Automatic driving automobile logic scene parameter space full coverage test method

Abstract

The invention belongs to the technical field of self-driving automobile testing, and particularly relates to a self-driving automobile logic scene parameter space full-coverage testing method. Comprising the following steps: step 1, calculating a scene parameter space radiation range; step 2, carrying out logic scene parameter space full-coverage solution; and step 3, scene parameter combination compression optimization. According to the scheme, the automatic driving automobile logic test scene parameter space is comprehensively covered through a small number of test samples, and the test efficiency, the test comprehensiveness and the test fairness are effectively guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving vehicle testing, and specifically provides a method for fully covering the parameter space of logical scenarios of autonomous driving vehicles. Background Art

[0002] Scenario-based testing methods are the mainstream means of verifying the safety of autonomous driving systems. As the main body of scenario library construction, logical scenarios use parameter spaces to describe the parameters of scenario elements. Theoretically, an infinite number of scenarios can be sampled in a continuous space. However, how companies or third-party testing agencies can cover the logical scenario parameter space by sampling a small number of samples has not been effectively solved. Existing scholars have conducted some research on the test scenario sampling generation process, such as the method and system for generating autonomous driving real-vehicle test scenarios (Patent No. CN202111444379.1), a framework and method for generating autonomous driving test scenarios (Patent No. CN202411644409.7), and an accelerated autonomous driving test method considering efficiency and coverage (Patent No. CN202111398403.2). However, most existing methods aim to generate key scenarios that are likely to pose challenges to the autonomous driving system under test and cannot achieve comprehensive coverage in the logical scenario parameter space. Summary of the Invention

[0003] To solve the above problems, the present invention provides a full coverage testing method for the logical scenario parameter space of an autonomous driving vehicle. Through a small number of test samples, the logical test scenario parameter space of an autonomous driving vehicle is fully covered, effectively ensuring test efficiency, test comprehensiveness and test fairness.

[0004] The technical solution of the present invention is described as follows in conjunction with the accompanying drawings:

[0005] The present invention provides a method for full coverage testing of logical scenario parameter space of an autonomous driving vehicle, comprising the following steps:

[0006] Step 1: Calculate the spatial radiation range of scene parameters;

[0007] Step 2: Solve the full coverage of the logical scenario parameter space;

[0008] Step 3: Compression optimization of scene parameter combinations.

[0009] Furthermore, the specific method of step one is as follows:

[0010] Set the spatial radiation range calculation function of different scene parameters as shown in formula (1):

[0011]

[0012] Where ri (j) is the radiation range of the i-th specific scene in the j-th dimension of the parameter space; ξ i is the scene danger index of the i-th scene; A i is the scene probability index of the i-th specific scene, ξ i and A i Obtained through statistical analysis of relevant data in natural driving data; λ is the risk factor adjustment index; δ is the scaling constant of the Gaussian probability density function; ε is the probability adjustment constant; β is the global influence range control coefficient; ω j is the weight coefficient of the j-th dimension.

[0013] Furthermore, the specific method of step 2 is as follows:

[0014] 21) The logical scene parameter space is discretized to generate a set of n-dimensional specific scene parameter combinations, as shown in formula (2):

[0015]

[0016] Where x is the position grid point after the parameter space is discretized. When x becomes the parameter combination point to be calculated, the radiation range is obtained through the scene parameter radiation model. where r i =[r i (1) ,r i (2) ,…,r i (n) ], is the distance that can be covered in each dimension;

[0017] 22) The greedy algorithm is used to preferentially generate the location parameters of regions with high scene heat. The higher the scene heat, the greater the probability of generation. The calculation process of scene heat h is shown in formula (3):

[0018] h0(x i )=exp(-α·dist(x i ,Ψ0)) (3)

[0019]

[0020] Where x i is the uncovered parameter point in the parameter space; Ψ0 is the real driver capability boundary predefined by the test data; x i ' is the distance x on Ψ0 i The nearest parameter point, subscript i

