Method for determining driving ability recovery time in highway automatic driving takeover process

Abstract

The application discloses a method for determining driving ability recovery time in a highway automatic driving takeover process, and specifically comprises the following steps: constructing an automatic driving takeover experiment platform; designing and building a typical highway automatic driving takeover scene; recruiting drivers to conduct a takeover simulation experiment and collecting vehicle trajectory data in the experiment; extracting and screening indicators reflecting driving states; constructing a driving state recognition method based on a Gaussian mixture model; and combining the screened driving state representation indicators to analyze the driving state recognition and determine the driving ability recovery time of the drivers. The method can accurately determine the driving ability recovery time of each driver in the takeover process, and provides a theoretical basis for the optimized design of an automatic driving system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, in particular to a driving ability recovery time determination method for an automatic driving takeover process on a highway. BACKGROUND

[0002] Conditional automatic driving vehicles can independently monitor road traffic environments and perform lateral and longitudinal kinematic control within the designed operating domain, allowing drivers to participate in non-driving related secondary tasks. When encountering traffic situations that the system cannot handle, the system will prompt the driver to take over, and the driver needs to take over the vehicle within a specified time. When automatic driving, the driver participating in non-driving related secondary tasks will reduce his perception, decision-making and driving ability, and thus affect the driver's control state of the vehicle. After taking over, the driver's driving ability will continuously recover to a stable normal driving state with the increase of environmental information perception and continuous manual operation. The length of the driving ability recovery time will affect the driver's takeover performance, and understanding how long the driver needs to recover his driving ability to a normal driving state after taking over will help the optimization design of the automatic driving system and the human-machine interface, and improve the driver's takeover performance.

[0003] Current research mainly focuses on determining the length of time it takes for the driver to take over the vehicle control during the automatic driving takeover process and the influence of different factors on the takeover performance. Few studies focus on how long it takes for the driver to recover his driving ability to a normal state after taking over. Therefore, it is necessary to use driving simulation experiments to design automatic driving takeover simulation scenarios considering the influence of different factors, recruit drivers to participate in simulator experiments, collect vehicle trajectory data during the experiment, and propose a driving ability recovery time determination method for the automatic driving takeover process on a highway to determine the driving ability recovery time of each driver. SUMMARY

[0004] The present application aims to provide a driving ability recovery time determination method for an automatic driving takeover process on a highway, to provide a theoretical basis for the optimization design of the human-machine interface of the automatic driving system, and to improve the driver's takeover performance.

[0005] To achieve the above functions, the present application designs a driving ability recovery time determination method for an automatic driving takeover process on a highway. For a vehicle performing automatic driving on a highway, the takeover process from an automatic driving state to a driver driving state is performed as follows: step S1 to step S6 to determine the driving ability recovery time of the driver.

[0006] Step S1: Based on a driving simulator, an automatic driving takeover simulation platform is constructed, and a takeover simulation test scene is built considering the influence of different experimental factors.

[0007] Step S2: based on the built takeover simulation test scene, a takeover simulation experiment is performed for different drivers, the takeover simulation experiment includes a preset takeover trigger event, when the preset takeover trigger event occurs, the vehicle driven by the driver is switched from an automatic driving state to a driver driving state, and vehicle trajectory data of the vehicle driven by each driver in the takeover simulation experiment is collected;

[0008] Step S3: according to the collected vehicle trajectory data in the takeover simulation experiment, indexes reflecting the driving state are extracted;

[0009] Step S4: according to a preset rule, the indexes reflecting the driving state are combined, the effect of each index combination reflecting the driving state is evaluated and quantified, the effect of each index combination reflecting the driving state is sorted, and the index combination reflecting the best driving state effect is selected;

[0010] Step S5: a driving state recognition method based on a machine learning algorithm is constructed, based on the index combination reflecting the best driving state effect obtained in step S4, the driving state of the driver after taking over the vehicle is determined to be stable or unstable;

[0011] Step S6: the driving state recognition method is applied to analyze and determine the driving ability recovery time of each driver.

