Driving behavior safety analysis method based on driving mode primitives

Abstract

The present invention discloses a driving behavior safety analysis method based on driving mode primitives, collects the driver's driving data to extract effective vehicle-following driving events, uses the good clustering characteristics of HDP and the hierarchical sharing principle to provide priors for the state transition distribution, observation probability distribution and duration distribution parameters of the HSMM model, and uses the HSMM double random process to model the driving behavior. The constructed HDP-HSMM can accurately segment and identify the potential basic driving modes in the time series driving behavior data, and effectively extract the driving mode primitives. The K-means clustering method is used to cluster the data of each driving mode primitive, determine the frequency distribution of driving behavior, and mark the driving behavior for risk. Based on the risk marking of driving events, a driving behavior characterization index set is established, and a driving risk prediction and evaluation model is built based on the SVM algorithm to predict the safety of driving behavior.

Description

Technical Field

[0001] The present invention relates to the technical field of driver's driving behavior evaluation or prediction, and in particular to a driving behavior safety analysis method based on driving mode primitives. Background Art

[0002] At present, most of the analysis and evaluation research on driving behavior safety is based on statistical indicators such as the mean and standard deviation of driving behavior data or the frequency of emergency events such as sudden acceleration and deceleration to cluster and evaluate driving safety performance. These indicators can show the static characteristics of driving behavior but are difficult to describe the dynamic and random characteristics of driving behavior, which is not conducive to the in-depth understanding and analysis of driving behavior. In terms of research purposes, it is also mainly focused on driver classification, and there is relatively little analysis of the driving behavior characteristics of individual drivers, which makes it very difficult to standardize and guide individual driving behavior and characterize the safety of individual driving behavior.

[0003] Therefore, there is an urgent need for a technology that can deeply understand the dynamic process of driving behavior and effectively characterize the characteristics of individual driving behavior to help detect and supervise high-risk drivers and improve individual driving behavior. Summary of the invention

[0004] In view of the shortcomings of the existing technology, the present invention proposes a driving behavior safety analysis method based on driving mode primitives, which can characterize the driving behavior characteristics of individual drivers to help detect and supervise high-risk drivers and improve the driving behavior of individual drivers. The specific technical solution is as follows:

[0005] In the first aspect, a driving behavior safety assessment method based on driving mode primitives is provided, comprising:

[0006] Processing the acquired natural driving data of the driver to obtain driving time series data;

[0007] The HDP-HSMM model is used to identify driving time series data and extract the driver's driving mode primitives;

[0008] The K-Means algorithm is used to cluster the extracted driving mode primitives to determine the frequency distribution of the driving mode primitives;

[0009] According to the frequency distribution and risk level of various driving mode primitives, a risk index for evaluating the safety of driving behavior is calculated.

[0010] In combination with the first aspect, in a first implementable manner of the first aspect, the driving time series data is driving time series data corresponding to a valid vehicle following event.

[0011] In combination with the first aspect, in a second implementable manner of the first aspect, the HDP-HSMM model is used to identify driving time series data, including:

[0012] Extracting driving behavior characterization parameters from driving time series data;

[0013] Standardize the extracted driving behavior representation parameters;

[0014] Based on the standardized driving behavior representation parameters, the HDP-HSMM model is used to extract driving mode primitives.

[0015] In combination with the second implementable manner of the first aspect, in a third implementable manner of the first aspect, the driving behavior characterization parameters include headway time, headway time change rate, and host vehicle acceleration.

[0016] In combination with the second implementable manner of the first aspect, in a fourth implementable manner of the first aspect, a Z-score method is used to standardize the driving behavior characterization parameters.

[0017] In combination with the first aspect, in a fifth possible implementation manner of the first aspect, clustering the extracted driving mode primitives using a K-Means algorithm includes:

[0018] The K-means clustering algorithm is used to cluster the driving behavior representation parameters of the same driving mode primitive, and the representation parameter clustering data of each driving mode primitive is obtained;

[0019] Match the representation parameter clustering data with the belonging threshold ranges of different semantic spaces to determine the semantic space to which each driving mode primitive belongs;

[0020] The number of driving mode primitives corresponding to each semantic space is normalized to determine the frequency distribution of the driving mode primitives.