[0021] The value range is 1, 2, 3...; α is a positive number that controls the heat decay rate of the boundary neighborhood. By changing the value, the distinction between the boundary position is enhanced or weakened. α = 1;

[0022] The H-GCO algorithm ensures that all position grid points are covered by at least one scene parameter combination, as shown in formula (5):

[0023]

[0024] Where, N(x,r i ) is centered at point x and has a radiation radius of r i The discrete neighborhood set of ;

[0025] y is the position grid point to be covered in the parameter space. If the distance between y and x in all dimensions does not exceed the corresponding radius threshold, it is considered to be covered by the parameter combination x. All the position grid points y together constitute the neighborhood N of x. j is the value of the center point x in the j dimension; y j To calculate the value of point y in dimension j;

[0026] In each round of greedy algorithm parameter combination selection, based on the current heat value h k (x i ), calculate the surrounding heat sum for all uncovered candidate specific scene parameters, as shown in formula (6):

[0027]

[0028] Where r est Used to calculate the range of heat and take r est =5; k is the round of the greedy algorithm; c(y)∈{0,1} is the coverage indicator function, which assigns 1 to the covered points in the parameter space; the algorithm selects the point with the largest gain as the scene generation point for this round, as shown in formula (7):

[0029]

[0030] After completing the parameter combination selection, use x * Create a test scenario and obtain the scenario risk results through the proxy model, and combine the probability density parameters of the corresponding points to obtain the representativeness of the real scenario of the test point r * , and update the parameter point coverage indicator state within the real coverage range accordingly, that is, let c * (y) = 1;

[0031] 23) Design a boundary correction method driven by a structural function. By constructing a continuous structural function based on the original boundary, combined with the newly obtained test points and real risk parameters in the iteration, the local difference judgment of the pre-given boundary function is performed and updated in real time.

[0032] Furthermore, the specific method of 23) is as follows:

[0033] 231) Based on the preset driver capability boundary, a symbolic distance function is constructed, as shown in formula (8):

[0034]

[0035] Where z is the sampling point on the capability boundary;

[0036] 232) Based on the distance function, the structure function is established as shown in formula (9):

[0037] B(x)=(1+exp(-σ·d(x))) -1 (9)

[0038] Where σ is the smoothing adjustment coefficient, which controls the degree of change of the structure function near the boundary, and is set to σ = 1; when d(x) = 0, that is, the parameter point is located on the capacity boundary, the structure function B(x) = 0.5 is the representation of the boundary function;

[0039] 233) Construct a risk-structure mapping function R(ξ) to map the observed real-scene hazard index ξ to the expected value in the structure space:

[0040] R(ξ)=exp(-τ(1-ξ)) (10)

[0041] Where τ is a positive number that controls the nonlinearity of the risk level mapping;

[0042] 234) In the k+1th iteration of the greedy algorithm, the parameter point x k+1 The structural error of is shown in formula (11):

[0043] △ k+1 =B k (x k+1 )-R(ξ k+1 ) (11)

[0044] Where B k (x k+1 ) is the structure value judgment of the parameter point by the current structure function; R(ξ k+1 ) is the expected structural value of the real-scene hazard index;

[0045] 235) Introduce x k+1 Gaussian kernel function centered at :

[0046]

[0047] Where γ is a positive number of bandwidth, and γ=1; K(x,x k+1 ) is the intensity of the impact of the test point correction on the point x on the capability boundary;

[0048] 236) The modified expression of the structure function is:

[0049] B k+1 (x) = B k (x)-λ·△ k+1 ·K(x,x k+1 ) (13)

[0050] Where λ is the correction step size parameter, which controls the adjustment amplitude of each iteration, and λ=0.5;

[0051] 237) The boundary function is corrected by updating the structure function, as shown in formula (14):

[0052]

[0053] Thus, the scene heat of each uncovered parameter point is calculated, and a new round of iteration of greedy search is completed until all parameter points in the logical scene parameter space are covered.

