Driving behavior primitive clustering method considering data change trend

Abstract

The invention relates to a driving behavior primitive clustering method considering a data change trend, and the method comprises the following steps: 1, collecting natural driving data, and selecting a driving behavior analysis variable, 2, performing driving behavior primitive extraction on the # imgabs1 # by using a BMASS method to obtain a driving behavior primitive set P = {Pj; j = 1: N; 3, performing preliminary discretization on the driving behavior primitive data; 4, judging the data change trend of the driving behavior primitive according to the derivative and data fluctuation condition of the t point and the neighborhood of the t point in the Pj; 5, integrating the data distribution item and the data change trend item to obtain a discretization result of the driving behavior primitives, and clustering the driving behavior primitives based on an LDA topic model; according to the method, the clustering accuracy of the driving behavior primitives can be improved, accurate expression of semantic information of the driving behavior primitives is realized, and a driver is helped to accurately and intuitively understand the driving behavior of the driver.

Description

Technical Field

[0001] The present invention relates to the technical field of driving behavior primitive clustering, and in particular to a driving behavior primitive clustering method taking into account data change trends. Background Art

[0002] Driving behavior primitives are the basic components of driving behavior and are the smallest driving behavior data segments with physical meaning. Although multiple independent driving behavior primitives can be obtained in the primitive extraction stage, the types of driving behavior primitives are limited. In order to intuitively understand driving behavior primitives, it is necessary to cluster driving behavior primitives.

[0003] The LDA topic model is often used for text clustering. To use the LDA model for primitive clustering, the continuous driving behavior data must first be discretized. At the same time, because the LDA topic model uses a bag-of-words model and does not consider the order in which words appear, when using this model to cluster driving behavior primitives, its temporal characteristics will be ignored, resulting in unreasonable results. This makes it impossible to accurately cluster driving behavior primitives, which leads to incorrect judgments on the driver's behavior status.

[0004] Therefore, the present invention provides a driving behavior primitive clustering method taking into account the changing trend of driving behavior data. After obtaining the data distribution item of driving behavior data based on the semantic space, a data changing trend item is added to make up for the deficiency of the LDA model caused by the bag-of-words model in clustering time series data, so as to achieve accurate clustering of driving behavior primitives and thus accurately identify the driving status of the driver. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a driving behavior primitive clustering method considering data change trends, the method comprising the following steps: step 1: collecting natural driving data of M drivers, selecting longitudinal driving behavior parameter speed v, longitudinal acceleration ax, lateral driving behavior parameter lateral acceleration ay, yaw angle r as driving behavior analysis variables, using a sliding average to filter the driving behavior analysis variable data to obtain pre-processed driving behavior data; derivatizing the pre-processed driving behavior data to obtain driving behavior derivative data;

[0006] The driving behavior data set of M drivers is obtained by preprocessing the driving behavior data and driving behavior derivative data. Said is the driving behavior data of the i-th driver, NM iis the data length of the i-th driver, the pre-processed driving behavior data includes speed v, longitudinal acceleration ax, lateral acceleration ay, and yaw angle r, and the driving behavior derivative data includes speed v derivative dv, longitudinal acceleration ax derivative dax, lateral acceleration ay derivative day, and yaw angle r derivative dr;

[0007] Step 2: Based on the pre-processed driving behavior data, use the BMASS method to collect the driving behavior data Extract driving behavior primitives and obtain a driving behavior primitive set P = {P j ;} j＝1:N ; N is the total number of primitives; is the jth driving behavior primitive, NN j is the data length of the jth driving behavior primitive;

[0008] Step 3: Construct the semantic space, preliminarily discretize the driving behavior primitive data, and obtain the data distribution item output code of the driving behavior primitive;

[0009] Step 4: Based on driving behavior primitive P j The data change trend of the driving behavior primitive is judged by the derivative and data fluctuation of the middle t point and its neighborhood, and the output code of the data change trend item of the driving behavior primitive is obtained;

[0010] The data at point t includes four driving behavior variables: speed v, longitudinal acceleration ax, lateral acceleration ay, and yaw angle r;

[0011] Step 5: The data distribution item output code and the data change trend item output code of the driving behavior primitive are integrated to obtain the discretization result of the driving behavior primitive, wherein the discretization result includes the distribution item output code and the change trend item output code, and the driving behavior primitive is clustered based on the LDA topic model.

[0012] Compared with the prior art, the present invention has the following beneficial effects: by filtering the driving data collected by the sensor through the sliding average, the sensor noise can be effectively eliminated and the data reliability can be improved; the driving behavior data is divided based on the data structure using the BMASS method, thereby obtaining the driving behavior primitive set P = {P j ;} j＝1:N, which can effectively identify the state transition points of the driver's driving process and facilitate the analysis of the driver's complex driving behavior. When discretizing the driving behavior primitives, the data percentile characteristics are integrated based on the empirical information to encode the data distribution items of speed and lateral acceleration, avoiding the problem that the distribution of the collected driving data is quite different from the standard data set due to the driver's driving habits. The steering behavior encoding of the lateral acceleration and yaw angle based on the actual road linear label is conducive to the accurate identification of the driver's lateral and longitudinal driving behavior. The existence of the data change trend item can realize the capture of the time series of the driving behavior primitives, improve the clustering effect, and avoid the occurrence of primitives with similar data distribution. The clustering errors caused by the data point and its neighborhood data fluctuations and their derivative information are combined to obtain the change trend of the data point, which can effectively avoid the misjudgment caused by too frequent data fluctuations and realize the accurate judgment of the change trend of driving behavior primitive data. The discretization results of driving behavior primitives are obtained through the data distribution items and data change trend items of four variables, and the perplexity and consistency scores are used to determine the most reasonable number of driving behavior primitive categories, which can make up for the shortcomings of the LDA model caused by the bag of words model in clustering time series data, improve the clustering accuracy of driving behavior primitives, realize the accurate expression of the semantic information of driving behavior primitives, and help drivers accurately and intuitively understand their driving behavior.

