CN112215487B

CN112215487B - Vehicle running risk prediction method based on neural network model

Info

Publication number: CN112215487B
Application number: CN202011076552.2A
Authority: CN
Inventors: 胡宏宇; 王�琦; 杜来刚; 鲁子洋
Original assignee: Jilin University
Current assignee: Jilin University
Priority date: 2020-10-10
Filing date: 2020-10-10
Publication date: 2023-05-23
Anticipated expiration: 2040-10-10
Also published as: CN112215487A

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 history data; step two: extracting features of the vehicle history data by adopting a context time window to form statistical features; step three: extracting the statistical features, and step four: dividing the extraction result data in the third step, and in the fifth step: constructing a neural network; constructing an LSTM encoder-1 DCNN-LSTM decoder network architecture; step six: the unlabeled dataset is denoted as { X ] _U -a }; dividing the labeled dataset into a training set and a test set, wherein the training set is denoted as S _L ＝{X _L ，Y _L -a }; using labeled dataset S _L Pre-training the neural network; then enter a self-learning stage for the set Y _PL Predicting, and directly taking the predicted value as a real label; after the completion, all the labels without label data and the trained network model are obtained.

Description

Vehicle running risk prediction method based on neural network model

Technical Field

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

Background

According to statistics of the world health organization in 2018, 135 ten thousand people lose life due to traffic accidents each year, and the loss caused by road traffic collision can reach 3% of the total national production value of most countries. In addition, in traffic-related death accidents, about 94% are caused by drivers, and improper driving behavior of the drivers becomes a major factor in traffic accidents. These driving behaviors are often caused by poor perception of the surrounding environment by the driver, misleading or aggressive decisions and decisions, improper driving maneuvers of the vehicle. The driving risk assessment is based on analysis of various driving characteristics (including driver, vehicle and surrounding environment) at present and past time, and gives possibility of collision or other traffic accidents of the present vehicle. And (3) evaluating and predicting the running safety of the vehicle, and timely feeding back to a driver so as to improve the running safety of the vehicle and further reduce traffic accidents. Therefore, it is indispensable to evaluate and predict the running risk of the vehicle.

However, in driving risk assessment tasks, tagging driving data with risk is a challenging task. If the data is classified by using an unsupervised learning method, the obtained result may not be classified strictly according to the level of risk, and it is difficult to achieve a satisfactory effect in terms of accuracy. In addition, the driving risk assessment task needs to face massive high-dimensional, time-sequence-carrying and class-unbalanced vehicle driving data. Finally, the running risk assessment has a high requirement for accuracy, and it is difficult to accept the case where a high risk is determined as risk-free. In summary, there are great challenges in comprehensively, accurately, and efficiently assessing driving risk.

Through retrieval, chinese patent No. 201711234967.6 discloses a driving risk warning method and device, wherein a pre-established BP network is used for classifying corresponding road sections into high-risk road sections or low-risk road sections; sending out warning information to the vehicle so as to control the vehicle to send out warning when driving to a high-risk road section; CN201910574565.3 discloses a vehicle illegal running risk analysis method based on a Beidou positioning system, by accurately calculating the risk score of the illegal running of the vehicle, a user generates a risk analysis report of the illegal running according to the risk score of the illegal running of the vehicle, reminds and urges a driver to improve driving behaviors, and plays a role in early warning and checking the driving behaviors of the driver. However, the method is insufficient in consideration of massive high-dimensionality vehicle running data with time sequence and unbalanced categories in the vehicle running process, and is difficult to achieve fine running risk prediction, so that the accuracy is poor.

