v_pLongitudinal coordinate x_p. THW headway, namely the time interval when two continuous vehicle headwaters pass through a certain section in a vehicle queue running on the same lane; the DHW vehicle-head distance refers to the distance interval between the vehicle head ends of two continuous vehicles passing through a certain section in a vehicle queue running on the same lane; TTC collision time, namely the collision time of the two vehicles when the rear vehicle and the front vehicle both run at the current speed. The calculation formulas of THW, DHW, and TTC are as follows.

DHW＝(x_p-x_e)

However, taking the TTC as an example, the following two situations may occur when the TTC is directly applied: when the speed of the rear vehicle is lower than that of the front vehicle, namely the rear vehicle can never catch up with the front vehicle at the current relative speed, and the TTC is a negative value; the TTC is a large positive value if the rear vehicle speed is only a little faster than the front vehicle speed, i.e., it takes a long time for the rear vehicle to catch up with the front vehicle at the current relative speed. Therefore, the TTC range is theoretically (- ∞, + ∞), and the real high-risk TTC range is very small. Therefore, directly using TTC as input may result in a decrease in model accuracy, and furthermore, the high-risk TTC interval may be compressed again by the subsequent feature normalization process. Therefore, the inverse time to collision TTC is adopted^-1The following formula is shown below.

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

wherein, THW^-1 _max＝10；

Wherein, DHW^-1 _max＝200；

After processing, the TTC value relative to safety is compressed to be between 0 and a very small value, the TTC value with high risk is amplified, and the value range is also amplified, so that the accuracy of the model is improved. Similarly, THW and DHW are converted into inverse headway THW^-1And reverse headwear distance DHW^-1To enlarge the range of high risk THW, DHW. And finally, selecting the maximum value, the mean value and the variance as output characteristics, wherein the output characteristics are 9-dimensional.

When lane changing intentions occur, the perception of the driver of the lane to be changed is very important. If the lane and the lane to be changed are not well sensed, the lane changing operation is directly performed, and high driving risk may be brought. Therefore, whether vehicles running in parallel exist on the lane to be changed when the lane changing intention appears and the maximum TTC of the front vehicle and the rear vehicle of the lane to be changed are selected^-1And THW^-1These 3-dimensional features serve as input.

D) Interaction between vehicle and road

When the vehicle does not have the intention of changing the lane, the broken line is pressed for a long time, the line is pressed for a long time, and the vehicle is driven out of the solid line for a long time.

E) Local traffic situation descriptor

In order to better describe other running vehicles, roads, obstacles and the like in the running process of the vehicle, a corresponding descriptor is needed for description. First, the factors of the vehicle, the obstacle, and the like running around the target vehicle are considered. Local traffic density descriptors (LTD) based on gaussian weights are proposed. With reference to the target vehicle, 8 vehicles, i.e., a front vehicle, a left front vehicle, a right front vehicle, a left vehicle, a right vehicle, a left rear vehicle, a right rear vehicle, and a rear vehicle, are considered as the vehicle of interest (as shown in fig. 2, if there is no vehicle, it is denoted as 0), and the degree of contribution of the vehicle of interest to the flow density of the target vehicle is calculated.

The specific calculation formula is as follows:

wherein, X_j＝(x_j,y_j)^TFor the vehicle of interest, x_j,y_jRespectively the abscissa and the ordinate of the vehicle of interest; mu ═ 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 is₂To compensate for the factor, k is preferably₁＝0.625，k₂＝1.25。

To this end, each piece of data in each time window forms a corresponding local traffic density descriptor. In addition, the local traffic flow density is used as a weight, and the local speed difference is solved to obtain a local speed difference descriptor which is used for describing the running speed difference (LVD) of the surrounding vehicles by taking the target vehicle as a reference. Similarly, the Local Acceleration Difference (LAD) is solved. As shown in the following formula:

v_efor the target vehicle speed, v, mentioned above_jAs the speed of the vehicle of interest, a_eIs a target vehicle acceleration, a_jFor vehicle acceleration of interest, N_iThe maximum value for the number of interested vehicles is 8, and as shown in fig. 2, it is reduced if there is no corresponding vehicle, and is at least 0, i.e., none of the interested vehicles is present, i.e., there is no vehicle around.

