Driving behavior semantic understanding-based driving style recognition method

Abstract

The present application belongs to the field of personalized driving technology, and relates to a driving style recognition method in the process of driving based on semantic understanding of driving behavior. The present application considers driving operation in the process from multiple dimensions based on vehicle trajectory information, and quickly extracts and identifies potential driving style characteristics of drivers. First, semantic understanding rules of driving behavior are established, and continuous trajectory time series data is discretized into semantic understanding vocabulary of driving behavior from the decision layer, operation layer, mode layer and scene perception layer of driving operation. Second, multiple driving styles are defined according to the topic perplexity index. Finally, the driving style of each driver in the process is recognized as a combination of probabilities of different driving styles to reflect the current driving state of the driver. The driving style recognition method in the process proposed by the present application explains the phenomenon that the driving styles of the same driver in different driving environments are different, and improves the comprehensiveness and understandability of driving style recognition in the process.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of personalized driving, and particularly relates to a driving style recognition method in the process based on semantic understanding of driving behavior. BACKGROUND

[0002] In recent years, automatic driving technology has developed rapidly, but due to the complexity of the traffic environment and the influence of the intelligent vehicle industrialization process in China, fully automatic driving technology still needs a long time to be popularized. Therefore, in the next 10 years, the man-machine co-driving mode will continue to exist. The man-machine co-driving mode refers to the cooperative operation between the driver and the intelligent system in the vehicle partially automated driving mode. Correctly analyzing and understanding the driving style of the driver in the process can enhance the understanding of the driving behavior of the driver by the intelligent system in the man-machine co-driving mode, improve the online matching degree of the driver and the intelligent system, and thus improve the satisfaction of the driver and the passenger.

[0003] At present, the research methods for driving style mainly have two categories. One is from the subjective perspective, and the driving style is analyzed by designing a questionnaire, but due to the different understanding degrees of the driver to the questionnaire, the survey results may have deviations. The other method is from the objective perspective, and the driving data are obtained through the driving simulator or real vehicle test, and specific driving feature parameters are extracted from the data to analyze the driving style. The existing researches often use speed, acceleration and other feature parameters to identify the driving style as the inherent attribute of the driver, and rarely consider the heterogeneity and inconsistency of the driving style in the driving operation of the driver. The heterogeneity means that the driving styles of different drivers are different, and the inconsistency means that the driving style of the same driver may be different in different driving environments and scenes, such as the conservative driver may also make aggressive behaviors such as emergency braking.

[0004] It has been found in the existing researches that when the driving mode and the driving environment are different, the driver does not always show one driving style. Ma et al. analyzed the online car-hailing data and found that the driver's driving style changes when performing different driving tasks (such as cruising, taking orders), and the driving style also differs in the process of acceleration, deceleration and turning. Chen et al. analyzed the following behavior and found that the driver's internal long-term driving characteristics change in the short term after experiencing external stimuli, and based on this, a vehicle following model was proposed which fuses long-term and short-term driving styles. Therefore, it is difficult to reflect the driving style in the process by using only speed, acceleration and other feature parameters. In the process of driving style recognition, the semantic understanding of the driving behavior of the driver should be carried out, and the potential driving style in the process of the driver should be extracted from the semantic understanding vocabulary of the driving behavior.

[0005] Therefore, in the development of the man-machine co-driving system, the heterogeneity and inconsistency of the driving style should be considered, the information is introduced into the vehicle supervision system and the vehicle decision control system, and a method for identifying the driving style in real time online during the vehicle operation is established, so that the driving safety and comfort of different types of drivers and passengers are improved. SUMMARY

[0006] The application aims to provide a driving style recognition method in the process of driving based on a latent Dirichlet allocation (LDA) topic model and driving behavior semantic understanding, considering the heterogeneity and inconsistency of the driving style in the driving operation process. To achieve the above-mentioned purpose, the main content of the application is as follows: first, the driving behavior semantic understanding rules are established, and the continuous trajectory time series data is discretized into driving behavior semantic understanding vocabulary from the decision layer, operation layer, mode layer and scene perception layer of the driving operation; second, the number of driving style topics is divided according to the topic perplexity index, and different driving style topics are explained through topic analysis; finally, the driving style of each driver in the process is identified as a combination of the above-mentioned driving style probability.

