A driving risk quantification method, grading method and prediction method

Abstract

The present invention discloses a driving risk quantification method, a grading method and a prediction method. The driving risk quantification method includes the following steps: obtaining the driving basic data of a first vehicle and a second vehicle; according to the driving basic data, obtaining the severity of a potential collision accident, and determining a first index representing the driving risk according to the severity of the potential collision accident; calculating the TTC according to the driving basic data, and determining a second index representing the driving risk according to the TTC; calculating the THW according to the driving basic data, and determining a third index representing the driving risk according to the THW; and obtaining a driving risk index according to the first index, the second index and the third index. The driving risk quantification method of the present invention takes into account the influence of the severity of a potential collision accident on the driving risk, and quantifies the driving risk more carefully and reasonably.

Description

Technical Field

[0001] The present invention belongs to the technical field of safety risk assessment, and particularly relates to a method for quantifying driving risks, a method for grading driving risks, and a method for predicting driving risks. Background Art

[0002] The assessment of driving risks, as an important part of the vehicle driving safety system, comprehensively and timely assessing driving risks helps the vehicle make reasonable decisions as early as possible to reduce risks and ensure driving safety.

[0003] In the assessment of driving risks, a reasonable way to quantify risks is particularly important. The existing quantification of driving risks is mostly based on warning indicators such as TTC and safety distance, which can express the risks faced by the vehicle from the time dimension, and rarely assess risks from the perspective of the consequences caused by accidents. This leads to the traditional risk quantification method ignoring the impact of accident consequences on the quantification of driving risks. Summary of the Invention

[0004] In view of the above problems, the present invention proposes a method for quantifying driving risks, a method for grading driving risks, and a method for predicting driving risks, which can quantify driving risks from the perspective of the urgency of accident occurrence and the severity after the accident.

[0005] A method for quantifying driving risks provided by the present invention includes the following steps: obtaining the driving basic data of the first vehicle and the second vehicle; according to the driving basic data, obtaining the severity of a potential collision accident, and according to the severity of the potential collision accident, determining a first index representing driving risks; according to the driving basic data, calculating TTC, and determining a second index representing driving risks according to TTC, where TTC is the estimated time for the first vehicle and the second vehicle to continue driving in their current operating states until they collide; according to the driving basic data, calculating THW, and determining a third index representing driving risks according to THW, where THW is the estimated time for the first vehicle to continue driving in its current operating state until it reaches the current position of the second vehicle; according to the first index, the second index, and the third index, obtaining a driving risk index.

[0006] Preferably, obtaining the severity of a potential collision accident according to the driving basic data, and determining a first index representing driving risks according to the severity of the potential collision accident specifically includes:

[0007] S1. Establish a prediction model for the severity of collision accidents: Obtain collision accident samples, where the collision events in the collision accident samples are labeled with the severity of the collision. The collision accident samples are divided into several collision accident sub-samples based on the severity of the collision label. The basic characteristics of the collision accident samples include: the absolute speed of the involved vehicle, the longitudinal speed of the involved vehicle, the lateral speed of the involved vehicle, the mass of the involved vehicle, and the heading angle of the involved vehicle. Based on the basic characteristics of the collision accident samples, perform feature construction to obtain constructed features, screen the constructed features, and obtain representative features that are correlated with the severity of the collision. Based on the representative features, perform oversampling or undersampling on the collision accident sub-samples to obtain balanced collision accident samples. Based on the balanced collision accident samples, use the representative features as the input and the severity of the collision accident as the output to establish a prediction model for the severity of the collision.

[0008] S2. Based on the prediction model for the severity of the collision accident, obtain the severity of potential collision accidents according to the driving basic data: From the driving basic data, screen out the basic characteristics. Based on the screened basic characteristics, perform feature construction to obtain representative features. Input the obtained representative features into the prediction model for the severity of the collision, and the severity of potential collision accidents can be obtained.

[0009] S3. According to the obtained severity of potential collision accidents, assign a value to the first index representing driving risk, and that's it.

[0010] Preferably, in S1, the collision events involved in the collision accident samples are collision events of two vehicles colliding. Number the involved vehicles and denote them as i, where i = 1 or 2. Based on the basic characteristics of the collision accident samples, perform feature construction to obtain constructed features, specifically including: the change in speed of involved vehicle i before and after the collision the change in longitudinal vehicle speed of involved vehicle i before and after the collision the change in lateral vehicle speed of involved vehicle i before and after the collision The relative speed of the two vehicles Relative-V = abs(v i -v 3-i ), and the relative collision angle Relative-θ = abs(θ i -θ 3-i ), where v i is the absolute speed of involved vehicle i, v 3-i is the absolute speed of involved vehicle 3 - i, vx i is the longitudinal speed of involved vehicle i, vx 3-i is the longitudinal speed of involved vehicle 3 - i, vy i is the lateral speed of involved vehicle i, vy 3-iis the lateral speed of the involved vehicle 3 - i, m i is the mass of the involved vehicle i, m 3-i is the mass of the involved vehicle 3 - i, θ i is the heading angle of the involved vehicle i, θ 3-i is the heading angle of the involved vehicle 3 - i, α is the angle at which the two vehicles approach each other, α = θ i +θ 3-i .

[0011] Preferably, a second index characterizing driving risk is determined according to TTC, specifically U = e -TTC ; and / or, a third index characterizing driving risk is determined according to THW, specifically k = e -THW ; and / or, according to the first index, the second index, and the third index, a driving risk index is obtained, specifically R = k·S·U; where R represents the driving risk index, S represents the first index, U represents the second index, and k represents the third index.