[0021] In combination with the first aspect, in a sixth implementation manner of the first aspect, a hierarchical analysis method is used to determine the risk level of each driving mode primitive.

[0022] In combination with the first aspect, in a seventh possible implementation manner of the first aspect, the present invention further includes:

[0023] Calculate the frequency of emergency safety events during the driver's driving process based on driving time series data;

[0024] Conduct correlation analysis on the frequency of emergency safety incidents and risk index to verify the risk index.

[0025] In the second aspect, a driving behavior safety prediction method based on driving mode primitives is provided, including:

[0026] Obtain natural driving data from different drivers;

[0027] Using the driving behavior safety assessment method of the first aspect or any one of the first to sixth achievable methods of the first aspect, the risk index of each vehicle following event of different drivers is calculated, and the risk type of each vehicle following event is determined;

[0028] Extract the driving characteristic parameters of each vehicle-following event to perform classification training on the SVM model;

[0029] The driver's natural driving data is collected, and the driving characteristic parameters are extracted and input into the trained SVM model to predict the driving behavior risk.

[0030] In conjunction with the second aspect, in a first possible implementation of the second aspect, determining the risk type to which each vehicle following event belongs includes:

[0031] According to the risk index of each car-following event, the K-means clustering algorithm is used to cluster all car-following events into two risk types: safe and dangerous.

[0032] Beneficial effects: The driving behavior safety analysis method based on driving mode primitives of the present invention is adopted, and the good clustering characteristics and hierarchical sharing principle of the hierarchical Dirichlet process (HDP) are used to provide the prior and hidden state numbers for the hidden semi-Markov model (HSMM). The driving behavior is modeled using the HSMM double random process. The constructed HDP-HSMM model can accurately segment and identify the potential basic driving modes in the driving data, and effectively extract the driving mode primitives. This method is an unsupervised learning method and does not require prior knowledge of the driving mode. On this basis, the driving mode primitives extracted by identification are further analyzed, the safety level of each mode primitive is quantitatively analyzed, and the frequency distribution of the driving mode primitives is determined, so as to calculate the risk index of the driver's driving behavior and quantitatively evaluate the driver's driving safety level. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the specific implementation of the present invention, the following will briefly introduce the drawings required for use in the specific implementation. In all the drawings, each element or part is not necessarily drawn according to the actual scale.

[0034] Figure 1 A flowchart of a driving behavior safety assessment method provided by an embodiment of the present invention;

[0035] Figure 2 A flowchart of a method for identifying a driving mode primitive provided by an embodiment of the present invention;

[0036] Figure 3A flow chart of a method for determining the frequency distribution of a driving mode primitive provided by an embodiment of the present invention;

[0037] Figure 4 It is the directed graph model of HDP-HSMM;

[0038] Figure 5 It is a schematic diagram of semantic space;

[0039] Figure 6 Schematic diagram for calculating risk levels for stratification;

[0040] Figure 7 This is a flow chart of a driving behavior safety prediction method provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following embodiments of the technical solution of the present invention are described in detail in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are therefore only used as examples, and cannot be used to limit the protection scope of the present invention.

[0042] "Driving mode primitives" refer to simple, small, and basic driving modes in driving behavior. Driving behavior can be regarded as a dynamic process composed of the driver's response to the external environment. This dynamic process can be composed of a series of driving mode primitives. For example, in the process of following a car, the driver's following behavior can be simply divided into several driving mode primitives such as approaching, steady-state following, and driving away. By analyzing the driving mode primitives, the dynamic and random characteristics of driving behavior can be obtained, and then the safety of driving behavior can be evaluated from the perspective of the dynamic process of driving behavior.

