CN116279471A - A Vehicle Acceleration Prediction Method Considering Driving Behavior Characteristics in Car-following Scenarios - Google Patents

Abstract

The invention relates to a vehicle acceleration prediction method considering driving behavior characteristics in a following scene, which comprises the following steps: step 1) obtaining a prediction data set, and carrying out data processing and variable extraction: screening and reserving the obtained prediction data set based on a preconfigured rule, and extracting characteristic variables from the effective following fragment data; step 2) constructing a driving behavior semantic division model based on CHMM, taking a characteristic variable as input of the driving behavior semantic division model, dividing the characteristic variable into segments with different behavior characteristics, and performing similarity evaluation on driving behavior semantics; and 3) constructing a CNN-LSTM acceleration prediction model, dividing the effective following segment data based on the similarity evaluation result to construct an input-output sample set, training the CNN-LSTM acceleration prediction model, and predicting the vehicle acceleration based on the trained CNN-LSTM acceleration prediction model. Compared with the prior art, the method has the advantages of high prediction precision and the like.

Description

A Vehicle Acceleration Prediction Method Considering Driving Behavior Characteristics in Car-following Scenarios

technical field

The invention relates to the technical fields of vehicle intelligent driving and active safety, in particular to a vehicle acceleration prediction method considering driving behavior characteristics in a car-following scene.

Background technique

In recent years, with the rapid development of my country's economy and the continuous acceleration of urbanization, the trend of urban regional integration has been strengthened, and the number and scale of transportation vehicles have also accelerated and expanded. At the same time, various traffic problems also appear thereupon. The problem of traffic safety is becoming more and more prominent, seriously affecting the quality of traffic travel and social and economic development.

Driving behavior is one of the important factors affecting traffic safety. In the complex traffic environment composed of people, vehicles and roads, human factors are the most important factors leading to traffic accidents. Some studies have shown that more than 70% of traffic accidents are caused by human errors, especially for long-distance travel, and the driving environment is complex. Therefore, research on driving behavior is a research hotspot in the field of traffic safety; in addition, vehicle intelligence and personalized services It is an important trend in the development of vehicle technology in the future. Personalized prediction of drivers with different driving habits can help realize personalized modeling for different styles of driver behavior, capture driver behavior characteristics, and improve the prediction accuracy of driving behavior. Current research on driving behavior prediction is mainly divided into two categories: model-driven methods and data-driven methods.

The model-driven approach is a fixed-structure model based on specific assumptions, whose parameters are generally calculated from empirical data. Model-driven forecasting methods generally use less data and have limited complexity. Therefore, when dealing with large data sets, the effective information in large data sets cannot be fully mined, and the complexity of the model will also increase.

Data-driven methods are a class of models without fixed structure and fixed parameters. Data-driven models for driving behavior prediction can use multiple datasets from different sources to predict performance. However, most of these prediction methods do not consider the classification of driving behavior and cannot tap the potential heterogeneity of driving behavior. In addition, since driving behavior data is random and non-linear, it is difficult to accurately describe it using traditional linear models. In recent years, methods based on deep learning have been widely used in time series data prediction. However, the fitting ability of a single deep learning model is weak, and the fitting effect is poor for data with high dimensions and complex correlations.

Contents of the invention

The purpose of the present invention is to provide a vehicle acceleration prediction method that considers driving behavior characteristics in a car-following scene, and considers the driver's behavior characteristics to improve the accuracy of acceleration prediction.

The purpose of the present invention can be achieved through the following technical solutions:

A vehicle acceleration prediction method considering driving behavior characteristics in a car-following scene, comprising the following steps:

Step 1) Obtain the predicted data set and perform data processing and variable extraction: For the obtained predicted data set, filter and retain the effective car-following segments based on pre-configured rules, and extract feature variables from the effective car-following segment data. The prediction data set includes the vehicle data and the preceding vehicle data;

Step 2) Build a CHMM-based driving behavior semantic segmentation model, use feature variables as the input of the driving behavior semantic segmentation model, segment the feature variables into segments with different behavioral characteristics, and evaluate the similarity of driving behavior semantics;

Step 3) Construct a CNN-LSTM acceleration prediction model, divide the effective car-following segment data based on the similarity evaluation results to construct an input and output sample set, train the CNN-LSTM acceleration prediction model, and perform vehicle training based on the trained CNN-LSTM acceleration prediction model. Acceleration prediction.

Described step 1) comprises the following steps:

Step 11) Carry out data cleaning on the predicted data set, remove outliers or missing frames greater than 1s, and use linear interpolation to make up for missing segments less than 1s to obtain a valid data set;

Step 12) Extracting the effective car-following segment from the obtained valid data set, according to the relative position between the two vehicles, the speed of the target vehicle, whether there is an overtaking event, and the duration of the car-following event. carry out screening retention;

Step 13) Extract feature variables from the effective car-following segment data.

Described step 2) comprises the following steps:

Step 21) divide the feature variable into different grades according to the pre-configured segmentation threshold;

Step 22) standardize the feature variable sequence with the driver as the unit;

Step 23) According to the temporal nature of driving behavior and data distribution characteristics, construct a CHMM-based semantic segmentation model of driving behavior, adopt the method of unsupervised learning, use the standardized characteristic variables as the input of the semantic segmentation model of driving behavior, and convert the characteristic variable data Segmentation into segments with different behavioral characteristics;

Step 24) Based on the standardized occurrence frequency distribution and JS divergence calculation method, solve the optimal distribution of behavior semantic duration, behavior semantic frequency distribution, and JS divergence of characteristic variable distribution between each pair of drivers, and obtain the driving Quantitative index value of behavioral similarity.

