A high-risk driving scene identification method with global accelerated optimization search

Abstract

The present invention discloses a method for identifying high-risk driving scenarios with global accelerated optimization search, including: S1, dividing the global driving scenario into several scenario groups through a segmentation module; S2, establishing an optimization problem through an identification module, and collecting data for each scenario group, calculating the risk level of each scenario group, and identifying the risk scenario group according to the risk level value; S3, determining an exploration direction in the risk scenario group through an exploration module, and dividing the risk scenario group into several scenario subgroups along the exploration direction, and finding the high-risk scenarios therein, and then judging whether enough high-risk scenarios are found; when it is judged that enough high-risk scenarios are found, stop the high-risk scenario identification, and when it is judged that not enough high-risk scenarios are found, return to step S2. According to the present invention, it is possible to search the global driving scenario to ensure high coverage, accelerate the identification of high-risk scenarios, improve efficiency, and promote scenario-based autonomous driving testing.

Description

Technical Field

[0001] The present invention relates to the technical field of automobile safe driving, and in particular to a high-risk driving scene identification method with global accelerated optimization search. Background Art

[0002] Autonomous vehicles are able to significantly reduce the burden on drivers by performing tedious, driving-related tasks. Autonomous vehicles need to be able to operate safely in a variety of traffic flows, road environments, and weather conditions. Otherwise, risks such as accidents or collisions caused by autonomous vehicles may occur. In view of this, AVs must be thoroughly evaluated to ensure their driving capabilities before they can be operated on public roads.

[0003] Currently, only a small number of cars equipped with low-level (L0-L2) assisted driving functions can be driven on real public roads. For safety reasons, if a car equipped with high-level autonomous driving functions (L3-L5) is to be driven on the road, it must complete at least hundreds of millions of miles of testing and verification. It takes several years to test and verify the initial version of the autonomous driving prototype alone, not to mention that the evolution of autonomous driving cars to higher levels requires continuous iteration and updating, and each iteration requires the completion of the same huge amount of testing. Time-consuming and labor-intensive inefficient testing technology has greatly delayed the large-scale implementation of autonomous driving.

[0004] At present, in order to accelerate the testing of autonomous driving, scenario-based testing has attracted much attention. However, scenario-based testing also faces a large number of driving scenarios, and most of the scenarios are very common and pose no safety threats. Therefore, when only high-risk scenarios are selected for testing, the required time can be greatly reduced. However, the current high-risk driving scenario identification methods still have defects such as low efficiency and low coverage. Summary of the invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a high-risk driving scene identification method with global accelerated optimization search, which can search the global driving scene, ensure high coverage, accelerate the identification of high-risk scenes, improve efficiency, and promote scene-based autonomous driving testing. In order to achieve the above-mentioned purpose and other advantages of the present invention, a high-risk driving scene identification method with global accelerated optimization search is provided, comprising:

[0006] S1. Divide the global driving scene into several scene groups through the segmentation module;

[0007] S2. Establish an optimization problem through the identification module, collect data for each scenario group, calculate the risk level of each scenario group, and identify the risk scenario group according to the risk level value;

[0008] S3. An exploration direction is determined in the risk scenario group through the exploration module, and the risk scenario group is divided into several scenario subgroups along the exploration direction, and the high-risk scenarios are found, and then it is determined whether enough high-risk scenarios are found; when it is determined that enough high-risk scenarios are found, the high-risk scenario identification is stopped, and when it is determined that not enough high-risk scenarios are found, it returns to step S2.

[0009] Preferably, step S1 further comprises the following steps:

[0010] S11, discretizing the global driving scene into a scene set consisting of a single driving scene through a segmentation module;

[0011] S12, normalizing the discretized scene set;

[0012] S13: Divide the processed scene set into several scene groups of equal size.

[0013] Preferably, the global driving scene discretization is specifically as follows:

[0014]

[0015] in, is the initial state set of the global scene; when i∈[0,n], X i is a static scene element; when i∈[n+1,n+m], X i It is a dynamic scene element; For X i The domain of

[0016] is discretized into:

[0017]

[0018] Among them, α i is a constant representing the discretization resolution; X i，max For X i The maximum value of X i，min For X i The minimum value of .

