Heterogeneous driving behavior feature extraction and online identification method in man-machine cooperation scene

Abstract

The invention provides a heterogeneous driving behavior feature extraction and online identification method in a man-machine cooperation scene, and the method comprises the steps: obtaining traffic data in multiple man-machine cooperation driving scenes, and carrying out the research based on a data set; carrying out preprocessing methods such as target detection, multi-target tracking and acceleration smoothing on the data set, and extracting characteristic parameters in a car following process; a clustering index is calculated from the characteristic parameters, clustering analysis is carried out on the car following behaviors based on a K-Means clustering algorithm, and the car following behaviors are divided into three types according to driving behavior differences; on the basis of a factor analysis method, common factors are extracted from the characteristic parameters, and heterogeneous driving behavior car-following characteristic parameter extraction is achieved; according to the extracted common factors, an online identification model is constructed through an artificial neural network (ANN), three types of heterogeneous driving behaviors are classified in real time, and the problem that ACC cannot accurately identify the car following characteristics of the heterogeneous driving behaviors under man-machine cooperative driving can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of human-machine collaborative driving behavior analysis, in particular to a heterogeneous driving behavior feature extraction and online identification method in a human-machine collaborative scenario. BACKGROUND

[0002] With the continuous progress of modern science and technology, the development process of vehicle intelligence is also accelerating. As one of the core technologies in the current intelligent driving field, adaptive cruise control (ACC) has become a representative technology in the automobile auxiliary driving system. ACC not only has the function of traditional constant speed cruise control (CC) in the longitudinal control of autonomous vehicles, but also can automatically adjust the throttle and brake mechanism according to the vehicle external environment perception and its own motion state parameters, to realize the longitudinal motion control of maintaining a dynamic safe distance from the front vehicle.

[0003] In the prior art, there are still some inherent problems in the ACC control model. ACC cannot accurately identify the following characteristics of heterogeneous driving behavior in human-machine collaborative driving. The traditional safety distance model strategy cannot be applied to the safety driving preferences of different drivers, and cannot meet the psychological expectations of drivers. SUMMARY

[0004] The present application provides a heterogeneous driving behavior feature extraction and online identification method in a human-machine collaborative scenario to solve the problem that ACC cannot accurately identify the following characteristics of heterogeneous driving behavior in human-machine collaborative driving. By obtaining traffic data in a human-machine collaborative driving scenario, extracting longitudinal driving feature parameters, dividing driving behavior based on a K-Means clustering algorithm, extracting common factors and constructing an online identification model for real-time identification, the problem of ACC not being able to accurately identify the following characteristics of heterogeneous driving behavior in human-machine collaborative driving can be solved.

[0005] To solve the above technical problems, the technical solution adopted by the present application is:

[0006] The heterogeneous driving behavior feature extraction and online identification method in a human-machine collaborative scenario includes the following steps:

[0007] S1, obtain a traffic data set in a human-machine collaborative driving scenario, which includes an adaptive cruise control mode scenario in an urban road environment and an adaptive cruise control mode scenario in a highway environment;

[0008] S2, preprocess the traffic data set to generate the motion trajectory of each vehicle; filter ACC following behavior data, and extract longitudinal driving feature parameters in different human-machine collaborative driving scenarios using the following mode;

[0009] S3, based on the K-Means clustering algorithm, select the maximum value of the reciprocal of the collision time , the average of the headway time and the acceleration variance As a clustering index, clustering analysis is carried out, and the car-following behavior is divided into three types of aggressive type, general type and conservative type according to the difference of driving behavior;

[0010] S4, the factor analysis method is used for data dimension reduction, and the public factor capable of representing heterogeneous driving behavior is extracted from the longitudinal driving characteristic parameters;

[0011] S5, the public factor is used as the input of the artificial neural network model, an online identification model is constructed and trained, the three types of heterogeneous driving behavior are classified, and real-time identification and strategy adjustment are realized.