The standardization process in the step 22) is:

Among them, x represents the sequence of characteristic variables; m is the number of drivers, and M represents the number of drivers; l is the number of stable car-following events of driver m, and L represents the number of stable car-following events of the driver; μ _m and σ _m are respectively Indicates the mean and standard deviation of the characteristic variables of the driver m; after standardization, the mean of the characteristic variables is 0, and the standard deviation is equal to 1.

In the CHMM-based driving behavior semantic division model,

In HMM, there are two sequence types, where the observation sequence X={x ₁ , x ₂ ,...,x _T } can be observed, and the hidden state sequence S={s ₁ , s ₂ ,..., s _T } cannot be observed; HMM is represented by the state transition matrix A, the observation probability matrix B and the initial state probability distribution П: λ=(A, B, Π), the number of hidden states is I, and the number of all possible values of the observed variable for J;

Among them, the state transition matrix A＝[a _ij ] _I×I , the state transition probability is:

a _ij ＝P(s _t+1 ＝q _j |s _t ＝q _i )

Observation probability matrix B=[b _ij ] _I×J , the observation value is only determined by the hidden state at the current moment:

b _ij =P(x _t =x _j |s _t =q _i )

Initial state probability distribution ∏=[π _i ] _1×I , which represents the probability distribution of the hidden state at the initial moment:

π _i =P(s ₁ =q _i )

HMM由模型参数λ＝(A，B，∏)进行表示，其中，状态转移矩阵A＝{a_u，v}、观测值概率B＝b_u(X_t)和初始状态概率∏＝{π_u}的计算方法如下：The number of hidden states of each HMM chain in CHMM is set to N, then the number of hidden states of the CHMM model is N ^Q , where Q represents the number of characteristic variables; the combination of hidden states of the model at any time can be expressed as

HMM is represented by model parameters λ=(A, B, ∏), where state transition matrix A={a _{u, v} }, observation probability B= _bu (X _t ) and initial state probability ∏={π _u } is calculated as follows:

The hidden state layer of CHMM decodes the dependencies between feature variables. Considering that the feature variables are all continuous variables, a Gaussian distribution is introduced to calculate the observation probability:

Among them, μ _{c, u} and σ _{c, u} respectively represent the mean and standard deviation of the probability distribution when the hidden state in the HMM chain corresponding to the characteristic variable c is q _{c, u} .

Described step 24) comprises the following steps:

Step 241) The duration of each behavioral semantic segment extracted by the driver D _m is d _m , f(d _m ) represents the distribution of the duration of the behavioral semantics, and f(d _n ) represents the behavioral semantics extracted by the driver D _n The distribution of duration, by quantifying the closeness of the distributions f(d _m ) and f(d _n ), to measure the similarity of the driving pattern transfer laws of different drivers

Step 242) Measure the similarity of the driver D _m and the driver D _n in the driving mode selection preference by quantifying the proximity of the probability mass function

和/>

的接近程度后求均值，得到/>

Considering that under each type of car-following distance condition, the sum of the occurrence probability of behavioral semantics is equal to 1, respectively measure the probability mass function under the car-following distance condition S _Δd

and />

After calculating the mean value of the closeness, get />

和/>

采用正态分布/>

简化驾驶员D_m在区段d中特征变量x_k的概率密度函数；依次计算每一个持续时长区段内的特征变量的概率密度函数对的接近程度，求均值得到驾驶员D_m和驾驶员D_n驾驶操作激进程度的相似性/>

Step 243) Divide the behavioral semantic duration into j segments according to the pre-configured segment threshold, and count the mean and standard deviation of the characteristic variable data of each segmental behavioral semantic segment respectively. Driver D _m is in segment d The mean and standard deviation of the characteristic variable x _k in are recorded as

and />

Using a normal distribution />

Simplify the probability density function of the characteristic variable x _k of the driver D _m in the segment d; calculate the closeness of the probability density function pairs of the characteristic variables in each duration segment in turn, and calculate the average value to obtain the driver D _m and the driver D _n similarity in aggressiveness of driving maneuvers />

Among them, K represents the number of feature variables;

和/>

分别从三个维度量化了驾驶员D_m和驾驶员D_n风格的相似程度，通过加权平均得到驾驶风格相似性的综合评价指标η_m，n：step 244)

and />

The similarity degree of driver D _m and driver D _n style is quantified from three dimensions respectively, and the comprehensive evaluation index η _m,n of driving style similarity is obtained by weighted average:

分别表示三个驾驶风格相似性指标的权重；where w _i ∈ [0, 1] and

represent the weights of the three driving style similarity indicators respectively;

和/>

采用JS散度来度量两个概率分布的相似性：Step 245) given two discrete probability density distributions of standardized frequency distributions

and />

JS divergence is used to measure the similarity of two probability distributions:

和/>

的KL散度/>

的计算方法为：Among them, the discrete probability density distribution

and />

KL divergence />

The calculation method is:

和/>

分别表示在跟驰距离工况S_Δd下，驾驶员D_m和驾驶员D_n行为语义的标准化出现频率；

and />

Respectively represent the normalized occurrence frequency of driver D _m and driver D _n behavior semantics under the car-following distance condition S _Δd ;

The value range of JS divergence is [0,1]. When the two probability distributions are the same, the JS divergence is equal to 0. The greater the difference in probability distribution, the closer the JS value is to 1.