[0013] Furthermore, the data distribution item output code includes a speed distribution item output code, a longitudinal acceleration distribution item output code, and a steering behavior output code;

[0014] The speed distribution item output code includes numbers a, b, and c, where a is low speed, b is medium speed, and c is high speed;

[0015] The longitudinal acceleration distribution item output code includes numbers d, e, f, g, and h, wherein d represents a sudden deceleration, e represents a deceleration, f represents a uniform speed, g represents an acceleration, and h represents a sudden acceleration;

[0016] The steering behavior output code is obtained by integrating the lateral acceleration ay output code and the yaw angle r output code, the lateral acceleration ay output code includes numbers s and t, the yaw angle r output code includes numbers s and t, the s is straight driving, and the t is turning. When the lateral acceleration ay output code and the yaw angle r output code are both turning, the steering behavior output code outputs turning, otherwise it is recognized as straight driving.

[0017] The beneficial effect of adopting the previous step is that the output coding of the data distribution item can effectively define the driver's current driving behavior, which is convenient for understanding the driving behavior. Combining the actual road linear label to determine the turning behavior coding can realize the accurate division of the driver's lateral and longitudinal driving behavior.

[0018] Further, the data change trend item output code includes a speed change trend item output code, a longitudinal acceleration change trend item output code, a lateral acceleration change trend item output code, and a yaw angle change trend item output code;

[0019] The speed change trend item output code includes numbers: vu, vd, vl;

[0020] The longitudinal acceleration change trend item output code includes numbers axu, axd, axl;

[0021] The output codes of the lateral acceleration change trend items include numbers ayu, ayd, and ayl;

[0022] The output code of the yaw angle change trend item includes numbers ru, rd, rl;

[0023] The vu is the output code of the speed increasing trend item; the vd is the output code of the speed decreasing trend item; the vl is the output code of the speed steady trend item;

[0024] The axu is the output code of the longitudinal acceleration rising trend item; the axd is the output code of the longitudinal acceleration falling trend item; the axl is the output code of the longitudinal acceleration steady trend item;

[0025] The ayu is the output code of the lateral acceleration rising trend item; the ayd is the output code of the lateral acceleration falling trend item; the ayl is the output code of the lateral acceleration steady trend item;

[0026] The ru is the output code for the yaw angle rising trend item; the rd is the output code for the yaw angle falling trend item; and the rl is the output code for the yaw angle steady trend item.

[0027] The beneficial effect of adopting the previous step is that the changing trends of speed, longitudinal acceleration, lateral acceleration and yaw angle are characterized by output encoding of data change trend items, which can capture the temporal characteristics of driving behavior primitives, improve the shortcomings of LDA topic model in time series clustering, and thus improve the clustering accuracy of driving behavior primitives, realize the precise expression of semantic information of driving behavior primitives, and help drivers accurately and intuitively understand their driving behavior.

[0028] Furthermore, the preliminary discretization of the driving behavior primitive data includes preliminary discretization of the speed v;

[0029] The preliminary discretization of the speed v includes selecting the 20% quantile and the 90% quantile as the discretization thresholds of low speed, medium speed and high speed according to the empirical division information of low speed, medium speed and high speed. The empirical division information generally considers 0-40km / h as low speed, 40km / h-70km / h as medium speed, and above 70km / h as high speed. The 20% quantile is 11m / s, and the 90% quantile is 20m / s.

[0030] The beneficial effect of adopting the previous step is that by combining the speed experience division information, the speed is divided into low speed, medium speed and high speed, and 11m / s of the 20% quantile is selected as the discretization threshold of low speed and medium speed, and 20m / s of the 90% quantile is selected as the discretization threshold of medium speed and high speed. Based on one of the encodings of the output speed distribution item, the speed data is discretized from the data distribution perspective, and the problem of large differences between the collected speed data distribution and the standard data set due to the driving habits of the driver is avoided, making the judgment of driving behavior more accurate.

[0031] Furthermore, the preliminary discretization of the driving behavior primitive data includes preliminary discretization of the longitudinal acceleration ax;

[0032] The preliminary discretization of the longitudinal acceleration ax includes selecting 30%, 45%, 60%, and 70% quantiles as discretization thresholds of sudden deceleration, deceleration, uniform speed, acceleration, and sudden acceleration according to the empirical division information of sudden deceleration, deceleration, uniform speed, acceleration, and sudden acceleration. The empirical division information generally considers that when the absolute value of the acceleration is less than 0.05m / s 2 When the driver cannot sense the acceleration, the excessive acceleration is greater than 0.2m / s 2 It will cause physical and mental discomfort to the driver. The 30%, 45%, 60%, and 70% quantiles of the longitudinal acceleration are -0.18 m / s respectively. 2 、-0.05m / s 2 , 0.07m / s 2 、0.18m / s 2 .

[0033] The beneficial effect of the previous step is that by combining the longitudinal acceleration experience division information, the 30% fraction of -0.18m / s is selected. 2 As the discretization threshold for rapid deceleration and deceleration, -0.05m / s at the 45% fraction is selected. 2 As the discretization threshold for deceleration and uniform speed, 0.07m / s at the 60% fraction is selected. 2 As the discretization threshold for uniform speed and acceleration, the 70% quantile of 0.18 m / s is selected. 2The discrete thresholds for acceleration and sudden acceleration are output based on one of the codes, which realizes the discretization of longitudinal acceleration data from the perspective of data distribution. At the same time, it avoids the problem that the distribution of collected longitudinal acceleration data is significantly different from the standard data set due to the driver's driving habits, making the judgment of driving behavior more accurate.