[0043] like Figure 1 The flowchart of the driving behavior safety assessment method based on the driving mode primitive is shown, and the assessment method includes:

[0044] Step 1: Processing the acquired natural driving data of the driver to obtain driving time series data;

[0045] Step 2: Use the HDP-HSMM model to identify the driving time series data and extract the driver's driving mode primitives;

[0046] Step 3: cluster the extracted driving mode primitives using the K-Means algorithm to determine the frequency distribution of the driving mode primitives;

[0047] Step 4: Calculate the risk index for evaluating the safety of driving behavior based on the frequency distribution and risk level of various driving mode primitives.

[0048] Specifically, first, the natural driving data of the driver during driving can be collected and processed to obtain the driving time series data during driving. Then, the trained HDP-HSMM model is used to identify the driving mode primitives in the driving time series data. The driving mode primitive refers to the simplest and most basic driving mode in the driving process. The HDP-HSMM model can effectively divide the entire driving process into a series of driving mode primitives. After that, the K-Means algorithm can be used to cluster the extracted driving mode primitives to determine the frequency distribution of the driving mode primitives. Finally, according to the frequency distribution and risk level of each driving mode primitive, the risk index used to evaluate the safety of driving behavior is calculated.

[0049] Therefore, the evaluation method implemented in the present invention utilizes the good clustering characteristics of the hierarchical Dirichlet process and the hierarchical sharing principle to provide the prior and hidden state numbers for the hidden semi-Markov model, and utilizes the HSMM double random process to model the driving behavior. The constructed HDP-HSMM model can accurately segment and identify the potential basic driving modes in the driving data, and effectively extract the driving mode primitives to characterize the driving behavior characteristics of individual drivers.

[0050] Then, the extracted driving mode primitives are further analyzed using algorithms such as K-Means to determine the frequency distribution of the driving mode primitives. Finally, the risk index of driving behavior is calculated based on the frequency distribution of the driving mode primitives and the risk level, thereby quantitatively evaluating the driver's safety level. Among them, the risk index can quantitatively evaluate the driver's safety performance. The higher the risk index, the worse the driving safety performance.

[0051] The evaluation method implemented by the present invention will be described in detail below with reference to the accompanying drawings.

[0052] In step 1, the collected natural driving data includes vehicle operation state data and driving environment signal data. After collecting the natural driving data, data processing can be performed on the natural driving data, and the data processing includes sampling and filtering. It should be understood that in an embodiment of the present invention, the resample function can be used to downsample the natural driving data, and the Savitzky-Golay filter can be used to filter the sampled data, but the present invention is not limited to this.

[0053] In this embodiment, preferably, when the distance between the two vehicles is too large, the driving behavior of the main vehicle will no longer be affected by the front vehicle and will be in a free driving state, which is invalid driving time domain data. In order to avoid the influence of non-following scenes during driving, the driving time series data refers to the driving time series data corresponding to the effective following event. It should be understood that in the embodiment of the present invention, the following rules can be used to identify the following event during driving based on the sampled data, thereby extracting the driving time series data.

[0054] (1) There is a vehicle ahead of the main vehicle in its lane;

[0055] (2) The relative distance between the main vehicle and the vehicle in front is less than 120m, and the following distance is less than 6s. If it exceeds this range, the main vehicle is considered to be in a free driving state;

[0056] (3) The vehicle speed is greater than 18 km / h;

[0057] (4) The duration of the vehicle-following incident is greater than 15 seconds.

[0058] In this embodiment, preferably, Figure 2 As shown, in step 2, the HDP-HSMM model is used to identify the driving time series data, including:

[0059] Step 2-1, extracting driving behavior characterization parameters from driving time series data;

[0060] Step 2-2, standardizing the extracted driving behavior characterization parameters;

[0061] Step 2-3: Based on the standardized driving behavior characterization parameters, the HDP-HSMM model is used to extract the driving mode primitives.