[0012] As a preferred technical solution of the present application: the influence of the preset different experimental factors considered in step S1 includes automatic driving duration, takeover request time, automatic driving speed, and front vehicle speed;

[0013] The automatic driving duration is the time during which the automatic driving system independently controls the vehicle to run, or the time during which the driver performs a secondary task during automatic driving, wherein the secondary task is a behavior that distracts the driver, the takeover request time is the early warning time of the vehicle issuing a takeover request, the automatic driving speed is the running speed of the vehicle during automatic driving, and the front vehicle speed is the running speed of the front vehicle when the vehicle issues a takeover request.

[0014] As a preferred technical solution of the present application: the preset takeover trigger event in step S2 includes a front vehicle lane changing and deceleration event, specifically, the vehicle in automatic driving travels at a preset speed on the middle lane, and the front vehicle traveling on the left lane changes lanes to the middle lane and travels at a deceleration.

[0015] As a preferred technical solution of the present application: the collected vehicle trajectory data in step S2 includes the running trajectory data of the vehicle and the front vehicle after the driver takes over the vehicle, the length of the running trajectory data segment is 60s, and the collection frequency is 100Hz.

[0016] As a preferred embodiment of the present invention: the indicators reflecting the driving state extracted in step S3 include average longitudinal speed x1, standard deviation of longitudinal speed x2, average longitudinal acceleration x3, average following distance x4, standard deviation of following distance x5, mean of the speed difference between the front and rear vehicles x6, standard deviation of the speed difference between the front and rear vehicles x7, average headway x8, time after takeover x9, and average lateral speed x1. 10 lateral velocity standard deviation x 11 Mean lateral acceleration x 12 Average lane deviation x 13 Standard deviation of lane deviation x 14 The time after takeover is x9, with the moment the driver takes over the vehicle as the 0th moment of the timekeeping.

[0017] As a preferred technical solution of the present invention: In step S3, a combination optimization method is used to combine the indicators reflecting the driving state, and the bulldozing distance is used to evaluate the effect of each combination on reflecting the driving state. The bulldozing distance is the minimum cost required to transform one distribution into another, and its calculation method is as follows:

[0018] assumed Let p be the first distribution with m classes, where p i Let P be the i-th category. For p i The corresponding weights, i∈{1,2,…,m}; assume Let be another distribution with n classes, where q j For the j-th category of Q, For q j The corresponding weights, j∈{1,2,…,n}; D=[d ij ] is the distance matrix, d ij For category p i and category q j The distance between them, the total cost l required to transform distribution P into distribution Q. cost As shown in the following formula:

[0019]

[0020] In the formula, f i,j For category p i and category q j Given the quantity of soil between distributions, the minimum cost required to transform distribution P into distribution Q is obtained under the following constraints:

[0021] f i,j ≥0, 1≤i≤m, 1≤j≤n

[0022]

[0023]

[0024]

[0025] Solving the above linearization problem, the optimal The bulldozing distance is as follows:

[0026]

[0027] In the formula, EMD(P, Q) is the bulldozing distance converted from the distribution P to the distribution Q.

[0028] As a preferred technical solution of the present application: the steps of the combination optimization method are as follows:

[0029] Step S31: From the collected vehicle trajectory data within 60s after the driver takes over the vehicle in the takeover simulation experiment, and from the extracted indicators reflecting the driving state, perform normalization conversion;

[0030] Step S32: Select K indicators in the indicators reflecting the driving state, the K indicators do not contain the time x9 after taking over, and use the permutation and combination principle to combine them to obtain The index combination and the corresponding vehicle trajectory data, the time x9 after taking over is added to each index combination, and the vehicle trajectory data is numbered, and the combination of the K indicators and the time x9 after taking over is used as reference data;

[0031] Step S33: Use the t-SNE algorithm to reduce the high-dimensional vehicle trajectory data corresponding to each index combination obtained in step S32 to two-dimensional data, calculate the bulldozing distance between the two-dimensional data corresponding to each index combination and the two-dimensional data corresponding to the reference data, and arrange the values of each bulldozing distance in descending order;

[0032] Step S34: Take the index combination corresponding to the minimum bulldozing distance as the index combination reflecting the best driving state.

