An unsupervised driving style analysis method based on basic driving operation events

Abstract

The present invention provides an unsupervised driving style analysis method based on basic driving operation events, including: obtaining data and preprocessing; extracting basic driving operation events; constructing and extracting features for each event in the basic driving operation events to obtain event intensity features; clustering event intensity by k-means and labeling event intensity categories; obtaining a dynamic time window, constructing an event change curve with event intensity category labels that characterizes the driving style in the dynamic time window; clustering the curves of each dynamic time window based on a curve clustering algorithm fused with DTW, obtaining various types of time window curves and labeling driving style types. The present invention takes basic driving operation events as basic units, considers event intensity and event transfer features, takes event change curves with time as features that describe driving style, reflects dynamic decision information, data continuity and time characteristics of driving behavior, retains original data information, and improves the accuracy of driving style analysis.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic analysis, and in particular to an unsupervised driving style analysis method based on basic driving operation events. Background Art

[0002] With the development of social science and technology, driving style analysis has become a research hotspot. Driving style plays an important role in road safety, vehicle economy, vehicle insurance and smart car design: detecting driving style and providing real-time feedback to drivers can effectively reduce the occurrence of road traffic accidents; driving style greatly affects the fuel economy of vehicles. The more aggressive the driving behavior, the lower the fuel economy; vehicle insurance costs also depend on the driving style of vehicle users. Setting different premiums for drivers with different driving styles can maximize the interests of insurance companies; with the development of artificial intelligence, smart vehicles have become a new trend in the development of vehicle-related industries. The decision-making process of smart vehicles depends to a large extent on the analysis of driving behavior. In order to ensure the decision-making safety of smart vehicles and improve user acceptance, effective analysis of driving style is a key technology. Therefore, accurate and efficient realization of driving style recognition is the focus of current attention.

[0003] In the context of big data, facing a massive amount of natural driving data, each hour of natural driving data requires about 800 hours of calibration, so it is very necessary to use unsupervised methods to cluster analysis of driving styles. Cluster analysis is an unsupervised machine learning algorithm used for data mining. Cluster analysis can find the inherent distribution structure of data, cluster data with similar characteristics into one category, and have obvious differences between different types of data, expand the similarity of similar data and the difference between different types of data, so as to improve the interpretability and facilitate subsequent research and analysis.

[0004] However, existing driving style analysis still has the following shortcomings: (1) There is a lack of driving style analysis methods that start from basic driving operation events and consider both event intensity and event transition characteristics; (2) Most existing technologies use discrete feature points to characterize driving style, which cannot preserve the continuity of natural driving data and miss the temporal characteristics; (3) Existing driving style clustering algorithms are all clustering algorithms for discrete points, and lack clustering algorithms for continuous curves of different lengths. It is impossible to achieve efficient and accurate clustering analysis of driving styles for data samples of different lengths based on continuous curve features. Summary of the invention

[0005] The present invention provides an unsupervised driving style analysis method based on basic driving operation events, so as to solve the problems existing in the existing driving style analysis method, namely, not starting from the basic driving operation events, not considering the event intensity and event transfer characteristics; using discrete feature points to characterize the driving style, failing to retain the continuity of natural driving data, and omitting the time characteristics; lacking a clustering algorithm for continuous curves of different lengths, failing to implement efficient and accurate clustering analysis of driving styles for data samples of different lengths based on the characteristics of continuous curves.

[0006] The present invention provides an unsupervised driving style analysis method based on basic driving operation events, comprising the following steps: S1, obtaining natural driving data, and preprocessing the natural driving data to obtain effective natural driving data; S2, extracting basic driving operation events based on the effective natural driving data; S3, based on the effective natural driving data, constructing and extracting features for each event in the basic driving operation events through driving behavior characteristics to obtain event intensity features characterizing the intensity of each event; S4, based on the event intensity features, performing event intensity analysis on each event in the basic driving operation events according to straight-ahead events and turning events through a k-means clustering algorithm. Clustering, and according to the statistical value of the event intensity feature, labeling the various events obtained by clustering with event intensity category labels; S5, obtaining the road alignment based on the effective natural driving data, and obtaining the dynamic time window according to the straight section and the curved section based on the change of the road alignment; based on the event intensity category and event transfer feature in the dynamic time window, constructing the event change curve with time with the event intensity category label that characterizes the driving style in the dynamic time window; based on the curve clustering algorithm fused with DTW, clustering the curves of each dynamic time window according to the straight event and the turning event to obtain various time window curves; based on the event intensity feature and the event transfer feature, labeling the driving style type for various time window curves.

[0007] In some embodiments of the present invention, the natural driving data includes speed, longitudinal acceleration and lateral acceleration, and the preprocessing includes abnormal data removal and data smoothing.

[0008] In some embodiments of the present invention, the basic driving operation events include six types of driving operation events, and the six types of driving operation events are straight-moving acceleration, straight-moving constant speed, straight-moving deceleration, turning acceleration, turning constant speed and turning deceleration.

[0009] In some embodiments of the present invention, the method for extracting the basic driving operation event in step S2 is as follows: S21, calculating the driving trajectory radius based on the speed and lateral acceleration: Where R is the radius of the driving track, v is the speed, a yis the lateral acceleration; S22, based on the longitudinal acceleration, constructing a longitudinal acceleration waveform; S23, extracting basic driving operation events based on the driving track radius, the peak value and / or the trough value of the longitudinal acceleration waveform.

[0010] In some embodiments of the present invention, basic driving operation events are extracted by the following conditions: the driving trajectory radius threshold is set to 1000m, when the driving trajectory radius R>=1000m, it is extracted as a straight event; when the driving trajectory radius R<1000m, it is extracted as a turning event; the longitudinal acceleration waveform peak value threshold is set to 1.0m / s 2 , the longitudinal acceleration waveform valley value threshold is -1.0m / s 2 , when the peak value of the longitudinal acceleration waveform>1.0m / s 2 When -1.0m / s 2 ≤Longitudinal acceleration waveform peak value or trough value ≤1.0m / s 2 When the longitudinal acceleration waveform trough value is less than -1.0m / s 2 , it is extracted as a deceleration event.