[0054] Furthermore, the specific method of step three is as follows:

[0055] 31) Genetic algorithm is introduced to eliminate redundant test points;

[0056] All scene parameter combinations obtained by the hierarchical greedy algorithm are used as initialization input:

[0057]

[0058] Where m is the number of initial scene parameter combinations;

[0059] Let the chromosome be:

[0060] c=(c1,c2,…,c m ) c i ∈{0,1} (16)

[0061] Where c is a feasible test point subset configuration. If c i =1, then keep the corresponding test point q i , if c i =0, the corresponding test point is removed, and the initial population is generated by randomly removing some points in Q0;

[0062] 32) Set the objective function as shown in formula (17):

[0063]

[0064] In the formula, the fitness function f(c) is the gene length of the individual, which represents the number of test points actually retained. In order to ensure that all parameter space points are covered, a feasibility constraint needs to be introduced, requiring that any parameter point x in the set G generated by the discretization of the complete parameter space of the test scenario j At least one retained test point q i The specific constraints are as follows:

[0065]

[0066] If a chromosome does not meet the coverage constraint, it is directly removed;

[0067] 33) According to the fitness value of each individual in the current population, individual selection is performed through roulette; according to the preset crossover probability p c , perform a single-point crossover operation between the selected parents, and take p c =0.6; for two parent chromosomes c (1) 、c (2) , perform gene segment exchange at random position k∈{1,…,m-1} to generate two offspring:

[0068] c' (1) =(c1 (1) ,…c t (1) ,c t+1 (2) ,…c m (2) ) (19)

[0069] c' (2) =(c1 (2) ,…c t (2) ,c t+1 (1) ,…c m (1) ) (20)

[0070] The offspring after crossover further undergo mutation operation; let the mutation probability be p m , take p m =0.01, for each site c i Independent mutations are made according to the following rules:

[0071]

[0072] The feasibility of the new individual is checked, and repair is performed if the full spatial coverage constraint is not met. The repair strategy is based on the greedy backfill principle: the minimum number of original points required from the uncovered grid points is selected to fill in the gaps so that the feasibility constraint is met.

[0073] Finally, the individual with the best current fitness is retained in each generation of the population. The process is terminated when the maximum number of iterations is reached or the optimal solution remains unchanged for a long time, and the compressed optimization point set c corresponding to the current optimal chromosome is output. * As the extracted set of final test parameter combinations.

[0074] The beneficial effects of the present invention are:

[0075] 1) The present invention sets a quantitative evaluation index of scenario representativeness and uses a dual optimization method to generate a test scenario parameter combination that can fully cover the entire parameter space of the logical scenario;

[0076] 2) Since the coverage, fairness, and efficiency of scenario generation are taken into account, the proposed method of the present invention has important engineering significance for third-party testing organizations to establish standardized test scenarios. It can help testing organizations establish standardized testing processes and promote the performance verification process of autonomous vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0078] Figure 1 It is a schematic diagram of the process of the present invention;

[0079] Figure 2 Schematic diagram of the process of step 2;

[0080] Figure 3 A schematic diagram of the preset driver's ability boundaries;

[0081] Figure 4 Generate a result diagram for the preceding vehicle cut-in scenario. DETAILED DESCRIPTION

[0082] The present invention will be further described in detail below with reference to the accompanying drawings and examples. It will be understood that the specific embodiments described herein are intended only to illustrate the present invention and are not intended to limit the present invention. It should also be noted that, for ease of description, the accompanying drawings only illustrate portions relevant to the present invention, not all structures.