[0034] Furthermore, the preliminary discretization of the driving behavior primitive data includes preliminary discretization of the lateral acceleration ay and the yaw angle r; the preliminary discretization of the lateral acceleration ay and the yaw angle r includes maximizing the road alignment recognition accuracy based on the actual road alignment label to determine the straight-through turn discretization threshold, and the actual road alignment label shows that the straight section accounts for 62.80% and the turning section accounts for 37.20%; the maximum alignment recognition accuracy is 83.26%, and the corresponding lateral acceleration straight-through turn discrimination threshold is 0.35m / s 2 , the recognition accuracy of straight sections is 90.10%, the recognition accuracy of turning sections is 71.72%, and the -0.35m / s 2 ≤ay≤0.35m / s 2 For straight travel, ay<-0.35m / s 2 ,ay>0.35m / s 2 For turning;

[0035] The maximum linear recognition accuracy is 83.33%, and the corresponding yaw angle fluctuation straight-line turning judgment threshold is 0.04 rad. The straight-line section recognition accuracy is 88.80%, and the turning section recognition accuracy is 74.08%. The yaw angle data change range within one second is ≤0.04 rad for straight-line driving, and the yaw angle data change range within one second is >0.04 rad for turning.

[0036] The beneficial effect of adopting the previous step is that the road line shape discrimination threshold is 0.35m / s with a maximum line shape recognition accuracy of 83.26% through lateral acceleration. 2 , as the judgment standard for straight driving and turning, the straight driving and turning outputs of the lateral acceleration are encoded as s and t; the road line shape discrimination threshold of 0.04rad with a maximum line shape recognition accuracy of 83.33% is used through the yaw angle r changing range per second. As the judgment standard for straight driving and turning, the straight driving and turning outputs of the yaw angle are encoded as s and t; when the line shape encodings of both the lateral acceleration and the yaw angle are t, it will be identified as turning, otherwise it will be identified as straight driving. The encoding of the lateral acceleration and yaw angle for steering behavior based on the actual road line shape label realizes the accurate division of the driver's lateral and longitudinal driving behaviors.

[0037] Further, the data change trend judgment of the driving behavior includes the data change trend judgment of the X variable at the t point; the data change trend judgment process of the X variable at the t point is as follows: the X variable is any one of the four variables of speed v, longitudinal acceleration ax, lateral acceleration ay and yaw angle r;

[0038] The method of judging the data change trend at point t based on the derivative of point t and its neighborhood and the data fluctuation includes:

[0039] Step 1: Get data point x t Variable data SW_X in the neighborhood t and variable derivative data SW_dX t , the neighborhood range is [max(t-30,1), min(t+30,NN j )]; the NN j The length of the current driving behavior primitive data;

[0040] Step 2: If max(SW_X t )-min(SW_X t )>α·(max(X)-min(X)), α is the threshold setting ratio, with a value of 0.02. Based on the data point x t-1 、x t 、x t+1 The sign of the derivative at x t The trend of data changes;

[0041] Step 3: If max(SW_X t )-min(SW_X t )≤α·(max(X)-min(X)), then based on SW_dX t The symbol distribution of x t The data change trend.

[0042] The beneficial effect of the previous step is that by setting the neighborhood radius to 30, the neighborhood range [max(t-30,1), min(t+30,NN j )], get the variable data SW_X within the neighborhood t , variable derivative data SW_dX t The process of judging the data change trend at point t and obtaining the rising, falling or stable trend takes into account the time characteristics of the driving behavior primitive sequence and avoids misjudgment due to local noise;

[0043] When a data point corresponds to variable data SW_X t When the range exceeds the set data fluctuation threshold range, based on x t itself and x t The data x at the previous time pointt-1 and the data at the next time point x t+1 The derivative of x t The trend of data changes;

[0044] When a data point corresponds to variable data SW_X t When the range is within the set data fluctuation threshold range or is equal to the threshold range, based on SW_dX t The symbol distribution of x t The data change trend.

[0045] Furthermore, the max(SW_X t )-min(SW_X t )>α·(max(X)-min(X)), if |dx t |>α·min(|max(dX)|,|min(dX)|), then based on dx t The symbol output data point x t The changing trend (upward, downward), when dx t When it is greater than the derivative discrimination threshold, the data point x t The trend of change is upward, otherwise the data point x t The changing trend is downward;

[0046] If |dx t |≤α·min(|max(dX)|, |min(dX)|), then based on dx t-1 The symbol output data point x t The changing trend (increasing, decreasing, stable), when dx t-1 When the absolute value of is less than or equal to the derivative discrimination threshold, it is stable. t-1 If it is greater than the derivative judgment threshold, it is rising, otherwise it is falling. In particular, when t = 1, based on dx t+1 The symbol output data point x t The changing trend (increasing, decreasing, stable), when dx t+1 When the absolute value of is less than or equal to the derivative discrimination threshold, it is stable. t+1 If it is greater than the derivative discrimination threshold, it is rising, otherwise it is falling.

[0047] The beneficial effect of adopting the previous step is: by max(SW_X t )-min(SW_X t )>α·(max(X)-min(X)), dx t The absolute value of dx is outside the threshold range, so based on t The sign of dx determines the changing trend of the data at point t. tThe absolute value of x is within the threshold range, so based on t-1 or x t+1 The relationship between the derivative at point t and the derivative discrimination threshold is used to judge the changing trend of the data at point t. By setting the derivative discrimination threshold, the misjudgment of the data changing trend due to data noise can be effectively avoided.