[0062] In step 2-1, preferably, the driving behavior characterization parameters include headway, headway change rate and host vehicle acceleration, wherein;

[0063] Headway (THW) refers to the time difference between the leading vehicle and the main vehicle when they pass the same observation position. Headway can well reflect the driver's driving habits and preferences, and can also characterize the degree of danger of following the vehicle. Based on the driving time series data, the headway can be calculated using the following formula:

[0064]

[0065] △S=x 2 -x 1

[0066] Among them, V s is the speed of the main vehicle, x 1 、x 2They are the positions of the main vehicle and the front vehicle at the current moment respectively.

[0067] The headway time change rate (dTHW / dt) is the first-order derivative of the headway time. Its value directly reflects the speed of change of the headway time, and can well represent the dynamic process of the driver's headway time control based on his own perception during the following process. The headway time change rate can be calculated using the following formula:

[0068]

[0069] Among them, △V is the difference between the speed of the main vehicle and the front vehicle at the current moment.

[0070] Main vehicle acceleration (a s ) is a direct reflection of the driver's operation of the accelerator pedal and the brake pedal, representing the intensity of the driver's operation and can be used as a characterization indicator of driving style and driving behavior safety analysis.

[0071] It should be understood that in the embodiment of the present invention, the headway time, the headway time change rate and the host vehicle acceleration are optimally selected as driving behavior characterization parameters, but the present invention is not limited thereto.

[0072] In this embodiment, preferably, in step 2-2, the driving behavior characterization parameter may be standardized by using a Z-score method to eliminate the influence of the characteristic parameter dimension on the recognition of the driving mode primitive. The specific calculation formula is as follows:

[0073]

[0074] in, is the characterization parameter data of the mth vehicle following event, y = [THW, dTHW / dt, a s ], μ and σ are the mean and standard deviation of the driver’s characteristic parameter data respectively.

[0075] In this embodiment, in step 2-3, the driving behavior characterization parameters of each vehicle-following event after standardization can be input into the trained HDP-HSMM model to extract the potential driving mode primitives in each vehicle-following event.

[0076] It should be understood that in the embodiment of the present invention, the HDP-HSMM model is a Bayesian non-parametric inference model, and its directed graph is as follows: Figure 4 As shown, it can be described as:

[0077] β|γ~GEM(γ)

[0078] (θ i ,ω i )~H×G

[0079] πi |α, β~DP(α, β)

[0080]

[0081] D S ~g(ω zs )

[0082]

[0083]

[0084] Among them, α, γ are hyperparameters that need to be set in advance; β, π i is the weight coefficient of the Stick-breaking process, GEM refers to the Stick-breaking process; DP represents the Dirichlet process; G obeys the Dirichlet process, denoted as G~DP(α,H); H is a basis distribution on the measure space Θ; z S is the hidden state sequence; s represents the observation sequence; g(ω S ) is a specific state z S Duration D S The distribution of is the prior parameter of the distribution, F(·) is a parameter distribution family, i=1,2,.......,∞;s=1,2,.......,S.

[0085] It should be understood that in the embodiment of the present invention, three characteristic parameters are used to characterize the vehicle following process, and the dimension of the observation value is 3, but the present invention is not limited to this. It should also be understood that the present embodiment can use a ternary Gaussian distribution as the observation distribution, and the corresponding parameter vector θ i ={μ i ,∑ i}, set the prior distribution to parameter φ={μ 0 , κ 0 , ν 0 , S 0} is a normal-inverse-Wishart distribution, that is:

[0086]

[0087]

[0088] Among them, μ 0 is the prior mean, ν 0 ,△ 0 are the degrees of freedom and scale matrix of the IW distribution respectively.

[0089] In the embodiment of the present invention, Poisson distribution can be used as the duration distribution form of the HSMM model, and the parameters of the prior distribution are shown in the following table:

[0090]

[0091] Where N is the dimension of the input data, is the covariance matrix of the observed data.