[0083] Example 1

[0084] This embodiment provides a method for testing the full coverage of the logical scenario parameter space of an autonomous driving vehicle, including the following steps:

[0085] Step 1: Calculate the spatial radiation range of scene parameters, as follows:

[0086] Set the spatial radiation range calculation function of different scene parameters as shown in formula (1):

[0087]

[0088] Where r i (j) is the radiation range of the i-th specific scene in the j-th dimension of the parameter space; ξ i is the scene danger index of the i-th scene; A i is the scenario probability index of the i-th specific scenario, and the two are obtained by statistically analyzing the relevant data in natural driving data; λ is the risk factor adjustment index; δ is the scaling constant of the Gaussian probability density function; ε is the probability adjustment constant; β is the global influence range control coefficient; ω j is the weight coefficient of the j-th dimension;

[0089] Step 2: Refer to Figure 2 , the full coverage solution of the logical scenario parameter space is as follows:

[0090] This paper proposes a heat-based hierarchical greedy (H-GCO) algorithm to generate specific scenario combinations that can fully cover the logical scenario parameter space, as follows:

[0091] 21) In order to facilitate the calculation of parameter space coverage and reduce the computing power consumption during coverage calculation, the present invention discretizes the logical scene parameter space to generate a set of n-dimensional specific scene parameter combinations, as shown in formula (2):

[0092]

[0093] Where x is the position grid point after the parameter space is discretized. When x becomes the parameter combination point to be calculated, its radiation range is obtained through the scene parameter radiation model. where r i =[r i (1) ,r i (2) ,…,r i (n) ], is the distance that can be covered in each dimension;

[0094] 22) To guide the greedy algorithm in generating scene parameter combinations within the parameter space, this paper introduces the concept of scene heat. Scene heat refers to the distance between a specific scene and the actual driver's ability limit. Since the purpose of conventional testing is to discover the performance limit of the algorithm being tested, the greedy algorithm defined in this paper prioritizes generating parameters for locations with high scene heat. The calculation process of scene heat h is shown in formula (3):

[0095] h0(x i )=exp(-α·dist(x i ,Ψ0)) (3)

[0096]

[0097] Where x i is the uncovered parameter point in the parameter space, Ψ0 is the real driver ability boundary predefined by the test data, x i ' is the distance x on Ψ0 i The nearest parameter point, the subscript i takes values of 1, 2, 3...; α is a positive number that controls the heat decay rate of the boundary neighborhood. By changing its value, the distinction of the boundary position is enhanced or weakened. In this invention, α=1;

[0098] The H-GCO algorithm ensures that all position grid points are covered by at least one scene parameter combination, as shown in formula (5):

[0099]

[0100] Where, N(x,r i ) is centered at point x and has a radiation radius of r i The discrete neighborhood set of ;

[0101] y is the position grid point to be covered in the parameter space. If the distance between y and x in all dimensions does not exceed the corresponding radius threshold, then the point is considered to have been covered by the parameter combination x. All the position grid points y together constitute the neighborhood N of x. j is the value of the center point x in the j dimension; y j To calculate the value of point y in dimension j;

[0102] In each round of greedy algorithm parameter combination selection, based on the current heat value h k (x i ), calculate the heat sum of all uncovered candidate specific scene parameters around them, as shown in formula (6):

[0103]

[0104] Where r estFor the range used to calculate the heat sum, take r est =5; k is the round of the greedy algorithm; c(y)∈{0,1} is the coverage indicator function, which assigns 1 to the covered points in the parameter space; the algorithm selects the point with the largest gain as the scene generation point for this round, as shown in formula (7):

[0105]

[0106] After completing the parameter combination selection, use x * Create a test scenario and obtain the scenario risk results through the proxy model, and combine the probability density parameters of the corresponding points to obtain the representativeness of the real scenario at the test point r * , and accordingly update the parameter point coverage indicator state within its true coverage range, that is, let c * (y) = 1;