[0048] Furthermore, the max(SW_X t )-min(SW_X t )≤α·(max(X)-min(X)), if SW_dX t The number of data points between [-α·min(|max(dX)|, |min(dX)|), α·min(|max(dX)|, |min(dX)|)] is denoted as l, the number of data points greater than α·min(|max(dX)|, |min(dX)|) is denoted as m, and the number of data points less than -α·min(|max(dX)|, |min(dX)|) is denoted as n;

[0049] if Then the data point x t The trend of change is upward. If Then the data point x t The trend of change is downward, otherwise the data point x t The changing trend is stable.

[0050] The beneficial effect of the previous step is that if more than three-quarters of the derivative values ​​of the data points corresponding to a certain data point are greater than the derivative discrimination threshold, then the data point x t The trend of change is rising. If more than three quarters of the data points have derivative values ​​less than the negative derivative discrimination threshold, then the data point x t The trend of the change is downward, otherwise the data x t The changing trend of is stable. The changing trend of data points is judged based on the distribution of derivative data, which effectively avoids the misjudgment caused by data noise.

[0051] Furthermore, the discretization results are data distribution items and data change trend items of speed, longitudinal acceleration, lateral acceleration and yaw angle;

[0052] The discretization result is: one of the output codes of the velocity distribution item, one of the output codes of the velocity change trend item, one of the output codes of the longitudinal acceleration distribution item, one of the output codes of the longitudinal acceleration change trend item, one of the output codes of the steering behavior item, one of the output codes of the lateral acceleration change trend item, and one of the output codes of the yaw angle change trend item.

[0053] The beneficial effect of adopting the previous step is that the discretization results of driving behavior primitives are obtained through the speed, longitudinal acceleration, lateral acceleration and yaw angle data distribution items and data change trend items, which can achieve the matching of time series data and LDA topic model, and at the same time make up for the shortcomings of LDA model caused by bag of words model in clustering time series data, improve the clustering accuracy of driving behavior primitives, and achieve accurate expression of semantic information of driving behavior primitives, helping drivers to accurately and intuitively understand their driving behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Generate a data distribution item encoding table for discretization of driving behavior primitives for discretization thresholds of speed, longitudinal acceleration, lateral acceleration and yaw angle;

[0055] Figure 2 Generate a flow chart for the data change trend item of the present invention;

[0056] Figure 3 The LDA clustering results for the items with or without data change trend - entropy value and topic difference;

[0057] Figure 4 The LDA clustering results for the items with and without data change trend are shown below:

[0058] Figure 5 Visualize the LDA clustering results with and without data change trend items. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] Embodiment 1:

[0061] The present embodiment provides a driving behavior primitive clustering method considering data change trends, the method comprising the following steps: Step 1: collecting natural driving data of M drivers, selecting longitudinal driving behavior parameter speed v, longitudinal acceleration ax, lateral driving behavior parameter lateral acceleration ay, yaw angle r as driving behavior analysis variables, using a sliding average to filter the driving behavior analysis variable data to obtain pre-processed driving behavior data; derivating the pre-processed driving behavior data to obtain driving behavior derivative data; obtaining a driving behavior data set of M drivers through the pre-processed driving behavior data and the driving behavior derivative data Said is the driving behavior data of the i-th driver, NM i is the data length of the i-th driver; the pre-processed driving behavior data includes speed v, longitudinal acceleration ax, lateral acceleration ay, and yaw angle r, and the driving behavior derivative data includes speed v derivative dv, longitudinal acceleration ax derivative dax, lateral acceleration ay derivative day, and yaw angle r derivative dr;

[0062] Step 2: Based on the pre-processed driving behavior data, use the BMASS method to collect the driving behavior data Extract driving behavior primitives and obtain a driving behavior primitive set P = {P j ;} j＝1:N ; N is the total number of primitives; is the jth driving behavior primitive, NN j is the data length of the jth driving behavior primitive;

[0063] Step 3: Construct the semantic space, preliminarily discretize the driving behavior primitive data, and obtain the data distribution item output code of the driving behavior primitive;

[0064] The data distribution item output code includes a speed distribution item output code, a longitudinal acceleration distribution item output code, and a steering behavior output code; the speed distribution item output code includes numbers a, b, and c, where a is low speed, b is medium speed, and c is high speed; the longitudinal acceleration distribution item output code includes numbers d, e, f, g, and h, where d is a sudden deceleration, e is a deceleration, f is a uniform speed, g is an acceleration, and h is a sudden acceleration; the steering behavior output code is obtained by integrating the lateral acceleration ay output code and the yaw angle r output code, the lateral acceleration ay output code includes numbers s and t, the yaw angle r output code includes numbers s and t, s is straight driving, and t is turning, when the lateral acceleration ay output code and the yaw angle r output code are both turning, the steering behavior output code outputs turning, otherwise it is identified as straight driving; the data distribution item output code can effectively define the current driving behavior of the driver, which is convenient for understanding the driving behavior, and the output of the steering behavior code based on the actual road linear label can realize the accurate division of the driver's lateral and longitudinal driving behaviors;

[0065] The preliminary discretization of the driving behavior primitive data includes preliminary discretization of the speed v; the preliminary discretization of the speed v includes selecting the 20% quantile and the 90% quantile as the discretization thresholds of low speed, medium speed and high speed according to the experience division information of low speed, medium speed and high speed, the experience division information generally considers 0-40km / h as low speed, 40km / h-70km / h as medium speed, and 70km / h or more as high speed, the 20% quantile is 11m / s, and the 90% quantile is 20m / s. By combining the speed experience division information, the speed is divided into low speed, medium speed and high speed, 11m / s of the 20% quantile is selected as the discretization threshold of low speed and medium speed, and 20m / s of the 90% quantile is selected as the discretization threshold of medium speed and high speed, based on one of the encodings of the output speed distribution item, the speed data is discretized from the data distribution perspective, and the problem that the distribution of the collected speed data is greatly different from the standard data set due to the driving habits of the driver is avoided, so that the judgment of the driving behavior is more accurate;