[0011] In some embodiments of the present invention, the method for obtaining the event intensity feature characterizing the intensity of each event in step S3 is as follows: S31, using effective natural driving data as variables, constructing driving behavior features based on each variable, wherein the driving behavior features include mean, median, trimmed mean, standard deviation, interquartile range, absolute median, percentile, minimum, maximum, Shannon entropy and approximate entropy; S32, calculating the weight of each driving behavior feature of each variable by entropy method; S33, based on the weight of each driving behavior feature of each variable, calculating the intensity feature of each variable by the following polynomial: Among them, var_idx i Represents the intensity characteristics of the var variable of the i-th event, var = [effective natural driving data] = [speed, longitudinal acceleration, lateral acceleration], w_s vark represents the k-th driving behavior feature weight of the var variable, f ik is the kth driving behavior feature of the var variable of the i-th event, and K is the number of driving behavior features of each variable; S34, based on the intensity characteristics of each variable, the variance contribution rate and cumulative principal component contribution rate of each principal component are calculated by principal component analysis, and the corresponding principal component with cumulative principal component contribution rate ≥ 90% is selected as the effective principal component; S35, the weight of each effective principal component is calculated by the following formula: Among them, σ m Indicates the variance contribution rate corresponding to the selected mth effective principal component, w_p mrepresents the weight of the mth effective principal component, M represents the number of effective principal components; S36, based on the effective principal component weight w_p m And the corresponding effective principal component load a calculated by principal component analysis mvar , the event intensity feature characterizing the intensity of each event is calculated by the following formula: Among them, event_idx i Represents the event intensity feature of the i-th event.

[0012] In some embodiments of the present invention, in step S4, based on the event intensity feature, each event in the basic driving operation events is clustered according to the event intensity of the straight-ahead event and the turning event through the k-means clustering algorithm, and the clustering result is determined based on the silhouette coefficient and the Davidson-Botting index; the statistical value of the event intensity feature includes the average value, maximum value, minimum value and standard deviation of the event intensity feature of each type in the straight-ahead event and the turning event obtained after clustering, and based on the statistical value of the event intensity feature, each type of event obtained by clustering is labeled with an event intensity category label.

[0013] In some embodiments of the present invention, in step S5, based on the curve clustering algorithm integrated with DTW, the curves of each dynamic time window are clustered according to the straight-through event and the turning event, and the method for obtaining various types of time window curves is as follows: 1) The curves of each dynamic time window are recorded as a curve set C; 2) The distance between the two curves in the curve set C is calculated based on the DTW algorithm, and any curve of the two curves with the largest distance between the two curves is selected, recorded as curve a, and the curve similarity distance threshold T is set; 3) Curve a is included in the curve set C1 , and record the curve set C = C-C1; 4), calculate the distance between each curve in the curve set C and the curve set C1, obtain the curve b in the curve set C corresponding to the minimum distance, and record the curve set C1' = C1+b; 5), calculate the intra-class similarity D(C1') of the curve set C1', if D(C1')>T, then go to step 2), otherwise, classify the curve b into the curve set C1, record the curve set C1 = Cl+b, the curve set C = C-C1, and the algorithm goes to step 4); 6), when the algorithm runs to the curve set C is an empty set, the algorithm terminates.

[0014] In some embodiments of the present invention, the curve similarity distance threshold T is the percentile of the distance set between each pair of curves in the curve set C calculated based on the DTW algorithm, the percentile of the dynamic time window of the straight section is the 87th percentile, and the percentile of the dynamic time window of the curved section is the 50th percentile.

[0015] In some embodiments of the present invention, in step S5, the method of labeling each type of time window curve with a driving style type based on the event intensity category and the event transition feature is as follows: Step 1: Calculate the event transition probability of the z-th type of time window curve by the following formula: The z-th time window curve represents a time window curve of a certain type among the various time window curves obtained by clustering the curves of each dynamic time window based on the curve clustering algorithm fused with DTW, θ t represents the current event, θ t+1 Indicates the event at the next moment, event o Indicates the event intensity category corresponding to the S4 step clustering at the current moment, event l represents the event intensity category corresponding to the clustering of step S4 at the next moment, n(θ t =event o ,θ t+1 =event l ) indicates that a certain type of time window is from event o Transfer to event l The number of this transfer form, n(θ t+1 =event l ) indicates that time t+1 is an event l Step 2: Calculate the Shannon entropy of the event transition probability of the z-th time window curve by the following formula: Among them, event r is the number of event intensity categories involved in the z-th time window curve; Step 3, calculate the transfer tendency of the z-th time window curve by the following formula: Among them, o is the current event intensity category, is the transfer form weight, is the span difference of event intensity categories before and after the transfer of each transfer form of the z-th time window curve, O=N and O and N both represent the number of event intensity categories clustered in step S4, event z. is the event intensity category after transfer, is the event intensity category before the transfer; Step 4, calculate the average event intensity of all events included in the z-th time window curve by the following formula, as the event intensity characteristic of the z-th time window curve: Among them, s is the number of curves in the z-th time window curve, is the event intensity category corresponding to the tth moment of the sth curve, and T is the duration of the sth curve; step five, based on the Shannon entropy of the event transfer probability of the zth time window curve, the transfer tendency of the zth time window curve and the numerical value of the event intensity characteristic of the zth time window curve, each type of time window curve is labeled with cautious, normal and aggressive driving style in order from small to large.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] (1) The natural driving data obtained by the present invention is continuous time series data, which describes the complex driving behavior of the driver when completing the driving task. Direct analysis is not conducive to understanding the differences in driving styles from a micro level. Therefore, the present invention extracts basic driving operation events based on effective natural driving data, and uses the basic driving operation events as the basic units to describe the complex driving behavior of the driver. It can reflect the characteristics of driving behavior and the decision-making information of driving behavior from a micro level, which is conducive to driving style analysis. In addition, in addition to using the basic driving operation events as the basic units, the analysis method of the present invention also considers the event intensity category and event transfer characteristics, and uses the event change curve over time as the feature to describe the driving style, which reflects the dynamic decision-making information of driving behavior and the continuity and time characteristics of natural driving data, better retains the original information of natural driving data, and improves the interpretability and accuracy of the driving style analysis results.