[0107] 23) As greedy search iterations and proxy model testing proceed, the actual collision results of the driver model under the current parameter combination can be obtained. As the data prior capability boundary gradually fails to accurately describe the collision uncertainty of the scene, it is necessary to fine-tune the boundary function in combination with the actual risk parameters of a small number of known test points. This paper proposes a structure function-driven boundary correction method. By constructing a continuous structure function based on the original boundary, combined with the newly acquired test points and their actual risk parameters during the iteration, the prior boundary function is judged for local differences and updated in real time. The specific method is as follows:

[0108] 231) Based on the preset driver capability boundary, a symbolic distance function is constructed, as shown in formula (8):

[0109]

[0110] Where z is the sampling point on the capability boundary;

[0111] 232) Based on the distance function, the structure function is established as shown in formula (9):

[0112] B(x)=(1+exp(-σ·d(x))) -1 (9)

[0113] Where σ is the smoothing adjustment coefficient, which controls the degree of change of the structure function near the boundary, and is set to σ = 1. When d(x) = 0, that is, the parameter point is on the capability boundary, the structure function B(x) = 0.5 represents the boundary function. When the parameter point gradually moves away from the boundary, the structure function B(x) smoothly approaches the high-risk and safe scenario parameter representations, respectively.

[0114] 233) Construct a risk-structure mapping function R(ξ) to map the observed real-scene hazard index ξ to the expected value in the structure space:

[0115] R(ξ)=exp(-τ(1-ξ)) (10)

[0116] Where τ is a positive number that controls the nonlinearity of the risk level mapping;

[0117] 234) In the k+1th iteration of the greedy algorithm, the parameter point x k+1 The structural error of is shown in formula (11):

[0118] △ k+1 =B k (x k+1 )-R(ξ k+1 ) (11)

[0119] Where B k (x k+1 ) is the structural value judgment of the parameter point by the current structural function; R(ξ k+1 ) is the expected structure value of the real-world hazard index, a positive value indicates a high structure value, and a negative value indicates a low structure value;

[0120] 235) In order to make the structure function locally correct in space, we introduce x k+1 Gaussian kernel function centered at :

[0121] K(x,x k+1 )=exp(-γ·||xx k+1 || 2 ) (12)

[0122] Where γ is a positive number representing bandwidth, and γ = 1; K(x,x k+1 ) is the intensity of the impact of the test point correction on the point x on the capability boundary;

[0123] 236) The modified expression of the structure function is:

[0124] B k+1 (x) = B k (x)-λ·△ k+1 ·K(x,x k+1 ) (13)

[0125] Where λ is the correction step size parameter, which controls the adjustment amplitude of each iteration, and λ=0.5;

[0126] The update of the boundary is reflected in the change of the isosurface of the structure function, that is, the boundary function is corrected by the update of the structure function, as shown in formula (14):

[0127]

[0128] Thus, the scene heat of each uncovered parameter point is calculated, and a new round of iteration of greedy search is completed until all parameter points in the logical scene parameter space are covered.

[0129] Step 3: Compression optimization of scene parameter combination. The specific method is as follows:

[0130] 31) A heat-driven hierarchical greedy algorithm is used to generate an initial set of scenario parameter combinations covering the logical scenario parameter space. While this algorithm satisfies the constraint that all regions within the space are effectively covered, the resulting test point set often contains some redundancy due to the inherently local optimality of the greedy strategy. To further reduce testing costs and optimize point set configuration, it is necessary to compress the point set without compromising coverage. To this end, the present invention introduces a genetic algorithm to eliminate redundant test points, resulting in a smaller but more comprehensive test solution.