[0012] The second index U of the present invention can be considered as an index for quantifying driving risk from the perspective of accident urgency, and the third index k can be considered as the influence weight of the accident severity S on the following vehicle or the preceding vehicle, through which the risk and threat degree of the accident severity to the vehicle at different moments are adjusted. During the approaching collision process, the urgency of the accident should be increasing, until the moment of collision, the accident urgency U reaches the maximum value of 1, that is, the TTC is getting smaller and the sense of urgency of the accident is getting stronger. Therefore, TTC is negatively correlated with the accident urgency index; at the same time, the accident urgency does not change linearly with time. The smaller the collision time TTC, the greater the rising trend of the accident sense of urgency. When approaching collision and TTC is small, the accident urgency increases rapidly, and the overall change trend is first slow and then rapid. In summary, this patent draws on the change law of the exponential function with base e to obtain the expression of the second index. Similarly, for the third index k, the influence coefficient of the collision severity, during different stages of the approaching collision process, the risk brought by the potential accident severity increases as THW decreases. At the same time, the closer to the collision, the smaller THW is, the greater the risk caused by the accident severity. Therefore, the change law of the third index k during the approaching collision process is also expressed by an exponential function with base e.

[0013] Preferably, it further includes obtaining a driving risk index change rate according to the driving risk index, specifically: where t is a certain instantaneous moment, △t is a period of time after this instantaneous moment t, R(t) is the driving risk index corresponding to this instantaneous moment t, R(t + △t) is the driving risk index corresponding to the instantaneous moment (t + △t), and K R is the driving risk index change rate within [t, t + △t].[[]END]]

[0014] The present invention also provides a driving risk grading method, including the following steps: obtaining original driving samples, where the original driving samples include time series before several collision events occur, and the time series include the driving basic data of the first vehicle and the second vehicle at all times; obtaining the severity of a potential collision accident at the to-be-evaluated moment according to the driving basic data at the to-be-evaluated moment, and determining a first index characterizing the driving risk at the to-be-evaluated moment according to the severity of the potential collision accident; calculating the TTC at the to-be-evaluated moment according to the driving basic data at the to-be-evaluated moment, and determining a first index characterizing the driving risk at the to-be-evaluated moment according to the TTC, where the TTC is the estimated time for the first vehicle and the second vehicle to keep running at this moment until they collide; calculating the THW at the to-be-evaluated moment according to the driving basic data at the to-be-evaluated moment, and determining a third index characterizing the driving risk at the to-be-evaluated moment according to the THW, where the THW is the estimated time for the first vehicle to keep running at this moment until it reaches the position of the second vehicle at this moment; obtaining the driving risk index at the to-be-evaluated moment according to the first index, the second index, and the third index; clustering the driving risk indexes at each to-be-evaluated moment, and obtaining the driving risk level at each to-be-evaluated moment according to the clustering result; based on the driving risk level, a first driving sample with a driving risk level label can be obtained.

[0015] Preferably, clustering the driving risk indexes at each to-be-evaluated moment to obtain the driving risk level at each to-be-evaluated moment, specifically: judging whether the TTC at the to-be-evaluated moment is greater than a preset threshold. When the TTC is greater than the preset threshold, it is determined that the driving risk level at the to-be-evaluated moment is safe and risk-free. When the TTC is not greater than the preset threshold, the driving risk index at the to-be-evaluated moment is put into the R sample library for clustering, and the driving risk level at the to-be-evaluated moment is obtained according to the clustering result.

[0016] Preferably, it further includes the following steps: obtaining the change rate of the driving risk index in the time period where the to-be-evaluated moment is located according to the driving risk index, where the to-be-evaluated moment is the initial moment of its time period; performing unsupervised classification on the change rates of the driving risk indexes in the time periods where each to-be-evaluated moment is located, and obtaining the driving risk change trend in the time periods where each to-be-evaluated moment is located according to the clustering result; based on the driving risk change trend, a second driving sample with a driving risk trend label can be obtained.

[0017] Preferably, it further includes the following steps: clustering {LR, LK R}, and obtaining the driving risk patterns at each to-be-evaluated moment and its time period according to the clustering result, where LR is the driving risk level at a certain to-be-evaluated moment, and LK Ris the driving risk change trend in the time period where the moment to be evaluated is located; based on the driving risk pattern, the third driving sample with the driving risk pattern label can be obtained.

[0018] The present invention also provides a driving risk prediction method, including the following steps: predicting the driving trajectories of the host vehicle and other vehicles from the current moment to a future moment using a trajectory prediction method to obtain the online driving data of the host vehicle and other vehicles from the current moment to a future moment; predicting the online driving risk from the current moment to a future moment based on the online driving data and a driving risk assessment model; wherein, the establishment of the driving risk assessment model includes the following steps: using the method described in any one of claims 6-9 to obtain the driving samples after driving risk classification, and the driving samples after driving risk classification are one of the first driving sample, the second driving sample, and the third driving sample; using the driving samples after driving risk classification to construct a driving risk assessment model based on the HMM algorithm.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] (1) The present invention quantifies the driving risk from the perspectives of the urgency of accident occurrence and the severity after the accident through the severity of potential collision accidents (i.e., the severity of the accident if a collision occurs) and TTC, THW. This quantification method can take into account the impact of accident consequences on driving risk and is more detailed and reasonable in quantifying driving risk.

[0021] (2) The present invention obtains the change rate of the driving risk index according to the driving risk index, and the quantification of the driving risk is more comprehensive, and the risk change can also be included in the evaluation scope.

[0022] (3) After processing the driving samples through the driving risk classification method, the present invention can construct a driving risk assessment model using the obtained samples after risk classification, which can more intuitively display the driving risk.