[0007] The technical scheme of the application is as follows:

[0008] The driving style recognition method in the process of driving based on driving behavior semantic understanding is as follows:

[0009] (1) Data acquisition

[0010] Collecting vehicle-to-everything (V2X) data in the driving operation process, the data frequency is 1Hz-10Hz, the data collection time is not less than 10s, including vehicle position, longitudinal speed, lateral speed, longitudinal acceleration, lateral acceleration, time headway (THW), time to conclusion (TTC) and perception data (surrounding vehicle information, weather, road information).

[0011] (2) Data preprocessing

[0012] The trajectory data is preprocessed to obtain the data form required by the driving behavior semantic understanding, including vehicle longitudinal speed (V x ), lateral speed (V y ), longitudinal acceleration (a x ), longitudinal acceleration change rate (a xx ), time headway (THW) and time headway change rate (d(THW) / dt).

[0013] (3) Driving behavior semantic understanding

[0014] The driving operation main task mainly includes four levels: decision level, mode level, operation level and scene perception level. Considering the four levels of driving operation, the driving behavior scene semantic understanding division principle is established, and the trajectory time series data is converted into discrete driving behavior semantic understanding vocabulary. Specifically as follows:

[0015] The decision level semantic division is based on longitudinal velocity and longitudinal acceleration characteristic parameters, and the longitudinal velocity distribution and safety evaluation index are combined to establish the longitudinal velocity and longitudinal acceleration evaluation basis. The speed evaluation index is divided into very low (IV), low (III), high (II) and very high (I) in turn, and the speed range is [0, 30), [30, 70), [70, 120), [120, ∞) / (km·h -1 ) respectively. The acceleration evaluation index is safe acceleration, risk acceleration, high risk acceleration and strong risk acceleration in turn, and the acceleration range is [0, 1), [1, 2), [2, 3), [3, ∞) / (m·s -2 ) respectively.

[0016] The operation level semantic division is based on longitudinal acceleration and longitudinal acceleration rate of change parameters. When the longitudinal acceleration is greater than 0, if the absolute value of the longitudinal acceleration rate of change is in the range of [0, 0.2) / (g·s -1 ), the operation state is normal acceleration, and if the absolute value of the longitudinal acceleration rate of change is in the range of [0.2, ∞) / (g·s -1 ), the operation state is rapid acceleration. When the longitudinal acceleration is less than 0, if the absolute value of the longitudinal acceleration rate of change is in the range of [0, 0.8) / (g·s -1 ), the operation state is normal deceleration, and if the absolute value of the longitudinal acceleration rate of change is in the range of [0.8, ∞) / (g·s -1 ), the operation state is rapid deceleration. g = 9.8 / (m·s -2 ), which represents the acceleration of gravity.

[0017] The mode level semantic division divides the driving mode into 9 from the horizontal and vertical directions. The division process includes:

[0018] The horizontal dimension driving mode is divided according to the horizontal velocity v y . In the process of lane changing, the vehicle horizontal velocity will change suddenly, and the lane changing mode is set when the horizontal velocity exceeds 0.3 / (m·s -1 ). Otherwise, it is a non-lane changing mode. According to the headway, the lane changing mode is divided into free lane changing and restricted lane changing. If the headway THW>3s, the driving mode is free lane changing, that is, the lane changing mode when the distance between the ego vehicle and the front vehicle is greater than the interaction distance, and if the headway THW≤3s, the driving mode is restricted lane changing, that is, the lane changing mode when the distance between the ego vehicle and the front vehicle is less than the interaction distance.