[0131] The basic idea of the genetic algorithm is derived from natural selection and heredity, evolving a better solution through the "selection-crossover-mutation" operation between individuals. First, the combination of all scenario parameters obtained by the hierarchical greedy algorithm is used as the initial input;

[0132] All scene parameter combinations obtained by the hierarchical greedy algorithm are used as initialization input:

[0133]

[0134] Where m is the number of initial scene parameter combinations;

[0135] Let the chromosome be represented as:

[0136] c=(c1,c2,…,c m ) c i ∈{0,1} (16)

[0137] Where c represents a feasible test point subset configuration. If c i =1, then keep the corresponding test point q i , if c i =0, the corresponding test point is removed, and the initial population is generated by randomly removing some points in Q0;

[0138] 32) The optimization goal of the genetic algorithm is to reduce the number of retained points as much as possible. The objective function is set as shown in formula (17):

[0139]

[0140] In the formula, the fitness function f(c) is the "gene length" of the individual, which represents the number of test points actually retained. In order to ensure that all parameter space points are covered, a feasibility constraint needs to be introduced, requiring that any parameter point x in the set G generated by the discretization of the complete parameter space of the test scenario j At least one retained test point q i The specific constraints are expressed as:

[0141]

[0142] If a chromosome does not meet the coverage constraint, it is directly removed;

[0143] 33) According to the fitness value of each individual in the current population, individual selection is performed through roulette; according to the preset crossover probability p c , perform a single-point crossover operation between the selected parents, and take p c =0.6; for two parent chromosomes c (1) 、c (2) , perform gene segment exchange at random position k∈{1,…,m-1} to generate two offspring:

[0144] c' (1) =(c1 (1) ,…c t (1) ,c t+1 (2) ,…c m (2) ) (19)

[0145] c' (2) =(c1 (2) ,…c t (2) ,c t+1 (1) ,…c m (1) ) (20)

[0146] The offspring after crossover further undergo mutation operation; let the mutation probability be p m , take p m =0.01, for each site c i Independent mutations are made according to the following rules:

[0147]

[0148] Because crossover and mutation can disrupt coverage integrity, the new individuals are then tested for feasibility and repaired if they do not meet the full spatial coverage constraint. The repair strategy is based on the greedy backfilling principle: the minimum number of original points required to fill in the uncovered grid points is selected to meet the feasibility constraint.

[0149] Finally, the individual with the best current fitness is retained in each generation of the population to ensure the monotonic improvement of the solution. The algorithm terminates when the maximum number of iterations is reached or the optimal solution remains unchanged for a long time, and outputs the compressed optimization point set c corresponding to the current optimal chromosome. * As the extracted set of final test parameter combinations.

[0150] Example 2

[0151] This embodiment selects the front vehicle cutting-in scenario as a test case for experimental analysis. The scenario parameter space is selected as the relative speed Δv∈[-10,10](m / s) and the relative distance Δdis∈[5,75](m). Based on the test results under known approximate scenarios, the driver's ability boundary is preset as follows: Figure 3 shown.

[0152] The scene compression optimization extraction method proposed in this invention is used to select and extract specific scenes in the parameter space. The weight coefficients of relative speed and relative distance are set as 0.6 and 0.4 respectively in the process of obtaining the influence radius. The coverage effect of the method is then presented through the distribution of the neighborhood of all extracted points in the complete parameter space. Figure 4 shown.

[0153] The red dots in the figure represent specific scenario parameter combinations generated using the present invention. It can be seen that the present method can effectively achieve full coverage of the logical scenario parameter space. The extracted parameter points are clearly representative in the parameter space, with a significant advantage in sparse testing, with less repeated coverage areas (dark blue), effectively demonstrating the effectiveness of the present method in terms of test coverage.

[0154] The final number of test scenarios extracted was 141, saving over 91% of the testing cost compared to a full-space traversal test (1,491 times), demonstrating the effectiveness of the proposed method in terms of testing efficiency and cost. Furthermore, since the proposed method can output standardized scenario parameter combinations for the same logical scenario, it effectively ensures test fairness.

[0155] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for full coverage testing of the logical scenario parameter space of an autonomous driving vehicle, characterized in that: The following steps are involved: Step 1: Calculate the spatial radiation range of scene parameters; Step 2: Solve the full coverage of the logical scenario parameter space; Step 3: Compression optimization of scene parameter combinations.