Described step 3) comprises the following steps:

Step 31) According to the similarity evaluation results and the pre-configured similarity threshold, select several drivers who meet the similarity conditions as a group, and classify the effective car-following segments into the training set and the test set according to the pre-configured ratio of the drivers and validation set;

Step 32) Construct input and output samples, convert the data containing acceleration and other characteristic variables in the training set and test set into an N-dimensional array, each input sample is composed of characteristic variables at K moments, and the output label sample is the K+1th moment acceleration data;

Step 33) build CNN-LSTM vehicle acceleration prediction model;

Step 34) model training and hyperparameter optimization: use the verification set to evaluate the model in the training process in real time, check its training effect, use Adam optimizer to iterate, and repeat the training process to make the network parameters tend to be optimal;

Step 35) Predict the acceleration of the target vehicle based on the trained CNN-LSTM vehicle acceleration prediction model.

The CNN-LSTM vehicle acceleration prediction model includes an input layer, a CNN layer, an LSTM layer and a fully connected layer connected in sequence, wherein,

The input layer is used to receive input data;

The CNN layer is used to compress and extract the spatial features of the input data, and the CNN layer includes a convolution layer and a maximum pooling layer;

The LSTM layer is used to extract the time series features of CNN output data, and the LSTM network in the LSTM layer is many-to-one LSTM;

The fully connected layer is used to output acceleration prediction results;

Each layer in the model network is added with a Dropout layer, and the mean square error is selected as the evaluation of the model loss function and prediction accuracy.

The calculation formula of the convolutional layer is as follows:

为第k层、第j维的特征图，M_j为输入图的集合，/>

为卷积核，/>

为偏置。Among them, σ is the activation function,

is the feature map of the kth layer and the jth dimension, M _j is the set of input maps, />

is the convolution kernel, />

for the bias.

The forget gate parameter update formula of the LSTM network is:

f _t ＝σ(w _f h _t-1 +w _f x _t +b _f )

Among them, f _t represents the state of the forget gate at time t, w _f and b _f are the weight and bias of the forget gate, respectively, σ represents the sigmoid function, and the output value ranges from 0 to 1, h _t represents the state of the hidden layer, x _t represents the input value of the network at time t;

The input gate of the LSTM network is used to update the information in the input cell state:

i _t ＝σ(w _i x _t +w _i h _t-1 +b _i )

Among them, i _t represents the state of the input gate at time t, and w _i and b _i are the weight and bias of the input gate, respectively;

The LSTM network memory cell state c _t is updated by the information c _t-1 stored in the previous memory cell and the new candidate information:

是t时刻的候选细胞状态，c_t是更新后的细胞状态，w_c和b_c分别是细胞状态的权值和偏置，.代表点积；in,

is the candidate cell state at time t, c _t is the updated cell state, w _c and b _c are the weight and bias of the cell state respectively, . represents the dot product;

The output gate of LSTM is used to control the information output in the cell state, and the output hidden layer state h t is calculated according to the memory cell state c _t and the output gate state o _{t :}

o _t ＝σ(w _o x _t +w _o h _t-1 +b _o )

h _t ＝o _t ·tanh(c _t )

Compared with the prior art, the present invention has the following beneficial effects:

(1) The CHMM-based driving behavior semantic segmentation model of the present invention divides the feature variable data into segments with different behavioral characteristics, which is an extension of the HMM model, by establishing the correlation between the hidden state variables of multiple HMM chains , to realize the coupling of the HMM chain structure, so as to describe the comprehensive statistical characteristics of multi-characteristic variable data; in addition, CHMM also makes full use of the mathematical structure and probabilistic reasoning ability of HMM, and establishes the influence relationship between multi-channel hidden states, It is very suitable for the fusion of multivariate data information, which can mine the heterogeneity of driving behavior among different drivers and improve the prediction accuracy.

(2) The present invention uses the CNN-LSTM vehicle acceleration prediction model to carry out acceleration prediction on the grouped driver's car-following data, obtains the spatial features between the data, overcomes the limitations of the LSTM single model, and combines CNN in spatial feature extraction With the advantages of LSTM in time series learning, it fully captures the spatial and temporal features between data, and effectively improves the prediction accuracy of the model.

(3) The car-following scene proposed by the present invention considers the vehicle acceleration prediction method of driving behavior characteristics, which can reasonably divide the vehicle-following data into behavioral semantic segments, group homogeneous drivers into one group through similarity evaluation, and effectively capture Spatial and temporal characteristics of the data, so as to achieve high-precision acceleration prediction, promote the development of advanced driver assistance systems (ADAS), and improve traffic safety issues.

Description of drawings

Fig. 1 is method flowchart of the present invention;

FIG. 2 is a semantic segmentation diagram of driving behavior according to an embodiment of the present invention;

Fig. 3 is the driving behavior semantic similarity evaluation figure of the embodiment of the present invention;

Fig. 4 is the structure diagram of CNN-LSTM vehicle acceleration prediction model of the present invention;

Fig. 5 is a comparison chart of the vehicle acceleration prediction result and the real value according to the embodiment of the present invention.

Detailed ways

The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is carried out on the premise of the technical solution of the present invention, and detailed implementation and specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

This embodiment provides a vehicle acceleration prediction method considering driving behavior characteristics in a car-following scene, as shown in FIG. 1 , including the following steps:

Step 1) Obtain the predicted data set and perform data processing and variable extraction: For the obtained predicted data set, filter and retain the effective car-following segments based on pre-configured rules, and extract feature variables from the effective car-following segment data. The above prediction data set includes self-vehicle data and front-vehicle data.

Step 11) Perform data cleaning on the prediction data set, remove outliers or missing frames longer than 1s, and use linear interpolation to make up for missing segments less than 1s to obtain a valid data set.