[0018] (2) The present invention constructs features based on effective natural driving data, which can comprehensively represent data information and improve the accuracy of subsequent algorithms. It reduces the number of features through feature extraction, obtains more representative features and reduces computational complexity.

[0019] (3) The driving style analysis method of the present invention adopts two clustering algorithms based on k-means and DTW, and integrates k-means and DTW into the analysis method of the present invention to realize the analysis of driving style under unsupervised conditions, breaking through the limitations of manual calibration.

[0020] (4) The present invention determines a dynamic time window according to the road line shape, ensures that a sufficient number of events are included in the time window, reflects the dynamic decision of driving behavior, realizes real-time online clustering analysis of driving style, can effectively identify the driver's constantly changing real-time driving style, and explains the composition of the driver's long-term driving style based on the real-time driving style; for curve clustering, the present invention proposes a large-scale curve clustering method that integrates DTW with dynamic time warp similarity measurement that considers the global situation. The large-scale curve clustering algorithm based on DTW can cluster curves of different lengths, avoiding the local optimization caused by the hierarchical clustering algorithm used in the prior art. The present invention realizes clustering analysis of driving style in dynamic time window based on continuous characteristic curves, can capture the driver's constantly changing driving behavior, realizes real-time online analysis of driving style, can effectively capture the constantly changing dynamic driving style, and enables the clustering results to better retain the continuity and time information of natural driving data, accurately and effectively identify the driving style, and improve the interpretability and application scope of driving style clustering analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required to be used in the embodiments of the present invention will be described below.

[0022] Figure 1 A basic driving operation event curve diagram in an unsupervised driving style analysis method based on basic driving operation events according to an embodiment of the present invention;

[0023] Figure 2 A distribution histogram of straight-ahead events and turning events in basic driving operation events according to an embodiment of the present invention;

[0024] Figure 3 A driving behavior feature weight map constructed from variables in a straight-moving event and a turning event in an embodiment of the present invention;

[0025] Figure 4 A distribution diagram of the intensity characteristics of each variable in a straight-ahead event and a turning event according to an embodiment of the present invention;

[0026] Figure 5 A box plot showing the intensity characteristics of each variable in a straight-ahead event and a turning event according to an embodiment of the present invention;

[0027] Figure 6 A distribution diagram of the intensity characteristics of each event in a straight-ahead event and a turning event according to an embodiment of the present invention;

[0028] Figure 7 This is a diagram showing changes in event intensity clustering indicators of straight-ahead events and turning events according to an embodiment of the present invention;

[0029] Figure 8It is a schematic diagram of event intensity clustering results of straight-ahead events and turning events according to an embodiment of the present invention;

[0030] Fig. 9 A dynamic time window length distribution histogram according to an embodiment of the present invention;

[0031] Fig.10 This is a schematic diagram of the division result of the time window straight-ahead events and the turning events according to an embodiment of the present invention;

[0032] Fig.11 A graph showing changes in event intensity categories of straight-ahead events and turning events over time according to an embodiment of the present invention;

[0033] Fig.12 This is a schematic diagram of clustering results of a time window curve including straight-moving events and turning events according to an embodiment of the present invention;

[0034] Fig.13 A time window driving style category curve diagram including straight-moving events and turning events according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to make the objectives, technical solutions and advantages of the present invention more clear, various aspects of the present invention will be described in detail below in conjunction with specific embodiments. However, these specific embodiments are only used to illustrate the present invention and do not constitute any limitation on the scope of protection and essential content of the present invention.

[0036] This embodiment provides an unsupervised driving style analysis method based on basic driving operation events. Figure 1 A basic driving operation event curve diagram in the unsupervised driving style analysis method based on basic driving operation events of this embodiment is shown; Figure 2 shows a distribution histogram of straight-ahead events and steering events in the basic driving operation events of this embodiment; Figure 3 The driving behavior characteristic weight diagram constructed by various variables in the straight-ahead event and the turning event of this embodiment is shown; Figure 4 The intensity characteristic distribution diagram of each variable in the straight-ahead event and the turning event of this embodiment is shown; Figure 5 The box plots of the intensity characteristics of each variable in the straight-ahead event and the turning event of this embodiment are shown; Figure 6 The intensity characteristic distribution diagram of each event in the straight-ahead event and the turning event of this embodiment is shown; Figure 7 The diagram showing the change of event intensity clustering indexes of the straight-ahead event and the turning event of this embodiment is shown; Figure 8 A schematic diagram showing the event intensity clustering results of the straight-ahead event and the turning event of this embodiment is shown; Fig. 9 The dynamic time window length distribution histogram of this embodiment is shown; Fig.10A schematic diagram showing the division result of the time window straight-ahead events and turning events of this embodiment is shown; Fig.11 A graph showing the change in event intensity categories of straight-ahead events and turning events over time in this embodiment is shown; Fig.12 A schematic diagram of clustering results of time window curves including straight-ahead events and turn events in this embodiment is shown; Fig.13 A time window driving style category curve diagram including straight-moving events and turning events in this embodiment is shown.