[0019] The driving mode in the longitudinal dimension in the non-lane-changing mode is divided according to longitudinal parameter characteristics, including deceleration a x , time headway (THW), and time headway rate (dTHW / dt). When the vehicle acceleration is less than -3 / (m·s -2 ), the current driving mode is emergency braking, otherwise the vehicle is in normal driving mode. When the time headway THW is greater than 3s, the driving mode is free straight driving, that is, the headway distance is greater than the maximum interaction distance, the vehicle is not restricted by surrounding vehicles, and is in a free driving state. When the time headway THW is less than or equal to 3s, the vehicle is in a car-following mode, which is divided into close car-following, steady car-following, and gradual car-following according to the time headway rate. If the time headway rate is greater than 0.03 at this time, the driving mode is close car-following, that is, the vehicle speed is greater than the front vehicle speed and gradually shortens the distance with the front vehicle. If the time headway rate is less than -0.03 at this time, the driving mode is gradual car-following, that is, the vehicle speed is less than the front vehicle speed, which leads to the gradual increase of the relative distance between the two vehicles. If the time headway rate is in the range of [-0.03, 0.03] at this time, the current driving mode is steady car-following, that is, the relative speed of the two vehicles is small, and the relative stable driving is maintained. According to the time headway, the steady car-following is further divided into long-distance steady car-following, medium-distance steady car-following, and short-distance steady car-following. When the vehicle is in the steady car-following driving mode, if the time headway THW is less than 1.0s, the vehicle is in the short-distance steady car-following driving mode, if the time headway 1.0s≤THW<2.0s, the vehicle is in the medium-distance steady car-following driving mode, and if the time headway THW is greater than or equal to 2.0s, the vehicle is in the long-distance steady car-following driving mode.

[0020] The scene perception layer needs relevant information of the target vehicle surrounding vehicles, weather, and road information, etc. which have an impact on driving. The perception data can be obtained by relying on the vehicle-road cooperation system (V2X). Among them, the surrounding vehicles include the front vehicle, rear vehicle, side vehicle, oblique front vehicle, and oblique rear vehicle of the target vehicle. In the driving process, the front vehicle has the greatest impact on the driver, and the distance between the vehicle and the front vehicle is an important index in driving safety research. The side vehicle and the oblique front and rear vehicle have a greater impact on the vehicle in the lane-changing mode. Weather information includes sunny, cloudy, rainy, foggy, and snowy. Road information includes the number of lanes, lane width, traffic signs, and speed limit, etc.

[0021] (4) LDA topic model

[0022] The LDA topic model can be expressed as a hierarchical Bayesian model with parameters, as follows:

[0023]

[0024] z m,n |θ m ~ Multinomial(θ m ) (2)

[0025] θ m ~ Dirichlet(α) (3)

[0026]

[0027] where w m,n denotes the word, subject to a multinomial distribution with parameters in equation (1); denotes the word distribution of topic k, is an N m dimensional vector, and has k = 1, 2, 3, … K; for topic k ∈ [1, K], is obtained through a Dirichlet distribution with parameters β in equation (4); z m,n denotes the topic of the nth word in document m, subject to a multinomial distribution with parameters m in equation (2); θ m denotes the topic distribution of document m, is a K-dimensional vector, and has m = 1, …, M, and has n = 1, …, N m , θ m is obtained according to a Dirichlet distribution with parameters β in equation (3); α, β represent the hyperparameters of the document-topic distribution and the topic-word distribution, respectively; M represents the number of documents; N m denotes the words in the document, and K represents the number of topics.

[0028] The driver trajectory data file can be regarded as a document, the driving behavior semantic understanding vocabulary can be regarded as a word, and the driving style can be regarded as a topic. Wherein ω m , m ∈ [1, M] denotes the driving behavior semantic vocabulary of driver m, denotes the driving behavior semantic vocabulary distribution of topic k, θ m , m ∈ [1, M] denotes the mixed driving style distribution of driver m. The underlying structure of the driving style mined from the driving behavior semantic understanding vocabulary by using the LDA model is: the driver semantic understanding vocabulary is subject to the driving behavior semantic vocabulary distribution with parameters , corresponding to K driving style topics, and the on-the-go driving style of the driver can be defined as a probability combination of different driving style topics to describe the real-time driving state of the driver, i.e., the on-the-go driving style of the driver.

[0029] (5) Topic discovery

[0030] The application determines the number of topics in combination with a topic perplexity and a topic consistency index, uses a log perplexity index to evaluate the driving style classification results under different topic number conditions, the definition of the log perplexity is shown in formula (5), the topic consistency is an important index for explaining the potential topics generated by the topic model, the topic is easy to explain, and the topic words with high importance in the topic appear more frequently in the corresponding document, and the greater the topic consistency, the better the model effect.