The embodiment selects the data of 30 drivers with similar driving scenes in the SPMD data set as the prediction data set for analysis. The important variable information in the dataset is as follows:

Table 1 SPMD data description

Step 12) Extracting the effective car-following segment from the obtained valid data set, according to the relative position between the two vehicles, the speed of the target vehicle, whether there is an overtaking event, and the duration of the car-following event. Keep screening.

For the data set after data cleaning, this embodiment retains the valid car-following segments that meet the requirements according to the following conditions:

(1) The main vehicle and the vehicle in front are in the same lane;

(2) The relative distance along the lane is greater than 5m and less than 120m;

(3) The vehicle speed is greater than 5m/s;

(4) When the adjacent vehicle cuts in, that is, when the ID of the preceding vehicle changes, the car-following event is terminated;

(5) The duration of a single car-following event is longer than 50s.

Step 13) Extract feature variables from the effective car-following segment data.

In this embodiment, in order to study the car-following behavior characteristics of the main vehicle, three characteristic variables of the main vehicle acceleration, relative speed and relative distance are selected to describe the car-following scene:

(1) Acceleration of the main vehicle (a ₁ ): it can directly reflect the driving intention and the driver's behavior preference.

(2) Relative distance (Δd): The position difference between the main vehicle and the leading vehicle along the forward direction, Δd=x ₂ -x ₁ , always greater than 0.

(3) Relative speed (Δv): the speed difference between the main vehicle and the front vehicle along the forward direction, Δv=v ₂ -v ₁ , when the relative speed is greater than 0, the speed of the front vehicle is faster, and the two vehicles gradually move away; when the relative distance When it is less than 0, the speed of the rear vehicle is faster, and the rear vehicle is getting closer to the front vehicle.

Step 2) Construct a CHMM-based semantic segmentation model of driving behavior, use feature variables as the input of the semantic segmentation model of driving behavior, segment the feature variables into segments with different behavioral characteristics, and evaluate the similarity of driving behavior semantics.

Step 21) Divide the feature variable into different levels according to the pre-configured segmentation threshold.

According to the frequency distribution histogram and cumulative distribution map of the characteristic variables, the quantile points of the data are obtained, combined with the quantile points and the driver's perceived comfort threshold, the relative distance is divided into {long-distance LD, medium-distance ND, short-distance CD} three The relative speed and acceleration are divided into {fast away from RFB, slowly away from FB, keep KE, slowly close to CI, fast close to RCI} and {fast acceleration AA, slow acceleration GA, no acceleration NA, slow deceleration GD, rapid deceleration AD} five levels.

The segmentation information is shown in Table 2:

Table 2 Description of variable segmentation information

Step 22) standardize the sequence of feature variables with the driver as the unit:

Wherein, x={Δd, Δv, a} ^T represents relative distance, relative speed and acceleration sequence; m is the driver number, M represents the number of drivers, in the present embodiment, M=30; l is the stability of driver m number of car-following events, L represents the number of stable car-following events of the driver; μ _m and σ _m represent the mean and standard deviation of the characteristic variables of driver m respectively; the mean value of the characteristic variables after standardization is 0, and the standard deviation is equal to 1 .

Step 23) According to the temporal nature of driving behavior and data distribution characteristics, construct a CHMM-based semantic segmentation model of driving behavior, adopt the method of unsupervised learning, use the standardized characteristic variables as the input of the semantic segmentation model of driving behavior, and convert the characteristic variable data Segmentation into segments with different behavioral characteristics.

In HMM (Hidden Markov Model, Hidden Markov Model), there are two sequence types, where the observation sequence X={x ₁ ,x ₂ ,...,x _T } can be observed, and the hidden state sequence S={ s ₁ ,s ₂ ,...,s _T } cannot be observed. HMM can be represented by state transition matrix A, observation probability matrix B and initial state probability distribution Π, λ=(A, B, Π). The number of hidden states is I, and the number of all possible values of the observed variable is J.

Its state transition matrix A=[a _ij ] _I×I , and its state transition probability is:

a _ij ＝P(s _t+1 ＝q _j |s _t ＝q _i )

Observation probability matrix B=[b _ij ] _I×J , the observation value is only determined by the hidden state at the current moment:

b _ij =P(x _t =x _j |s _t =q _i )

The initial state probability distribution Π=[π _i ] _1×I represents the probability distribution of the hidden state at the initial moment:

π _i =P(s ₁ =q _i )

The CHMM-based driving behavior semantic division model is an improved model based on HMM, wherein,

HMM由模型参数λ＝(A,B,Π)进行表示，其中，状态转移矩阵A＝{a_u,v}、观测值概率B＝b_u(X_t)和初始状态概率Π＝{π_u}的计算方法如下：The number of hidden states of each HMM chain in CHMM is set to N, then the number of hidden states of the CHMM model is N ³ , where 3 represents the number of characteristic variables selected in this embodiment; the hidden state combination of the model at any time It can be expressed as

HMM is represented by model parameters λ=(A,B,Π), where state transition matrix A={a _u,v }, observation probability B= _bu (X _t ) and initial state probability Π={π _u } is calculated as follows:

The hidden state layer of CHMM decodes the dependencies between feature variables. Considering that the feature variables are all continuous variables, a Gaussian distribution is introduced to calculate the observation probability:

Among them, μ _{c, u} and σ _{c, u} respectively represent the mean value and standard deviation of the probability distribution when the hidden state in the HMM chain corresponding to the characteristic variable c is q _{c, u} .

In this embodiment, the representative trajectory segmentation results of driver No. 1 are shown in FIG. 2 . The abscissa is time, and the solid line, dot-dash line, and dotted line respectively show the original data information of the feature variables. The background color block is the extracted driving behavior sequence unit, and the same color indicates the same type of behavior semantics.