[0019] Preferably, the scene set normalization is specifically as follows:

[0020]

[0021] in, is a normalized global scene set; is the normalized initial state set of the global scene, specifically:

[0022]

[0023] in, is a normalized initial state of the scene; is a normalized scene element; is the normalized domain of a scene; g is and The normalized mapping between is one-to-one, specifically:

[0024]

[0025] Preferably, the scene set segmentation is specifically as follows:

[0026]

[0027] in, is a scene group; k max is the number of scene groups, Defined as:

[0028]

[0029] in, is the normalized initial state set of the scene group, specifically:

[0030]

[0031] in, is a normalized initial state in the normalized scene group; p is the index of a selected element; the domain of the pth scene element is divided into k max share, The kth slice cut from the domain of the pth scene element is:

[0032]

[0033] in, is the pth normalized scene element.

[0034] Preferably, the optimization problem in step S2 is to search for candidate high-risk scenarios, where the candidate high-risk scenarios have the highest risk level, and the search is performed in untested scenarios. Specifically, the optimization problem is:

[0035]

[0036]

[0037] in, The initial state of a scenario that has not been tested yet; A set of initial states for scenarios that have not yet been tested.

[0038] Preferably, data collection specifically includes:

[0039]

[0040] Among them, S (k) For the scene group A scene sampled from ; is the index set of the scene group that has not been sampled, specifically:

[0041]

[0042] in, It is the index set of all scene groups; is the index set of the sampled scene groups;

[0043] For S (k) The initial state is:

[0044]

[0045] in, is the i-th element X i The value of is calculated as:

[0046]

[0047] Among them, g -1 is the inverse function of the normalized mapping; is the value of the i-th normalized factor, specifically:

[0048]

[0049] in, For scene groups The domain of the i-th normalized element of ; for The median value of .

[0050] Preferably, the risk level of the scenario group is quantified by two indicators, including the approximate risk level and the expected risk level. The approximate risk level measures the risk level of a scenario group and is calculated using the collected data, specifically:

[0051]

[0052] in, For scene groups The approximate risk level, C(S (k) )for The sampled S(k) the degree of risk;

[0053] The expected critical level is obtained based on the approximate risk level, predicting the risk level of the unexplored area in a scene group. The expected risk level is defined as the upper limit of the risk level of the unexplored area, specifically:

[0054]

[0055] in, For scene groups The i-th normalized element of .

[0056] Preferably, the risk scenario group identification is specifically as follows:

[0057]

[0058] in, is a corresponding Lipschitz constant risk scenario groups; is the collection of all scene groups; is the feasible domain of the risk scenario group, specifically:

[0059]

[0060] Among them, C max is the highest risk value in the tested scenario; ∈ is a constant, and ∈>0, ∈ is used to ensure that the search will not fall into the local optimum; C max Specifically:

[0061] From the above formula, we can see that as the constant Changes in risk scenarios is also changing, in order to identify all risk scenario groups, The value from a set is enumerated, and the enumeration set With a set of key scenarios Related, so The combination of defines a convex hull so that all scenario groups are included in the convex hull. By constructing the convex hull, all risk scenario groups can be found.

[0062] Preferably, the exploration direction is determined as follows:

[0063]

[0064] in, For risk scenarios Medium element dimension The exploration value is as follows:

[0065]

[0066] in, For risk scenarios The i-th normalized element in The maximum value of is the minimum value;

[0067] The scene subgroup segmentation is as follows:

[0068]

[0069] in, is a scene subgroup; R+1 is the number of scene subgroups, and R is an even number.