[0092] When constructing the HDP-HSMM model, the driving data of different drivers can be collected through the driving simulation test platform. The collected data includes the driver's operating behavior, vehicle motion state, and driving environment signal data. The above data processing method is used for data processing to obtain the driving time series data of different drivers, and the driving behavior characterization parameters of all following events of different drivers are extracted from them, and a training set is constructed to train the HDP-HSMM model. During the training process, considering the complexity of the model and the sample data, based on the model prior settings in the above table, Gibbs sampling can be used to infer all parameters in the HDP-HSMM model to obtain a trained HDP-HSMM model.

[0093] In step 2-3, the driving behavior characterization parameters of the vehicle-following event obtained after the standardized processing are input into the trained HDP-HSMM model, so as to accurately extract the driving mode primitives of the vehicle-following event that occurs during the driving process.

[0094] In this embodiment, preferably, Figure 3 As shown, in step 3, the K-Means algorithm is used to cluster the extracted driving mode primitives, including:

[0095] Step 3-1, clustering the driving behavior representation parameters of the same driving mode primitive using a K-means clustering algorithm to obtain clustering data of the representation parameters of each driving mode primitive;

[0096] Step 3-2, matching the representation parameter clustering data with the attribution threshold ranges of different semantic spaces to determine the semantic space to which each driving mode primitive belongs;

[0097] Step 3-3: Normalize the number of driving mode primitives corresponding to each semantic space to determine the frequency distribution of the driving mode primitives.

[0098] In step 3-1, it should be understood that in an embodiment of the present invention, the K-means clustering algorithm is used to cluster the driving behavior characterization parameters of different driving mode primitives in the same effective vehicle following event, and obtain the characterization parameter clustering data of each driving mode, but the present invention is not limited to this.

[0099] In step 3-2, the characterization parameter clustering data corresponding to each driving mode primitive can be matched with the attribution threshold range of different semantic spaces to determine the semantic space to which each driving mode primitive belongs, wherein the attribution thresholds of different semantic spaces can be divided according to the driver's physiological and psychological perception thresholds.

[0100] In the study of longitudinal driving behavior safety, the headway is considered to be one of the most effective indicators of driving safety and has been widely used in related research. It can be seen intuitively that different headway correspond to different risk levels. For example, when following closely, due to the small headway, the driver has less time to react, so rear-end collisions are likely to occur. Therefore, close-range following is also considered to be one of the typical representatives of aggressive driving behavior. Compared with close-range following, the long-range following mode has a higher safety because it reserves a larger reaction time. It is a following distance mode that simply explains the driving mode primitive.

[0101] In the embodiment of the present invention, the vehicle-following behavior is divided into three levels: close-range, medium-range and long-range following, using the headway thresholds of 1s and 2s obtained from the driver's braking reaction time test in "Research on Driving Style Evaluation Based on Mode Shift and Control Characteristics" (Tsinghua University, Li Guofa, 2016).

[0102] The headway change rate is the first-order derivative of the headway change rate, which can well reflect the dynamic change process of the headway change process during the following process. At the same time, its value can also reflect the dynamic stability of the driver's driving of the vehicle. The larger its value, the more intense the driving process changes, the greater the headway change of the vehicle, and the greater the driving risk fluctuation caused. For example, in the close-following mode, when the driver operates the vehicle to get closer to the front vehicle with a larger headway change rate, the collision risk increases sharply. On the contrary, when following closely, if the driver moves away from the front vehicle with a larger headway change rate, the collision risk will drop rapidly.

[0103] The "Driving Style Evaluation Study Based on Mode Shift and Control Characteristics" uses 0.03 and -0.03 as thresholds to divide the following mode into three types: approaching, gradually moving away, and steady-state following. However, from the perspective of driving safety, when the absolute value of the headway change rate is greater than 0.1, the headway will change significantly and the safety will fluctuate greatly. Therefore, the thresholds of -0.1, -0.03, 0.03, and 0.1 are used to divide the changes in driving modes into five modes: fast approaching, approaching, keeping headway, leaving, and fast leaving.