[0066] The preliminary discretization of the driving behavior primitive data includes preliminary discretization of the longitudinal acceleration ax; the preliminary discretization of the longitudinal acceleration ax includes selecting 30%, 45%, 60%, and 70% quantiles as discretization thresholds of sudden deceleration, deceleration, uniform speed, acceleration, and sudden acceleration according to the empirical division information of sudden deceleration, deceleration, uniform speed, acceleration, and sudden acceleration, and the empirical division information generally considers that when the absolute value of the acceleration is less than 0.05m / s 2 When the driver cannot sense the acceleration, the excessive acceleration is greater than 0.2m / s 2 It will cause physical and mental discomfort to the driver. The 30%, 45%, 60%, and 70% quantiles of the longitudinal acceleration are -0.18 m / s respectively. 2 、-0.05m / s 2 , 0.07m / s 2 、0.18m / s 2 , by combining the longitudinal acceleration experience division information, the 30% fraction of -0.18m / s is selected 2 As the discretization threshold for rapid deceleration and deceleration, -0.05m / s at the 45% fraction is selected. 2 As the discretization threshold for deceleration and uniform speed, 0.07m / s at the 60% fraction is selected. 2 As the discretization threshold for uniform speed and acceleration, the 70% quantile of 0.18 m / s is selected. 2 The thresholds for acceleration and sudden acceleration are discretized. Based on this, one of the codes is output, which realizes the discretization of longitudinal acceleration data from the perspective of data distribution. At the same time, it avoids the problem that the distribution of collected longitudinal acceleration data is greatly different from the standard data set due to the driving habits of the driver, making the judgment of driving behavior more accurate.

[0067] The preliminary discretization of the driving behavior primitive data includes preliminary discretization of the lateral acceleration ay and the yaw angle r; the preliminary discretization of the lateral acceleration ay and the yaw angle r includes maximizing the road alignment recognition accuracy based on the actual road alignment label to determine the straight-through turn discretization threshold, the actual road alignment label shows that the straight section accounts for 62.80% and the turning section accounts for 37.20%; the maximum alignment recognition accuracy is 83.26%, and the corresponding lateral acceleration straight-through turn discrimination threshold is 0.35m / s 2 , the recognition accuracy of straight sections is 90.10%, the recognition accuracy of turning sections is 71.72%, and the -0.35m / s 2 ≤ay≤0.35m / s 2 For straight travel, ay<-0.35m / s 2 ,ay>0.35m / s 2 The maximum linear recognition accuracy is 83.33%, and the corresponding yaw angle fluctuation straight-line turning discrimination threshold is 0.04 rad. The straight section recognition accuracy is 88.80%, and the turning section recognition accuracy is 74.08%. The yaw angle data change range per second is ≤0.04 rad for straight driving, and the yaw angle data change range per second is >0.04 rad for turning. When both the lateral acceleration ay and the yaw angle r are within the turning discrimination threshold, it will be identified as turning, otherwise it will be identified as straight driving. The road linear discrimination threshold with a maximum linear recognition accuracy of 83.26% through lateral acceleration is 0.35 m / s 2 , as the judgment standard for straight driving and turning, the straight driving and turning outputs of the lateral acceleration are encoded as s and t; the road line shape discrimination threshold of 0.04rad with a maximum line shape recognition accuracy of 83.33% is used through the yaw angle r changing range per second. As the judgment standard for straight driving and turning, the straight driving and turning outputs of the yaw angle are encoded as s and t; when the line shape encodings of the two parameters of the lateral acceleration and yaw angle are both t, it will be identified as turning, otherwise it will be identified as straight driving. The encoding of the lateral acceleration and yaw angle for steering behavior based on the actual road line shape label realizes the accurate division of the driver's lateral and longitudinal driving behaviors.

[0068] Step 4: Based on driving behavior primitive P j The data change trend of the driving behavior primitive is judged by the derivative and data fluctuation of the middle t point and its neighborhood, and the output code of the data change trend item of the driving behavior primitive is obtained;

[0069] The data at point t includes four driving behavior variables: speed v, longitudinal acceleration ax, lateral acceleration ay, and yaw angle r;

[0070] The data change trend item output code includes a speed change trend item output code, a longitudinal acceleration change trend item output code, a lateral acceleration change trend item output code, and a yaw angle change trend item output code;

[0071] The speed change trend item output code includes numbers: vu, vd, vl;

[0072] The longitudinal acceleration change trend item output code includes numbers axu, axd, axl;

[0073] The output codes of the lateral acceleration change trend items include numbers ayu, ayd, and ayl;

[0074] The output code of the yaw angle change trend item includes numbers ru, rd, rl;

[0075] The vu is the output code of the speed increasing trend item; the vd is the output code of the speed decreasing trend item; the vl is the output code of the speed steady trend item;

[0076] The axu is the output code of the longitudinal acceleration rising trend item; the axd is the output code of the longitudinal acceleration falling trend item; the axl is the output code of the longitudinal acceleration steady trend item;

[0077] The ayu is the output code of the lateral acceleration rising trend item; the ayd is the output code of the lateral acceleration falling trend item; the ayl is the output code of the lateral acceleration steady trend item;

[0078] The ru is the output code of the yaw angle rising trend item; rd is the output code of the yaw angle falling trend item; rl is the output code of the yaw angle stable trend item;

[0079] The data change trend item output encoding is used to characterize the change trend of speed, longitudinal acceleration, lateral acceleration and yaw angle, which can capture the temporal characteristics of driving behavior primitives, improve the shortcomings of LDA topic model in time series clustering, and thus improve the clustering accuracy of driving behavior primitives, realize the precise expression of driving behavior primitive semantic information, and help drivers accurately and intuitively understand their driving behavior;

[0080] Step 5: The data distribution item output code and the data change trend item output code of the driving behavior primitive are integrated to obtain the discretization result of the driving behavior primitive, wherein the discretization result includes the distribution item output code and the change trend item output code, and the driving behavior primitive is clustered based on the LDA topic model.