[0033] As a preferred technical solution of the present application: the driving state recognition method in step S5 is based on a Gaussian mixture model, and its expression is as follows:

[0034]

[0035] λ={w i ,μ i ,Σ i},i=1,2,…,N

[0036]

[0037]

[0038] In the formula, p(x|λ) is a probability distribution of the Gaussian mixture model, λ is a parameter of the Gaussian mixture model, w i is a weight term of the i-th Gaussian distribution, x is a vector composed of D-dimensional continuous observation data, g(x|μ i ,Σ i ) is a probability density function of the i-th Gaussian distribution, μ i is a mean of the i-th Gaussian distribution, Σ i is a covariance matrix of the i-th Gaussian distribution, and N is the number of Gaussian distributions.

[0039] The posterior probability P(i|x t ,λ) of the i-th Gaussian distribution is as follows:

[0040]

[0041] In the formula, x t is a given T-dimensional training vector, x t ∈{x1,x2,...,x T}.

[0042] A preset posterior probability threshold is set, and the Gaussian distribution greater than or equal to the preset posterior probability threshold is determined as stable, and the Gaussian distribution less than the preset posterior probability threshold is determined as unstable.

[0043] As a preferred technical solution of the present application: in step S6, the hyperparameters in the Gaussian mixture model are set as follows: the number of partitions is 2, the maximum number of iterations is 1000, the covariance type is full covariance, and the k-means method is used to initialize the weight, mean and precision.

[0044] As a preferred technical solution of the present application: in step S6, the Gaussian mixture model is applied, and index combinations related to the horizontal and vertical directions are selected respectively, the driving state of the driver after taking over the vehicle is determined in the horizontal and vertical directions, the time for the driver to reach a stable state after taking over the vehicle in the horizontal and vertical directions is determined respectively, and the maximum value of the two is taken as the driving ability recovery time of the driver.

[0045] Advantages: Compared with the prior art, the present application has the following advantages:

[0046] 1. The present application is based on a driving simulator, an automatic driving takeover experimental platform is constructed, trajectory data in the takeover process is processed and analyzed, a large number of indexes reflecting the driving state are extracted and calculated, a combined optimization method is proposed to screen the indexes, a multi-index combination that best reflects the driving state is determined, the one-sidedness and vulnerability of using a single index to represent the driving state are made up, and the combined optimization method improves the objectivity and comprehensiveness of the evaluation index selection.

[0047] 2、The application uses a Gaussian mixture model in an unsupervised machine learning algorithm to identify and classify the driving state after takeover, improving the objectivity and reliability of the identification result;

[0048] 3、The driving ability recovery time determination method provided by the application can accurately determine the driving ability recovery time of each driver in different takeover scenarios, improving the calculation accuracy of the driving ability recovery time;

[0049] 4、The application reduces the risk and cost of real vehicle road experiment through simulation experiment, ensures the safety of the driver and improves the accuracy of the experimental result, and has good economic benefits. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a flowchart of a driving ability recovery time determination method for a highway automatic driving takeover process according to an embodiment of the application;

[0051] Figure 2 is a schematic diagram of an automatic driving takeover experiment platform according to an embodiment of the application;

[0052] Figure 3 is a flowchart of a takeover simulation experiment according to an embodiment of the application;

[0053] Figure 4 is a scene schematic diagram of a takeover triggering event according to an embodiment of the application;

[0054] Figure 5 (a)- Figure 5 (b) is the bulldozing distance between the data of different longitudinal index combinations and the reference data according to an embodiment of the application;

[0055] Figure 6 (a)- Figure 6 (l) is the longitudinal driving state classification result under different time intervals according to an embodiment of the application;

[0056] Figure 7 (a)- Figure 7 (l) is the lateral driving state classification result under different time intervals according to an embodiment of the application. DETAILED DESCRIPTION

[0057] The application will be further described below in conjunction with the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the application, and cannot be used to limit the protection scope of the application.

[0058] The driving ability recovery time determination method of the expressway automatic driving takeover process provided by the embodiment of the application is for a vehicle performing automatic driving on an expressway, a takeover process from an automatic driving state to a driver driving state, and refers to Figure 1 , performs the following steps S1-S6 to complete determination of the driving ability recovery time of the driver:

[0059] Step S1: An automatic driving takeover simulation platform is constructed based on a driving simulator, and refers to Figure 2 , the automatic driving takeover simulation platform includes a hardware input system, a data acquisition system, and a simulation scene rendering system, considers the influence of preset different experimental factors, uses the orthogonal experimental design principle, and uses a simulation software to build a takeover simulation test scene, and the vehicle controlled by the driver in the takeover simulation test scene is generated by a driving simulation software, such as a UC-win / Road simulation software;

[0060] The influence of the preset different experimental factors includes automatic driving duration, takeover request time, automatic driving speed, and front vehicle speed.