[0012] As preferred, in S2, the motion trajectory generation method of each vehicle is specifically,

[0013] The data in the data set is time-synchronized and converted in format; Apollo CNN segmentation algorithm is used to process the data, and target detection is carried out with the help of convolutional neural network to generate the bounding box of the detected vehicle; multi-target tracking algorithm is used to track the detected vehicle, and the motion trajectory of each vehicle is generated by combining Kalman filter and Hungarian algorithm.

[0014] As preferred, in S2, the longitudinal driving characteristic parameters include vehicle longitudinal acceleration, vehicle longitudinal speed, relative distance, front vehicle longitudinal acceleration, front vehicle longitudinal speed and minimum longitudinal distance.

[0015] As preferred, in S2, FIR filter is used to smooth the vehicle longitudinal acceleration and front vehicle longitudinal acceleration.

[0016] As preferred, in S2, all traffic data corresponding to the vehicle speed higher than 8.94 m / s are selected as ACC following behavior data.

[0017] As preferred, in S3, the collision time is the estimated time interval required for the collision with the front vehicle under the condition that the vehicle maintains the current motion state, and the calculation formula is:

[0018] (1)

[0019] The headway time is the amount of time required for the vehicle to travel to the rear position of the front vehicle at the current speed, and the calculation formula is:

[0020] (2)

[0021] Wherein, represents the distance between the vehicle and the front vehicle; and respectively represent the vehicle speed of the vehicle and the front vehicle.

[0022] As preferred, in S4, the factor analysis method is specifically:

[0023] S4.1, the initial data of the longitudinal driving characteristic parameters are standardized, and the standardization formula is:

[0024] (3)

[0025] wherein, is the initial data, is the data after standardization, is the average value of the jth sample variable, is the standard deviation of the jth sample variable, and the range of the processed data is between

[0026] S4.2, the correlation between the initial data after standardization is systematically analyzed, and the analysis formula is:

[0027] (4)

[0028] wherein, is the correlation coefficient, is the ith row and kth data, is the jth column and kth data, and are the average of the initial data of the ith row and the jth column, respectively;

[0029] S4.3, KMO test and Bartlett sphericity test are carried out, and the Bartlett sphericity test is based on the correlation coefficient matrix. When the KMO value is greater than 0.5 and the Bartlett sphericity test result is significant, the common factor is further extracted;

[0030] S4.4, the common factor of the longitudinal driving characteristic parameters is extracted based on the principal component analysis method;

[0031] S4.5, the factor rotation is carried out on the correlation coefficient matrix R constructed by the principal component analysis method by using the maximum variance rotation method;

[0032] S4.6, the factor score coefficient matrix is obtained by using the regression method , wherein is the contribution coefficient of the jth original variable to the ith common factor, and then the score expression of the common factor is obtained, and the common factor variable is calculated. The calculation formula of the heterogeneous driving behavior data as a characteristic parameter is:

[0033] (7)

[0034] ​As preferred, in S4.4, the extracted common factors include position factor, speed factor and acceleration factor.

[0035] As preferred, in S5, the construction method of the online recognition model is specifically:

[0036] S5.1, taking the score expression of each common factor as the input of the artificial neural network model , the form is The matrix, in the process of input layer data transformation to hidden layer, linear transformation is adopted to obtain the data of hidden layer ;

[0037] (8)

[0038] In the formula, is the linear change coefficient; is the bias, which provides a fixed input value for the neuron;

[0039] S5.2, the hidden layer adopts ReLu function as the activation function, when the hidden layer data is non-negative, the output is itself; when the hidden layer data is less than 0, the output is 0;

[0040] (9)

[0041] S5.3, from the hidden layer to the output layer, linear transformation is adopted to output the data as ,

[0042] (10)

[0043] The output value is processed by regularization to convert the output into probability score,

[0044] (11)

[0045] In the formula, is the probability of the category appearing, is the total number of categories, is the output value of the category ;

[0046] S5.4, label processing is carried out by adopting one-hot encoding to convert the discrete classification value into binary vector, the original data is divided into test set, validation set and training set according to the proportion, and the online recognition model is trained;

[0047] S5.5. The gensim function in Matlab is used to realize the online modular construction and application of the online identification model, so as to support the real-time identification and strategy adjustment during the system operation.