[0023] (4) The present invention combines the above driving risk assessment model with the trajectory prediction method, can comprehensively and meticulously predict the driving risk, and can give the vehicle more sufficient decision-making time, which helps the vehicle to make reasonable decisions in advance and reduce the driving risk. Description of the Drawings

[0024] Figure 1 is a schematic diagram of the driving risk quantification method of the present invention;

[0025] Figure 2 is the clustering result of the driving risk index in Embodiment 1 of the present invention;

[0026] Figure 3 is the clustering result of the change rate of the driving risk index in Embodiment 1 of the present invention;

[0027] Figure 4 For the clustering result of {LR, LK R} in Embodiment 1 of the present invention;

[0028] Figure 5 It is a diagram showing an effect of predicting the driving risk of Scenario 1 in Embodiment 1 of the present invention;

[0029] Figure 6 It is another diagram showing an effect of predicting the driving risk of Scenario 1 in Embodiment 1 of the present invention;

[0030] Figure 7 It is a diagram showing an effect of predicting the driving risk of Scenario 2 in Embodiment 1 of the present invention;

[0031] Figure 8 It is another diagram showing an effect of predicting the driving risk of Scenario 2 in Embodiment 1 of the present invention;

[0032] Figure 9 It is a diagram showing an effect of predicting the driving risk of Scenario 3 in Embodiment 1 of the present invention;

[0033] Figure 10 It is another diagram showing an effect of predicting the driving risk of Scenario 3 in Embodiment 1 of the present invention. Detailed implementation manners

[0034] The technical solutions in the embodiments of the present invention will be described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0035] Embodiment 1

[0036] This embodiment provides a driving risk quantification method, including the following steps:

[0037] Step 1: Obtain the driving basic data of the first vehicle and the second vehicle.

[0038] Specifically, the driving basic data of the first vehicle and the second vehicle include the absolute speed, longitudinal speed, lateral speed, mass, heading angle, relative position (or distance), etc. of the first vehicle and the second vehicle.

[0039] Step 2: Determine the first index, second index, and third index characterizing the driving risk according to the driving basic data.

[0040] Among them, determining the second indicator specifically includes: calculating TTC according to the driving basic data, where TTC is the estimated time for the first vehicle and the second vehicle to keep driving in the current running state until the two vehicles collide; determining the second indicator U representing driving risk according to TTC, specifically U = e -TTC 。

[0041] Determining the third indicator specifically includes: calculating THW according to the driving basic data, where THW is the estimated time for the first vehicle to keep driving in the current running state until it reaches the current position of the second vehicle; determining the third indicator k representing driving risk according to THW, specifically k = e -THW 。

[0042] Determining the first indicator specifically includes: S1. Establishing a prediction model for the severity of collision accidents; S2. Based on the prediction model for the severity of collision accidents, obtaining the severity of potential collision accidents according to the driving basic data; S3. Assigning a value to the first indicator representing driving risk according to the obtained severity of potential collision accidents, that's all.

[0043] Furthermore,

[0044] S1. Establishing a prediction model for the severity of collision accidents specifically includes the following steps:

[0045] (1) Obtaining collision accident samples, where the collision events in the collision accident samples are labeled with the severity of collision, and the collision accident samples are divided into several collision accident sub-samples based on the severity of collision label.

[0046] The collision accident samples can be collected by oneself, or obtained by using third-party websites or databases, etc. As a preferred solution in this embodiment, the collision accident samples are obtained from the NHTSA collision sample library in the United States. This database is an open-source database with rich collision cases, large amounts of information, and all are collision events with severity of collision labels. The severity of collision labels severity carried by the collision events in the collision accident samples obtained in this embodiment can be divided into three categories, namely severity1 representing minor collisions, severity2 representing moderate collisions, and severity3 representing severe collisions. Based on the severity of collision label, the collision accident samples can be divided into three collision accident sub-samples, namely the minor collision accident sub-sample crash1, the moderate collision accident sub-sample crash2, and the severe collision accident sub-sample crash3.

[0047] As a preferred solution, this embodiment obtains collision accident samples based on the collision type. Considering that the characterization features of the severity of collision accidents corresponding to collision events of different collision types are slightly different, in order to establish a more accurate and better prediction model, the collision events involved in the collision accident samples are collision events of the same collision type, all of which are same-direction collisions or all of which are reverse collisions, wherein the same-direction collisions and reverse collisions include vehicle collisions in the same lane (rear-end collisions in same-direction collisions or frontal collisions in reverse collisions) and vehicle collisions in adjacent lanes (side collisions in the same driving direction or side collisions in opposite driving directions). For example, this embodiment establishes a prediction model for the research scenario of multiple lanes in the same direction, and the collision events involved in the collision accident samples obtained are all same-direction collisions, and the same-direction collisions include rear-end collisions and side collisions in the same driving direction.

[0048] (2) Based on the basic characteristics of the collision accident sample, feature construction is performed to obtain construction features, and the construction features are screened to obtain characterization features that are correlated with the severity of the collision.

[0049] In this embodiment, as a preferred solution, the collision event involved in the collision accident sample is a collision event between two vehicles. The vehicles involved are numbered and recorded as i, i=1 or 2, where 1 is the active vehicle and 2 is the passive vehicle.

[0050] Among the factors that affect the severity of a collision accident, the factors related to the vehicle involved mainly include vehicle speed, mass, and collision position. To predict the severity of a collision, the basic features corresponding to these factors in the collision accident sample are first analyzed. Those skilled in the art should know that when studying the severity of a collision accident, the basic features of the collision accident sample refer to the basic parameters of the vehicle at the time of the collision. For example, if the collision process is a time period of t0-t1, the time of the collision is t0.

[0051] The basic characteristics of the collision accident sample include: the absolute speed v of the vehicles involved (including the absolute speed v1 of the vehicle 1 involved and the absolute speed v2 of the vehicle 2 involved), the longitudinal speed vx of the vehicles involved (including the longitudinal speed vx1 of the vehicle 1 involved and the longitudinal speed vx2 of the vehicle 2 involved), the lateral speed vy of the vehicles involved (including the lateral speed vy1 of the vehicle 1 involved and the lateral speed vy2 of the vehicle 2 involved), the mass m of the vehicles involved (including the mass m1 of the vehicle 1 involved and the mass m2 of the vehicle 2 involved), and the heading angle θ of the vehicles involved (including the heading angle θ1 of the vehicle 1 involved and the heading angle θ2 of the vehicle 2 involved).