As for the obstacle, it can be regarded as a vehicle whose running speed is zero. And finally, respectively calculating the statistical indexes of the 3 descriptors, such as the mean value, the maximum value, the minimum value, the standard deviation and the like, and obtaining 12-dimensional characteristics.

The following table is a characteristic statistic:

TABLE 1 statistical input characteristics

And step 3: dividing all the counted data, randomly extracting no more than 5% of the data from the data to label, and using the rest data for label-free training and testing of semi-supervised learning. And (3) evaluating the risk score of the current moment according to an upper quantile value of 2-5% (namely the data exceeds 95-98% of the value of all the data), wherein the evaluation is marked with values of 0 (good), 1 (general), 2 (poor) and 3 (poor). The data is labeled by a labeling method, and the final result is averaged and rounded to an integer.

And 4, step 4: and constructing a neural network. The LSTM encoder-1 DCNN-LSTM decoder network architecture is constructed as shown in the figure. Wherein, the constraint is 1D-CNN, Max Pooling is the maximum pooling layer, Dropout is the discarding layer, FC is the full link layer, and Softmax is the active layer. All convolutional layers and fully-connected layer activation functions except the last layer are ReLU.

The method comprises the following specific steps: and constructing a neural network. An LSTM encoder-1 DCNN-LSTM decoder network architecture is constructed as shown. First, the statistical data is embedded and the data is mapped to the LSTM encoder through an embedding layer. The number of embedding layer input nodes is the feature dimension and the output is 128. The LSTM encoder has 128 hidden units. And finally, taking the tensor of the last hidden unit for activation and deformation to obtain a one-dimensional tensor. The tensor is subjected to three times of one-dimensional convolution, activation, one-dimensional pooling and random discarding to obtain the tensor after convolution. The convolution process mainly extracts deep patterns among hidden features under different time sequences, and the patterns can better reflect risk information. The number of the convolution channels of the three convolutional layers is 64, 128 and 256 respectively. After the output of the last layer is expanded, the first branch is connected with the two full-connection layers and the Softmax layer to obtain the current risk score, and the number of the hidden units is 128, 64 and 4 respectively. The other branch is connected with an LSTM decoder to decode the current hidden features to obtain the prediction of the future risk value, and the number of the hidden units is 128 LSTM units of the decoder and the full connection layers 128, 64 and 4 respectively. The environment used by the computer was Win10, using the software name python3.7, the deep learning framework keras2.2.4, and the backskend Tensorflow1.14

The LSTM encoder encodes the input, and a Dropout method is adopted, namely in the deep learning network training process, for the neural network unit, the neural network unit is randomly discarded from the network according to a certain probability to reduce overfitting, and a certain probability is preferably 0.2; followed by 1DCNN (one-dimensional convolutional neural network) to reduce the error introduced by the convolution between 2DCNN pair features. Features that contribute significantly to risk values can be selected in conjunction with max pooling Maxpooling. After three times of convolution-pooling-Dropout, the first branch is connected with two full connection layers and the Softmax layer to obtain the current risk score. The other branch accesses an LSTM decoder to decode the current high-level features to obtain a prediction of future risk values.

And 5: the network is embedded into a semi-supervised learning architecture, as shown in fig. 4. 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}. First, a tagged data set S is applied_LThe network is pre-trained. And an Early-Stopping method is adopted in the training process, namely when the network performance is not improved after certain monitoring value is iterated for times through the probability, the network is automatically terminated, and the network weight with the best performance before Stopping is returned. The monitoring value is selected as the loss, namely the loss does not decrease after the probability times. The probability choice is relatively large here because the network is given sufficient learning for a small amount of data.