[0048] As a preferred option, in S5.4, the K-Means clustering results are used as the classification of real heterogeneous driving behaviors. During training, forward propagation calculation and backpropagation optimization are continuously performed to gradually reduce the cross-entropy value of the neural network identification model until convergence.

[0049] The present invention has the following beneficial effects:

[0050] 1. This invention improves the adaptive cruise control system by extracting key vehicle motion and risk perception parameters for cluster analysis, accurately classifying three types of heterogeneous driving behaviors, including aggressive driving, and providing a clear classification framework for driving behavior research. This helps to deeply understand the characteristics of different driving behaviors and lays the foundation for subsequent targeted research.

[0051] 2. This invention uses factor analysis to reduce dimensionality and constructs an online identification model using artificial neural networks. The accuracy rates on urban roads and highways reach 86.99% and 89.39% respectively, which can quickly and accurately identify heterogeneous driving behaviors and provide a reliable basis for adaptive control strategies.

[0052] 3. The present invention designs an effective online identification strategy. Based on the accurate identification results, it can adaptively match the corresponding control strategy, effectively cope with the differences in heterogeneous driving behaviors in terms of scene changes and risk perception, and improve the safety and stability of human-machine collaborative driving. Attached Figure Description

[0053] Figure 1 This is a flowchart of the method of the present invention.

[0054] Figure 2 A schematic diagram of a Simulink module for online identification of heterogeneous driving behaviors.

[0055] Figure 3 This is a comparison chart showing the data before and after smoothing.

[0056] Figure 4 This is a schematic diagram showing the clustering results of urban roads and highways.

[0057] Figure 5 This is a schematic diagram of the correlation matrix of longitudinal driving characteristic parameters.

[0058] Figure 6 This is a graph showing the mean square error results for urban roads and highways. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings and specific preferred embodiments.

[0060] In the description of this invention, it should be understood that the terms "left side," "right side," "upper part," "lower part," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. "First," "second," etc., do not indicate the importance of the components, and therefore should not be construed as a limitation of this invention. The specific dimensions used in this embodiment are only for illustrating the technical solution and do not limit the scope of protection of this invention.

[0061] like Figures 1-6 As shown, a method for extracting and identifying heterogeneous driving behavior features in a human-machine collaborative driving environment includes the following steps.

[0062] S1: Obtain traffic datasets for human-machine cooperative driving scenarios, including adaptive cruise control mode scenarios in urban road environments and adaptive cruise control mode scenarios in highway environments.

[0063] In this embodiment, the research is carried out based on a traffic dataset equipped with ADAS function for human-machine cooperative driving scenarios. This dataset has multi-mode and multi-scenario characteristics, which can realistically present the actual situation of drivers driving on the road with the help of the assistance system, covering typical modes of adaptive cruise control (ACC) in urban road environment and highway environment.

[0064] The Central Ohio dataset is a traffic data collection project undertaken by the Federal Highway Administration (FHWA) in 2021. This dataset covers complex driving scenarios, including two highways (I-270 and I-750) and three urban roads (US23, US33, and US315). The Central Ohio dataset includes data on vehicle trajectories, speeds, accelerations, lane changes, vehicle following, and intersection crossings. This dataset includes two data collection methods: single-vehicle and two-vehicle, referring to data collected from a single ADAS vehicle and two RI-ADAS vehicles, respectively. One part is used for heterogeneous driving behavior classification, while the single-vehicle dataset is specifically used to filter ACC following behavior parameters.

[0065] S2. Preprocess the traffic dataset to generate the motion trajectory of each vehicle; filter the ACC following behavior data and extract longitudinal driving feature parameters under different human-machine cooperative driving scenarios using the following mode.

[0066] In this embodiment, the specific method is as follows.

[0067] The traffic dataset was time-synchronized and format-converted. The Apollo CNN segmentation algorithm was used to process the data, and a Convolutional Neural Network (CNN) was employed for object detection to generate bounding boxes for detected vehicles. Subsequently, a multi-object tracking algorithm was used to track the detected vehicles, generating motion trajectories for each vehicle by combining a Kalman filter and the Hungarian algorithm. Then, the vehicle position information was converted to the Frenet coordinate system, and all traffic data corresponding to vehicle speeds higher than 8.94 m / s were selected to obtain higher-quality car-following data.