Disclosure of Invention

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

Another object of the invention is a neural network model training method for vehicle driving risk prediction that can manually label only a small portion 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,

step one: collecting vehicle running data to form vehicle history data;

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

step three: extracting the statistical features, including: the type of vehicle, length and width of the vehicle; steering entropy value; parameter time to reverse collision TTC ^-1 Time interval THW of inverse head ^-1 Distance between head of inverse vehicle and DHW ^-1 The method comprises the steps of carrying out a first treatment on the surface of the The time length of the broken line is when the vehicle does not have the lane change intention, the time length of the solid line is when the vehicle is pressed, and the time length of the vehicle is when the vehicle is driven out of the solid line; local traffic flow density, local speed difference, and local acceleration difference.

Step four: dividing the extracted result data in the third step, randomly extracting no more than 5% of data from the data to carry out labeling, and forming a labeled data set; the rest data are label-free data sets, and are 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: the unlabeled dataset is denoted as { X ] _U -a }; dividing the labeled dataset into a training set and a test set, wherein the training set is denoted as S _L ＝{X _L ，Y _L -a }; using labeled dataset S _L Pre-training the neural network;

then enter a self-learning stage to make the unlabeled dataset { X } _U Generating pseudo tags { Y } using a pre-trained network _P -a }; for each generated pseudo tag, a certain confidence epsilon is carried out, and the confidence epsilon is matched with a threshold epsilon _th Comparing, the set greater than the threshold is denoted as S _P ^t ＝{X _Uh ^t ,Y _Ph ^t Sets smaller than the threshold are denoted Y _PL ^t T is the iteration number; for the set { Y _Ph ^t -considering the pseudo tag, i.e. the real tag, according to manifold assumptions; will be set S _L And S is equal to _P ^t Merging to form a new set S _L ^t Then the training method is used for training the network; for { X ] _U ^t Performing pseudo tag regeneration by using the retrained network; for the data set { X } which is not yet labeled in the last stage of self-learning _U ^mst Predicting to take the predicted value as the real label; after the completion, all the labels without label data and the trained network model are obtained.

As a preference, 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 ^t N (0, k) = |s in|, OBC (k) _L,k ⁰ I (I); n is the number of data in mini-batch, m is the category number, y _ik To be a true value of the value,

is a predicted value.

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

where ev is the limit value.

As a preference, the steering entropy value SRE:

preferably, the reverse time to collision TTC ^-1

As one preferable, the specific calculation formula of the local traffic flow density is as follows:

wherein X is _j ＝(x _j ,y _j ) ^T For a vehicle of interest, μ= (x, y) ^T As the center coordinates of the target vehicle,

wherein sigma _x And sigma (sigma) _y Defined by the formula:

σ _x ＝|v _x |+k ₁ L

σ _y ＝|v _y |+k ₂ W

wherein v is _x ,v _y For transverse and longitudinal speed, k, of the vehicle ₁ And k is equal to ₂ Is a compensation factor.

Preferably, the local velocity difference is calculated as follows:

as one preferable, the local acceleration difference is calculated as follows:

the beneficial effects of the invention are as follows:

the method firstly extracts the characteristics of the running of the target vehicle, the interaction of the vehicle and the road, the local traffic condition and the like. And extracting features of massive vehicle driving data, and manually labeling a small part of the vehicle driving data, so that a small data set with labels and a large data set without labels are obtained. A convolutional neural network combining a one-dimensional convolutional neural network (1D-CNN) and a long and short time memory network (LSTM) is built, and a self-adaptive over-balanced cross entropy is adopted as a loss function of the neural network. Embedding the neural network into a semi-supervised learning framework, and pre-training, self-learning and fine-tuning the two data sets to obtain a final label result and a trained network. Finally, a limit penalty module is added to refine the model.

And a cost-sensitive semi-supervised deep learning method is adopted to analyze the vehicle driving data to obtain current and future driving risk scores, wherein the scores are continuous values ranging from 0 to 3, and the risk assessment is finer. 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 by only using a small amount of tag data, and the tag problem of a large amount of non-tag data is greatly solved. The self-adaptive over-balanced cross entropy loss function is adopted, so that the training performance of the semi-supervised deep learning with unbalanced categories 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 for describing traffic conditions, speed and acceleration differences around a target vehicle, and a descriptor is provided, so that the description of surrounding scenes is simplified. The description may also be used in other similar fields.