[0104] The acceleration of the main vehicle mainly reflects the intensity of the driver's operation. The larger the value, the stronger the driver's reaction to the traffic environment. Existing driving behavior studies have shown that external stimuli can only be perceived by the driver when they exceed a certain threshold. In related studies, 0.05m / s and 0.2m / s are used as thresholds for the driver's vestibular kinesthetic perception and for making the driver feel more comfortable with the vehicle acceleration. Therefore, -0.2m / s, -0.05m / s, 0.05m / s and 0.2m / s are used as thresholds to divide the acceleration into 5 levels, namely rapid deceleration, deceleration, no acceleration or deceleration, acceleration and rapid acceleration.

[0105] In summary, in the embodiment of the present invention, the attribution threshold ranges corresponding to different semantic spaces are specifically shown in the following table:

[0106]

[0107]

[0108] According to the levels of the threshold parameters in the above table, 75 (3×5×5) semantic spaces can be divided. The sentence pattern used in the semantic space is "the driver [rapidly decelerates, decelerates, does not accelerate or decelerate, accelerates, and accelerates rapidly] with the vehicle in front of him [rapidly approaches, approaches, maintains the distance between vehicles, drives away, and drives away rapidly] in the following state of [close distance, medium distance, and long distance]" to describe each semantic space. The semantic space is as follows: Figure 5 shown.

[0109] In step 3-3, the number of driving mode primitives corresponding to each semantic space is normalized to determine the frequency distribution of the driving mode primitives. The specific calculation formula is as follows:

[0110]

[0111] Among them, i∈[long distance, medium distance, close distance], j∈[rapid approaching, approaching, keeping distance, leaving, fast leaving], N (i , j) is the number of driving mode primitives in the semantic space.

[0112] Driving mode primitives can help understand the dynamic process of driving behavior and can easily perform safety analysis. By normalizing the number of driving mode primitives in different semantic spaces, the frequency distribution of driving mode primitives can be determined. Based on the frequency distribution, a frequency distribution diagram of the driver's driving mode primitives can be drawn. The frequency distribution diagram of the driving mode primitives can be used to intuitively and qualitatively analyze the driver's driving habits and preferences. By combining it with the risk level of the driving mode primitive, the driver's safety performance in the car-following incident can be quantitatively evaluated, that is, the risk index is calculated. The specific calculation method of the risk index is as follows:

[0113] Score=∑f(i,j)×R (i，j)

[0114] Among them, R (i，j) is the risk level corresponding to different driving mode primitives. In this embodiment, optionally, a hierarchical analysis method can be used to determine the risk level of each driving mode primitive. Specifically, Figure 6 As shown, at the first level, considering that the headway can be used as the most direct indicator of driving safety, and the collision risks when following a vehicle at long, medium, and close distances vary greatly, based on the analysis in step 3.1, scores of 2, 6, and 10 can be assigned to long, medium, and close distances, respectively.

[0115] At the second level, the impact of different headway change rates on the risk level is divided into two directions. The direction of rapid departure and departure is negative. Considering the different degrees of impact, they can be assigned -2 ​​and -1 respectively. Similarly, the direction of rapid approach and approach impact is positive, and they can be assigned +2 and +1 respectively. Distance maintenance has no impact and is assigned +0. The higher the risk level, the greater the degree of danger. The hierarchical calculation process is shown in the figure.

[0116] It should be understood that in the embodiment of the present invention, a hierarchical analysis method is used to determine the risk level of the driving mode primitive in each semantic space, but the present invention is not limited thereto.

[0117] In this embodiment, optionally, it further includes:

[0118] Calculate the frequency of emergency safety events during the driver's driving process based on driving time series data;

[0119] Conduct correlation analysis on the frequency of emergency safety incidents and risk index to verify the risk index.