[0037] An unsupervised driving style analysis method based on basic driving operation events in this embodiment includes the following steps:

[0038] S1. Acquire natural driving data, and pre-process the natural driving data to obtain effective natural driving data;

[0039] S2. Extracting basic driving operation events based on effective natural driving data;

[0040] S3. Based on the effective natural driving data, feature construction and feature extraction are performed on each event in the basic driving operation events through driving behavior characteristics to obtain event intensity features that characterize the intensity of each event;

[0041] S4. Based on the event intensity feature, each event in the basic driving operation event is clustered according to the straight-ahead event and the turning event through the k-means clustering algorithm, and the event intensity category labels are given to each type of event obtained by clustering according to the statistical value of the event intensity feature;

[0042] S5. Obtain the road alignment based on effective natural driving data, and obtain the dynamic time window according to the straight section and the curved section based on the change of the road alignment; construct an event change curve with time that has an event intensity category label and represents the driving style in the dynamic time window based on the event intensity category and event transfer characteristics in the dynamic time window; cluster the curves of each dynamic time window according to the straight-ahead event and the turning event based on the curve clustering algorithm fused with DTW to obtain various time window curves; label various time window curves with driving style types based on the event intensity category and event transfer characteristics.

[0043] In this embodiment, the acquisition method of the natural driving data in step S1 is not restricted, and those skilled in the art can select the acquisition method according to actual conditions. In this embodiment, preferably, the natural driving data in step S1 is specifically collected using the RADS type 8-DOF panoramic driving simulation system developed by the Highway Science Research Institute (RIOS) of the Ministry of Transport. The driving simulation system performs traffic simulation, and the test driver drives freely and changes lanes autonomously according to the simulated traffic (pre-set road section route) according to personal driving habits. The driving simulation system collects the natural driving data of the test driver. In this embodiment, the natural driving data includes speed, longitudinal acceleration and lateral acceleration. During the driving simulation experiment, due to operational errors of the test driver or the equipment itself, the collected natural driving data has abnormalities or defects, and the natural driving data needs to be preprocessed. The preprocessing of this embodiment includes abnormal data removal and data smoothing. In this embodiment, the specific method of abnormal data removal and data smoothing is not limited, and those skilled in the art can make reasonable selections according to actual needs. For example, a sliding average method can be used for data smoothing. The method for removing abnormal data can be as follows: remove abnormal values ​​according to the normal data threshold in the art, and the longitudinal acceleration and lateral acceleration greater than 10m / s 2 and less than -10m / s 2 The data point with a speed greater than 36 m / s is replaced by the average of the first 3 data points and the last 3 data points. The data point with a speed greater than 36 m / s is replaced by the average of the first 3 data points and the last 3 data points.

[0044] In this embodiment, the basic driving operation events extracted in step S2 include six types of driving operation events, such as Figure 1-2 As shown, the six types of driving operation events are straight-ahead acceleration, straight-ahead constant speed, straight-ahead deceleration, turning acceleration, turning constant speed, and turning deceleration.

[0045] In this embodiment, the method for extracting basic driving operation events in step S2 is specifically as follows:

[0046] S21. Calculate the radius of the driving trajectory based on the speed and lateral acceleration: Where R is the radius of the driving track, v is the speed, a y is the lateral acceleration;

[0047] S22. Based on the longitudinal acceleration, construct a longitudinal acceleration waveform (in this embodiment, the complete waveform of the longitudinal acceleration is a process in which the longitudinal acceleration increases (decreases) from 0 to a certain value and then decreases (increases) to 0);

[0048] S23: extracting basic driving operation events based on the driving track radius, the peak value and / or the trough value of the longitudinal acceleration waveform.

[0049] In this embodiment, step S23 extracts basic driving operation events through the following conditions: setting the driving trajectory radius threshold to 1000m, when the driving trajectory radius R>=1000m, extracting it as a straight event; when the driving trajectory radius R<1000m, extracting it as a turning event; setting the longitudinal acceleration waveform peak value threshold to 1.0m / s 2 , the longitudinal acceleration waveform valley value threshold is -1.0m / s 2 , when the peak value of the longitudinal acceleration waveform>1.0m / s 2 When -1.0m / s 2 ≤Longitudinal acceleration waveform peak value or trough value ≤1.0m / s 2 When the longitudinal acceleration waveform trough value is <-1.0m / s 2 It should be noted that in the actual driving process, since the driving behavior is directly controlled by the driver, whether it is accelerating driving, constant speed driving or decelerating driving, it will be accompanied by changes in acceleration to a certain extent. For example, acceleration behavior is generally not accelerated at a constant acceleration, deceleration behavior is generally not decelerated at a constant acceleration, and constant speed behavior generally cannot guarantee that the acceleration is always zero (it will have more or less slight acceleration changes, but under the human perception of the passengers or external observation, the acceleration is small and can be ignored, so it presents a constant speed behavior). In order to more accurately judge the basic driving operation events and prevent the accompanying The uniform speed behavior with small acceleration changes is misjudged as acceleration behavior or deceleration behavior. This embodiment constructs a longitudinal acceleration waveform based on the longitudinal acceleration in the effective natural driving data collected and preprocessed, and judges acceleration, uniform speed and deceleration events based on the preset longitudinal acceleration waveform peak value threshold, the preset longitudinal acceleration waveform trough value threshold and the peak value and trough value of the constructed longitudinal acceleration waveform. Taking into account the influence of acceleration changes brought about by acceleration, uniform speed or deceleration driving when the driver manually controls the driving behavior, it can more accurately judge acceleration, uniform speed and deceleration events to prevent event misjudgment. This embodiment comprehensively considers the driving trajectory radius threshold, the longitudinal acceleration waveform peak value threshold and the longitudinal acceleration waveform trough value threshold, and can automatically extract six types of driving operation events, namely, straight acceleration, straight uniform speed, straight deceleration, steering acceleration, steering uniform speed and steering deceleration.