[0061] The automatic driving duration is the time during which the automatic driving system independently controls the vehicle to run, or the time during which the driver performs a secondary task (60s, 180s, 300s) during automatic driving, wherein the secondary task is an action that distracts the driver, such as playing a game; the takeover request time is the early warning time of the vehicle sending a takeover request (4s, 5s, 6s), the automatic driving speed is the running speed of the vehicle during automatic driving (80km / h, 85km / h, 90km / h), and the front vehicle speed is the running speed of the front vehicle when the vehicle sends a takeover request (55km / h, 60km / h, 65km / h).

[0062] Step S2: Based on the built takeover simulation test scene, a takeover simulation experiment is performed for different drivers, and refers to Figure 3 , the takeover simulation experiment includes a preset takeover trigger event, when the preset takeover trigger event occurs, the vehicle driven by the driver is switched from the automatic driving state to the driver driving state, and vehicle trajectory data of each driver driven vehicle in the takeover simulation experiment is collected;

[0063] Refers to Figure 4 , the preset takeover trigger event includes a front vehicle lane changing and decelerating event, specifically, the vehicle performing automatic driving travels on the middle lane at a preset speed, the front vehicle traveling on the left lane changes lanes to the middle lane and travels at a decelerated speed.

[0064] The collected vehicle trajectory data includes the running trajectory data of the vehicle and the front vehicle after the driver takes over the vehicle, the length of the running trajectory data segment is 60s, and the collection frequency is 100Hz.

[0065] Step S3: Based on the vehicle trajectory data collected in the takeover simulation experiment, extract various indicators reflecting the driving state; the calculation period for each extracted indicator reflecting the driving state is 1 second.

[0066] The extracted indicators reflecting driving status include average longitudinal velocity x1 (m / s), standard deviation of longitudinal velocity x2, and average longitudinal acceleration x3 (m / s²). 2 ), average following distance x4 (m), standard deviation of following distance x5, mean of speed difference between front and rear vehicles x6 (m / s), standard deviation of speed difference between front and rear vehicles x7, average headway x8 (s), time after takeover x9 (s), average lateral speed x 10 (m / s), standard deviation of lateral velocity x 11 Mean lateral acceleration x 12 (m / s 2 ), average lane deviation x 13 (m), Standard deviation of lane deviation x 14 The time after takeover is x9, with the moment the driver takes over the vehicle as the 0th moment of the timekeeping.

[0067] In step S3, a combinatorial optimization method is used to combine various indicators reflecting driving conditions. The bulldozing distance is used to evaluate the effectiveness of each combination in reflecting driving conditions. The bulldozing distance is the minimum cost required to transform one distribution into another; this cost is the product of the amount of "soil" directly moved and the moving distance. Its calculation method is as follows:

[0068] assumed Let p be the first distribution with m classes, where p i Let P be the i-th category. For p i The corresponding weights, i∈{1,2,…,m}; assume Let be another distribution with n classes, where q j For the j-th category of Q, For q j The corresponding weights, j∈{1,2,…,n}; D=[d ij ] is the distance matrix, d ij For category p i and category q j The distance between them, the total cost l required to transform distribution P into distribution Q. cost As shown in the following formula:

[0069]

[0070] In the formula, f i,j For category p i and category q jThe number of earth between the two, according to the following constraints to obtain the minimum cost required to convert the distribution P to the distribution Q:

[0071] f i,j ≥0,1≤i≤m,1≤j≤n

[0072]

[0073]

[0074]

[0075] Solving the above linear problem, the optimal The bulldozing distance is as follows:

[0076]

[0077] In the formula, EMD(P, Q) is the bulldozing distance from the distribution P to the distribution Q.