[0031]

[0032] wherein p (ω m |M) is the topic perplexity, omega m ,m [1, M] represents the driving behavior semantic vocabulary of the driver m, represents the driving behavior semantic vocabulary distribution of the topic k, theta m ,m [1, M] represents the mixed driving style distribution of the driver m.

[0033] (6) Topic analysis

[0034] The topic analysis is visualized by using a pyLDAvis visualization tool, the proportions of different topics and the correlation between the topics are analyzed, and the importance and rarity of different topic words for the topic are displayed; the current driver state represented by each driving style is defined and analyzed.

[0035] (7) In-route driving style recognition

[0036] The in-route driving style of the driver is determined as a probability combination of different driving style topics in step (5) by using the in-route driving data of steps (1) and (2), so as to reflect the current driving state of the driver, and the sum of probabilities of different driving style topics is 1, the in-route driving style can reflect the current driving state of the driver and the influence of the driving environment on the current driving style of the driver in real time, and the heterogeneity and inconsistency of the driving style are fully considered.

[0037] The in-route driving style recognition method provided by the application can consider the heterogeneity and inconsistency of the driver in the driving operation process, explain the phenomenon that the driving styles of the same driver are different under different driving environments, improve the comprehensiveness and understandability of the in-route driving style recognition, reflect the driving style characteristics more carefully, meet the computing power requirements of the online driving style recognition, make real-time judgments on the current state of the driver, and can be applied to the development of the man-machine co-driving system and the intelligent driving system to improve the driving safety and comfort. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1is the driving style recognition flowchart of the present application.

[0039] Figure 2 is the driving behavior semantic understanding schematic diagram in step (3) of the present application.

[0040] Figure 3 is the mode layer semantic understanding division rule in step (3) of the present application.

[0041] Figure 4 is the surrounding vehicle distribution schematic diagram in the scene perception layer in step (3) of the present application.

[0042] Figure 5 is the LDA topic model bottom structure diagram in the driving style recognition on the way in step (4) of the present application.

[0043] Figure 6 is the driving style topic perplexity change curve in step (5) of the present application.

[0044] Figure 7 is the driving style topic consistency change curve in step (5) of the present application.

[0045] Figure 8 is the driving style distribution diagram of the driver on the way in step (7) of the present application. DETAILED DESCRIPTION

[0046] The specific embodiments of the present application are described in detail below in combination with examples, and the implementation effects of the present application are simulated.

[0047] As shown in Figure 1 , the specific process of the present embodiment is as follows:

[0048] (1) Data acquisition

[0049] The highD vehicle trajectory open source data set is used in the embodiment case of the present application, and through data screening, finally 483 vehicle data are selected as the research object.

[0050] (2) Data preprocessing

[0051] The trajectory data are preprocessed to obtain the data form required by the driving behavior semantic understanding, including the vehicle longitudinal speed (V x ), lateral speed (V y ), longitudinal acceleration (a x ), longitudinal acceleration change rate (a xx ), time headway (THW) and time headway change rate (d(THW) / dt). The vehicle part data are shown in Table 1.

[0052] Table 1 vehicle part data preprocessing results

[0053]

[0054] (3) Driving behavior semantic understanding

[0055] According to Figure 2 The driving behavior semantic understanding method proposed discretizes continuous trajectory data into driving behavior semantic understanding data in four dimensions, including the decision layer, the mode layer, the operation layer, and the scene perception layer, comprehensively considers the four levels of driving operations, establishes driving behavior scene semantic understanding division principles, and converts trajectory time series data into discretized driving behavior semantic understanding vocabulary.

[0056] The decision layer semantic division is based on speed and acceleration characteristic parameters, and establishes speed and acceleration evaluation criteria combined with speed distribution and safety evaluation indexes. The speed evaluation index is divided into very low (IV), low (III), high (II), and very high (I) in turn, with a speed range of [0, 30), [30, 70), [70, 120), [120, ∞) / (km·h -1 ). The acceleration evaluation index is safe acceleration, risk acceleration, high-risk acceleration, and strong-risk acceleration in turn, with an acceleration range of [0, 1), [1, 2), [2, 3), [3, ∞) / (m·s -2 ).