[0070] Compared with the prior art, the present invention has the following beneficial effects: by dividing the global driving scene into several scene groups to simplify the scene identification work, and by establishing an optimization problem, the goal is to find high-risk scenes, and at the same time, based on the risk level of each scene group, identify the risk scene group therein, and further explore the risk scene group to find the high-risk scene therein, the method can search the global driving scene, ensure high coverage, and accelerate the identification of high-risk scenes, thereby improving efficiency. By setting a segmentation module, an identification module and an exploration module, the segmentation module is used to divide the global driving scene into scene groups, the identification module is used to identify and find the risk scene group in the scene group, and the exploration module is used to explore the risk scene group to find the high-risk scene, thereby in the global driving scene, by establishing an optimization problem for the search target, the high-risk scene is accurately identified, the accelerated identification of high-risk driving scenes is achieved, and the scene-based autonomous driving test is promoted. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A schematic diagram of the process of the high-risk driving scene identification method of the global accelerated optimized search according to the present invention. DETAILED DESCRIPTION

[0072] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0073] Reference Figure 1A high-risk driving scene identification method with global accelerated optimization search includes: a segmentation module, an identification module connected to the segmentation module signal, and an exploration module connected to the identification module signal. The entire high-risk driving scene identification method constructs an optimization problem and finds high-risk scenes by solving it, thereby searching the global driving scene, ensuring high coverage, and accelerating the identification of high-risk scenes to improve efficiency. The specific scene identification method includes the following steps:

[0074] S1. Divide the global driving scene into several scene groups through the segmentation module; the segmentation module divides the global driving scene, and the segmentation module can be divided into three sub-steps. First, in the first sub-step, the segmentation module discretizes the global driving scene into a scene set consisting of a single driving scene. Specifically, the global driving scene is discretized into:

[0075]

[0076] in, is the initial state set of the global scene; when i∈[0,n], X i is a static scene element; when i∈[n+1,n+m], X i It is a dynamic scene element; For X i The domain of definition.

[0077] can be discretized into:

[0078]

[0079] Among them, α i is a constant representing the discretization resolution; X i，max For X i The maximum value of X i，min For X i The minimum value of .

[0080] In the second sub-step, the segmentation module normalizes the discretized scene set. Specifically, the scene set is normalized as follows:

[0081]

[0082] in, is a normalized global scene set; is the normalized initial state set of the global scene, specifically:

[0083]

[0084] in, is a normalized initial state of the scene; is a normalized scene element; is the normalized domain of a scene; g is and The normalized mapping between is one-to-one, specifically:

[0085]

[0086] In the third sub-step, the segmentation module segments the standardized scene set into a number of scene groups of equal size. Specifically, the scene set is segmented into:

[0087]

[0088] in, is a scene group; k max is the number of scene groups, Specifically:

[0089]

[0090] in, is the normalized initial state set of the scene group, which can be written as:

[0091]

[0092] in, is a normalized initial state in the normalized scene group; p is the index of a selected element; the domain of the pth scene element is divided into k max share, The kth slice cut from the domain of the pth scene element is:

[0093]

[0094] in, is the pth normalized scene element.

[0095] S2. Establish an optimization problem through the identification module, collect data for each scenario group, calculate the risk level of each scenario group, and identify the risk scenario group according to the risk level value; the optimization problem is to search for candidate high-risk scenarios, which are scenarios with the highest risk level. The search is performed in untested scenarios. The optimization problem is:

[0096]

[0097]

[0098] in, The initial state of a scenario that has not been tested yet; A set of initial states for scenarios that have not yet been tested.

[0099] In the second sub-step, the identification module collects data from the scene group. The collected data is to help assess the risk level of the scene group. Specifically, the data is collected as follows:

[0100]

[0101] Among them, S (k) For the scene group A scene sampled from ; is the index set of the scene group that has not been sampled, specifically:

[0102]

[0103] in, It is the index set of all scene groups; It is the index set of the sampled scene groups.

[0104] For S (k) The initial state is:

[0105]

[0106] in, is the i-th element X i The value of is calculated as:

[0107]

[0108] Among them, g -1 is the inverse function of the normalized mapping; is the value of the i-th normalized factor, specifically:

[0109]

[0110] in, For scene groups The domain of the i-th normalized element of ; for The median value of .