[0078] The steps of the combined optimization method are as follows:

[0079] Step S31: From the collected vehicle trajectory data within 60s after the driver takes over the vehicle in the takeover simulation experiment, and from which each index reflecting the driving state is extracted, and normalized conversion is performed;

[0080] Step S32: Select K indicators in each index reflecting the driving state, the K indicators do not contain the time x9 after the takeover, and use the permutation and combination principle to combine them to obtain The index combination and the corresponding vehicle trajectory data, the time x9 after the takeover is added to each index combination, and the vehicle trajectory data is numbered, and the combination of the K indicators and the time x9 after the takeover is taken as the reference data;

[0081] In one embodiment, the indicators x1, x2, x3, x4, x5, x6, x7, x8 reflecting the driving state are used to determine the longitudinal motion state of the vehicle, and then the index combination in 254 can be obtained And the vehicle trajectory data corresponding to different index combinations. Since the driving ability recovery time is related to time, the index time x9 after the takeover is added to the above 254 index combinations, and then the vehicle trajectory data is numbered. The indicators contained in the reference data are x1, x2, x3, x4, x5, x6, x7, x8, x9.

[0082] Step S33: The t-SNE (t-distributed stochastic neighbor embedding) algorithm is used to reduce the dimensionality of the high-dimensional vehicle trajectory data corresponding to the 255 sets of index combinations obtained in Step S32 to two-dimensional data. The bulldozing distances between the two-dimensional data corresponding to the 254 sets of index combinations and the two-dimensional data corresponding to the reference data are calculated respectively. The bulldozing distance values ​​are then sorted in descending order. The bulldozing distances between the data of different vertical index combinations and the reference data are referenced. Figure 5 (a)- Figure 5 (b);

[0083] Step S34: The combination of indicators corresponding to the minimum bulldozing distance is the best combination to reflect the driving state.

[0084] Step S4: Combine the indicators reflecting driving status according to the preset rules, evaluate and quantify the effect of each indicator combination on the driving status, rank the effect of each indicator combination on the driving status, and select the indicator combination that best reflects the driving status.

[0085] Depend on Figure 4 As can be seen, in this embodiment, the bulldozing distances of data numbered 253 and 254 are 0.056 and 0.054 respectively, which is very small. Data numbered 254 contains 7 variables: x1, x2, x3, x5, x6, x7, x9, while data numbered 253 contains 5 variables: x1, x2, x3, x6, x9, fewer than data numbered 254. Therefore, data numbered 253 was used for driving state identification research.

[0086] Step S5: Construct a driving state identification method based on machine learning algorithms. Based on the best combination of indicators that reflect the driving state obtained in step S4, determine whether the driving state after the driver takes over the vehicle is stable or unstable.

[0087] The driving state identification method in step S5 is based on a Gaussian mixture model, which is a probabilistic model and belongs to unsupervised machine learning classification algorithms. The expectation-maximization algorithm is typically used to solve the Gaussian mixture model. The Gaussian mixture model assumes that all data points are a mixture of a finite number of Gaussian distributions containing unknown parameters, and can be expressed as a weighted function of a finite number of Gaussian distributions, as shown in the following equation:

[0088]

[0089] λ={w i ,μ i ,Σ i}, i = 1, 2, ..., N

[0090]

[0091]

[0092] where p(x|λ) is the probability distribution of the Gaussian mixture model, λ is the parameter of the Gaussian mixture model, w i is the weight term of the i-th Gaussian distribution, x is a vector composed of D-dimensional continuous observation data, g(x|μ i ,Σ i ) is the probability density function of the i-th Gaussian distribution, μ i is the mean of the i-th Gaussian distribution, Σ i is the covariance matrix of the i-th Gaussian distribution, and N is the number of Gaussian distributions. In the present application, the driving state is divided into two categories: stable and unstable, so N is equal to 2.

[0093] The posterior probability P(i|x t ,λ) of the i-th Gaussian distribution is as follows:

[0094]

[0095] where x t is a given T-dimensional training vector, and x t ∈{x1,x2,...,x T}.

[0096] The posterior probability threshold is set to 0.5, and the Gaussian distribution greater than or equal to 0.5 is determined as stable, and the Gaussian distribution less than 0.5 is determined as unstable.

[0097] Step S6: applying the driving state recognition method, analyzing and determining the driving ability recovery time of each driver.