[0057] The operation layer semantic division is based on acceleration and jerk parameters. When the acceleration is greater than 0, if the jerk range is [0, 0.2) / (g·s -1 ), the operation state is normal acceleration, and if the jerk range is [0.2, ∞) / (g·s -1 ), the operation state is rapid acceleration. When the acceleration is less than 0, if the jerk range is [0, 0.8) / (g·s -1 ), the operation state is normal deceleration, and if the jerk range is [0.8, ∞) / (g·s -1 ), the operation state is rapid deceleration.

[0058] The mode layer semantic division is divided into 9 driving modes from the longitudinal and lateral dimensions, and the division flowchart is as shown in Figure 3 , and the details are as follows:

[0059] The driving mode in the longitudinal dimension is divided according to the longitudinal parameter characteristics, including deceleration a x , time headway (THW, Time Headway), time headway change rate (dTHW / dt), and lateral speed a y . When the vehicle deceleration exceeds 3 / (m·s -2When the vehicle deceleration does not exceed 3 m / s, the current driving mode is emergency braking. -2 When the vehicle is in normal driving mode, if the headway (THW) is greater than 3 seconds, the driving mode is free straight driving, meaning the headway is greater than the maximum interaction distance, and the vehicle is not constrained by surrounding vehicles and is in a free driving state. If the headway (THW) is greater than 3 seconds, the vehicle is in car-following mode, which is divided into approaching car-following, steady-state car-following, and gradually distancing car-following based on the rate of change of headway. Approaching car-following refers to a driving mode where the vehicle's speed is greater than the speed of the vehicle in front, thus gradually shortening the distance between the two vehicles. Steady-state car-following refers to a driving mode where the relative speeds of the two vehicles are relatively low, maintaining a relatively stable driving mode. Gradually distancing car-following refers to a driving mode where the vehicle's speed is less than the speed of the vehicle in front, resulting in a gradually increasing relative distance between the two vehicles. This study further divides steady-state car-following into long-distance steady-state car-following, medium-distance steady-state car-following, and short-distance steady-state car-following based on the headway. When the vehicle is in steady-state car-following driving mode, if the distance to the front of the vehicle (THW) is less than 1.0s, the vehicle is in close-range steady-state car-following driving mode; if the distance to the front of the vehicle (THW) is 1.0s ≤ THW < 2.0s, the vehicle is in medium-range steady-state car-following driving mode; and if the distance to the front of the vehicle (THW) is ≥ 2.0s, the vehicle is in long-range steady-state car-following driving mode.

[0060] Lateral driving mode is based on lateral speed a y During lane changes, the vehicle's lateral speed will change abruptly. In follow-the-car mode and before and after lane changes, the vehicle's lateral speed must not exceed 0.3 m / s. -1 The lane-changing modes are categorized into free lane changing and restricted lane changing based on the distance between the vehicles. Free lane changing refers to a lane-changing mode where the distance between the vehicle and the vehicle in front is greater than the interaction distance, while restricted lane changing refers to a lane-changing mode where the distance between the vehicle and the vehicle in front is less than the interaction distance.

[0061] The scene perception layer requires information affecting driving operations, such as surrounding vehicles, weather, and road conditions, which can be obtained using vehicle-to-everything (V2X) perception data. Surrounding vehicles include those in front of, behind, to the side of, diagonally in front of, and diagonally behind the target vehicle. Figure 4 As shown, during driving operations, the vehicle in front has the greatest impact on the driver, and the distance between the vehicle and the vehicle in front is an important indicator in driving safety research. Vehicles to the side, as well as vehicles diagonally in front and behind, have a significant impact on the vehicle during lane changing. Weather information includes sunny, cloudy, rainy, foggy, and snowy weather, while road information includes the number of lanes, lane width, traffic signs, and speed limits. In this embodiment, the scene perception layer only includes information about surrounding vehicles, and only analyzes the impact of surrounding vehicles. If weather and road condition information is included, it can also be input into the model. Table 2 shows some data examples.