[0120] Specifically, the risk index of the driver in the following car event calculated by the above method can be correlated with the frequency of the driver's emergency safety time in the following car event calculated based on the driving time series data to verify whether the calculated risk index is reasonable. Specifically, it includes:

[0121] In an embodiment of the present invention, the safety of the driver's driving behavior can be evaluated and marked based on emergency safety events (SCEs), and the drivers can be divided into safe drivers and unsafe drivers. The definition of emergency safety events uses the vehicle state parameter related threshold and the sudden acceleration event threshold. The specific values ​​of the thresholds are shown in the following table:

[0122] name Trigger threshold Low collision time events TTC<1.5s Emergency braking incident Longitudinal acceleration <-0.3g Maximum longitudinal deceleration event Longitudinal acceleration <-5.5g Rapid acceleration events Longitudinal acceleration>0.3g Extreme jerk events Longitudinal jerk <1g

[0123] Where TTC is the time to collision and g is the acceleration due to gravity.

[0124] Specifically, based on the threshold standards in the above table, the number of emergency safety events that occur during the driver's driving process can be determined by traversing the driving data, thereby calculating the frequency of emergency safety events during the driver's driving process, and then evaluating the driver's driving safety performance. The higher the incidence of emergency safety events, the lower the safety of the driver's driving behavior.

[0125] like Figure 7 The flowchart of the driving behavior safety prediction method based on the driving mode primitive is shown, and the prediction method includes:

[0126] S1. Obtaining natural driving data of different drivers;

[0127] S2. Using the above-mentioned evaluation method, the risk index of each vehicle following incident of different drivers is calculated, and the risk type of each vehicle following incident is determined;

[0128] S3, extracting driving characteristic parameters of each vehicle following event to perform classification training on the SVM model;

[0129] S4. Collect the driver's natural driving data, extract the driving characteristic parameters and input them into the trained SVM model to predict the driving behavior risk.

[0130] Specifically, first of all, the natural driving data of different drivers can be collected through the driving simulation test platform. The natural driving data includes the driver's operating behavior, vehicle motion status, driving environment signal data, etc.

[0131] Then, the same process as the above evaluation method is used to process all the collected natural driving data to obtain the driving time series data during the driving process. The trained HDP-HSMM model is used to identify the driving mode primitives in the driving time series data, and the entire driving process is divided into a series of driving mode primitives. The K-Means algorithm is then used to cluster the extracted driving mode primitives to determine the frequency distribution of the driving mode primitives. According to the frequency distribution and risk level of each driving mode primitive, the risk index of different drivers in different following events is calculated, and the risk type of each following event of the driver is determined based on the risk index.

[0132] Afterwards, the driving characteristic parameters of different drivers in different car-following events are extracted, a training set is constructed, and the SVM model is classified and trained to obtain a trained SVM model. Finally, the natural driving data of the driver during driving is collected, and the driving characteristic parameters are extracted and input into the trained SVM model to predict the driving behavior risk.

[0133] In this embodiment, optionally, in S2, determining the risk type to which each car-following event belongs includes: clustering all car-following events into two risk types, safe and dangerous, using a K-means clustering algorithm according to a risk index of each car-following event. It should be understood that in the embodiment of the present invention, the K-means clustering algorithm is used to determine the risk type of the car-following event, but the present invention is not limited thereto.

[0134] In S3, in order to ensure the accuracy and reliability of the driving behavior risk prediction and assessment model, four parameters, namely, following distance, relative speed, main vehicle speed, and main vehicle acceleration, can be used as characteristic parameters of the model, and 16 basic statistical features are used to fully obtain parameter characteristics. The 16 basic statistical features are as follows: mean, standard deviation, coefficient of variation (quotient of standard deviation and mean), median, maximum value, minimum value, range, difference between maximum value and mean, difference between mean and minimum value, initial and final difference, difference between maximum value and median, root mean square, Shannon entropy, 25th percentile, 50th percentile, and 75th percentile. The calculation formula of Shannon entropy is as follows:

[0135]

[0136] Here, x represents the possible value of the random variable X.