[0050] In this embodiment, the method for obtaining the event intensity feature characterizing the intensity of each event in step S3 is specifically as follows:

[0051] S31. Using effective natural driving data as variable parameters, construct the driving behavior characteristics of Table 1 below based on each variable to represent the operating characteristics of the driving behavior on each variable; wherein the driving behavior characteristics include mean, median, trimmed mean, standard deviation, interquartile range, absolute median, percentile, minimum, maximum, Shannon entropy and approximate entropy;

[0052] Table 1 Variable parameters and constructed driving behavior characteristics

[0053]

[0054] S32. Calculate the weight of each driving behavior feature of each variable by entropy method (in this embodiment, entropy method is a commonly used method for calculating the weight of underlying indicators. It measures the discreteness of variable data according to the entropy value. The smaller the entropy value of the variable, the greater the discreteness of the variable, and the greater the weight of the variable. The entropy method is a well-known method in the art and will not be described in detail here), such as Figure 3 As shown, the driving behavior feature weights constructed by various variables in the straight-ahead event and the turning event obtained in this embodiment are shown;

[0055] S33, based on the weight of each driving behavior feature of each variable, the intensity feature of each variable is calculated by the following polynomial (such as Figure 4 FIG. 4 shows the intensity characteristic distribution diagram of each variable in the straight-ahead event and the turning event obtained in this embodiment; Figure 5 The box plot of the intensity characteristics of each variable in the straight-ahead event and the turning event obtained in this embodiment is shown:

[0056]

[0057] Among them, var_idx i Represents the intensity characteristics of the var variable of the i-th event, var = [effective natural driving data] = [speed, longitudinal acceleration, lateral acceleration], w_s vark represents the k-th driving behavior feature weight of the var variable, f ik is the kth driving behavior feature of the ith event var variable, and K is the number of driving behavior features of each variable;

[0058] S34. Based on the strength characteristics of each variable, the variance contribution rate and cumulative principal component contribution rate of each principal component are calculated by principal component analysis, and the corresponding principal component with a cumulative principal component contribution rate ≥ 90% is selected as the effective principal component (in this embodiment, the principal component analysis method is a top-level indicator weight calculation method. The principal component analysis method is a well-known method in the art and will not be described in detail here);

[0059] S35. Calculate the weight of each effective principal component using the following formula:

[0060]

[0061] Among them, σ m Indicates the variance contribution rate corresponding to the selected mth effective principal component, w_p m represents the weight of the mth effective principal component, and M represents the number of effective principal components;

[0062] S36, based on the effective principal component weight w_p m And the corresponding effective principal component load a calculated by principal component analysis mvar , the event intensity feature characterizing the intensity of each event is calculated by the following formula (such as Figure 6 The intensity characteristic distribution diagram of each event in the straight-ahead event and the turning event obtained in this embodiment is shown:

[0063]

[0064] Among them, event_idx i Represents the event intensity feature of the i-th event.

[0065] In this embodiment, in step S4, based on the event intensity feature, the k-means clustering algorithm is used to cluster the event intensity of each event in the basic driving operation event according to the straight-ahead event and the turning event, and the events with similar intensity are clustered into one category to obtain the event intensity category (in this embodiment, the K-means algorithm is a classic clustering algorithm with iterative solution, which will not be described in detail here). The number of clusters is selected in sequence as 1 to 10, and the clustering results are analyzed based on the silhouette coefficient SC and the Davidson-Boding index DBI to determine the optimal number of clusters (in this embodiment, the silhouette coefficient SC and the Davidson-Boding index DBI are commonly used internal indicators for clustering performance measurement, and the clustering results are evaluated by calculating the intra-class compactness and inter-class dispersion. The silhouette coefficient SC threshold is [-1,1]. The closer the value is to 1, the better the clustering effect, and the smaller the value of the Davidson-Boding index DBI, the better the clustering effect.). Figure 7 The event intensity clustering index change diagram of straight-through events and turning events is shown. It can be seen that for straight-through events, the clustering effect is best when the number of clusters is 3; while for turning events, the clustering effect is best when the number of clusters is 2. Figure 7 The optimal number of clusters is Figure 8A schematic diagram of the event intensity clustering results of straight-ahead events and turning events is shown. In the present embodiment, the statistical values ​​of the event intensity features include the average value, maximum value, minimum value and standard deviation of the intensity features of each type of events in the straight-ahead events and turning events obtained after clustering. Based on the statistical values ​​of the event intensity features, event intensity category labels are assigned to each type of events obtained by clustering. Specifically, the change trends of the average value, maximum value and minimum value of the intensity features of each type of events in the straight-ahead events and turning events obtained after clustering are consistent, and the standard deviations are not much different. The larger the average value, maximum value and minimum value of the intensity features of each type of events in the straight-ahead events and turning events obtained after clustering, the stronger the event intensity is, and the labels are used to characterize high intensity when labeling; the smaller the average value, maximum value and minimum value of the intensity features of each type of events in the straight-ahead events and turning events obtained after clustering, the lower the event intensity is, and the labels are used to characterize low intensity when labeling. The standard deviation of each category is small, indicating that the distribution of the intensity features of each category of events is relatively concentrated. In the present embodiment, preferably, each category label of the straight-ahead events and turning events can be mapped to a category value with physical meaning. The higher the intensity, the larger the mapping value (for example, Figure 8 The numerical values ​​mapped in: the event intensity categories of straight events include category 1, category 2, and category 3; the event intensity categories of turning events include category 1 and category 2).

[0066] In this embodiment, step S5 obtains the road alignment based on effective natural driving data, and obtains the dynamic time window according to the straight road segment and the curved road segment based on the change of the road alignment. Specifically, the road alignment is determined according to the driving trajectory radius R, and the road alignment is used as the basis for determining the length of the dynamic time window. When the driving trajectory radius R>1000m and the duration t>4s, it is a dynamic time window containing straight-line events, and when the driving trajectory radius R<1000m and the duration t>4s, it is a dynamic time window containing turning events. Fig. 9 The dynamic time window length distribution histogram is shown as Fig.10 The schematic diagram of the division result of the straight-line event and the turning event in the time window is shown. In this embodiment, the straight road section is the road section including the straight-line event (straight-line uniform speed, straight-line deceleration, straight-line acceleration), and the curved road section is the road section including the turning event (turning uniform speed, turning deceleration, turning acceleration).