Drawings

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

FIG. 2 is a schematic illustration of a vehicle of interest with respect to a target vehicle in accordance with the present invention.

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

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

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

Detailed Description

The present invention is described in further detail below with reference to the drawings to enable those skilled in the art to practice the invention by referring to the description.

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 the running of the target vehicle, the interaction of the vehicle and the road, the local traffic condition and the like. And extracting features of massive vehicle driving data, and manually labeling a small part of the vehicle driving data, so that a small data set with labels and a large data set without labels are obtained. A convolutional neural network combining a one-dimensional convolutional neural network (1D-CNN) and a long and short time memory network (LSTM) is built, and a self-adaptive over-balanced cross entropy is adopted as a loss function of the neural network. Embedding the neural network into a semi-supervised learning framework, and pre-training, self-learning and fine-tuning the two data sets to obtain a final label result and a trained network. As a preference, finally, a limit value penalty module is added to refine the model. The method can only manually label a small part of data, automatically learn potential characteristics and predict risk values in a future period of time, so that timely feedback is provided for a driver.

Step 1: and converting and counting the vehicle driving 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. A fixed window is extracted with a contextual time window (as shown in fig. 1). The width of the selection window is W _w The sampling frequency per second is not less than 10Hz =3 seconds. A total of 5 seconds of history data is selected, and each time window overlap ratio ov=66.67%, i.e., 2 seconds. 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 altogether, and comprise vehicle basic information, vehicle running and interaction information, vehicle surrounding environment information and the like. Each statistic has mean, maximum, minimum, variance statistics, so each data has 4 statistics.

A) Basic information of the vehicle:

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

B) Target vehicle travel characteristics:

the longitudinal travel direction of the vehicle is defined as the x-axis and the lateral travel direction as the y-axis. And selecting the speed and acceleration data of the vehicle in the transverse direction and the longitudinal direction, extracting the average value, the maximum value, the minimum value and the variance of the speed and the acceleration data of the vehicle in the transverse direction and the longitudinal direction in each time window as input features, and summing up 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 under fatigue, distraction and the like. Steering entropy is used to quantitatively indicate the steering characteristics of the driver to reflect steering smoothness and driving safety.

Heading angle in timing window is theta= (theta) ₁ ,θ ₂ ,θ ₃ ,L,θ _m ) M is the number of data in the time window. The next corner is predicted using taylor second-order expansion for a given time, as shown in the following equation:

namely:

error function definition:

in the method, in the process of the invention,

is the nth time theta _n Is a predicted value of (a). Setting an alpha value which is preferably 0.035; the prediction error of 90% of the data falls between-alpha and alpha. Dividing the prediction error interval into 9 segments, i.e. 9 intervals, and then finding the distribution frequency of each segment, i.e. p _i I=1, 2, ·9. The steering entropy value SRE is calculated using:

c) Interaction characteristics between vehicles:

the target vehicle speed is noted as v _e Longitudinal coordinate x _e The speed of the front vehicle is recorded as v _p Longitudinal coordinate x _p . The THW headway, i.e. the train of vehicles travelling on the same lane, is the passage of two successive vehicle headway endsA time interval of a certain section; the DHW head space refers to the distance interval between two continuous vehicle head ends passing through a certain section in a vehicle queue running on the same lane; TTC collision time, namely the collision time of two vehicles when the rear vehicle and the front vehicle both run according to the current speed. The calculation formula of THW, DHW, TTC is as follows.