[0081] The sliding average is used to filter the driving data collected by the sensor, which can effectively eliminate sensor noise and improve data reliability. The BMASS method is used to divide the driving behavior data based on the data structure, thereby obtaining the driving behavior primitive set P = {P j ;} j＝1:N , which can effectively identify the state transition points of the driver's driving process and facilitate the analysis of the driver's complex driving behavior. When discretizing the driving behavior primitives, the data percentile characteristics are integrated based on the empirical information to encode the data distribution items of speed and lateral acceleration, avoiding the problem that the distribution of the collected driving data is quite different from the standard data set due to the driver's driving habits. The steering behavior encoding of the lateral acceleration and yaw angle based on the actual road linear label is conducive to the accurate identification of the driver's lateral and longitudinal driving behavior. The existence of the data change trend item can realize the capture of the time series of the driving behavior primitives, improve the clustering effect, and avoid the occurrence of primitives with similar data distribution. The clustering errors caused by the data point and its neighborhood data fluctuations and their derivative information are combined to obtain the change trend of the data point, which can effectively avoid the misjudgment caused by too frequent data fluctuations and realize the accurate judgment of the change trend of driving behavior primitive data. The discretization results of driving behavior primitives are obtained through the data distribution items and data change trend items of four variables, and the perplexity and consistency scores are used to determine the most reasonable number of driving behavior primitive categories, which can make up for the shortcomings of the LDA model caused by the bag of words model in clustering time series data, improve the clustering accuracy of driving behavior primitives, realize the accurate expression of the semantic information of driving behavior primitives, and help drivers accurately and intuitively understand their driving behavior.

[0082] Further, the data change trend judgment of the driving behavior includes the data change trend judgment of the X variable at the t point; the data change trend judgment process of the X variable at the t point is as follows: the X variable is any one of the four variables of speed v, longitudinal acceleration ax, lateral acceleration ay and yaw angle r; the data change trend judgment at the t point based on the derivative of the t point and its neighborhood and the data fluctuation includes: step 1: obtaining the data point x t Variable data SW_X in the neighborhood t and variable derivative data SW_dX t , the neighborhood range is [max(t-30,1), min(t+30,NN j )]; the NN j The length of the current driving behavior primitive data;

[0083] Step 2: If max(SW_X t )-min(SW_X t )>α·(max(X)-min(X)), α is the threshold setting ratio, with a value of 0.02. Based on the data point xt-1 、x t 、x t+1 The sign of the derivative at x t Data change trend; Step 3: If max(SW_X t )-min(SW_X t )≤α·(max(X)-min(X)), then based on SW_dX t The symbol distribution of x t By setting the neighborhood radius to 30, the neighborhood range [max(t-30,1), min(t+30,NN j )], get the variable data SW_X within the neighborhood t , variable derivative data SW_dX t The process of judging the data change trend at point t and obtaining the rising, falling or stable trend takes into account the time characteristics of the driving behavior primitive sequence and avoids misjudgment due to local noise; when a data point corresponds to the variable data SW_X t When the range exceeds the set data fluctuation threshold range, based on x t itself and x t The data x at the previous time point t-1 and the data at the next time point x t+1 The derivative of x t The data change trend; when a data point corresponds to the variable data SW_X t When the range is within the set data fluctuation threshold range or is equal to the threshold range, based on SW_dX t The symbol distribution of x t The data change trend.

[0084] The max(SW_X t )-min(SW_X t )>α·(max(X)-min(X)), if |dx t |>α·min(|max(dX)|,|min(dX)|), then based on dx t The symbol output data point x t The changing trend (upward, downward), when dx t When it is greater than the derivative discrimination threshold, the data point x t The trend of change is upward, otherwise the data point x t The trend of change is downward; if |dx t |≤α·min(|max(dX)|, |min(dX)|), then based on dx t-1 The symbol output data point x t The changing trend (increasing, decreasing, stable), when dxt-1 When the absolute value of is less than or equal to the derivative discrimination threshold, it is stable. t-1 If it is greater than the derivative judgment threshold, it is rising, otherwise it is falling. In particular, when t = 1, based on dx t+1 The symbol output data point x t The changing trend (increasing, decreasing, stable), when dx t+1 When the absolute value of is less than or equal to the derivative discrimination threshold, it is stable. t+1 If it is greater than the derivative judgment threshold, it is rising, otherwise it is falling. t )-min(SW_X t )>α·(max(X)-min(X)), dx t The absolute value of dx is outside the threshold range, so based on t The sign of dx determines the changing trend of the data at point t. t The absolute value of x is within the threshold range, so based on t-1 or x t+1 The relationship between the derivative at point t and the derivative discrimination threshold is used to judge the changing trend of the data at point t. By setting the derivative discrimination threshold, the misjudgment of the data changing trend due to data noise can be effectively avoided.