[0062] Table 2 Vehicle partial driving behavior semantic understanding data

[0063]

[0064] (4) LDA topic model

[0065] The driving behavior semantic understanding vocabulary is input into the LDA topic model. The underlying structure of the driving style is mined from the driving behavior semantic understanding vocabulary by using the LDA model, as shown in Figure 5

[0066] (5) Topic discovery

[0067] Figure 6 and Figure 7 are respectively the topic perplexity variation curve and the topic consistency variation curve drawn according to different numbers of topics. If the number of topics is too small, it is difficult to distinguish different driving styles. If the number of topics is infinitely increased, the model perplexity will certainly always decrease, but too many numbers of topic classifications will cause model overfitting. Therefore, the number of topics is set to 1-9 in the process of drawing the topic perplexity curve and the topic consistency curve, and the number of driving styles of the LDA topic model is determined to be four.

[0068] (6) Topic analysis

[0069] The proportions of the four driving styles obtained by the topic model are: DS1 (28.4%), DS2 (26.8%), DS3 (24.9%), and DS4 (19.8%). In order to analyze the four driving styles, the importance and rarity of the driving behavior semantic understanding vocabulary are comprehensively considered, and the first five driving behavior semantic understanding vocabularies in the four driving styles are extracted for topic analysis, as shown in Table 3.

[0070] Table 3 Driving style topic word distribution

[0071]

[0072] The four driving styles are defined by the driving style topic word distribution as follows:

[0073] DS1 is defined as stable. The driver of this driving style tends to drive at low risk acceleration, the operation mode is normal deceleration, the driving mode is gradually far following, free straight driving, free lane changing, and other relatively stable driving operations. At this time, the driver's driving environment is less restricted, and the surrounding vehicles have less impact on him.

[0074] ​DS2 is defined as moderate type, and the driver of this driving style is more likely to drive with medium risk acceleration, and the operation mode is rapid acceleration, and the driving mode is medium distance following, long distance following, and close following, which is more conservative than the driving behavior of DS1. At this time, the driver is greatly affected by the driving environment, and the surrounding vehicles have more restrictions on the driver.

[0075] DS3 is defined as conservative type, and the driver of this driving style is more likely to drive at low speed, and the operation mode is normal acceleration, and the driving mode is long distance following, gradually far following, and free straight driving.

[0076] DS4 is defined as aggressive type, and the driver of this driving style is more likely to drive with high risk acceleration, and the operation mode is rapid acceleration, and the driving mode is close following, close following, and restricted lane changing. Compared with the previous three driving styles, the driver of DS4 driving style is more aggressive.

[0077] (7) In-Route Driving Style Analysis

[0078] The in-route driving style of the driver is the probability combination of the four driving styles, and the probability sum is 1, Figure 8 The four driving style distribution probabilities of the extracted 30 drivers are the probabilities of the driver showing a certain driving style, which may be related to the driving environment of the driver.

[0079] Taking drivers 1, 5, 7, and 29 as examples, the in-route driving style recognition results are analyzed. The probability of driver 29 showing DS1 is 0.96, indicating that the driver is in free driving mode most of the time and is less affected by the external environment, so the driving style is stable. Driver 1 has a 0.19 probability of showing a stable type, which may be related to the fact that the surrounding vehicles have less impact on him in the driving environment. When the surrounding vehicles have a greater impact on him, driver 1 tends to show a moderate type. The probability of driver 5 showing DS4 is 0.78, and the probabilities of showing DS2 and DS3 are both less than 0.1, indicating that the surrounding vehicles have some restrictions on driver 5 in the current driving environment, and the driver tends to show an aggressive driving style. The probability of driver 7 showing DS3 is 0.68, and the probability of showing DS1 is 0.18, indicating that the driver is less affected by the surrounding vehicles at some times, but when the surrounding vehicles have a greater impact on him, the driver tends to show a conservative type.

[0080] The in-route driving style recognition method can explain the driving behavior of the same driver individual in different driving environments, can meet the requirements of short-time online recognition in computing power, can infer the current in-route driving style according to the short-term trajectory data of the current stage of the driver, and has important role in improving the user travel safety and driving experience. Future research will further optimize the encoding process of driving behavior semantic understanding vocabulary, and the research results can be applied in the field of driver behavior decision analysis, and the driving behavior of the driver in different driving environments is analyzed and predicted, and the understanding of the intelligent system to the driving behavior preference of the driver under the condition of man-machine co-driving is improved.