DHW＝(x _p -x _e )

However, taking TTC as an example, two cases may occur when 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 TTC is a negative value; there is a case that 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, TTC is a very large positive value. The TTC range is theoretically (- +. ++ infinity a) of the above-mentioned components, while the truly high risk TTC interval is very small. Taking TTC directly as input may result in a decrease of the model accuracy, and furthermore, the high risk TTC interval may be compressed again by the subsequent feature normalization process. Therefore, the reverse time to collision TTC is adopted ^-1 As shown in the following formula.

Furthermore, for negative TTCs, all negative TTCs are assigned to one sufficiently large positive TTC in a unified manner for simplicity of the feature ^-1 _max (typically 50 seconds) that the target vehicle will collide with the lead vehicle after a long time. The method for eliminating and replacing the irrelevant values can reduce negative confusion influence on the model and improve training accuracy. The expression is as follows:

wherein THW is ^-1 _max ＝10；

Wherein, DHW ^-1 _max ＝200；

By processing, the TTC value is compressed to a small value between 0 and a small value relative to the safety, and the TTC value with high risk is amplified, and the value range is also amplified, so that the accuracy of the model is improved. In the same way, converting THW and DHW into inverse headway THW ^-1 Distance between head of inverse vehicle and DHW ^-1 To amplify the range of high risk THW, DHW. And finally, selecting the maximum value, the mean value and the variance as output characteristics, and taking the maximum value, the mean value and the variance as 9 dimensions.

When the lane change intention occurs, the driver's perception of the lane to be changed is very important. If the lane is not well perceived with the lane to be changed, the lane changing operation is directly carried out, and higher running risk can be brought. Therefore, the maximum TTC between the vehicles running in parallel on the lane to be changed and the front and rear vehicles of the lane to be changed is selected when the lane to be changed is intended ^-1 And THW ^-1 This 3-dimensional feature serves as an input.

D) Interaction between a vehicle and a road

The time length of the broken line is when the vehicle does not have the lane change intention, the time length of the solid line is when the vehicle is pressed, and the time length of the vehicle is when the vehicle is driven out of the solid line.

E) Local traffic situation descriptor

In order to better describe other vehicles, roads, obstacles, etc. running during the running process of the vehicle, corresponding descriptors are needed for description. First, factors such as a vehicle traveling around a target vehicle, an obstacle, and the like are considered. Local traffic density descriptors (LTDs) based on gaussian weights are presented. Taking the target vehicle as a reference, taking 8 vehicles, namely 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, as interested vehicles (shown in fig. 2, if the vehicles are not used, the vehicles are marked as 0), and calculating the contribution degree of the interested vehicles to the vehicle flow density of the target vehicle.

The specific calculation formula is as follows:

wherein X is _j ＝(x _j ,y _j ) ^T For a vehicle of interest, x _j ,y _j The abscissa and ordinate of the vehicle of interest, respectively; mu= (x, y) ^T As the center coordinates of the target vehicle,

wherein sigma _x And sigma (sigma) _y Defined by the formula:

σ _x ＝|v _x |+k ₁ L

σ _y ＝|v _y |+k ₂ W

wherein v is _x ,v _y For transverse and longitudinal speed, k, of the vehicle ₁ And k is equal to ₂ As a compensation factor, k is preferably ₁ ＝0.625，k ₂ ＝1.25。

Thus, each piece of data in each time window forms a corresponding local traffic density descriptor. In addition, the local velocity difference is solved by taking the local traffic density as a weight, and a local velocity difference descriptor is obtained and used for describing the running velocity difference (LVD) of the surrounding vehicles taking the target vehicle as a reference. Similarly, the Local Acceleration Difference (LAD) is solved. The following formula is shown:

v _e for target vehicle speed, let us say, v _j For the speed of the vehicle of interest, a _e For target vehicle acceleration, a _j For vehicle acceleration of interest, N _i For the number of vehicles of interest, the maximum value is 8, and the number is reduced if no corresponding vehicle exists, as shown in fig. 2, and the minimum value is 0, namely, no vehicle of interest exists in one vehicle, namely, no surrounding vehicles exist.