[0052] For the collision accident samples of this embodiment, the correlation between the basic characteristics and the collision severity label severity is analyzed, and the Pearson correlation coefficient r is used to represent it as shown in Table 1:

[0053] Pearson correlation coefficients between the basic features and the collision severity labels in Table 1

[0054] r <![CDATA[v1]]> <![CDATA[vx1]]> <![CDATA[vy1]]> <![CDATA[v2]]> <![CDATA[vx2]]> <![CDATA[vy2]]> <![CDATA[m1]]> <![CDATA[m2]]> <![CDATA[θ1]]> <![CDATA[θ2]]> severity 0.127 0.115 0.071 -0.017 -0.011 0.009 -0.189 0.062 -0.018 -0.032

[0055] Generally, it is considered that if the absolute value of the Pearson correlation coefficient r is above 0.8, there is a strong correlation between the two; if it is between 0.3 and 0.8, there is a weak correlation between the two; if it is below 0.3, there is no correlation between the two.

[0056] It can be seen from Table 1 that there is no correlation between each basic feature of the collision accident samples and the collision severity labels, and it is impossible to predict the severity of potential vehicle collision accidents through the basic features. Thus, it can be seen that the three factors of mass, speed, and impact position do not affect the severity individually, but jointly affect the collision severity, and it is impossible to achieve the purpose of predicting the severity of potential collision accidents only by the basic features.

[0057] In order to obtain the characteristic features that are correlated with the collision severity labels so as to predict the severity of potential collision accidents, in this embodiment, based on the above basic features, feature construction is carried out. When a vehicle collides, energy transfer occurs, that is, kinetic energy is converted into other energies, causing vehicle deformation and personnel injuries. Therefore, in this embodiment, considering the change in kinetic energy, the following features are selected for construction. The constructed features include: the change in speed ΔV before and after the collision of the involved vehicles (including the change in speed ΔV1 before and after the collision of involved vehicle 1 and the change in speed ΔV2 before and after the collision of involved vehicle 2), the change in longitudinal vehicle speed ΔVx before and after the collision of the involved vehicles (including the change in longitudinal vehicle speed ΔVx1 before and after the collision of involved vehicle 1 and the change in longitudinal vehicle speed ΔVx2 before and after the collision of involved vehicle 2), the change in lateral vehicle speed ΔVy before and after the collision of the involved vehicles (including the change in lateral vehicle speed ΔVy1 before and after the collision of involved vehicle 1 and the change in lateral vehicle speed ΔVy2 before and after the collision of involved vehicle 2), the relative speed of the two vehicles Relative-V, and the collision relative angle Relative_θ.

[0058] However, since the collision process time is short, it is relatively difficult to directly obtain the change process of the vehicle's basic parameters. Therefore, when carrying out feature construction in this embodiment, from the perspective of the conservation of momentum during the collision process, the collision process of the vehicle is described, and feature construction is carried out based on the above basic features to construct features that can represent the severity of the accident after the collision. Specifically, it includes:

[0059] The change in speed of involved vehicle i before and after the collision

[0060] The change in longitudinal vehicle speed of involved vehicle i before and after the collision

[0061] Lateral vehicle speed change of involved vehicle i before and after collision

[0062] Relative speed of two vehicles Relative-V = abs(v i - v 3-i ),

[0063] Relative collision angle Relative-θ = abs(θ i - θ 3-i ),

[0064] wherein, v i is the absolute speed of involved vehicle i, v 3-i is the absolute speed of involved vehicle 3 - i,

[0065] vx i is the longitudinal speed of involved vehicle i, vx 3-i is the longitudinal speed of involved vehicle 3 - i,

[0066] vy i is the lateral speed of involved vehicle i, vy 3-i is the lateral speed of involved vehicle 3 - i,

[0067] m i is the mass of involved vehicle i, m 3-i is the mass of involved vehicle 3 - i,

[0068] abs refers to the absolute value function,

[0069] θ i is the heading angle of involved vehicle i, θ 3-i is the heading angle of involved vehicle 3 - i,

[0070] α is the angle at which the two vehicles approach each other, α = θ i + θ 3-i .

[0071] For the collision accident samples of this embodiment, analyze the correlation between the above structural features and the collision severity label severity, and represent it with the Pearson correlation coefficient r as shown in Table 2:

[0072] Table 2 Pearson correlation coefficient between structural features and collision severity label

[0073] r <![CDATA[DeltaV1]]> <![CDATA[DeltaV2]]> <![CDATA[DeltaVx1]]> <![CDATA[DeltaVx2]]> <![CDATA[DeltaVy1]]> <![CDATA[DeltaVy2]]> Relative-V Relative_θ severity 0.877 0.695 0.872 0.696 0.125 0.097 0.401 -0.237

[0074] As can be seen from Table 2, there are features in the structural features that are correlated with the collision severity label. Select the structural features that are correlated with the collision severity label as the characteristic features, and that's it.

[0075] In this embodiment, as a preferred solution, in order to reduce the complexity of the model and find more suitable characterization features for representing the severity of the accident, the structural features are screened to obtain characterization features that are correlated with the collision severity labels. It further includes: calculating the importance of the structural features through the Gini coefficient, and selecting the structural features with high importance; analyzing the correlation of the structural features through the Pearson correlation coefficient, and deleting redundant features; screening out the characterization features from the structural features according to the results of the importance analysis and the correlation analysis. The specific method for this step is described in the previous patent application with the application number 2022101760035 of this unit, and will not be elaborated here. For the collision accident samples obtained in this embodiment, after screening the structural features, the characterization features of the collision accident severity are finally determined to be DeltaV1 and Relative-V. If the obtained collision accident sample is a rear-end collision, the method for constructing its features is the same as that of this embodiment, that is, the obtained structural features are the same as those of this embodiment. However, when screening out the characterization features, due to the slightly different sorting of the importance of the structural features, the finally obtained characterization features may be different from those of the collision accident samples of the head-on collision.