The behavioral semantic segmentation results of CHMM are basically consistent with the data fluctuations, and can accurately extract effective behavioral semantic sequence units based on the potential behavioral characteristics of feature variable data. At the same time, the model can effectively overcome the interference of data noise and avoid too short behavioral semantic fragments.

Step 24) Based on the standardized occurrence frequency distribution and JS divergence calculation method, solve the optimal distribution of behavior semantic duration, behavior semantic frequency distribution, and JS divergence of characteristic variable distribution between each pair of drivers, and obtain the driving Quantitative index value of behavioral similarity.

Step 241) The duration of each behavioral semantic segment extracted by the driver D _m is d _m , f(d _m ) represents the distribution of the duration of the behavioral semantics, and f(d _n ) represents the behavioral semantics extracted by the driver D _n The distribution of duration, by quantifying the closeness of the distributions f(d _m ) and f(d _n ), to measure the similarity of the driving pattern transfer laws of different drivers

Step 242) Measure the similarity of the driver D _m and the driver D _n in the driving mode selection preference by quantifying the proximity of the probability mass function

和/>

的接近程度后求均值，得到/>

Considering that under each type of car-following distance condition, the sum of the occurrence probability of behavioral semantics is equal to 1, respectively measure the probability mass function under the car-following distance condition S _Δd

and />

After calculating the mean value of the closeness, get />

和/>

Step 243) Divide the semantic duration of the behavior into 8 segments according to the pre-configured segment division threshold: {<1, [1,5), [5,10), [10,15), [15,20) _{. _} The mean and standard deviation are recorded as

and />

简化驾驶员D_m在区段d中特征变量x_k的概率密度函数。同一个持续时长区段内的相同特征变量的分布规律的相似程度具有可比性，因此，本发明依次计算每一个持续时长区段内三个特征变量的概率密度函数对的接近程度，求均值得到驾驶员D_m和驾驶员D_n驾驶操作激进程度的相似性/>

Since most of the distributions can be approximated by the normal distribution when the amount of data is relatively large, the present invention adopts the normal distribution

Simplify the probability density function of the characteristic variable x _k of the driver D _m in the section d. The similarity of the distribution laws of the same characteristic variable in the same duration segment is comparable, therefore, the present invention calculates the closeness of the probability density function pairs of the three characteristic variables in each duration segment in turn, and calculates the mean value to obtain The similarity between the aggressiveness of driver D _m and driver D _n driving maneuvers />

Among them, K represents the number of feature variables.

和/>

分别从三个维度量化了驾驶员D_m和驾驶员D_n风格的相似程度，通过加权平均得到驾驶风格相似性的综合评价指标η_m,n：step 244)

and />

The similarity degree of driver D _m and driver D _n style is quantified from three dimensions respectively, and the comprehensive evaluation index η _m,n of driving style similarity is obtained by weighted average:

分别表示三个驾驶风格相似性指标的权重，可以根据应用场景的需要进行调整。本实施例认为三个维度相似程度的重要性一致，设定

where w _i ∈ [0,1] and

Respectively represent the weights of the three driving style similarity indicators, which can be adjusted according to the needs of the application scenario. This embodiment considers that the importance of the similarity of the three dimensions is the same, setting

和/>

采用JS散度来度量两个概率分布的相似性：Step 245) given two discrete probability density distributions of standardized frequency distributions

and />

JS divergence is used to measure the similarity of two probability distributions:

和/>

的KL散度/>

的计算方法为：Among them, the discrete probability density distribution

and />

KL divergence />

The calculation method is:

和/>

分别表示在跟驰距离工况S_Δd下，驾驶员D_m和驾驶员D_n行为语义的标准化出现频率。

and />

Respectively represent the normalized occurrence frequency of driver D _m and driver D _n behavior semantics under the car-following distance condition S _Δd .

The JS divergence is symmetric, and the value range is [0,1]. When the two probability distributions are the same, the JS divergence is equal to 0, and the greater the difference between the probability distributions, the closer the JS value is to 1.

In this embodiment, the JS divergence results of the three-aspect similarity analysis are integrated, and the weighted average is calculated to obtain the multi-dimensional and comprehensive quantitative evaluation results of the driving style similarity. The heat map is used to visually display the results, as shown in Figure 3. The JS divergence value of the dark area is larger, and the difference is more significant; the JS divergence value of the diagonal line is equal to 0, drawn with a white background.

As shown in Figure 3, the driving style of driver #16 is quite different from most drivers, and the driving style of driver #16 is particularly conservative, the frequency of driving mode switching is low, and the expected car-following distance is relatively large. The driving styles of drivers #24 and #25 are significantly different from those of the four drivers, and these drivers have significant differences in driving mode selection preferences and driving aggressiveness.

Step 3) Construct the CNN-LSTM acceleration prediction model, the structure of which is shown in Figure 4. Based on the similarity evaluation results, divide the effective car-following segment data to construct an input and output sample set, and train the CNN-LSTM acceleration prediction model. CNN-LSTM acceleration prediction model for vehicle acceleration prediction.

Step 31) According to the similarity evaluation results and the pre-configured similarity threshold, select 10 drivers who meet the similarity conditions as a group, and put the effective car-following segments into the training according to the ratio of 7:2:1 drivers set, test set, and validation set.

Step 32) Construct input and output samples, convert the data containing acceleration and other characteristic variables in the training set and test set into an N-dimensional array, each input sample is composed of characteristic variables at K moments, and the output label sample is the K+1th moment acceleration data.