Claims

1. A method for identifying driving style in transit based on semantic understanding of driving behavior, characterized in that, The steps are as follows: (1) Data acquisition Collecting car-road cooperation data during driving operation, the data frequency is 1Hz-10Hz, the data collection time is not less than 10s, including vehicle position, longitudinal speed, lateral speed, longitudinal acceleration, lateral acceleration, headway, expected collision time and perception data, the perception data including surrounding vehicle information, weather and road information; (2) Data preprocessing The trajectory data is preprocessed to obtain the data form required for driving behavior semantic understanding, including vehicle longitudinal speed, lateral speed, longitudinal acceleration, longitudinal acceleration change rate, headway and headway change rate; (3) Driving behavior semantic understanding The driving operation main task contains four levels: decision level, mode level, operation level and scene perception level. Considering the four levels of driving operation, the driving behavior scene semantic understanding division principle is established to convert the trajectory time series data into discrete driving behavior semantic understanding vocabulary; the specific steps are as follows: The decision layer semantic division understanding is divided according to longitudinal speed and longitudinal acceleration characteristic parameters, longitudinal speed distribution and safety evaluation indexes are combined to establish longitudinal speed and longitudinal acceleration evaluation basis; the speed evaluation indexes are divided into very low IV, low III, high II and very high I in sequence, and the speed ranges are [0, 30), [30, 70), [70, 120) and [120, ∞) / (km·h -1 ); the acceleration evaluation indexes are safe acceleration, risk acceleration, high risk acceleration and strong risk acceleration in sequence, and the acceleration ranges are [0, 1), [1, 2), [2, 3) and [3, ∞) / (m·s -2 ); The operation layer semantic division is based on longitudinal acceleration and longitudinal acceleration change rate parameters; when the longitudinal acceleration is greater than 0, if the longitudinal acceleration change rate absolute value range is [0, 0.2) / (g·s -1 ), the operation state is normal acceleration, if the longitudinal acceleration change rate absolute value range is [0.2, ∞) / (g·s -1 ), the operation state is rapid acceleration; when the longitudinal acceleration is less than 0, if the longitudinal acceleration change rate absolute value range is [0, 0.8) / (g·s -1 ), the operation state is normal deceleration, if the longitudinal acceleration change rate absolute value range is [0.8, ∞) / (g·s -1 ), the operation state is rapid deceleration; g = 9.8 / (m·s -2 ) represents the gravitational acceleration; The mode level semantic division divides the driving mode into 9 categories from the lateral and longitudinal directions, and the division process includes: The lateral dimension driving mode is divided according to lateral speed, in the process of lane changing, the lateral speed of the vehicle will change suddenly, and the lane changing mode is set when the lateral speed exceeds 0.3 / (m·s -1 ) and the non-lane changing mode is set otherwise; the lane changing mode is divided into free lane changing and limited lane changing according to the time headway; if the time headway THW>3s, the driving mode is free lane changing, that is, the lane changing mode when the distance between the ego vehicle and the preceding vehicle is greater than the interaction distance, and if the time headway THW≤3s, the driving mode is limited lane changing, that is, the lane changing mode when the distance between the ego vehicle and the preceding vehicle is less than the interaction distance; The driving mode in the longitudinal dimension in the non-lane-changing mode is divided according to longitudinal parameter characteristics, including deceleration, headway and headway rate of change; when the vehicle acceleration is less than -3 / (m·s -2 ), the current driving mode is emergency braking, otherwise the vehicle is in normal driving mode; when the headway THW is greater than 3s, the driving mode is free straight driving, that is, the headway is greater than the maximum interaction distance, the vehicle is not restricted by surrounding vehicles and is in a free driving state; when the headway THW is less than or equal to 3s, the vehicle is in a car-following mode, which is divided into close car-following, steady car-following and distant car-following according to the headway rate of change; if the headway rate of change is greater than 0.03 at this time, the driving mode is close car-following, that is, the vehicle speed is greater than the front vehicle speed and gradually shortens the distance with the front vehicle; if the headway rate of change is less than -0.03 at this time, the driving mode is distant car-following, that is, the vehicle speed is less than the front vehicle speed, which leads to the gradual increase of the relative distance between the two vehicles; if the headway rate of change is in the range of [-0.03, 0.03] at