[0085] The max(SW_X t )-min(SW_X t )≤α·(max(X)-min(X)), if SW_dX t The number of data points between [-α·min(|max(dX)|, |min(dX)|), α·min(|max(dX)|, |min(dX)|)] is denoted as l, the number of data points greater than α·min(|max(dX)|, |min(dX)|) is denoted as m, and the number of data points less than -α·min(|max(dX)|, |min(dX)|) is denoted as n; if Then the data point x t The trend of change is upward. If Then the data point x t The trend of change is downward, otherwise the data point x t The trend of change is stable. If more than three-quarters of the derivative values ​​of the data points corresponding to a data point are greater than the derivative discrimination threshold, then the data point x t The trend of change is rising. If more than three quarters of the data points have derivative values ​​less than the negative derivative discrimination threshold, then the data point x t The trend of the change is downward, otherwise the data x tThe changing trend of is stable. The changing trend of data points is judged based on the distribution of derivative data, which effectively avoids the misjudgment caused by data noise.

[0086] The discretization results are data distribution items and data change trend items of speed, longitudinal acceleration, lateral acceleration and yaw angle; the discretization results are: one of the output codes of speed distribution item, one of the output codes of speed change trend item, one of the output codes of longitudinal acceleration distribution item, one of the output codes of longitudinal acceleration change trend item, one of the output codes of steering behavior, one of the output codes of lateral acceleration change trend item, and one of the output codes of yaw angle change trend item. The discretization results of driving behavior primitives are obtained through the data distribution items and data change trend items of speed, longitudinal acceleration, lateral acceleration and yaw angle, which can achieve the matching of time series data and LDA topic model, and at the same time make up for the deficiency of LDA model in clustering time series data caused by bag of words model, improve the clustering accuracy of driving behavior primitives, achieve the precise expression of semantic information of driving behavior primitives, and help drivers to accurately and intuitively understand their driving behavior.

[0087] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field should understand that they can still modify the technical solutions recorded in the above embodiments, or replace some of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A driving behavior primitive clustering method considering data change trends, characterized in that: The method comprises the following steps: Step 1: Collect natural driving data of M drivers, select longitudinal driving behavior parameters velocity v, longitudinal acceleration ax, lateral driving behavior parameters lateral acceleration ay, yaw angle r as driving behavior analysis variables, use sliding average to filter the driving behavior analysis variable data to obtain pre-processed driving behavior data; derive the pre-processed driving behavior data to obtain driving behavior derivative data; The driving behavior data set of M drivers is obtained by preprocessing the driving behavior data and driving behavior derivative data. Said is the driving behavior data of the i-th driver, NM i is the data length of the i-th driver, the pre-processed driving behavior data includes speed v, longitudinal acceleration ax, lateral acceleration ay, and yaw angle r, and the driving behavior derivative data includes speed v derivative dv, longitudinal acceleration ax derivative dax, lateral acceleration ay derivative day, and yaw angle r derivative dr; Step 2: Based on the pre-processed driving behavior data, use the BMASS method to collect the driving behavior data Extract driving behavior primitives and obtain a driving behavior primitive set P = {P j ;} j＝1:N ; N is the total number of primitives; is the jth driving behavior primitive, NN j is the data length of the jth driving behavior primitive; Step 3: Construct the semantic space, preliminarily discretize the driving behavior primitive data, and obtain the data distribution item output code of the driving behavior primitive; Step 4: Based on driving behavior primitive P j The data change trend of the driving behavior primitive is judged by the derivative and data fluctuation of the middle t point and its neighborhood, and the output code of the data change trend item of the driving behavior primitive is obtained; The data at point t includes four driving behavior variables: speed v, longitudinal acceleration ax, lateral acceleration ay, and yaw angle r; Step 5: The data distribution item output code and the data change trend item output code of the driving behavior primitive are integrated to obtain the discretization result of the driving behavior primitive, wherein the discretization result includes the distribution item output code and the change trend item output code, and the driving behavior primitive is clustered based on the LDA topic model.

2. The driving behavior primitive clustering method considering data change trend according to claim 1, characterized in that: The data distribution item output code includes a speed distribution item output code, a longitudinal acceleration distribution item output code, and a steering behavior output code; The speed distribution item output code includes numbers a, b, and c, where a is low speed, b is medium speed, and c is high speed; The longitudinal acceleration distribution item output code includes numbers d, e, f, g, and h, wherein d represents a sudden deceleration, e represents a deceleration, f represents a uniform speed, g represents an acceleration, and h represents a sudden acceleration; The steering behavior output code is obtained by integrating the lateral acceleration ay output code and the yaw angle r output code. The lateral acceleration ay output code includes numbers s and t, and the yaw angle r output code includes numbers s and t. The s represents straight driving, and the t represents turning. When the lateral acceleration ay output code and the yaw angle r output code are both turning, the steering behavior output code outputs turning, otherwise it is recognized as straight driving.

3. The driving behavior primitive clustering method considering data change trend according to claim 1, characterized in that: The data change trend item output code includes a speed change trend item output code, a longitudinal acceleration change trend item output code, a lateral acceleration change trend item output code, and a yaw angle change trend item output code; The speed change trend item output code includes numbers: vu, vd, vl; The longitudinal acceleration change trend item output code includes numbers axu, axd, axl; The output codes of the lateral acceleration change trend items include numbers ayu, ayd, and ayl; The output code of the yaw angle change trend item includes numbers ru, rd, rl; The vu is the output code of the speed rising trend item; The vd is the output code of the speed decreasing trend item; the vl is the output code of the speed steady trend item; The axu is the output code of the longitudinal acceleration rising trend item; The axd is the output code of the longitudinal acceleration downward trend item; Ax1 is the output code of the longitudinal acceleration steady trend item; The ayu is the output code of the lateral acceleration rising trend item; The ayd is the output code of the lateral acceleration downward trend item; The ayl is the output code of the lateral acceleration steady trend item; The ru is the output code of the yaw angle rising trend item; rd is the output code of the downward trend term of the yaw angle; rl is the output code of the yaw angle steady trend term.