[0068] The ACC following behavior data is smoothed. Since directly calculating acceleration from speed data will introduce noise, a Finite Impulse Response Filter (FIR) is used to smooth the vehicle acceleration data. Then, longitudinal driving feature parameters are extracted in different human-machine cooperative driving scenarios, such as urban road environments and highway environments, using the following mode.

[0069] This embodiment uses preprocessing methods such as target detection, multi-target tracking, and acceleration smoothing to process the dataset. The acceleration data smoothing result is as follows: Figure 3 As shown, based on the processed Central Ohio dataset, six longitudinal driving feature parameters of the driver during the following process are extracted, namely vehicle longitudinal acceleration (acc_sv), vehicle longitudinal speed (speed_sv), relative distance (headway), longitudinal acceleration of the preceding vehicle (acc_av), longitudinal speed of the preceding vehicle (speed_av), and minimum longitudinal distance (gap).

[0070] S3. Based on the K-Means clustering algorithm, select the maximum value of the reciprocal of the collision time. Average headway and acceleration variance As a clustering indicator, cluster analysis was conducted, and following behavior was divided into three types: aggressive, general, and conservative, based on differences in driving behavior.

[0071] In this embodiment, the specific method is as follows.

[0072] This invention employs the K-Means clustering algorithm to classify heterogeneous driving behaviors in vehicle-following data, thereby uncovering the intrinsic relationships and potential features among the samples. The K-Means clustering algorithm can be divided into the following four key steps: (a) randomly select k samples from the initial data and set them as initial cluster centers; (b) calculate the Euclidean distance between each sample and each initial cluster center, and assign the sample to the cluster set with the smallest Euclidean distance; (c) subsequently, for each category, use the mean vector of the sample points in that category as the new cluster center, and continuously update the cluster centers through iterative optimization based on the objective function minimization criterion; (d) stop the operation and output the final result until the calculation results converge.

[0073] In selecting the clustering index, a comprehensive consideration was made from the time, space, and dynamic dimensions, and the maximum value of the reciprocal of the collision time was selected. ), average headway ( ) and acceleration variance ( () is used as a clustering index. Among them, collision time refers to the estimated time interval required for the vehicle to collide with the vehicle in front while maintaining its current motion state, and its value is determined by formula (1); headway represents the amount of time required for the vehicle to travel at its current speed to the rear of the vehicle in front, and is determined by formula (2):

[0074] (1) (2)

[0075] In the formula, Indicates the distance between your vehicle and the vehicle in front of you; and These represent the speeds of the vehicle and the vehicle in front, respectively. To avoid a denominator of 0 when the speeds of the vehicle and the vehicle in front are equal, the reciprocal of the collision time is used. Based on the overall numerical differences of iTTC, thmean, and avar (e.g., significantly higher iTTCmax and lower thmean for aggressive drivers) and the cluster centers of each category, heterogeneous driving behaviors are divided into three categories: aggressive, general, and conservative.

[0076] In this embodiment, the clustering result is as follows: Figure 4 (a) Figure 4 (b) and Table 1 are shown. Analysis revealed that in the highway scenario, A larger value indicates that the driver is more sensitive to potential risks when driving at high speeds; Smaller following distances indicate that drivers tend to maintain smaller following distances in high-speed scenarios. This clustering result aligns with the differences in risk perception among various driver groups.

[0077] Table 1 Clustering Results and Classification of Heterogeneous Driving Behaviors

[0078]

[0079] S4. Factor analysis is used to reduce the dimensionality of the data and extract common factors that can characterize heterogeneous driving behavior from the longitudinal driving characteristic parameters. The common factors include position factors, speed factors and acceleration factors to solve the data redundancy problem.

[0080] In this embodiment, the specific method is as follows.

[0081] S4.1 Standardize the initial data to ensure that the analysis is carried out on the same scale. The standardization formula is determined by equation (3).

[0082] (3)

[0083] In the formula, For initial data, For standardized data, Let j be the average of the j-th sample variable. Let be the standard deviation of the j-th sample variable, and the range of the processed data be within... between.