[0076] (3) Based on the characterization features, oversampling or undersampling is performed on the collision accident sub-samples to obtain balanced collision accident samples.

[0077] In the field of accident safety, there is a common and real phenomenon: collision events with a high severity level are relatively rare, while minor collision events are relatively common. This results in an imbalance in the number of collision samples at different severity levels, which will affect the training of the subsequent collision severity prediction model and make the prediction result tend to the severity level with a larger number of samples. Therefore, it needs to be balanced before use.

[0078] In this embodiment, as a preferred solution, it specifically includes: when the collision accident sub-sample is a majority class sample, deleting the in-class outliers in the collision accident sub-sample to achieve undersampling of the collision accident sub-sample; when the collision accident sub-sample is a minority class sample, oversampling the collision accident sub-sample based on the core seed clusters of the collision accident sub-sample. The further specific method is described in the previous patent application with the application number 2022101760035 of this unit, and will not be elaborated here. Of course, other processing methods in the prior art can also be used to balance the collision accident samples.

[0079] (4) Based on the balanced collision accident samples, using the characterization features as the input and the collision accident severity as the output, a collision severity prediction model is established.

[0080] In this embodiment, as a preferred solution, training samples are obtained based on the balanced collision accident samples. Using the characterization features as the input and the severity of the collision accident as the output, a collision severity prediction model is established based on the XGBoost algorithm; and the grid search algorithm is used to adjust the parameters of the collision severity prediction model, and the performance of the collision severity prediction model is tested based on one or more of the indicators of accuracy, recall, precision, and f1_score. Specifically as follows:

[0081] The balanced collision accident samples are divided into training samples (train) and test samples (test). The input uses the characterization feature group of the accident severity selected above: {Deltav1, relative_v}, and the output is the severity of the collision accident: {severity1, severity2, severity3}. In python, the initial XGBoost multi-classification model is constructed using the XGBoost toolkit with default parameters. To improve the model effect, the grid search (GridSearchCV) algorithm is used to adjust the hyperparameters of the model.

[0082] Based on the common classification indicators accuracy, recall, precision, and f1_score, the performance of the model is tested. For the balanced collision accident samples obtained in this embodiment, a collision severity prediction model is established based on the XGBoost algorithm, and the performance of the model is shown in Table 3:

[0083] Table 3 Model Performance

[0084] accuracy precision recall f1_score Model performance 0.9068 0.9078 0.9074 0.9074

[0085] For the prediction of the severity of collision accidents, high requirements are placed on accuracy and timeliness. As can be seen from Table 3, the performance of the prediction model obtained based on the XGBoost algorithm is excellent.

[0086] S2. Based on the collision severity prediction model, the severity of potential collision accidents is obtained according to the driving basic data, which specifically includes the following steps:

[0087] (1) Select the basic features from the driving basic data.

[0088] Referring to the basic characteristics of the collision accident sample, data corresponding to the basic characteristics are screened from the driving basic data as the basic characteristics of the first vehicle and the second vehicle, specifically including: the absolute speed v1' of the first vehicle, the absolute speed v2' of the second vehicle, the longitudinal speed vx1' of the first vehicle, the longitudinal speed vx2' of the second vehicle, the lateral speed vy2' of the first vehicle, the lateral vehicle speed vy2' of the second vehicle, the mass m1' of the first vehicle, the mass m2' of the second vehicle, the heading angle θ1' of the first vehicle, and the heading angle θ2' of the second vehicle.

[0089] (2) Based on the screened basic characteristics, feature construction is performed to obtain the characterizing features.

[0090] When predicting the severity of a head-on collision based on the collision severity prediction model established in this embodiment, the characterizing features required to be constructed include the relative speed relative_v' = abs(v1' - v2') between the first vehicle and the second vehicle. If it is predicted that the first vehicle is the active vehicle, the characterizing feature that also needs to be constructed is the speed change amount of the first vehicle before and after the collision. If it is predicted that the second vehicle is the active vehicle, the characterizing feature that also needs to be constructed is the speed change amount of the second vehicle before and after the collision. Among them, α' is the approaching angle between the first vehicle and the second vehicle, and α' = θ1' + θ2'.

[0091] (3) Input the obtained characterizing features into the collision severity prediction model, and the severity of the potential collision accident can be obtained.

[0092] S3. According to the severity of the obtained potential collision accident, assign a value to the first index characterizing the driving risk, specifically:

[0093] According to the severity of the predicted potential collision accident, assign a value to the first index. The more severe the severity of the predicted potential collision accident, the greater the value assigned to the first index characterizing the driving risk. For the convenience of calculation, according to the severity of the potential collision accident from low to high, the first index is sequentially assigned values of 1, 2, 3...

[0094] For the collision severity prediction model of this embodiment, the severities of the collision accidents that can be predicted are respectively three categories: severity1, severity2, and severity3. Then the corresponding first index is sequentially assigned values of 1, 2, and 3.

[0095] Step three. According to the first index, the second index, and the third index, obtain the driving risk index R, specifically R = k·S·U.

[0096] Step 4. As a preferred solution of this embodiment, it further includes obtaining the change rate of the driving risk index according to the driving risk index, specifically: where t is a certain instantaneous moment, △t is a period of time after this instantaneous moment t, R(t) is the driving risk index corresponding to this instantaneous moment t, and R(t + △t) is the driving risk index corresponding to the instantaneous moment (t + △t), and K R is the change rate of the driving risk index within [t, t + △t].