[0111] In the third sub-step, the identification module will quantify the risk level of a scene group based on the collected information. Specifically, the risk level of a scene group is quantified by two indicators, including the approximate risk level and the expected risk level. The approximate risk level measures the risk level of a scene group and is calculated using the collected data as:

[0112]

[0113] in, For scene groups The approximate risk level, C(S (k) )for The sampled S (k) degree of risk.

[0114] The expected critical level is obtained based on the approximate risk level, predicting the risk level of the unexplored area in a scenario group. The expected risk level is defined as the upper limit of the risk level of the unexplored area, specifically:

[0115]

[0116] in, For scene groups The i-th normalized element of .

[0117] In the fourth sub-step, the identification module will determine which scenario groups are risk scenario groups based on the risk level values ​​of the scenario groups. These risk scenario groups may contain high-risk scenarios. Specifically, the risk scenario groups are identified as follows:

[0118]

[0119] in, is a corresponding Lipschitz constant risk scenario groups; is the collection of all scene groups; is the feasible domain of the risk scenario group, specifically:

[0120]

[0121] Among them, C max is the highest risk value in the tested scenario; ∈ is a constant, and ∈>0, ∈ is used to ensure that the search will not fall into the local optimum, C max Specifically:

[0122]

[0123] The above formula shows that with the constant Changes in risk scenarios In order to identify all risk scenario groups, The value should be from a set Enumeration set With a set of key scenarios Therefore, The combination of defines a convex hull so that all scenario groups are included in the convex hull. By constructing this convex hull, all risk scenario groups can be found.

[0124] S3. The exploration module determines an exploration direction in the risk scenario group, divides the risk scenario group into several scenario subgroups along the exploration direction, finds out the high-risk scenarios, and then determines whether enough high-risk scenarios are found; when it is determined that enough high-risk scenarios are found, the high-risk scenario identification is stopped; when it is determined that not enough high-risk scenarios are found, the step returns to step S2. This step includes four sub-steps: In the first sub-step, the exploration module determines an exploration direction in each risk scenario group. This exploration direction is the most valuable scenario element dimension worth exploring. Specifically, the exploration direction is determined as:

[0125]

[0126] in, For risk scenarios Medium element dimension The exploration value can be written as:

[0127]

[0128] in, For risk scenarios The i-th normalized element in The maximum value of is the minimum value.

[0129] In the second sub-step, the exploration module will divide each risk scenario group into several scenario sub-groups along the exploration direction. Specifically, the scenario sub-groups are divided into:

[0130]

[0131] in, is a scene subgroup; R+1 is the number of scene subgroups, and R is an even number.

[0132] In the third sub-step, the exploration module will search and identify high-risk scenarios in each scenario subgroup. Specifically, the sampling method of high-risk scenarios is consistent with the data collection method. If the risk level of the identified scenario is greater than the high-risk level threshold, the scenario is a high-risk scenario.

[0133] In the fourth sub-step, the exploration module will determine whether enough high-risk scenes are found. If so, the high-risk scene identification will be stopped. If not, the scene subgroup will be output to the identification module.

[0134] In summary, the present invention aims to find high-risk scenarios through optimization problems. At the same time, based on the risk level of each scenario group, it identifies the risk scenario groups, and further explores the risk scenario groups to find the high-risk scenarios. This method can search the entire driving scene, ensuring high coverage, while accelerating the identification of high-risk scenarios, improving efficiency, and promoting scenario-based autonomous driving testing.

[0135] The number of devices and processing scales described here are used to simplify the description of the present invention, and the application, modification and variation of the present invention will be obvious to those skilled in the art.