Then entering a self-learning stage, and setting the unlabeled data set { X_UApply the pretrained network to generate a pseudo label (Y)_P}. For each generated pseudo label, there is a certain confidence coefficient epsilon (the pre-trained network is obtained directly after passing through the Softmax layer). 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 (t is the number of iterations) less than this threshold is denoted as Y_PL ^tAnd t is the iteration times. For the set { Y_Ph ^tAnd according to the manifold assumption, the pseudo label is considered as a real label. Will gather S_LAnd S_P ^tMerge to form a new set S_L ^tAnd then used in the training of the network. For { X_U ^tAnd (an unlabeled data set which is not accepted by an iterative process) carrying out pseudo label regeneration by using the trained network again. The process is shown by the following formula:

s represents a training data set, L represents labeled data, t represents iteration times, X represents input characteristics, Y represents a label, U represents an unlabeled data, P represents a pseudo label, h represents a value greater than a threshold epsilon_thAnd l represents less than a threshold value ε_th，m_stIs the total number of iterations. Corresponding to, S_L ^tThe t-th iteration is the labeled training data set. The same process is carried out for the rest.

The threshold number of iterations is successively reduced as the value of i is smaller, since the lower the confidence, the easier it is to introduce noise. Similarly, the number of times of preference is selected to decrease as the confidence level decreases.

As a preference, a fine tuning of the model can also be performed. The fine-tuning process sets all CNN and LSTM layers to be untraceable, and only fine-tunes the fully-connected layers. The reason is that the feature extraction layer is trained, potential features of the unlabeled data set are learned, and only the weight of the full connection layer needs to be adjusted. For data set { X) not labeled in the final stage of self-learning_U ^mstAnd (6) predicting, and directly taking the predicted value as a real label of the predicted value. After the completion, all the labels of the label-free data and the trained network model are obtained.

Step 6: a neural network loss function is set. This can lead to poor network performance due to class imbalance, i.e., high risk data is always much less than non-risk data. The loss function is set to compensate for the class imbalance. Selecting a multi-class cross entropy penalty function (CE) as the underlying penalty function as a multi-class penalty function as follows:

wherein E is_yi:PAs a random variable y having a distribution P_iAt y_i,k

Mathematical expectation under distribution P; 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. The above loss function is weighted by an over weight (OBC), i.e.:

wherein,

however, in the process of continuously iterating semi-supervised learning, the amount of labeled data can be changed continuously, but each category is not changed uniformly. Therefore, it is necessary to adapt this point as shown in the following formula:

N(t,k)＝|S_L ^t|

AOBC(0,k)＝OBC(k)

wherein N ═ S in obc (k)_L ⁰|。

In the formula, | | represents the number of elements in the set; AOBC (t, k) is the weight lost by class k as the number of iterations t changes.

As a preference, step 7: and adding a limit value punishment module to the network. In order to compensate for the serious classification error of the neural network, a limit value punishment module needs to be introduced. The module uses fuzzy logic, based on penalties that are higher the closer a certain value is to the parameter at the moment of collision. As shown in the following formula:

wherein ev is a limit value; data above 99.99% of all values are generally taken as limits.

As a preference, step 8: and (4) pre-training, self-learning and fine-tuning the network. The optimizer of the network is selected as an Adam optimizer, and the learning rate is 10^-3Attenuation of 10^-6，ε_thThe selection is 0.999999, 0.99999, 0.9999, 0.999, 0.99, 0.95, 0.9. And obtaining all data labels and the trained network after fine adjustment. Each output label is associated with a confidence level, and the current risk score can be obtained by multiplying the confidence level by the value of all labels. The trained network can then be used directly (i.e., to evaluate new data, rather than in a semi-supervised fashion). As shown in the overall framework of fig. 5.

While embodiments of the invention have been described above, it is not limited to the applications set forth in the description and the embodiments, which are fully applicable in various fields of endeavor to which the invention pertains, and further modifications may readily be made by those skilled in the art, it being understood that the invention is not limited to the details shown and described herein without departing from the general concept defined by the appended claims and their 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