4. The driving behavior primitive clustering method considering data change trend according to claim 1, characterized in that: The preliminary discretization of the driving behavior primitive data includes preliminary discretization of the speed v; The preliminary discretization of the speed v includes selecting the 20% quantile and the 90% quantile as the discretization thresholds of low speed, medium speed and high speed according to the empirical division information of low speed, medium speed and high speed. The empirical division information generally considers 0-40km / h as low speed, 40km / h-70km / h as medium speed, and above 70km / h as high speed. The 20% quantile is 11m / s, and the 90% quantile is 20m / s.

5. The driving behavior primitive clustering method considering data change trend according to claim 1, characterized in that: The preliminary discretization of the driving behavior primitive data includes preliminary discretization of the longitudinal acceleration ax; The preliminary discretization of the longitudinal acceleration ax includes selecting 30%, 45%, 60%, and 70% quantiles as the discretization thresholds of sudden deceleration, deceleration, uniform speed, acceleration, and sudden acceleration according to the empirical division information of sudden deceleration, deceleration, uniform speed, acceleration, and sudden acceleration, and the empirical division information is -0.2m / s 2 、-0.5m / s 2 , 0.05m / s 2 , 0.2m / s 2 The longitudinal acceleration quantiles of 30%, 45%, 60%, and 70% are -0.18 m / s respectively. 2 、-0.05m / s 2 , 0.07m / s 2 、0.18m / s 2 .

6. The driving behavior primitive clustering method considering data change trend according to claim 1, characterized in that: The preliminary discretization of the driving behavior primitive data includes preliminary discretization of the lateral acceleration ay and the yaw angle r; the preliminary discretization of the lateral acceleration ay and the yaw angle r includes maximizing the road line recognition accuracy based on the actual road line label to determine the straight turn discretization threshold, and the lateral acceleration straight turn discrimination threshold is 0.35m / s 2 , -0.35m / s 2 ≤ay≤0.35m / s 2 For straight travel, ay<-0.35m / s 2 ,ay>0.35m / s 2 For turning; The straight-moving and turning discrimination threshold of the yaw angle fluctuation is 0.04 rad. If the yaw angle data change range within one second is ≤0.04 rad, it is straight-moving; if the yaw angle data change range within one second is >0.04 rad, it is turning.

7. The driving behavior primitive clustering method considering data change trend according to claim 1, characterized in that: The data change trend judgment of the driving behavior primitive includes the data change trend judgment of the X variable at the t point; the data change trend judgment process of the X variable at the t point is as follows: the X variable is any one of the four variables of speed v, longitudinal acceleration ax, lateral acceleration ay and yaw angle r; The method of judging the data change trend at point t based on the derivative of point t and its neighborhood and the data fluctuation includes: Step 1: Get data point x t Variable data SW_X in the neighborhood t and variable derivative data SW_dX t , the neighborhood range is [max(t-30,1), min(t+30,NN j )]; NNj is the length of the current driving behavior primitive data; Step 2: If max(SW_X t )-min(SW_X t )>α·(max(X)-min(X)), α is the threshold setting ratio, with a value of 0.

02. Based on the data point x t-1 、x t 、x t+ The derivative sign at 1 determines x t The trend of data changes; Step 3: If max(SW_X t )-min(SW_X t )≤α·(max(X)-min(X)), then based on SW_dX t The symbol distribution of x t The data change trend.

8. The driving behavior primitive clustering method considering data change trend according to claim 7, characterized in that: Said max(SW_X t )-min(SW_X t )>α·(max(X)-min(X)), if |dx t |>α·min(|max(dX)|,|min(dX)|), then based on dx t The symbol output data point x t The changing trend (upward, downward), when dx t When it is greater than the derivative discrimination threshold, the data point x t The trend of change is upward, otherwise the data point x t The changing trend is downward; If |dx t |≤α·min(|max(dX)|, |min(dX)|), then based on dx t-1 The symbol output data point x t The changing trend (increasing, decreasing, stable), when dx t-1 When the absolute value of is less than or equal to the derivative discrimination threshold, it is stable. t-1 If it is greater than the derivative judgment threshold, it is rising, otherwise it is falling. In particular, when t = 1, based on dx t+1 The symbol output data point x t The changing trend (increasing, decreasing, stable), when dx t+1 When the absolute value of is less than or equal to the derivative discrimination threshold, it is stable. t+1 If it is greater than the derivative discrimination threshold, it is rising, otherwise it is falling.

9. The driving behavior primitive clustering method considering data change trend according to claim 7, characterized in that: max(SW_X t )-min(SW_X t )≤α·(max(X)-min(X)), if SW_dX t The number of data points between [-α·min(|max(dX)|, |min(dX)|), α·min(|max(dX)|, |min(dX)|)] is denoted as l, the number of data points greater than α·min(|max(dX)|, |min(dX)|) is denoted as m, and the number of data points less than -α·min(|max(dX)|, |min(dX)|) is denoted as n; if Then the data point x t The trend of change is upward. If Then the data point x t The trend of change is downward, otherwise the data point x t The changing trend is stable.

10. The driving behavior primitive clustering method considering data change trend according to claim 1, characterized in that: The discretization results are data distribution items and data change trend items of speed, longitudinal acceleration, lateral acceleration and yaw angle; The discretization result is: one of the output codes of the velocity distribution item, one of the output codes of the velocity change trend item, one of the output codes of the longitudinal acceleration distribution item, one of the output codes of the longitudinal acceleration change trend item, one of the output codes of the steering behavior item, one of the output codes of the lateral acceleration change trend item, and one of the output codes of the yaw angle change trend item.

Images