[0136] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and implementation modes. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A high-risk driving scene identification method with global accelerated optimization search, characterized in that: The following steps are involved: S1. Divide the global driving scene into several scene groups through the segmentation module; S2. Establish an optimization problem through the identification module, collect data for each scenario group, calculate the risk level of each scenario group, and identify the risk scenario group according to the risk level value; The optimization problem in step S2 is to search for candidate high-risk scenarios. The candidate high-risk scenarios have the scenarios with the highest risk level. The search is performed in untested scenarios. The optimization problem is specifically: ； in, The initial state of a scenario that has not been tested yet; is the initial state set of scenarios that have not been tested yet; S3, determining an exploration direction in the risk scenario group through the exploration module, dividing the risk scenario group into several scenario subgroups along the exploration direction, finding out the high-risk scenarios therein, and then judging whether enough high-risk scenarios are found; when it is judged that enough high-risk scenarios are found, stopping the high-risk scenario identification, and when it is judged that not enough high-risk scenarios are found, returning to step S2; The specific exploration directions are as follows: ； in, For risk scenarios Medium element dimension The exploration value is as follows: ； in, For risk scenarios Middle Normalized elements The maximum value of is the minimum value; The scene subgroup segmentation is as follows: ； in, is a scene subgroup; is the number of scene subgroups, and is an even number.

2. The high-risk driving scene identification method of global accelerated optimization search according to claim 1, characterized in that: Step S1 also includes the following steps: S11, discretizing the global driving scene into a scene set consisting of a single driving scene through a segmentation module; S12, normalizing the discretized scene set; S13: Divide the processed scene set into several scene groups of equal size.

3. The high-risk driving scene identification method of global accelerated optimization search according to claim 2, characterized in that: The global driving scenario is discrete as follows: ； in, is the initial state set of the global scene; when hour, is a static scene element; when , It is a dynamic scene element; for The domain of is discretized into: ； in, is a constant representing the discretization resolution; for The maximum value of for The minimum value of .

4. The high-risk driving scene identification method of global accelerated optimization search according to claim 3, characterized in that: The scene set normalization is as follows: ； in, is a normalized global scene set; is the normalized initial state set of the global scene, specifically: ； in, is a normalized initial state of the scene; is a normalized scene element; is the normalized domain of a scene; for and The normalized mapping between is one-to-one, specifically: 。 5. The high-risk driving scene identification method of global accelerated optimization search according to claim 4, characterized in that: The scene set segmentation is as follows: ； in, For a scene group; is the number of scene groups, Defined as: ； in, is the normalized initial state set of the scene group, specifically: ； in, is a normalized initial state in the normalized scene group; is the index of a selected feature; The definition domain of each scene element is divided into share, For the The definition domain of the scene elements is cut into Slices, specifically: ； in, For the A normalized scene element.

6. The high-risk driving scene identification method of global accelerated optimization search according to claim 1, characterized in that: The data collected are: ； in, For the scene group A scene sampled from ; is the index set of the scene group that has not been sampled, specifically: ； in, It is the index set of all scene groups; is the index set of the sampled scene groups; for The initial state is: ； in, For the Elements The value of , is calculated as: ； in, is the inverse function of the normalized mapping; For the The values ​​of the normalized factors are: ； in, For scene groups No. The domain of the normalized elements; for The median value of .

7. The high-risk driving scene identification method of global accelerated optimization search according to claim 6, characterized in that: The risk level of a scenario group is quantified by two indicators, including the approximate risk level and the expected risk level. The approximate risk level measures the risk level of a scenario group and is calculated using the collected data. Specifically: ； in, For scene groups The approximate risk level of for The sampled the degree of risk; The expected critical level is obtained based on the approximate risk level, predicting the risk level of the unexplored area in a scene group. The expected risk level is defined as the upper limit of the risk level of the unexplored area, specifically: ； in; is the Lipschitz constant and ; for The degree of unexplored, specifically: ； in, For scene groups No. A normalization factor.

8. The high-risk driving scene identification method of global accelerated optimization search according to claim 7, characterized in that: The specific identification of risk scenario groups is as follows: ； in, is a corresponding Lipschitz constant risk scenario groups; is the collection of all scene groups; is the feasible domain of the risk scenario group, specifically: ； in, It is the highest risk level value among the tested scenarios; is a constant, and , It is used to ensure that the search does not fall into a local optimum; Specifically: ； With the constant Changes in risk scenarios is also changing, in order to identify all risk scenario groups, The value from a set is enumerated, and the enumeration set With a set of key scenarios Related, so The combination of defines a convex hull so that all scenario groups are included in the convex hull. By constructing the convex hull, all risk scenario groups can be found.

Images