[0067] In this embodiment, step S5 constructs a time-varying curve of events with event intensity category labels that characterize the driving style in the dynamic time window based on the event intensity category and event transfer features in the dynamic time window, such as Fig.11As shown, a graph showing the change of event intensity categories of straight-ahead events and turning events over time is shown, in which the horizontal axis of the curve is the time axis and the vertical axis is the event intensity category value corresponding to the time point. In this embodiment, the event transfer feature represents the feature of the current event intensity category at the current moment being transferred to the next event intensity category at the next moment.

[0068] In this embodiment, in step S5, based on the curve clustering algorithm integrated with DTW, the curves of each dynamic time window are clustered according to the straight-through events and the turning events to obtain various time window curves (such as Fig.12 The schematic diagram of the clustering result of the time window curve containing the straight-through events and the turning events is shown. Fig.12 In , categories A, B, and C represent the types of various time window curves obtained by clustering the curves of each dynamic time window based on the curve clustering algorithm fused with DTW, which reflects the driving style corresponding to each time window obtained after clustering, indicating that the clustering results are 3 categories, namely, category A, category B, and category C, that is, there are 3 types of driving styles, which can ensure the interpretability and acceptability of the clustering results. ) The specific method is as follows:

[0069] 1) Record the curves of each dynamic time window as curve set C;

[0070] 2) Calculate the distance between two curves in the curve set C based on the DTW algorithm, select any one of the two curves with the largest distance between the two curves, record it as curve a, and set the curve similarity distance threshold T;

[0071] 3) Classify curve a into curve set C1, and record curve set C = C-C1;

[0072] 4) Calculate the distance between each curve in the curve set C and the curve set C1, obtain the curve b in the curve set C corresponding to the minimum distance, and record the curve set C1'=C1+b;

[0073] 5) Calculate the intra-class similarity D(C1') of the curve set C1'. If D(C1')>T, go to step 2), otherwise classify curve b into the curve set C1, and record curve set C1=C1+b and curve set C=C-C1 respectively. The algorithm goes to step 4);

[0074] 6) When the algorithm runs to the point where the curve set C is an empty set, the algorithm terminates.

[0075] In this embodiment, the calculation method of step 4) in step S5 is as follows:

[0076] Definition 1: Curve L p (any curve in the curve set C) and the curve set C1(L 1 ,L 2 ,...,Lm ) is defined as the distance of the curve L p The average distance between all curves in the curve set C1, that is,

[0077]

[0078] Where m represents the number of curves in the curve set C1, L z represents the zth curve in the curve set C1, L p represents any curve in the curve set C.

[0079] In this embodiment, the calculation method of step 5) in step S5 is as follows:

[0080] Definition 2: Curve set C1'(L 1 ,L 2 ,...,L n ) is defined as the maximum distance between any two curves in the curve set C1', that is,

[0081]

[0082] Where n represents the number of curves in the curve set C1', L i represents the i-th curve in the curve set C1', L j represents the jth curve in the curve set C1′.

[0083] In this embodiment, in step S5, based on the curve clustering algorithm fused with DTW, the curves of each dynamic time window are clustered according to the straight-through events and the turning events, and the method for obtaining the curves of various time windows is stepwise performed according to the above steps 1)-6). It should be noted that the term appearing in the next step is the term processed in the previous step. If the term does not appear in the previous step, the term is the term processed in the previous step, and so on, until the term appears most recently (for example, the term "curve set C": step 5) and meets D(C1')>T, then the transition is performed. To step 2), at this time, the curve set C appears in step 2), and the previous step of step 2) is step 5) that meets D(C1')>T, in which the curve set C does not appear. Then, the previous step of step 5), step 4), appears at this time, and step 4) does not define the curve set C. Then, the previous step of step 4), step 3), processes the curve set C. It can be seen that when step 5) meets D(C1')>T, the curve set C used when turning to step 2) is the curve set C processed in step 3). In this embodiment, the meaning of curve set C=C-C1 in step 3) is: the curve set C processed in step 3) represents the curve set obtained by removing the curve set C1 from the curve set C in the previous step, i.e., step 2); the meanings of curve set C1'=C1+b in step 4), curve set C1=C1+b in step 5), and curve set C=C-C1 can be understood in a similar corresponding manner.

[0084] In this embodiment, the curve similarity distance threshold T set in step 2) of step S5 is the percentile of the distance set between each pair of curves in the curve set C calculated based on the DTW algorithm, the percentile of the dynamic time window of the straight section is the 87 percentile, and the percentile of the dynamic time window of the curved section is the 50 percentile, that is, the curve similarity distance threshold T of the dynamic time window of the straight section is the 87 percentile, and the curve similarity distance threshold T of the dynamic time window of the curved section is the 50 percentile.

[0085] In this embodiment, in step S5, the method of labeling each type of time window curve with a driving style type based on the event intensity category and event transition characteristics is as follows:

[0086] Step 1: Calculate the event transfer probability of the z-th time window curve using the following formula:

[0087]

[0088] The z-th time window curve represents a time window curve of a certain type among the various time window curves obtained by clustering the curves of each dynamic time window based on the curve clustering algorithm fused with DTW, θ t represents the current event, θ t+1Indicates the event at the next moment, event o Indicates the event intensity category corresponding to the S4 step clustering at the current moment, event l represents the event intensity category corresponding to the clustering of step S4 at the next moment, n(θ t =event o ,θ t+1 =event l ) indicates that a certain type of time window is from event o Transfer to event l The number of this transfer form, n(θ t+1 =event l ) indicates that time t+1 is an event l the number of

[0089] Step 2: Calculate the Shannon entropy of the event transition probability of the z-th time window curve using the following formula:

[0090]

[0091] Among them, event r is the number of event intensity categories involved in the z-th time window curve;

[0092] Step 3: Calculate the transfer tendency of the z-th time window curve using the following formula:

[0093]