this time, the current driving mode is steady car-following, that is, the relative speed of the two vehicles is small and the relative stable driving is maintained; and the steady car-following is further divided into distant steady car-following, medium-distance steady car-following and close steady car-following according to the headway; when the vehicle is in the steady car-following driving mode, if the headway THW is less than 1.0s, the vehicle is in the close steady car-following driving mode, if the headway 1.0s≤THW<2.0s, the vehicle is in the medium-distance steady car-following driving mode, and if the headway THW is greater than or equal to 2.0s, the vehicle is in the distant steady car-following driving mode. The scene perception layer needs the relevant information of the target vehicle surrounding vehicles, weather and road information, etc. which has an impact on driving operation, and can rely on the perception data obtained from the car-road cooperation system; among them, the surrounding vehicles include the front, rear, side, diagonal front and diagonal rear vehicles of the target vehicle; during driving operation, the front vehicle has the greatest impact on the driver; the side and diagonal front and rear vehicles have a great impact on the vehicle in the lane changing mode; weather information includes sunny, overcast, rainy, foggy and snowy; road information includes the number of lanes, lane width, traffic signs and speed limit; (4) LDA topic model The LDA topic model is represented by a hierarchical Bayesian model with parameters, as follows: z m,n |θ m ~ Multinomial(θ m ) (2) θ m ~ Dirichlet (a) (3) where w m,n represents the word, subject to a multinomial distribution with parameters in equation (1); represents the word distribution of topic k, is an N m dimensional vector, and has k = 1, 2, 3, … K; for topic k ∈ [1, K], is obtained by a Dirichlet distribution with parameters β in equation (4); z m,n represents the topic of the nth word in the document m, subject to a multinomial distribution with parameters m in equation (2); θ m represents the topic distribution of the document m, is a K-dimensional vector, and has m = 1, …, M, and has n = 1, …, N m , θ m is obtained according to a Dirichlet distribution with parameters β in equation (3); α, β represent the hyperparameters of the document-topic distribution and the topic-word distribution, respectively; M represents the number of documents; N m represents the number of words in the document, and K represents the number of topics. The driver trajectory data file can be regarded as a document, the driving behavior semantic understanding vocabulary can be regarded as a word, and the driving style can be regarded as a topic; wherein ω m , m∈[1, M] represents the driving behavior semantic vocabulary of the driver m, , represents the driving behavior semantic vocabulary distribution of the topic k, θ m , m∈[1, M] represents the mixed driving style distribution of the driver m; the underlying structure of the driving style mined from the driving behavior semantic understanding vocabulary by using the LDA model is that the driver semantic understanding vocabulary is subject to the parameter The driving behavior semantic vocabulary distribution corresponds to K driving style topics, and the in-route driving style of the driver is defined as a probability combination of different driving style topics to describe the real-time driving state of the driver, that is, the in-route driving style of the driver. (5) Topic discovery The number of topics is determined by combining the topic perplexity and topic consistency indicators, and the log perplexity indicator is used to evaluate the driving style classification results under different topic numbers. The definition of log perplexity is shown in formula (5). The topic is easy to explain, so the topic words with high importance in the topic appear more frequently in the corresponding document, and the greater the topic consistency, the better the model effect; where p(ω m |M) is the topic confusion degree, ω m ,m∈[1,M] represents the driving behavior semantic vocabulary of driver m, represents the driving behavior semantic vocabulary distribution of topic k, θ m ,m∈[1,M] represents the mixed driving style distribution of driver m; (6) Topic analysis Using pyLDAvis visualization tool for topic analysis, analyzing the proportion of different topics and the correlation between topics, and showing the importance and rarity of different topic words to the topic; defining and analyzing the current driver state represented by each driving style; (7) In-route driving style recognition Using the in-route driving data of steps (1) and (2), the in-route driving style of the driver is determined as the probability combination of different driving style topics in step (5) to reflect the current driving state of the driver, and the sum of the probabilities of different driving style topics is 1. The in-route driving style can reflect the current driving state of the driver and the influence of the driving environment on the current driving style of the driver, and fully considers the heterogeneity and inconsistency of driving style.

Images