[0098] When applying the driving state recognition method, the hyperparameters in the Gaussian mixture model are set as follows: the number of components is 2, the maximum number of iterations is 1000, the covariance type is full covariance, and the k-means method is used to initialize the weight, mean and precision.

[0099] The Gaussian mixture model is applied, and the index combination related to the horizontal and vertical directions is selected respectively. The driving state of the driver after taking over the vehicle is determined in the horizontal and vertical directions, and the time for the driver to reach a stable state after taking over the vehicle in the horizontal and vertical directions is determined respectively. The maximum value of the two is taken as the driving ability recovery time of the driver.

[0100] The specific process is as follows:

[0101] Step S61: the time length of each vehicle trajectory data is 60s, and the time interval is 1s, i.e. each vehicle trajectory data has 60 trajectory points.

[0102] Step S62: Based on the method described in step S3, indicators reflecting the lateral and longitudinal driving state of the vehicle are extracted from each trajectory point, and based on the method described in step S4, the best indicator combination reflecting the lateral and longitudinal driving state is selected;

[0103] Step S63: Using the selected indicator combination corresponding to the vehicle trajectory data, combined with the Gaussian mixture model, it is identified whether the vehicle driving state is stable or unstable in the lateral and longitudinal directions every second;

[0104] Step S64: Based on the driving state identification result of step S63 and the vehicle trajectory data used for driving state identification, it is determined whether the vehicle driving state is stable or unstable every second.

[0105] Step S65: After the driver takes over the vehicle, the driving state is unstable, and then evolves from unstable to stable state. The moment when the unstable state ends is the moment when the stable state starts. This moment is the driving state stability time.

[0106] In this embodiment, longitudinal driving state identification analysis is first performed, and the identification result is as shown in Figure 6 (a)-6(l), Figure 6 The dark points in (a)-6(l) represent unstable state, and the light gray points represent stable state. It is obvious that Figure 6 (a) corresponding data distribution (distribution 1) and Figure 6 (j) corresponding data distribution (distribution 2) are two completely different distributions, and distribution 1 gradually transitions and evolves into distribution 2 over time, indicating that the use of Gaussian mixture model for driving state identification research is completely reasonable. In addition, the first 20s after the driver takes over the vehicle, the driving state is unstable (see Figure 6 (a) to 6(d)); 46 to 60s after taking over, the driving state is stable; 21 to 45s after taking over, the driving state transitions from unstable to stable state. According to the driving state identification result, combined with the original vehicle trajectory data, the longitudinal driving state stability time of each driver can be calculated. The longitudinal driving state stability time is subject to a normal distribution with a mean of 27.23s and a standard deviation of 3.67.

[0107] Subsequently, lateral driving state identification research is performed, and the variables considered include x 10 ,x 11 ,x 12 ,x 13 ,x 14 and x9, according to the above process, it is found that the data containing variables x9,x 11 ,x 13 ,x 14 and reference data (variables containing x9,x10 ,x 11 ,x 12 ,x 13 ,x 14 The distance between the two sides is the smallest, so it is used for lateral driving state identification analysis. The identification results are as follows: Figure 7 As shown in (a)-7(l), the driving state is unstable from 1 to 10 seconds after the driver takes over the vehicle, stable from 31 to 60 seconds, and transitions from unstable to stable between 11 and 30 seconds. Based on this driving state identification result, combined with the original vehicle trajectory data, the lateral driving state stabilization time for each driver can be calculated. The longitudinal stabilization time of the driving state follows a normal distribution with a mean of 17.37 seconds and a standard deviation of 3.13.

[0108] Finally, for the same driver, the maximum value between the lateral and longitudinal driving state stabilization times was selected as the driver's driving ability recovery time. By comparison, the driving ability recovery times of all drivers in this example can be obtained. Ultimately, the driving ability recovery times of all drivers follow a normal distribution with a mean of 27.25s and a standard deviation of 3.67.