[0084] S4.2. Conduct a systematic analysis of the correlation between the initial data after standardization, and determine it using equation (4):

[0085] (4)

[0086] In the formula, The correlation coefficient, For the k-th data in the i-th row, For the k-th data in the j-th column, and These are the averages of the i-th row and j-th column of the initial data, respectively.

[0087] S4.3. Conduct the Kaiser-Meyer-Olkin (KMO) test and Bartlett's test of sphericity to verify the suitability of the dataset for factor analysis. A KMO value greater than 0.5 indicates that the sample is suitable for factor analysis. The Bartlett's test of sphericity uses the correlation coefficient matrix as a basis to test whether it is significantly different from the identity matrix. If the test value is significant, it indicates that there is a correlation between the variables, making it suitable for further extraction of common factors.

[0088] S4.4 Extracting common factors describing vehicle motion parameters based on principal component analysis. First, for the initial feature parameters... The transformation yields several independent composite parameters. It is determined by formula (5):

[0089] (5)

[0090] In the formula, It is a comprehensive parameter extracted from the linear combination of initial characteristic variables through the maximum variance criterion, and sorted in order of their variance contribution from high to low. This refers to the loadings of the initial feature parameters in the common factors. Next, the correlation coefficient matrix R of the initial data x is solved, and the eigenvalues ​​of the initial data are calculated. With feature vectors The amount of information contained in each principal component reflects its explanatory power and available contribution rate to the original variables. This is indicated by equation (6):

[0091] (6)

[0092] S4.5 The initial factor loadings are relatively concentrated on the feature parameters, resulting in insufficient discrimination and limited representational ability of each factor in characterizing car-following style. Therefore, the variance-maximum rotation method is used to rotate the correlation coefficient matrix R to enhance the identification effect and explanatory power of the factors, so that the factor loadings are concentrated at both ends of 0 and 1 in the horizontal direction, thereby improving the interpretability and representativeness of the factors.

[0093] S4.6 Obtain the factor score coefficient matrix using regression method. ,in Let j be the contribution coefficient of the j-th original variable to the i-th common factor, and then obtain the score expression of the common factor, and calculate the common factor variables. Heterogeneous driving behavior data, which serves as a characteristic parameter, is determined by formula (7).

[0094] (7)

[0095] In this embodiment, to ensure that the data meets the requirements for factor analysis to extract common factors, the data is first standardized and subjected to correlation testing. The correlation matrix of the feature parameters is as follows: Figure 5 As shown in the figure. The KMO and Bartlett test results show that the KMO value is greater than 0.7 and the significance level of the Bartlett test is 0, indicating that there is a significant correlation between the data, which meets the prerequisites for factor analysis.

[0096] Common factors were extracted using principal component analysis. Considering the cumulative variance contribution rate of the characteristic parameters, the top three principal factors were selected as common factors. The eigenvalues ​​of these three factors are all greater than 1, and their cumulative variance contribution rate exceeds 85%, thus effectively preserving the information of the initial characteristic parameters and possessing strong interpretability.

[0097] Analysis of the factor rotation results revealed that: relative distance and minimum vehicle distance have high load coefficients on factor 1, indicating that these two variables are closely related to the position state of the ACC vehicle, with factor 1 serving as the position factor; the speed of the preceding vehicle and the speed of the current vehicle have high load coefficients on factor 2, indicating that vehicle speed is an important factor affecting the ACC vehicle, with factor 2 serving as the speed factor; the acceleration of the preceding vehicle and the acceleration of the current vehicle have high load coefficients on factor 3, indicating that these two variables correspond to the impact of acceleration on the ACC vehicle, with factor 3 serving as the acceleration factor.

[0098] S5. Using position factor, velocity factor and acceleration factor as inputs to an artificial neural network model, an online recognition model is constructed to classify three types of heterogeneous driving behaviors and achieve real-time recognition and strategy adjustment.

[0099] In this embodiment, the specific method is as follows.