As for the obstacle, it can be regarded as a vehicle whose running speed is zero. Finally, the 3 descriptors are respectively calculated to obtain statistical indexes such as mean value, maximum value, minimum value, standard deviation and the like, and the total is 12-dimensional characteristics.

The following table is a feature statistic:

table 1 statistical input features

Step 3: all the counted data are divided, no more than 5% of the data are randomly extracted from the data for labeling, and the rest data are used for non-labeling training and testing of semi-supervised learning. The risk score at the current moment is evaluated according to an upper-level value of 2-5% (i.e. the data exceeds 95-98% of the values of all data), and the labeling values are 0 (good), 1 (general), 2 (bad) and 3 (very bad). The label method is to label the data, average the final result and round it into integer.

Step 4: and constructing a neural network. The LSTM encoder-1 DCNN-LSTM decoder network architecture is constructed as shown. Wherein Convolume is 1D-CNN, maxPooling is the largest pooling layer, dropout is the discarding layer, FC is the fully connected layer, and Softmax is the activating layer. All other convolutional and fully-connected layer activation functions except the last layer are ReLU.

The method comprises the following steps: and constructing a neural network. An LSTM encoder-1 DCNN-LSTM decoder network architecture is constructed as shown. First, the statistics are embedded and the data is mapped to the LSTM encoder by an embedding layer. The number of embedded layer input nodes is the characteristic dimension, and the output is 128. The LSTM encoder has 128 concealment units in total. 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-dimensional convolution, activation, one-dimensional pooling and random discarding, and the tensor after convolution is obtained. 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 convolution channels for the three convolution layers is 64, 128, 256, respectively. After the output of the last layer is unfolded, 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 hidden units is 128, 64 and 4 respectively. And the other branch is connected with an LSTM decoder to decode the current hidden characteristic to obtain the prediction of the future risk value, and the number of hidden units is 128 LSTM units of the decoder and the full connection layers 128, 64 and 4 respectively. The environment adopted by the computer is Win10, the software name Python3.7 is used, the deep learning framework is Keras2.2.4, and the background is Tensorflow1.14

The LSTM coder encodes the input, and adopts a Dropout method, namely, in the process of training a deep learning network, neural network units are randomly discarded from the network according to a certain probability to reduce the overfitting, and the probability is preferably 0.2; the following is a 1DCNN (one-dimensional convolutional neural network) to reduce the error caused by the convolution between the 2DCNN pair features. Features that contribute significantly to risk values can be selected in conjunction with maximum pooling Maxpooling. After three convolutions-pooling-dropouts, the first branch is connected with two full connection layers and a Softmax layer to obtain the current risk score. The other branch is connected to an LSTM decoder to decode the current advanced feature to obtain the prediction of future risk value.

Step 5: the above network is embedded in a semi-supervised learning architecture as shown in fig. 4. The unlabeled dataset is denoted as { X ] _U }. Dividing the labeled dataset into a training set and a test set, wherein the training set is denoted as S _L ＝{X _L ，Y _L }. First, use is made of a labeled dataset S _L The network is pre-trained. In the training process, an Early-stop method is adopted, namely, when a certain monitoring value is iterated for a plurality of times, the network performance is not improvedWhen the time rises, the network automatically terminates and returns the best performing network weight before stopping. The monitored value selects loss, i.e. loss does not drop after passing the parity times. The choice of parameters is relatively large here because the network is allowed to learn a small amount of data adequately.