2. The method for full coverage testing of the logical scenario parameter space of an autonomous driving vehicle according to claim 1 is characterized in that: The specific method of step one is as follows: Set the spatial radiation range calculation function of different scene parameters as shown in formula (1): Where r i (j) is the radiation range of the i-th specific scene in the j-th dimension of the parameter space; ξ i is the scene danger index of the i-th scene; A i is the scene probability index of the i-th specific scene, ξ i and A i Obtained through statistical analysis of relevant data in natural driving data; λ is the risk factor adjustment index; δ is the scaling constant of the Gaussian probability density function; ε is the probability adjustment constant; β is the global influence range control coefficient; ω j is the weight coefficient of the j-th dimension.

3. The method for full coverage testing of the logical scenario parameter space of an autonomous driving vehicle according to claim 1 is characterized in that: The specific method of step 2 is as follows: 21) The logical scene parameter space is discretized to generate a set of n-dimensional specific scene parameter combinations, as shown in formula (2): Where x is the position grid point after the parameter space is discretized. When x becomes the parameter combination point to be calculated, the radiation range is obtained through the scene parameter radiation model. where r i =[r i (1) ,r i (2) ,…,r i (n) ], is the distance that can be covered in each dimension; 22) The greedy algorithm is used to preferentially generate the location parameters of regions with high scene heat. The higher the scene heat, the greater the probability of generation. The calculation process of scene heat h is shown in formula (3): h0(x i )=exp(-α·dist(x i ,Ψ0)) (3) Where x i is the uncovered parameter point in the parameter space; Ψ0 is the real driver capability boundary predefined by the test data; x i ' is the distance x on Ψ0 i The nearest parameter point, the subscript i takes the value of 1, 2, 3...; α is a positive number that controls the heat decay rate of the boundary neighborhood. By changing the value, the distinction of the boundary position is enhanced or weakened, and α = 1; The H-GCO algorithm ensures that all position grid points are covered by at least one scene parameter combination, as shown in formula (5): Where, N(x,r i ) is centered at point x and has a radiation radius of r i The discrete neighborhood set of ; y is the position grid point to be covered in the parameter space. If the distance between y and x in all dimensions does not exceed the corresponding radius threshold, it is considered to be covered by the parameter combination x. All the position grid points y together constitute the neighborhood N of x. j is the value of the center point x in the j dimension; y j To calculate the value of point y in dimension j; In each round of greedy algorithm parameter combination selection, based on the current heat value h k (x i ), calculate the surrounding heat sum for all uncovered candidate specific scene parameters, as shown in formula (6): Where r est Used to calculate the range of heat and take r est =5; k is the round of the greedy algorithm; c(y)∈{0,1} is the coverage indicator function, which assigns 1 to the covered points in the parameter space; the algorithm selects the point with the largest gain as the scene generation point for this round, as shown in formula (7): After completing the parameter combination selection, use x * Create a test scenario and obtain the scenario risk results through the proxy model, and combine the probability density parameters of the corresponding points to obtain the representativeness of the real scenario of the test point r * , and update the parameter point coverage indicator state within the real coverage range accordingly, that is, let c * (y) = 1; 23) Design a boundary correction method driven by a structural function. By constructing a continuous structural function based on the original boundary, combined with the newly obtained test points and real risk parameters in the iteration, the local difference judgment of the pre-given boundary function is performed and updated in real time.