[0109] This invention utilizes a driving simulator to build an autonomous driving takeover experimental platform. Considering the influence of various experimental factors, an autonomous driving takeover scenario was designed and constructed. Drivers were recruited to conduct simulated takeover experiments. Multiple indicators characterizing driving states were extracted and calculated based on vehicle trajectory data after takeover completion. A method for selecting the combination of indicators that best reflects driving states was proposed. After selecting the indicators, a driving state identification method based on machine learning algorithms was constructed. Based on the selected multi-indicator combination data that best reflects driving states, lateral and longitudinal driving state stability identification studies were conducted. Combining the driving state identification results with the original data, the lateral and longitudinal driving state stabilization times for each driver were determined. Finally, through comparison, the driving ability recovery times for all drivers were determined.

[0110] This invention proposes a combinatorial optimization method to screen indicators representing driving states, determining the most representative combination of multiple indicators. This overcomes the limitations and fragility of using a single indicator to represent driving states, and the combinatorial optimization method improves the objectivity and comprehensiveness of indicator selection. Furthermore, this invention constructs a Gaussian mixture model based on machine learning algorithms to identify and classify driving states after takeover, improving the objectivity and reliability of the identification results. The proposed method for determining driving ability recovery time can accurately calculate the recovery time for each driver under different takeover scenarios, improving the accuracy of recovery time calculation. The research results provide a theoretical basis for the optimized design of human-machine interfaces in autonomous driving systems, improving driver takeover performance and possessing significant reference value.

[0111] The embodiments of the present application are described in detail above with reference to the accompanying drawings, but the present application is not limited to the above-described embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present application.

Claims

1. A method for determining a driving ability recovery time for a highway automated takeover process, characterized in that, For the vehicle performing automatic driving on the highway, the takeover process from the automatic driving state to the driver driving state, the following steps S1-S6 are performed to determine the driving ability recovery time of the driver: Step S1: Based on the driving simulator, an automatic driving takeover simulation platform is constructed, considering the influence of different preset experimental factors, and a takeover simulation test scene is built; Step S2: Based on the built takeover simulation test scene, a takeover simulation experiment is performed for different drivers, including a preset takeover trigger event. When the preset takeover trigger event occurs, the vehicle driven by the driver is switched from the automatic driving state to the driver driving state. Vehicle trajectory data of each driver driving the vehicle in the takeover simulation experiment is collected; Step S3: According to the collected vehicle trajectory data in the takeover simulation experiment, indicators reflecting the driving state are extracted; In step S3, a combination optimization method is used to combine the indicators reflecting the driving state, and a bulldozing distance is used to evaluate the effect of each combination reflecting the driving state, wherein the bulldozing distance is the minimum cost required to convert one distribution to another distribution, and the calculation method is as follows: Assume is a first distribution with m classes, where p i is the i-th class of P, is p i the corresponding weight, i ∈ {1, 2, …, m}; assume is another distribution with n classes, where q j is the j-th class of Q, is q j the corresponding weight, j ∈ {1, 2, …, n}; D = [d ij ] is a distance matrix, d ij is the distance between class p i and class q j , and l cost is the total cost of converting distribution P to distribution Q as follows: where f i,j is the number of categories p i and q j between the soil and the minimum cost required to convert the distribution P to the distribution Q is obtained according to the following constraint: f i,j ≥ 0,1 < i < m, 1 < j < n Solving the above linearized problem, the optimal The bulldozing distance is given by the equation: In the formula, EMD(P, Q) is the bulldozing distance from distribution P to distribution Q; Step S4: According to the preset rule, the indicators reflecting the driving state are combined, the effect of each indicator combination reflecting the driving state is evaluated and quantified, the effect of each indicator combination reflecting the driving state is sorted, and the indicator combination reflecting the best driving state is selected; Step S5: A driving state recognition method based on machine learning algorithm is constructed, and based on the best indicator combination reflecting the driving state obtained in step S4, the driving state of the driver after taking over the vehicle is determined to be stable or unstable; The driving state recognition method in step S5 is based on a Gaussian mixture model, and its expression is as follows: λ = {w i , μ i , Σ i}, i = 1, 2, …, N where p(x|λ) is a probability distribution of the Gaussian mixture model, λ is a parameter of the Gaussian mixture model, w i is a weight term of the i-th Gaussian distribution, x is a vector composed of D-dimensional continuous observation data, g(x|μ i ,Σ i ) is a probability density function of the i-th Gaussian distribution, μ i is a mean of the i-th Gaussian distribution, Σ i is a covariance matrix of the i-th Gaussian distribution, and N is the number of Gaussian distributions. P(i | x, λ) = (1 / Z) * exp(-0.5 * (x - μί)Η(χ - μί)) (1) t where Z is the normalization constant. where x t is a given T-dimensional training vector, x t ∈ {x1,x2,...,x T} A preset posterior probability threshold is used to determine that the Gaussian distribution greater than or equal to the preset posterior probability threshold is stable, and the Gaussian distribution less than the preset posterior probability threshold is unstable; Step S6: Apply the driving state recognition method to analyze and determine the driving ability recovery time of each driver.