Then enter a self-learning stage to make the unlabeled dataset { X } _U Generating pseudo tags { Y } using a pre-trained network _P }. For each pseudo tag generated, there is a certain confidence epsilon (the pre-trained network is obtained directly after passing through the Softmax layer). The confidence is calculated and the threshold epsilon is calculated _th Comparing, the set greater than the threshold is denoted as S _P ^t ＝{X _Uh ^t ,Y _Ph ^t The set with (t being the number of iterations) less than the threshold is denoted as { Y } _PL ^t And t is the iteration number. For the set { Y _Ph ^t Based on manifold assumptions, the pseudo tag is considered to be a real tag. Will be set S _L And S is equal to _P ^t Merging to form a new set S _L ^t And then is used for training of the network. For { X ] _U ^t Pseudo tag regeneration using a retrained network (unlabeled dataset not accommodated by an iterative process). The process is shown in the following formula:

/>

s represents a training data set, L represents a labeled, t represents the number of iterations, X represents an input feature, Y represents a label, U represents an unlabeled, P represents a pseudo-label, and h represents a value greater than a threshold ε _th L represents less than a threshold epsilon _th ，m _st Is the total number of iterations. Correspondingly, S _L ^t Is the training data set with labels at the t-th iteration. The rest is the same.

The number of iterations of the threshold decreases progressively as the value of i is smaller, because the lower the confidence level, the more likely noise is introduced. Similarly, the number of choice of parameters decreases with decreasing confidence.

Preferably, fine tuning of the model is also possible. The trimming process sets all CNN and LSTM layers to be untrainable, trimming only the fully connected layers. The reason is that the feature extraction layer is trained well, and potential features of the unlabeled dataset are learned, only the weights of the full connection layer need to be adjusted. For the data set { X } which is not yet labeled in the last stage of self-learning _U ^mst Prediction is performed with the predicted value directly as its real label. After the completion, all the labels without label data are obtained and the trained network model is obtained.

Step 6: a neural network loss function is set. The performance of the network is degraded due to class imbalance, i.e. high risk data is always far less than risk free data. The loss function is set to compensate for class imbalance. A 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 equation:

wherein E is _yi:P As random variable y with distribution P _i At y _i,k

Mathematical expectation under distribution P; n is the number of data in mini-batch, m is the category number, y _ik Is true value +.>

Is a predicted value. Adding an over-weight (OBC) to the above-mentioned loss function, namely:

wherein, the liquid crystal display device comprises a liquid crystal display device,

however, in the course of constantly iterating semi-supervised learning, the amount of tagged data will constantly change, but each category will not change uniformly. It is necessary to adapt this to the following conditions:

N(t,k)＝|S _L ^t |

AOBC(0,k)＝OBC(k)

wherein n= |s in OBC (k) _L ⁰ |。

The expression I represents the number of elements in the collection; AOBC (t, k) is the weight lost by the kth class as a function of the number of iterations t.

Preferably, step 7: and adding a limit value punishment module to the network. In order to compensate for very specific severe classification errors of the neural network, a limit value penalty module needs to be introduced. The module employs fuzzy logic, with higher penalties being based as certain values are closer to the parameters of the moment of collision. The following formula is shown:

wherein ev is the limit value; data exceeding 99.99% of all values are generally taken as limit values.

Preferably, step 8: the network is pre-trained, self-learned and fine-tuned. The network optimizer is selected as an Adam optimizer, and the learning rate is 10 ^-3 Attenuation of 10 ^-6 ，ε _th Selected to be 0.999999, 0.99999, 0.9999, 0.999, 0.99, 0.95, 0.9. And after fine tuning, obtaining all the data labels and the trained network. Each label output corresponds to a confidence level, and the current risk score can be obtained by multiplying the confidence level by the values of all labels. The trained network can then be used directly (i.e., to evaluate new data without semi-supervision). As shown in the overall frame of fig. 5.

Although embodiments of the present invention have been disclosed above, it is not limited to the details and embodiments shown and described, it is well suited to various fields of use for which the invention would be readily apparent to those skilled in the art, and accordingly, the invention is not limited to the specific details and illustrations shown and described herein, without departing from the general concepts defined in the claims and their equivalents.