[0100] S5.1. Based on the extracted common factors, an online identification model is constructed using an artificial neural network (ANN). This model consists of a two-layer fully connected neural network, comprising an input layer, a hidden layer, and an output layer. The common factor scores obtained through factor analysis are used as the input to the ANN network model. Its form is The matrix ( For data volume, (where the number of common factors is used), during the process of transforming the input layer data into the hidden layer, a linear transformation is employed to obtain the hidden layer data. As shown in formula (8):

[0101] (8)

[0102] In the formula, The coefficient of linear variation; As a bias, it provides a fixed input value to the neuron.

[0103] S5.2 The hidden layer uses the ReLU function as the activation function. When the hidden layer data... When non-negative, output It itself; when hidden layer data When the value is less than 0, the output is 0, as shown in equation (9).

[0104] (9)

[0105] S5.3, From the hidden layer to the output layer, a linear transformation is used to output the data. As shown in equation (10):

[0106] (10)

[0107] And output value The output needs to be regularized to transform it into a probability distribution, the form of which is determined by equation (11):

[0108] (11)

[0109] In the formula, It is a category The probability of occurrence The total number of categories, It is a category The output value.

[0110] S5.4 Next, one-hot encoding is used for label processing, converting discrete classification values ​​into binary vectors. The original data is divided into test, validation, and training sets according to proportions for offline training of the ANN neural network model, and K-Means clustering results are used as the classification of real heterogeneous driving behaviors. During training, forward propagation calculations and backpropagation optimizations are continuously performed to gradually reduce the cross-entropy value of the neural network recognition model until convergence.

[0111] S5.5. Using the gensim function in Matlab, online modular construction and application of neural networks are implemented to support real-time recognition and strategy adjustment during system operation. The model is as follows: Figure 2 As shown, a state transition diagram is designed using Stateflow to divide heterogeneous driving behavior regions. In a real-world scenario, when the sensors detect a target obstacle ahead and the vehicle begins to move, the system starts collecting driving parameters. It identifies interactions with the vehicle, and a recognition cycle is considered valid if the recognition time exceeds 5 seconds. The result with the highest recognition frequency is taken as the driver's driving behavior result for that particular driving action.

[0112] In this embodiment, the score representations of position factor, velocity factor, and acceleration factor obtained based on factor analysis are used as input to the artificial neural network model. This input format is as follows: The matrix was used to construct a complete online recognition model. The Central Ohio dataset was divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. The training results are as follows: Figure 6As shown, the model achieved an accuracy rate of 89.39% in identifying heterogeneous driving behaviors in relatively stable highway scenarios, and 86.99% in urban road scenarios. Notably, in highway scenarios, the model achieved 100% accuracy in identifying both aggressive and conservative drivers.

[0113] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for feature extraction and online identification of heterogeneous driving behavior in a human-machine collaborative scenario, characterized in that, The method comprises the following steps: S1, acquiring traffic data sets in a human-machine collaborative driving scene, the human-machine collaborative driving scene including a self-adaptive cruise control mode scene in an urban road environment and a self-adaptive cruise control mode scene in a highway environment; S2, preprocessing the traffic data sets to generate motion trajectories of each vehicle; S3, screening ACC following behavior data, and extracting longitudinal driving feature parameters in different human-machine collaborative driving scenes by using a following mode; S3, Select the maximum value of the reciprocal of the collision time based on the K-Means clustering algorithm , Mean of headway and acceleration variance As a clustering index, clustering analysis is carried out, and the following three types are divided according to the differences in driving behavior: aggressive, general and conservative. S4, using a factor analysis method to reduce the dimension of data, and extracting common factors capable of representing heterogeneous driving behaviors from the longitudinal driving feature parameters; S5, taking the common factors as inputs of an artificial neural network model, constructing an online identification model, and training the online identification model to classify three types of heterogeneous driving behaviors, thereby realizing real-time identification and strategy adjustment.

2. The method of claim 1, wherein the method further comprises: In S2, the method for generating the motion trajectories of each vehicle is specifically as follows: The data in the data set is time-synchronized and converted in format; the data is processed by using an Apollo CNN segmentation algorithm, and target detection is carried out by using a convolutional neural network to generate a bounding box of a detected vehicle; a multi-target tracking algorithm is used to track the detected vehicle, and a Kalman filter and a Hungarian algorithm are combined to generate the motion trajectories of each vehicle.