[0097] This embodiment also provides a driving risk grading method, including the following steps:

[0098] Step 1. Obtain the original driving samples, where the original driving samples include time series before the occurrence of a number of collision events, and the time series include the driving basic data of the first vehicle and the second vehicle at all times.

[0099] The original driving samples can be data samples collected during the driving process by oneself, or obtained by using third-party websites or databases, etc. Those skilled in the art should know that the first vehicle and the second vehicle mentioned here are the vehicles involved in the same collision event that collided.

[0100] As a preferred solution of this embodiment, the source of the original driving samples is traffic accident cases after 2015 published by the National Highway Traffic Safety Administration (NTHSA) of the United States, and time series of the near-collision process of 52 collision events are selected.

[0101] Step 2. According to the driving basic data at the moment to be evaluated, obtain the first index, the second index, and the third index characterizing the driving risk at this moment.

[0102] Specifically, it includes: according to the driving basic data at the moment to be evaluated, obtain the severity of the potential collision accident at this moment to be evaluated, and according to the severity of the potential collision accident, determine the first index characterizing the driving risk at this moment to be evaluated; according to the driving basic data at the moment to be evaluated, calculate the TTC at this moment to be evaluated, and according to the TTC, determine the first index characterizing the driving risk at this moment to be evaluated, where TTC is the estimated time for the first vehicle and the second vehicle to continue driving at the running state at this moment to be evaluated until the two vehicles collide; according to the driving basic data at the moment to be evaluated, calculate the THW at this moment to be evaluated, and according to the THW, determine the third index characterizing the driving risk at this moment to be evaluated, where THW is the estimated time for the first vehicle to continue driving at the running state at this moment to be evaluated until it reaches the position of the second vehicle at this moment.

[0103] The methods for obtaining the first, second, and third indicators characterizing the driving risk at the moment to be evaluated can be referred to the above-mentioned driving risk quantification method, which will not be elaborated here.

[0104] Step 3: Obtain the driving risk index at the moment to be evaluated according to the first, second, and third indicators. The specific calculation method can be referred to the above-mentioned driving risk quantification method, which will not be elaborated here.

[0105] Step 4: Cluster the driving risk indexes at each moment to be evaluated, and obtain the driving risk level at each moment to be evaluated according to the clustering result; based on the driving risk level, the first driving sample with the driving risk level label can be obtained.

[0106] As a preferred solution in this embodiment, when performing risk classification, the three-stage warning threshold of the AEB system is referred to. Therefore, a preset threshold of TTC is also set in this embodiment, and the preset threshold is preferably 2.6 s. Specifically: determine whether the TTC at the moment to be evaluated is greater than the preset threshold. When the TTC is greater than the preset threshold, it is determined that the driving risk level at this moment to be evaluated is safe without risk; when the TTC is not greater than the preset threshold, the driving risk index at this moment to be evaluated is put into the R sample library for clustering, and the driving risk level at this moment to be evaluated is obtained according to the clustering result.

[0107] For this embodiment, for the time series with TTC ≤ 2.6 s, the unsupervised clustering K-means technology is used to cluster the driving risk indexes, and the clustering result is as Figure 2 shown. The driving risk level can be divided into: low risk (LR1), medium risk (LR2), and high risk (LR3). The horizontal lines used to indicate each driving risk level in the figure are only approximate boundaries for easy observation and explanation, and the actually obtained clustering results are not neat boundaries.

[0108] The driving risk level at each moment to be evaluated is the driving risk level of the sample point corresponding to each moment to be evaluated in the original driving sample. The new sample obtained after screening out these sample points is the first driving sample with the driving risk level label, that is, each sample point in the first driving sample has a corresponding driving risk level.

[0109] Step 5: Obtain the change rate of the driving risk index in the time period where the moment to be evaluated is located according to the driving risk index; perform unsupervised classification on the change rates of the driving risk indexes in the time periods where each moment to be evaluated is located, and obtain the driving risk change trend in the time period where each moment to be evaluated is located according to the clustering result; based on the driving risk change trend, the second driving sample with the driving risk change trend label can be obtained.

[0110] For the specific calculation method of the change rate of the driving risk index, refer to the above driving risk quantification method, which will not be elaborated here.

[0111] As a preferred solution in this embodiment, in order to facilitate subsequent model construction for predicting driving risks, the moment to be evaluated is the initial moment of its corresponding time period. For example, if the moment to be evaluated is t1, then the time period where this moment to be evaluated is located is [t1, t1 + △t]. At the same time, in order to obtain a more detailed risk change rate and also to make the most of the original driving samples as much as possible for subsequent third clustering, when dividing the time periods in this embodiment, the width of each time period is 10 sample points and the step size is 1 sample point, that is, each time period includes 10 sample points and the interval between adjacent time periods is 1 sample point.

[0112] For this embodiment, using the unsupervised clustering K - means technique, after clustering the change rate of the driving risk index, the clustering results are as Figure 3 shown. The driving risk change trend can be divided into: the risk index decreases (LK R 1), the risk index is stable (LK R 2), and the risk index increases (LK R 3). The horizontal lines used to indicate each driving risk change trend in the figure are only approximate boundaries for easy observation and explanation, and the actual clustering results are not neat boundaries.

[0113] The driving risk change trend of each time period where the moment to be evaluated is located is the driving risk change trend of the sample point sequence corresponding to this time period in the original driving samples. The new samples obtained after screening out these sample point sequences are the second driving samples with driving risk change trend labels, that is, each sample point sequence in the second driving samples has a corresponding driving risk change trend.

[0114] Step Six: Cluster {LR, LK R}, and based on the clustering results, obtain the driving risk patterns of each moment to be evaluated and its corresponding time period, where LR is the driving risk level of a certain moment to be evaluated, and LK R is the driving risk change trend of the time period where this moment to be evaluated is located; based on the driving risk patterns, the third driving samples with driving risk pattern labels can be obtained.