Claims

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

step one: collecting vehicle running data to form vehicle history data;

step three: extracting the statistical features, including: the type of vehicle, length and width of the vehicle; steering entropy value; parameter time to reverse collision TTC ^-1 Time interval THW of inverse head ^-1 Distance between head of inverse vehicle and DHW ^-1 The method comprises the steps of carrying out a first treatment on the surface of the The time length of the broken line is when the vehicle does not have the lane change intention, the time length of the solid line is when the vehicle is pressed, and the time length of the vehicle is when the vehicle is driven out of the solid line; local traffic flow density, local speed difference and local acceleration difference;

then enter a self-learning stage to make the unlabeled dataset { X } _U Generating pseudo tags { Y } using a pre-trained network _P -a }; for each generated pseudo tag, a certain confidence epsilon is carried out, and the confidence epsilon is matched with a threshold epsilon _th Comparing, the set greater than the threshold is denoted as S _P ^t ＝{X _Uh ^t ,Y _Ph ^t Sets smaller than the threshold are denoted Y _Pl ^t T is the iteration number; for the set { Y _Ph ^t -considering the pseudo tag, i.e. the real tag, according to manifold assumptions; will be set S _L And S is equal to _P ^t Merging to form a new set S _L ^t Then the training method is used for training the network; for { X ] _U ^t Performing pseudo tag regeneration by using the retrained network; for the data set { X } which is not yet labeled in the last stage of self-learning _U ^mst Predicting to take the predicted value as the real label; after the completion, all labels without label data and a trained network model are obtained; predicting the running risk of the vehicle according to the trained network model; s represents a training data set, L represents a labeled, t represents the number of iterations, X represents an input feature, Y represents a label, U represents an unlabeled, P represents a pseudo-label, and h represents a value greater than a threshold ε _th L represents less than a threshold epsilon _th ，m _st Is the total iteration number; correspondingly, S _L ^t A training data set with a label at the t-th iteration;

the steering entropy value SRE:

dividing the prediction error interval into 9 segments, i.e. 9 intervals, and then obtaining the distribution frequency of each segment, i.e. p _i ，i＝1,2,..9；

The time to reverse collision TTC ^-1

TTC collision time, namely the collision time of two vehicles when the rear vehicle and the front vehicle both run at the current speed; the TTC of all negative values is uniformly assigned to a sufficiently large TTC of positive value ^-1 _max ；TTC ^-1 _max Taking 50 seconds;

wherein THW is ^-1 _max ＝10；

Wherein, DHW ^-1 _max ＝200；/>

The THW headway refers to the time interval that two continuous vehicle headway ends pass through a certain section in a vehicle queue running on the same lane; the DHW head space refers to the distance interval between two continuous vehicle head ends passing through a certain section in a vehicle queue running on the same lane;

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

wherein sigma _x And sigma (sigma) _y Defined by the formula:

σ _x ＝|v _x |+k ₁ L

σ _y ＝|v _y |+k ₂ W

wherein v is _x ,v _y For transverse and longitudinal speed, k, of the vehicle ₁ And k is equal to ₂ Is a compensation factor; length L and width W of the vehicle; n (N) _i Is the number of vehicles of interest; x is x _j ,y _j The abscissa and ordinate of the vehicle of interest, respectively;

the local acceleration difference is calculated as follows:

a _e for target vehicle acceleration, a _j Acceleration for a vehicle of interest;

the local velocity difference is calculated as follows:

v _e for target vehicle speed, let us say, v _j Is the speed of the vehicle of interest.

2. The vehicle running risk prediction method based on a neural network model 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)＝OBC(k)

Wherein N (t, k) = |s in AOBC (k) _L,k ^t I, since AOBC (0, k) =obc (k), N (0, k) = |s in OBC (k) _L,k ⁰ I (I); n is the number of data in mini-batch, m is the category number, y _ik To be a true value of the value,

is a predicted value; iteration times t; in the above expression, || represents the number of elements in the collection.

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

where ev is the limit value.