2. The method of claim 1, wherein The effects of different preset experimental factors considered in step S1 include automatic driving duration, takeover request time, automatic driving speed, and front vehicle speed; The automatic driving duration is the time during which the automatic driving system independently controls the vehicle, or the time during which the driver performs a secondary task while driving, wherein the secondary task is a behavior that distracts the driver. The takeover request time is the advance warning time for the vehicle to issue a takeover request. The automatic driving speed is the running speed of the vehicle during automatic driving. The front vehicle speed is the running speed of the vehicle in front of the vehicle when the vehicle issues a takeover request.

3. The method of claim 1, wherein the driving ability recovery time determination method of the highway automatic driving takeover process is characterized by, The preset takeover trigger event in step S2 includes a front vehicle lane changing and deceleration event, specifically, the vehicle driving in the middle lane at a preset speed, and the front vehicle driving on the left lane changes to the middle lane and decelerates.

4. The method of claim 1, wherein, The vehicle trajectory data collected in step S2 includes the operation trajectory data of the vehicle and the preceding vehicle after the driver takes over the vehicle, the length of the segment of the operation trajectory data is 60s, and the collection frequency is 100Hz.

5. The method of claim 4, wherein the driving ability recovery time determination method of the highway automatic driving takeover process is characterized by, The indexes reflecting the driving state extracted in step S3 include average longitudinal speed x1, standard deviation of longitudinal speed x2, average longitudinal acceleration x3, average following distance x4, standard deviation of following distance x5, average speed difference between front and rear vehicles x6, standard deviation of speed difference between front and rear vehicles x7, average headway x8, time after takeover x9, average lateral speed x 10 , standard deviation of lateral speed x 11 , average lateral acceleration x 12 , average lane offset x 13 , standard deviation of lane offset x 14 ; wherein the time after takeover x9 is timed from the time when the driver takes over the vehicle as the 0 time.

6. The method of claim 1, wherein, The steps of the combined optimization method are as follows: Step S31: vehicle trajectory data in the range of 60s after the driver takes over the vehicle in the collected takeover simulation experiment is obtained, and each index reflecting the driving state is extracted therefrom and normalized; Step S32: select K indicators from the indicators reflecting the driving state, the K indicators do not contain the time x9 after the takeover, and the K indicators are combined by using the permutation and combination principle to obtain The group index combination and the corresponding vehicle trajectory data, the time x9 after the takeover is added to each group index combination, and the vehicle trajectory data is numbered, and the combination of the K indicators and the time x9 after the takeover is taken as the reference data; Step S33: the high-dimensional vehicle trajectory data corresponding to each group of index combinations obtained in step S32 is reduced to two-dimensional data by using the t-SNE algorithm, the bulldozing distance between the two-dimensional data corresponding to each group of index combinations and the reference data is calculated, and the numerical values of the bulldozing distances are arranged in descending order; Step S34: the index combination corresponding to the minimum bulldozing distance is taken as the index combination reflecting the best driving state.

7. The method of claim 1, wherein, In step S6, the hyperparameters in the Gaussian mixture model are set as follows: the number of components is 2, the maximum number of iterations is 1000, the covariance type is full covariance, and the k-means method is used to initialize the weight, mean and precision.

8. The method of claim 7, wherein, In step S6, the Gaussian mixture model is applied, and the index combinations related to the horizontal and vertical directions are selected to determine the driving state of the driver after taking over the vehicle in the horizontal and vertical directions, respectively, and the maximum value of the time at which the driver reaches a stable state after taking over the vehicle in the horizontal and vertical directions is taken as the driving ability recovery time of the driver.

Images