[0137] In an embodiment, optionally, when constructing a training set, a full set of candidate features can be first constructed based on the extracted feature parameters, and then the constructed full set of candidate features can be optimized through a joint mutual information maximization feature selection algorithm (JMIM) based on information theory, and the contribution is sorted, and then the feature with the largest contribution to the driving risk prediction assessment is selected to construct a training set for training the SVM model, so as to reduce the redundancy of the training set.

[0138] In S3, all the collected driving characteristic parameter data can be divided into a test set and a training set, where the test set accounts for 20%. The SVM model is a classification model, and its support vector is as follows:

[0139]

[0140] Among them, α is the Lagrange multiplier, (x, y i )∈D, D is the training set sample, b is the model parameter, κ(x, y i ) is a kernel function, and the present invention adopts a Gaussian kernel, which is as follows:

[0141]

[0142] Here, σ>0 is the bandwidth of the Gaussian kernel.

[0143] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.

Claims

1. A driving behavior safety prediction method based on driving mode primitives, characterized in that: include: Obtain natural driving data from different drivers; The driving behavior safety assessment method is used to calculate the risk index of each car-following incident for different drivers, and determine the risk type of each car-following incident; The driving behavior safety assessment method comprises: The acquired natural driving data of the driver is processed to obtain driving time series data; The HDP-HSMM model is used to identify driving time series data and extract the driver's driving mode primitives; The K-Means algorithm is used to cluster the extracted driving mode primitives to determine the frequency distribution of the driving mode primitives; According to the frequency distribution and risk level of various driving mode primitives, a risk index for evaluating the safety of driving behavior is calculated; Determining the risk type of each vehicle following event includes: According to the risk index of each car-following event, the K-means clustering algorithm is used to cluster all car-following events into two risk types: safe and dangerous. Extract the driving characteristic parameters of each vehicle-following event to perform classification training on the SVM model; The driver's natural driving data is collected, and the driving characteristic parameters are extracted and input into the trained SVM model to predict the driving behavior risk.

2. The driving behavior safety prediction method based on driving mode primitives according to claim 1 is characterized in that: The driving time sequence data is driving time sequence data corresponding to a valid vehicle following event.

3. The driving behavior safety prediction method based on driving mode primitives according to claim 1 is characterized in that: The HDP-HSMM model is used to identify driving time series data, including: Extracting driving behavior characterization parameters from driving time series data; Standardize the extracted driving behavior representation parameters; Based on the standardized driving behavior representation parameters, the HDP-HSMM model is used to extract driving mode primitives.

4. The driving behavior safety prediction method based on driving mode primitives according to claim 3 is characterized in that: The driving behavior characterization parameters include headway, headway change rate and host vehicle acceleration.

5. The driving behavior safety prediction method based on driving mode primitives according to claim 3 is characterized in that: The Z-score method is used to standardize the driving behavior characterization parameters.

6. The driving behavior safety prediction method based on driving mode primitives according to claim 1 is characterized in that: The method of clustering the extracted driving mode primitives using the K-Means algorithm includes: The K-means clustering algorithm is used to cluster the driving behavior representation parameters of the same driving mode primitive, and the representation parameter clustering data of each driving mode primitive is obtained; Match the representation parameter clustering data with the belonging threshold ranges of different semantic spaces to determine the semantic space to which each driving mode primitive belongs; The number of driving mode primitives corresponding to each semantic space is normalized to determine the frequency distribution of the driving mode primitives.

7. The driving behavior safety prediction method based on driving mode primitives according to claim 1 is characterized in that: The analytic hierarchy process is used to determine the risk level of each driving mode primitive.

8. The driving behavior safety prediction method based on driving mode primitives according to claim 1 is characterized in that: Also includes: Calculate the frequency of emergency safety events during the driver's driving process based on driving time series data; Conduct correlation analysis on the frequency of emergency safety incidents and risk index to verify the risk index.

Images