4. The method for full coverage testing of the logical scenario parameter space of an autonomous driving vehicle according to claim 2 is characterized in that: The specific method of 23) is as follows: 231) Based on the preset driver capability boundary, a symbolic distance function is constructed, as shown in formula (8): Where z is the sampling point on the capability boundary; 232) Based on the distance function, the structure function is established as shown in formula (9): B(x)=(1+exp(-σ·d(x))) -1 (9) Where σ is the smoothing adjustment coefficient, which controls the degree of change of the structure function near the boundary, and is set to σ = 1; when d(x) = 0, that is, the parameter point is located on the capacity boundary, the structure function B(x) = 0.5 is the representation of the boundary function; 233) Construct a risk-structure mapping function R(ξ) to map the observed real-scene hazard index ξ to the expected value in the structure space: R(ξ)=exp(-τ(1-ξ)) (10) Where τ is a positive number that controls the nonlinearity of the risk level mapping; 234) In the k+1th iteration of the greedy algorithm, the parameter point x k+1 The structural error of is shown in formula (11): △ k+1 =B k (x k+1 )-R(ξ k+1 ) (11) Where B k (x k+1 ) is the structure value judgment of the parameter point by the current structure function; R(ξ k+1 ) is the expected structural value of the real-scene hazard index; 235) Introduce x k+1 Gaussian kernel function centered at : K(x,x k+1 )=exp(-γ·||x-x k+1 || 2 ) (12) Where γ is a positive number of bandwidth, and γ=1; K(x,x k+1 ) is the intensity of the impact of the test point correction on the point x on the capability boundary; 236) The modified expression of the structure function is: B k+1 (x)=B k (x)-λ·△ k+1 ·K(x,x k+1 ) (13) Where λ is the correction step size parameter, which controls the adjustment amplitude of each iteration, and λ=0.5; 237) The boundary function is corrected by updating the structure function, as shown in formula (14): Thus, the scene heat of each uncovered parameter point is calculated, and a new round of iteration of greedy search is completed until all parameter points in the logical scene parameter space are covered.

5. The method for full coverage testing of the logical scenario parameter space of an autonomous driving vehicle according to claim 1 is characterized in that: The specific method of step three is as follows: 31) Genetic algorithm is introduced to eliminate redundant test points; All scene parameter combinations obtained by the hierarchical greedy algorithm are used as initialization input: Where m is the number of initial scene parameter combinations; Let the chromosome be: c=(c1,c2,…,c m ) c i ∈{0,1} (16) Where c is a feasible test point subset configuration. If c i =1, then keep the corresponding test point q i , if c i =0, the corresponding test point is removed, and the initial population is generated by randomly removing some points in Q0; 32) Set the objective function as shown in formula (17): In the formula, the fitness function f(c) is the gene length of the individual, which represents the number of test points actually retained. In order to ensure that all parameter space points are covered, a feasibility constraint needs to be introduced, requiring that any parameter point x in the set G generated by the discretization of the complete parameter space of the test scenario j At least one retained test point q i The specific constraints are as follows: If a chromosome does not meet the coverage constraint, it is directly removed; 33) According to the fitness value of each individual in the current population, individual selection is performed through roulette; according to the preset crossover probability p c , perform a single-point crossover operation between the selected parents, and take p c =0.6; for two parent chromosomes c (1) 、c (2) , at random position k∈{1,…,m - 1} Perform gene segment exchange to generate two offspring: c '(1) =(c1 (1) ,…c t (1) ,c t+1 (2) ,…c m (2) ) (19) c '(2) =(c1 (2) ,…c t (2) ,c t+1 (1) ,…c m (1) ) (20) The offspring after crossover further undergo mutation operation; let the mutation probability be p m , take p m =0.01, for each site c i Independent mutations are made according to the following rules: The feasibility of the new individual is checked, and repair is performed if the full spatial coverage constraint is not met. The repair strategy is based on the greedy backfill principle: the minimum number of original points required from the uncovered grid points is selected to fill in the gaps so that the feasibility constraint is met. Finally, the individual with the best current fitness is retained in each generation of the population. The process is terminated when the maximum number of iterations is reached or the optimal solution remains unchanged for a long time, and the compressed optimization point set c corresponding to the current optimal chromosome is output. * As the extracted set of final test parameter combinations.

Images