3. The method of claim 1, wherein the method further comprises: In S2, the longitudinal driving feature parameters include vehicle longitudinal acceleration, vehicle longitudinal speed, relative distance, front vehicle longitudinal acceleration, front vehicle longitudinal speed, and minimum longitudinal distance.

4. The method of claim 3, wherein the method further comprises: In S2, a FIR filter is used to smooth the vehicle longitudinal acceleration and the front vehicle longitudinal acceleration.

5. The method of claim 1, wherein the method further comprises: In S2, all traffic data corresponding to a vehicle speed higher than 8.94 m / s are screened as ACC following behavior data.

6. The method of claim 1, wherein the method further comprises: In S3, the collision time is an estimated time interval required for a collision to occur between the host vehicle and the front vehicle under the condition that the host vehicle maintains the current motion state, and the calculation formula is as follows: (1)； The headway time is the time required for the host vehicle to travel to the rear position of the front vehicle at the current speed, and the calculation formula is as follows: (2)； wherein, represents the inter-vehicle distance between the ego vehicle and the preceding vehicle; and respectively represent the ego vehicle speed and the preceding vehicle speed.

7. The method of claim 1, wherein the method further comprises: In S4, the factor analysis method is specifically as follows: S4.1, the initial data of the longitudinal driving feature parameters are standardized, and the standardization formula is as follows: (3)； wherein, is the initial data, is the standardized data, is the mean value of the jth sample variable, is the standard deviation of the jth sample variable, the range of the processed data is between and S4.2, the correlation between the standardized initial data is systematically analyzed, and the analysis formula is as follows: (4)； wherein is the correlation coefficient, is the i-th row and k-th data, is the j-th column and k-th data, and are the average of the initial data i-th row and j-th column, respectively; S4.3, KMO test and Bartlett sphericity test are carried out, and the Bartlett sphericity test is based on the correlation coefficient matrix; when the KMO value is greater than 0.5 and the Bartlett sphericity test result is significant, the common factors are further extracted; S4.4, the common factors of the longitudinal driving feature parameters are extracted based on the principal component analysis method; S4.5, the variance maximum rotation method is used to rotate the correlation coefficient matrix R constructed by the principal component analysis method; S4.6, obtain the factor score coefficient matrix by regression method wherein is the contribution coefficient of the jth original variable to the ith common factor, and then obtain the score expression of the common factor, and calculate the common factor variable Heterogeneous driving behavior data as a characteristic parameter, The calculation formula is: (7)。 8.The method of claim 7, wherein, In S4.4, the extracted common factors include a position factor, a speed factor, and an acceleration factor. 9.The method of claim 1, wherein, In S5, the construction method of the online identification model is specifically as follows: S5.1, expressing the score expression of each public factor as an input of an artificial neural network model , in the form of a matrix, in the process of transforming the input layer data to the hidden layer, the data of the hidden layer is obtained by linear transformation ;​ (8)； wherein is a linear change coefficient; is a bias, providing a fixed input value to the neuron; S5.2, the hidden layer adopts ReLu function as the activation function, when the hidden layer data Non-negative, output Itself; when the hidden layer data Less than 0, the output is 0; (9)； S5.3, the hidden layer to the output layer, using a linear variation to output the data as , (10)； Output value The output is converted into a probability distribution by regularization. (11)； wherein is the class probability of occurrence, is the total number of classes, is the class output value; S5.4, label processing is carried out by using one-hot encoding, discrete classification values are converted into binary vectors, the online identification model is obtained, and the original data is divided into a test set, a validation set, and a training set according to a proportion, and the online identification model is trained. S5.5, using the gensim function in Matlab, the online module of the online identification model is constructed and applied to support the real-time identification and strategy adjustment of the system during operation. 10.The method of claim 9, wherein, In S5.4, the K-Means clustering result is used as the real heterogeneous driving behavior classification. In the training, the forward propagation calculation and the back propagation optimization are continuously performed to gradually reduce the cross-entropy value of the neural network identification model until convergence.

Images