[0115] As a preferred solution in this embodiment, the moment to be evaluated is the initial moment of its corresponding time period. For example, if the moment to be evaluated is t1, then the time period where this moment to be evaluated is located is [t1, t1 + △t], that is, the LR at the t1 moment corresponds to the LK R of the time period [t1, t1 + △t].

[0116] For this embodiment, the unsupervised clustering K-means technique is adopted to cluster {LR, LK R}, and the clustering result is as shown in Figure 4 . The driving risk modes can be divided into: safe and risk-free (R-L1), low risk (R-L2), medium risk (R-L3), medium risk with a rapid increase (R-L4), and high risk with a rapid increase (R-L5). The horizontal lines used to indicate each driving risk mode in the figure are only approximate boundaries for easy observation and explanation, and the actual clustering results are not neat boundaries.

[0117] The driving risk mode of each moment to be evaluated and its corresponding time period is the driving risk mode of the sample point sequence corresponding to that moment and its time period in the original driving samples. The new samples obtained after screening out these sample point sequences are the third driving samples with driving risk mode labels, that is, each sample point sequence in the third driving samples has a corresponding driving risk mode.

[0118] This embodiment also provides a driving risk prediction method, including the following steps:

[0119] Step 1: Use the trajectory prediction method to predict the driving trajectories of the host vehicle and other vehicles from the current moment to a future moment, so as to obtain the online driving data of the host vehicle and other vehicles from the current moment to a future moment.

[0120] Step 2: Based on the online driving data and the driving risk assessment model, predict the online driving risk from the current moment to a future moment.

[0121] Among them, the establishment of the driving risk assessment model includes the following steps:

[0122] 1. Adopt the above driving risk grading method to obtain the driving samples after driving risk grading.

[0123] Here, the driving samples after driving risk grading are one of the first driving samples, the second driving samples, and the third driving samples. As a preferred solution in this embodiment, the driving samples after driving risk grading are the third driving samples with driving risk modes.

[0124] 2. Use the driving samples after driving risk grading to construct a driving risk assessment model based on the HMM algorithm.

[0125] Apply the above driving risk prediction method to predict the driving risks in three simulation scenarios:

[0126] Simulation scenario 1: A collision occurs after a successful lane change, and its prediction result is as shown in Figure 5-6 ;

[0127] Simulation scenario 2: A collision occurs at the end of the lane change process, and its prediction results are as follows Figure 7-8 shown;

[0128] Simulation scenario 3: No collision occurs, and its prediction results are as follows Figure 9-10 shown;

[0129] Among them, Figure 5-10 the displayed is the driving risk of the vehicle that has not changed lanes (straight vehicle).

[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and 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 covered by the scope of the claims and the description of the present invention.

Claims

1. A driving risk quantification method, characterized in that Including the following steps: Obtain the driving basic data of the first vehicle and the second vehicle; Establish a prediction model for the severity of a collision accident, input the characterization features constructed according to the driving basic data into the prediction model for the severity of a collision accident, obtain the potential severity of a collision accident, where the potential severity of a collision accident is the collision severity label severity, and determine a first index characterizing the driving risk according to the potential collision accident severity label severity; Calculate the TTC according to the driving basic data, and determine a second index characterizing the driving risk according to the TTC, where the TTC is the estimated time for the first vehicle and the second vehicle to continue driving in their current operating states until the two vehicles collide; Calculate the THW according to the driving basic data, and determine a third index characterizing the driving risk according to the THW, where the THW is the estimated time for the first vehicle to continue driving in its current operating state until it reaches the current position of the second vehicle; Obtain a driving risk index according to the first index, the second index, and the third index; The establishment of the prediction model for the severity of a collision accident includes: Obtain collision accident samples, where the collision events in the collision accident samples carry collision severity labels, and the collision accident samples are divided into several collision accident sub-samples based on the collision severity labels; The basic characteristics of the collision accident samples include: the absolute speed of the involved vehicles, the longitudinal speed of the involved vehicles, the lateral speed of the involved vehicles, the mass of the involved vehicles, and the heading angle of the involved vehicles; Based on the basic characteristics of the collision accident samples, perform feature construction to obtain constructed features, and screen the constructed features to obtain characterization features that are correlated with the collision severity; Based on the characterization features, perform oversampling or undersampling processing on the collision accident sub-samples to obtain balanced collision accident samples; Based on the balanced collision accident samples, establish a collision severity prediction model with the characterization features as the input and the collision accident severity as the output; Based on the collision severity prediction model, obtaining the potential severity of a collision accident according to the driving basic data includes: Select the basic characteristics from the driving basic data; Based on the selected basic characteristics, perform feature construction to obtain characterization features; Input the obtained characterization features into the collision severity prediction model to obtain the potential severity of a collision accident.

2. The driving risk quantification method according to claim 1, wherein In S1, the collision events involved in the collision accident samples are collision events of two vehicles colliding. Number the involved vehicles and denote them as i, where i = 1 or 2. Based on the basic characteristics of the collision accident samples, perform feature construction to obtain constructed features, specifically including: The change in speed of the vehicle i involved in the incident before and after the collision , Longitudinal vehicle speed change of the involved vehicle i before and after the collision , Lateral vehicle speed change of involved vehicle i before and after the collision , Relative speed of two vehicles , Collision relative angle , where v i is the absolute speed of the vehicle involved in the incident i, and v 3-i is the absolute speed of the vehicle 3 - i involved in the incident vx i is the longitudinal speed of the vehicle involved in the incident i, vx 3-i is the longitudinal speed of the vehicle 3 - i involved in the incident vy i is the lateral velocity of the vehicle i involved in the incident, vy 3-i is the lateral velocity of the vehicle 3 - i involved in the incident m i is the mass of the vehicle involved in the incident i, m 3-i is the mass of the vehicle 3 - i involved in the incident θ i is the heading angle of the involved vehicle i, θ 3-i is the heading angle of the involved vehicle 3 - i, α is the angle at which the two vehicles approach each other, .