[0094] Among them, o is the current event intensity category, is the transfer form weight, is the difference in event intensity category span before and after the transfer of each transfer form of the z-th time window curve (i.e., the difference in category values ​​with physical meaning mapped by the event intensity category label before and after the transfer), O = N (O = N means that O and N have the same meaning), and O and N both represent the number of event intensity categories clustered in step S4, event z. is the event intensity category after transfer, is the category of intensity of the event before the transfer;

[0095] Step 4: Calculate the average event intensity of all events included in the z-th time window curve using the following formula as the event intensity characteristic of the z-th time window curve:

[0096]

[0097] Among them, s is the number of curves in the z-th time window curve, is the event intensity category corresponding to the tth moment of the sth curve, and T is the duration of the sth curve;

[0098] Step 5. Based on the numerical values ​​of the Shannon entropy of the event transfer probability of the z-th time window curve, the transfer tendency of the z-th time window curve, and the event intensity characteristic of the z-th time window curve, label each type of time window curve with cautious, normal, and aggressive driving style labels in order from small to large. In this embodiment, the three indicators of the Shannon entropy of the event transfer probability, the transfer tendency, and the event intensity characteristic are all positive indicators. The smaller the values ​​of the three indicators, the more cautious the driving style is, and the cautious label is assigned; the larger the values ​​of the three indicators, the more aggressive the driving style is, and the aggressive label is assigned; if the values ​​of the three indicators are in the middle, it means that the driving style is more normal, and the normal label is assigned. Fig.13 A driving style category curve diagram of a time window including straight-moving events and turning events is shown, from which it can be seen that: there is no obvious difference in the lateral acceleration of the time window including the straight-moving events; the speeds of the cautious and normal types are relatively low compared to the aggressive type; the longitudinal acceleration of the cautious type has no obvious change and basically remains near zero, so the cautious type maintains a constant speed at a relatively low speed; relative to the cautious type, the normal type has obvious acceleration and deceleration behaviors within the example time range, and the operations are more diverse and frequent; for the aggressive type, its speed is relatively large, and there is an obvious change in longitudinal acceleration, so the aggressive type is more inclined to high-intensity and high-difficulty operations. For the time window including the turning event, when the lateral acceleration is roughly the same, it can be seen that the cautious driver has a lower turning speed. In the early stage of the turning process, he enters the curve at a lower speed and decelerates slightly. As the turning process progresses, he continues to accelerate until the turn is completed. This is consistent with the common sense that cautious drivers first enter the curve at a slow speed when turning to reduce driving risks and slowly accelerate as the turning process progresses until the turn is completed. The driving risk stimulation is low and the comfort is strong. The aggressive driver has a higher speed when turning and has obvious acceleration and deceleration behavior, which shows that the aggressive driving behavior is still Maintaining strong operating characteristics, constantly accelerating and decelerating and maintaining a high speed during the turn, this shows that this type of driver pursues driving excitement and is more inclined to constantly adjust his driving strategy as the surrounding environment changes. Therefore, this type of driving behavior is less comfortable, and because of the constant change in acceleration, it also creates a strong sense of excitement, which is consistent with people's common sense of aggressive driving style; while the normal type passes the curve at a constant speed, the operation is neither too cautious nor too frequent like the aggressive type, and passes the curve in a relatively gentle and efficient driving behavior operation mode, which is in line with people's common sense. Fig.13 As can be seen from the above analysis, the results of the driving style analysis method of this embodiment are consistent with the actual characteristics of cautious, normal and aggressive driving, which shows that the driving style analysis method of this embodiment is effective and accurate.

[0099] The present invention has been described above in conjunction with specific embodiments. These specific embodiments are merely exemplary and cannot be used to limit the scope of protection of the present invention. Those skilled in the art may make various modifications, changes or substitutions without departing from the essence of the present invention. Therefore, various equivalent changes made according to the present invention still fall within the scope covered by the present invention.

Claims

1. An unsupervised driving style analysis method based on basic driving operation events, characterized in that: The following steps are involved: S1. Acquire natural driving data, and pre-process the natural driving data to obtain effective natural driving data; S2. extracting basic driving operation events based on the effective natural driving data, wherein the basic driving operation events include six types of driving operation events, namely, straight driving acceleration, straight driving at a constant speed, straight driving deceleration, turning acceleration, turning at a constant speed, and turning deceleration; S3, based on the effective natural driving data, performing feature construction and feature extraction on each event in the basic driving operation events through driving behavior characteristics to obtain event intensity features characterizing the intensity of each event; The method of determining the event intensity feature of the event intensity is specifically as follows: Use effective natural driving data as variables and construct driving behavior characteristics based on each variable; The weight of each driving behavior feature of each variable is calculated by entropy method; Based on the intensity characteristics of each variable, the variance contribution rate and cumulative principal component contribution rate of each principal component were calculated by principal component analysis, and the corresponding principal component with cumulative principal component contribution rate ≥ 90% was selected as the effective principal component; The weight of each effective principal component is calculated by the following formula: in, It represents the variance contribution rate corresponding to the selected mth effective principal component. represents the weight of the mth effective principal component, and M represents the number of effective principal components; Based on effective principal component weights And the corresponding effective principal component loadings calculated by principal component analysis , the event intensity feature characterizing the intensity of each event is calculated by the following formula: in, represents the event intensity characteristics of the i-th event; S4. Based on the event intensity feature, each event in the basic driving operation event is clustered according to the straight-ahead event and the turning event by using the k-means clustering algorithm, and event intensity category labels are given to each type of event obtained by clustering according to the statistical value of the event intensity feature; S5. Acquire the road alignment based on the effective natural driving data, and acquire the dynamic time window according to the straight section and the curved section based on the change of the road alignment; construct a curve of the event change over time with the event intensity category label that characterizes the driving style in the dynamic time window based on the event intensity category and the event transfer feature in the dynamic time window; cluster the curves of each dynamic time window according to the straight-ahead event and the turning event based on the curve clustering algorithm fused with DTW to obtain various types of time window curves; and label various types of time window curves with driving style types based on the event intensity category and the event transfer feature.