The input of the model in this embodiment is the three characteristic variables of main vehicle acceleration, relative distance and relative speed, and the output is the main vehicle acceleration at a future moment. Through debugging, the selected time window length is 80, and the prediction length is 1. The training set, verification set, and test set are divided into samples. As shown in Table 3, the input samples and label samples of the model are obtained.

Table 3 Data set usage and quantity

Step 33) Construct a CNN-LSTM vehicle acceleration prediction model.

The CNN-LSTM vehicle acceleration prediction model includes an input layer, a CNN layer, an LSTM layer and a fully connected layer connected in sequence, wherein,

The input layer is used to receive input data;

The CNN layer is used to compress and extract the spatial features of the input data, and the CNN layer includes a convolution layer and a maximum pooling layer;

The LSTM layer is used to extract the time series features of the CNN output data, the LSTM network in the LSTM layer is a many-to-one LSTM, and the hidden layer contains several LSTM units;

The fully connected layer is used to output acceleration prediction results;

Each layer in the model network is added with a dropout layer.

The calculation formula of the convolutional layer is as follows:

为第k层、第j维的特征图，M_j为输入图的集合，/>

为卷积核，/>

为偏置。Among them, σ is the activation function,

is the feature map of the kth layer and the jth dimension, M _j is the set of input maps, />

is the convolution kernel, />

for the bias.

The forget gate parameter update formula of the LSTM network is:

f _t ＝σ(w _f h _t-1 +w _f x _t +b _f )

Among them, f _t represents the state of the forget gate at time t, w _f and b _f are the weight and bias of the forget gate, respectively, σ represents the sigmoid function, and the output value ranges from 0 to 1, h _t represents the state of the hidden layer, x _t represents the input value of the network at time t;

The input gate of the LSTM network is used to update the information in the input cell state:

i _t ＝σ(w _i x _t +w _i h _t-1 +b _i )

Among them, i _t represents the state of the input gate at time t, and w _i and b _i are the weight and bias of the input gate, respectively;

The LSTM network memory cell state c _t is updated by the information c _t-1 stored in the previous memory cell and the new candidate information:

是t时刻的候选细胞状态，c_t是更新后的细胞状态，w_c和b_c分别是细胞状态的权值和偏置，.代表点积；in,

is the candidate cell state at time t, c _t is the updated cell state, w _c and b _c are the weight and bias of the cell state respectively, . represents the dot product;

The output gate of LSTM is used to control the information output in the cell state, and the output hidden layer state h t is calculated according to the memory cell state c _t and the output gate state o _{t :}

o _t ＝σ(w _o x _t +w _o h _t-1 +b _o )

h _t ＝o _t ·tanh(c _t )

The activation function of the convolutional layer in this embodiment uses ReLU to realize the function that if the input is greater than 0, the output is equal to the input, otherwise the output is 0. The pooling layer chooses the maximum pooling, which can reduce the calculation amount of the upper layer by eliminating the non-maximum value. At the same time, it is beneficial to extract the local dependencies of different regions, retain the most significant information, and use the obtained region vector as the input of the LSTM network. The LSTM layer uses the activation function tanh, and the fully connected layer is used as the model output layer. The optimizer of this embodiment selects the Adam optimizer, and Adam combines the gradient descent algorithm of the adaptive learning rate and the momentum gradient descent algorithm, so as to adapt to the problem of sparse gradient and gradient oscillation.

Step 34) Model training and hyperparameter optimization: use the verification set to evaluate the model in the training process in real time, check the training effect, use the Adam optimizer to iterate, and repeat the training process to make the network parameters tend to be optimal.

Aiming at the possible overfitting problem in CNN-LSTM, the present invention adds Dropout to each layer in the network. Dropout is a regularization method that randomly shields neurons, and the shielded neurons will not be considered as part of the network, that is, they will not participate in the forward propagation operation.

The present invention uses Mean Square Error (Mean Square Error, MSE) as the evaluation of the model loss function and prediction accuracy: the smaller the value of MSE, the better the accuracy of model prediction.

In this embodiment, the hyperparameters that need to be adjusted for the CNN-LSTM model include batch size, training rounds, learning rate, number of LSTM neurons, and the like. Select the hyperparameter combination with the smallest loss function after the model converges as the optimal hyperparameter combination. After adjustment, the hyperparameter settings of the CNN-LSTM vehicle acceleration prediction model are obtained as shown in Table 4:

Table 4 hyperparameter setting table

Step 35) Predict the acceleration of the target vehicle based on the trained CNN-LSTM vehicle acceleration prediction model.

Based on the predicted value and the true value of the predicted sample, the present embodiment selects mean squared error (mean squared error, MSE), root mean squared error (root mean squared error, RMSE), mean absolute error (mean absolute error, MAE) to measure the prediction accuracy. Given below are the calculation formulas for RMSE and MAE.

为预测值。Among them, Y _i is the real value,

for the predicted value.

In order to demonstrate the prediction effect of the model, the acceleration prediction results of the driver group with similar style and the remaining 20 driver groups (not necessarily similar in style) are compared, and the CNN-LSTM model is compared with the traditional LSTM, and the results are as follows:

Table 5 Comparison table of prediction results

It can be seen that the prediction results of the two models of group 1 with similar driving styles are better than those of group 2 with not necessarily similar driving styles, and the prediction error of group 1 of the CNN-LSTM model is 47.5% lower than that of group 2, which shows that the method used in the present invention Models have significant advantages for improving model training and predictive performance. Secondly, the CNN-LSTM model used in the present invention is better than the traditional LSTM method in the prediction tasks of the two groups. In the acceleration prediction of group 1, there is not much difference between the two models, because the similarity between the car-following segments of group 1 is relatively high, and the basic LSTM model can also achieve a better fitting effect. In panel 2, the CNN-LSTM model achieves 47.1% higher accuracy than LSTM, indicating that CNN-LSTM performs better than LSTM in handling complex prediction tasks.