3. According to the driving risk quantification method described in claim 1, characterized in that Determine a second indicator characterizing driving risk according to TTC, specifically as ; And / or, determine a third indicator characterizing driving risk according to THW, specifically ; And / or, obtain a driving risk index according to the first index, the second index, and the third index, specifically R = k·S·U; Wherein, R represents the driving risk index, S represents the first index, U represents the second index, and k represents the third index.

4. The driving risk quantification method according to claim 1, characterized in that, It further includes obtaining a change rate of the driving risk index according to the driving risk index, specifically: ， Among them, \(t\) is a certain instantaneous moment, \(\Delta t\) is the length of a period of time after this instantaneous moment \(t\), \(R(t)\) is the driving risk index corresponding to this instantaneous moment \(t\), and \(R(t + \Delta t)\) is the driving risk index corresponding to the instantaneous moment \((t + \Delta t)\), and \(K\) R is the change rate of the driving risk index within \([t, t + \Delta t]\).

5. A driving risk grading method, characterized in that, Including the following steps: Obtain the original driving samples, where the original driving samples include time series before several collision events occur, and the time series include the basic driving data of the first vehicle and the second vehicle at all times; Based on the characteristic features of the basic driving data at the moment to be evaluated, and based on the collision accident severity prediction model, obtain the potential collision accident severity at this moment to be evaluated. The potential collision accident severity is the collision severity label severity. According to the potential collision accident severity label severity, determine the first index characterizing the driving risk at this moment to be evaluated; The collision accident severity prediction model includes: Obtain collision accident samples, where the collision events in the collision accident samples are with collision severity labels, and the collision accident samples are divided into several collision accident sub-samples based on the collision severity labels; The basic characteristics of the collision accident samples include: the absolute speed of the involved vehicles, the longitudinal speed of the involved vehicles, the lateral speed of the involved vehicles, the mass of the involved vehicles, and the heading angle of the involved vehicles; Based on the basic characteristics of the collision accident samples, perform feature construction to obtain constructed features, and screen the constructed features to obtain the characteristic features that are correlated with the collision severity; Based on the characteristic features, perform oversampling or undersampling on the collision accident sub-samples to obtain balanced collision accident samples; Based on the balanced collision accident samples, use the characteristic features as input and the collision accident severity as output to establish a collision accident severity prediction model; Based on the collision accident severity prediction model, obtain the potential collision accident severity according to the basic driving data: Select the basic features from the basic driving data; Based on the selected basic features, perform feature construction to obtain characteristic features; Input the obtained characteristic features into the collision severity prediction model to obtain the potential collision accident severity; According to the obtained potential collision accident severity, assign a value to the first index characterizing the driving risk, and that's it; According to the basic driving data at the moment to be evaluated, calculate the TTC at this moment to be evaluated. According to the TTC, determine the first index characterizing the driving risk at this moment to be evaluated, where TTC is the estimated time for the first vehicle and the second vehicle to continue driving in the operating state at this moment to collide; According to the basic driving data at the moment to be evaluated, calculate the THW at this moment to be evaluated. According to the THW, determine the third index characterizing the driving risk at this moment to be evaluated, where THW is the estimated time for the first vehicle to continue driving in the operating state at this moment to reach the position of the second vehicle at this moment; According to the first index, the second index, and the third index, obtain the driving risk index at the moment to be evaluated; Cluster the driving risk indices at each moment to be evaluated, and obtain the driving risk levels at each moment to be evaluated according to the clustering results; Based on the driving risk levels, the first driving samples with driving risk level labels can be obtained.

6. The driving risk classification method according to claim 5, characterized in that Cluster the driving risk indices at each moment to be evaluated to obtain the driving risk levels at each moment to be evaluated, specifically: Judge whether the TTC at the moment to be evaluated is greater than a preset threshold. When the TTC is greater than the preset threshold, it is determined that the driving risk level at this moment to be evaluated is safe and risk-free. When the TTC is not greater than the preset threshold, put the driving risk index at this moment to be evaluated into the R sample library for clustering, and obtain the driving risk level at this moment to be evaluated according to the clustering result.

7. The driving risk classification method according to claim 5, characterized in that, It also includes the following steps: According to the driving risk index, obtain the change rate of the driving risk index in the time period where the moment to be evaluated is located, where the moment to be evaluated is the initial moment of its time period. Perform unsupervised classification on the change rates of the driving risk indices in the time periods where each moment to be evaluated is located, and obtain the driving risk change trend in the time periods where each moment to be evaluated is located according to the clustering result. Based on the driving risk change trend, a second driving sample with a driving risk change trend label can be obtained.

8. The driving risk grading method according to claim 7, wherein It also includes the following steps: Cluster {LR, LK R}, and obtain the driving risk patterns of each moment to be evaluated and the time periods they belong to according to the clustering results, where LR is the driving risk level at a certain moment to be evaluated, and LK R is the driving risk change trend during the time period when this moment to be evaluated is located; Based on the driving risk pattern, a third driving sample with a driving risk pattern label can be obtained.

9. A driving risk prediction method, characterized in that, It includes the following steps: Use the trajectory prediction method to predict the driving trajectories of the host vehicle and other vehicles from the current moment to a future moment to obtain the online driving data of the host vehicle and other vehicles from the current moment to a future moment. Based on the online driving data, predict the online driving risk from the current moment to a future moment based on the driving risk assessment model. Among them, the establishment of the driving risk assessment model includes the following steps: Adopt the method described in any one of claims 5-8 to obtain the driving samples after driving risk classification, and the driving samples after driving risk classification are one of the first driving sample, the second driving sample, and the third driving sample. Use the driving samples after driving risk classification to construct a driving risk assessment model based on the HMM algorithm.

Images