2. The method according to claim 1, characterized in that The natural driving data includes speed, longitudinal acceleration and lateral acceleration, and the preprocessing includes abnormal data elimination and data smoothing processing.

3. The method according to claim 2, characterized in that The method for extracting basic driving operation events in step S2 is as follows: S21. Calculate the radius of the driving trajectory based on the speed and lateral acceleration: , where R is the radius of the driving track, v is the speed, and a y is the lateral acceleration; S22, constructing a longitudinal acceleration waveform based on the longitudinal acceleration; S23: extracting basic driving operation events based on the driving track radius, the peak value and / or the trough value of the longitudinal acceleration waveform.

4. The method according to claim 3, characterized in that Extract basic driving operation events through the following conditions: The driving trajectory radius threshold is set to 1000m. When the driving trajectory radius R>=1000m, it is extracted as a straight-moving event; when the driving trajectory radius R<1000m, it is extracted as a turning event. Set the longitudinal acceleration waveform peak threshold to 1.0m / s 2 , the longitudinal acceleration waveform valley value threshold is -1.0m / s 2 , when the peak value of the longitudinal acceleration waveform>1.0m / s 2 When -1.0m / s 2 ≤Longitudinal acceleration waveform peak value or trough value ≤1.0m / s 2 When the longitudinal acceleration waveform trough value is <-1.0m / s 2 , it is extracted as a deceleration event.

5. The method according to claim 1, characterized in that The method for obtaining the event intensity feature characterizing the intensity of each event in step S3 is specifically as follows: In S3, a driving behavior feature is constructed based on each variable, wherein the driving behavior feature includes a mean, a median, a trimmed mean, a standard deviation, an interquartile range, an absolute median, a percentile, a minimum value, a maximum value, a Shannon entropy, and an approximate entropy; Based on the weight of each driving behavior feature of each variable, the intensity feature of each variable is calculated by the following polynomial: in, Represents the intensity characteristics of the var variable of the i-th event, var = [effective natural driving data] = [speed, longitudinal acceleration, lateral acceleration], represents the k-th driving behavior feature weight of the var variable, is the kth driving behavior feature of the ith event var variable, and K is the number of driving behavior features of each variable.

6. The method according to claim 1, characterized in that In step S4, based on the event intensity feature, each event in the basic driving operation event is clustered according to the straight-ahead event and the turning event by using a k-means clustering algorithm, and the clustering result is determined based on the silhouette coefficient and the Davidson-Botting index; The statistical values ​​of the event intensity features include the average value, maximum value, minimum value and standard deviation of each type of event intensity features in the straight-ahead events and turning events obtained after clustering. Based on the statistical values ​​of the event intensity features, event intensity category labels are assigned to each type of event obtained by clustering.

7. The method according to claim 1, characterized in that In step S5, based on the curve clustering algorithm integrated with DTW, the curves of each dynamic time window are clustered according to the straight-through events and the turning events, and the method for obtaining the curves of each time window is as follows: 1) Record the curves of each dynamic time window as a curve set C; 2) Calculate the distance between two curves in the curve set C based on the DTW algorithm, select any one of the two curves with the largest distance between the two curves, record it as curve a, and set the curve similarity distance threshold T; 3) Classify curve a into curve set C1, and record curve set C=C-C1; 4) Calculate the distance between each curve in the curve set C and the curve set C1, obtain the curve b in the curve set C corresponding to the minimum distance, and record the curve set C1'=C1+b; 5) Calculate the intra-class similarity D(C1') of the curve set C1'. If D(C1')>T, go to step 2), otherwise classify curve b into the curve set C1, and record curve set C1=C1+b and curve set C=C-C1 respectively. The algorithm goes to step 4); 6) When the algorithm runs to the point where the curve set C is an empty set, the algorithm terminates.

8. The method according to claim 7, characterized in that The curve similarity distance threshold T is the percentile of the distance set between each pair of curves in the curve set C calculated based on the DTW algorithm, the percentile of the dynamic time window of the straight section is the 87th percentile, and the percentile of the dynamic time window of the curved section is the 50th percentile.

9. The method according to claim 8, characterized in that In step S5, based on the event intensity category and event transition characteristics, the method of labeling each type of time window curve with a driving style type is as follows: Step 1: Calculate the event transfer probability of the z-th time window curve using the following formula: The z-th time window curve represents a certain type of time window curve among various types of time window curves obtained by clustering the curves of each dynamic time window based on the curve clustering algorithm fused with DTW. Indicates the current event. Indicates the event at the next moment. Indicates the event intensity category corresponding to the clustering in step S4 at the current moment, Indicates the event intensity category corresponding to the clustering in step S4 at the next moment, Indicates that within a certain time window Transfer to The number of this transfer form, Indicates that time t+1 is the number of Step 2: Calculate the Shannon entropy of the event transition probability of the z-th time window curve using the following formula: in, is the number of event intensity categories involved in the z-th time window curve; Step 3: Calculate the transfer tendency of the z-th time window curve using the following formula: Among them, o is the current event intensity category, is the transfer form weight, is the span difference of event intensity categories before and after the transfer of each transfer form of the z-th time window curve, O=N, and O and N both represent the number of event intensity categories clustered in step S4, is the event intensity category after transfer, is the event intensity category before the transfer; Step 4: Calculate the average event intensity of all events included in the z-th time window curve using the following formula as the event intensity characteristic of the z-th time window curve: Among them, s is the number of curves in the z-th time window curve, is the event intensity category corresponding to the tth moment of the sth curve, and T is the duration of the sth curve; Step 5. Based on the Shannon entropy of the event transfer probability of the z-th time window curve, the transfer tendency of the z-th time window curve, and the numerical value of the event intensity characteristic of the z-th time window curve, each type of time window curve is labeled with cautious, normal, and aggressive driving styles in order from small to large.

Images