Figure 5 shows the comparison between the actual value of part of the car-following segment of group 1 and the predicted value of the CNN-LSTM model. It can be seen from the figure that the prediction result of the present invention better fits the trend of vehicle acceleration variation, and the prediction effect is good.

The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make many modifications and changes according to the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experiments on the basis of the prior art shall be within the scope of protection defined in the claims.

Claims (10)

1. a vehicle acceleration prediction method considering driving behavior characteristics under a car-following scene, is characterized in that, comprises the following steps: Step 1) Obtain the predicted data set and perform data processing and variable extraction: For the obtained predicted data set, filter and retain the effective car-following segments based on pre-configured rules, and extract feature variables from the effective car-following segment data. The prediction data set includes the vehicle data and the preceding vehicle data; Step 2) Build a CHMM-based driving behavior semantic segmentation model, use feature variables as the input of the driving behavior semantic segmentation model, segment the feature variables into segments with different behavioral characteristics, and evaluate the similarity of driving behavior semantics; Step 3) Construct a CNN-LSTM acceleration prediction model, divide the effective car-following segment data based on the similarity evaluation results to construct an input and output sample set, train the CNN-LSTM acceleration prediction model, and perform vehicle training based on the trained CNN-LSTM acceleration prediction model. Acceleration prediction. 2. the vehicle acceleration prediction method considering driving behavior characteristics under a kind of car-following scene according to claim 1, is characterized in that, described step 1) comprises the following steps: Step 11) Carry out data cleaning on the predicted data set, remove outliers or missing frames greater than 1s, and use linear interpolation to make up for missing segments less than 1s to obtain a valid data set; Step 12) Extracting the effective car-following segment from the obtained valid data set, according to the relative position between the two vehicles, the speed of the target vehicle, whether there is an overtaking event, and the duration of the car-following event. carry out screening retention; Step 13) Extract feature variables from the effective car-following segment data. 3. the vehicle acceleration prediction method considering driving behavior characteristics under a kind of car-following scene according to claim 1, is characterized in that, described step 2) comprises the following steps: Step 21) divide the feature variable into different grades according to the pre-configured segmentation threshold; Step 22) standardize the feature variable sequence with the driver as the unit; Step 23) According to the temporal nature of driving behavior and data distribution characteristics, construct a CHMM-based semantic segmentation model of driving behavior, adopt the method of unsupervised learning, use the standardized characteristic variables as the input of the semantic segmentation model of driving behavior, and convert the characteristic variable data Segmentation into segments with different behavioral characteristics; Step 24) Based on the standardized occurrence frequency distribution and JS divergence calculation method, solve the optimal distribution of behavior semantic duration, behavior semantic frequency distribution, and JS divergence of characteristic variable distribution between each pair of drivers, and obtain the driving Quantitative index value of behavioral similarity. 4. the vehicle acceleration prediction method considering driving behavior characteristics under a kind of car-following scene according to claim 3, is characterized in that, the standardization process in described step 22) is:

Among them, x represents the sequence of characteristic variables; m is the number of drivers, and M represents the number of drivers; l is the number of stable car-following events of driver m, and L represents the number of stable car-following events of the driver; μ _m and σ _m are respectively Indicates the mean and standard deviation of the characteristic variables of the driver m; after standardization, the mean of the characteristic variables is 0, and the standard deviation is equal to 1. 5. the vehicle acceleration prediction method considering driving behavior characteristics under a kind of car-following scene according to claim 3, is characterized in that, in the described CHMM-based driving behavior semantic division model, In HMM, there are two sequence types, where the observation sequence X={x ₁ ,x ₂ ,...,x _T } is obtained by observation, and the hidden state sequence S={s ₁ ,s ₂ ,...,s _T } cannot be observed; HMM is represented by the state transition matrix A, the observation probability matrix B and the initial state probability distribution Π: λ=(A,B,Π), the number of hidden states is I, and the number of all possible values of the observed variable for J; Among them, the state transition matrix A＝[a _ij ] _I×I , the state transition probability is: a _ij ＝P(s _t+1 ＝q _j |s _t ＝q _i ) Observation probability matrix B=[b _ij ] _I×J , the observation value is only determined by the hidden state at the current moment: b _ij =P(x _t =x _j |s _t =q _i ) The initial state probability distribution Π=[π _i ] _1×I represents the probability distribution of the hidden state at the initial moment: π _i =P(s ₁ =q _i ) The number of hidden states of each HMM chain in CHMM is set to N, then the number of hidden states of the CHMM model is N ^Q , where Q represents the number of characteristic variables; the combination of hidden states of the model at any time can be expressed as

HMM is represented by model parameters λ=(A,B,Π), where state transition matrix A={a _u,v }, observation probability B= _bu (X _t ) and initial state probability Π={π _u } is calculated as follows:

The hidden state layer of CHMM decodes the dependencies between feature variables. Considering that the feature variables are all continuous variables, a Gaussian distribution is introduced to calculate the observation probability:

Among them, μ _{c, u} and σ _{c, u} respectively represent the mean value and standard deviation of the probability distribution when the hidden state in the HMM chain corresponding to the characteristic variable c is q _{c, u} . 6. the vehicle acceleration prediction method considering driving behavior characteristics under a kind of car-following scene according to claim 3, is characterized in that, described step 24) comprises the following steps: Step 241) The duration of each behavioral semantic segment extracted by the driver D _m is d _m , f(d _m ) represents the distribution of the duration of the behavioral semantics, and f(d _n ) represents the behavioral semantics extracted by the driver D _n The distribution of duration, by quantifying the closeness of the distributions f(d _m ) and f(d _n ), to measure the similarity of the driving pattern transfer laws of different drivers

Step 242) Measure the similarity of the driver D _m and the driver D _n in the driving mode selection preference by quantifying the proximity of the probability mass function

Considering that under each type of car-following distance condition, the sum of the occurrence probability of behavioral semantics is equal to 1, respectively measure the probability mass function under the car-following distance condition S _Δd

and />

After calculating the mean value of the closeness, get />

S _Δd ∈ {LD,ND,CD} Step 243) Divide the behavioral semantic duration into j segments according to the pre-configured segment threshold, and count the mean and standard deviation of the characteristic variable data of each segmental behavioral semantic segment respectively. Driver D _m is in segment d The mean and standard deviation of the characteristic variable x _k in are recorded as

and />

Using a normal distribution />

Simplify the probability density function of the characteristic variable x _k of the driver D _m in the segment d; calculate the closeness of the probability density function pairs of the characteristic variables in each duration segment in turn, and calculate the average value to obtain the driver D _m and the driver D _n similarity in aggressiveness of driving maneuvers />

Among them, K represents the number of feature variables; step 244)

and />

The similarity between driver D _m and driver D _n is quantified from three dimensions respectively, and the comprehensive evaluation index η _m,n of driving style similarity is obtained by weighted average:

where w _i ∈ [0,1] and

represent the weights of the three driving style similarity indicators respectively; Step 245) given two discrete probability density distributions of standardized frequency distributions

and />

JS divergence is used to measure the similarity of two probability distributions:

Among them, the discrete probability density distribution

and />

KL divergence />

The calculation method is:

and />

Respectively represent the normalized occurrence frequency of driver D _m and driver D _n behavior semantics under the car-following distance condition S _Δd ; The value range of JS divergence is [0,1]. When the two probability distributions are the same, the JS divergence is equal to 0. The greater the difference in probability distribution, the closer the JS value is to 1. 7. the vehicle acceleration prediction method considering driving behavior characteristics under a kind of car-following scene according to claim 1, is characterized in that, described step 3) comprises the following steps: Step 31) According to the similarity evaluation results and the pre-configured similarity threshold, select several drivers who meet the similarity conditions as a group, and classify the effective car-following segments into the training set and the test set according to the pre-configured ratio of the drivers and validation set; Step 32) Construct input and output samples, convert the data containing acceleration and other characteristic variables in the training set and test set into an N-dimensional array, each input sample is composed of characteristic variables at K moments, and the output label sample is the K+1th moment acceleration data; Step 33) build CNN-LSTM vehicle acceleration prediction model; Step 34) model training and hyperparameter optimization: use the verification set to evaluate the model in the training process in real time, check its training effect, use Adam optimizer to iterate, and repeat the training process to make the network parameters tend to be optimal; Step 35) Predict the acceleration of the target vehicle based on the trained CNN-LSTM vehicle acceleration prediction model. 8. the vehicle acceleration prediction method considering driving behavior characteristics under a kind of car-following scene according to claim 7, is characterized in that, described CNN-LSTM vehicle acceleration prediction model comprises input layer, CNN layer, LSTM layer connected successively and a fully connected layer, where, The input layer is used to receive input data; The CNN layer is used to compress and extract the spatial features of the input data, and the CNN layer includes a convolution layer and a maximum pooling layer; The LSTM layer is used to extract the time series features of CNN output data, and the LSTM network in the LSTM layer is many-to-one LSTM; The fully connected layer is used to output acceleration prediction results; Each layer in the model network is added with a Dropout layer, and the mean square error is selected as the evaluation of the model loss function and prediction accuracy. 9. the vehicle acceleration prediction method considering driving behavior characteristics under a kind of car-following scene according to claim 8, is characterized in that, the calculation formula of described convolution layer is as follows:

Among them, σ is the activation function,

is the feature map of the kth layer and the jth dimension, M _j is the set of input maps, />

is the convolution kernel, />

for the bias. 10. the vehicle acceleration prediction method considering driving behavior characteristics under a kind of car-following scene according to claim 8, is characterized in that, the forget gate parameter update formula of described LSTM network is: f _t ＝σ(w _f h _t-1 +w _f x _t +b _f ) Among them, f _t represents the state of the forget gate at time t, w _f and b _f are the weight and bias of the forget gate, respectively, σ represents the sigmoid function, and the output value ranges from 0 to 1, h _t represents the state of the hidden layer, x _t represents the input value of the network at time t; The input gate of the LSTM network is used to update the information in the input cell state: i _t ＝σ(w _i x _t +w _i h _t-1 +b _i ) Among them, i _t represents the state of the input gate at time t, and w _i and b _i are the weight and bias of the input gate, respectively; The LSTM network memory cell state c _t is updated by the information c _t-1 stored in the previous memory cell and the new candidate information:

in,

is the candidate cell state at time t, c _t is the updated cell state, w _c and b _c are the weight and bias of the cell state respectively, . represents the dot product; The output gate of LSTM is used to control the information output in the cell state, and the output hidden layer state h t is calculated according to the memory cell state c _t and the output gate state o _{t :} o _t ＝σ(w _o x _t +w _o h _t-1 +b _o ) h _t =o _t ·tanh(c _t ).

Images