An adaptive automatic driving dynamic scene general generation method for high and low dimensional evaluation scenes

Abstract

The application discloses a kind of high and low dimensional evaluation scene-oriented adaptive automatic driving dynamic scene general generation method, its method is: first, based on the evaluation scene element dimension optimization of chromatography analysis method;Second, evaluation scene space and key long tail function construction;Third, low-dimensional evaluation scene-oriented adaptive dynamic scene generation;Fourth, high-dimensional evaluation scene-oriented adaptive dynamic scene generation;Fifth, high and low dimensional evaluation scene-oriented adaptive dynamic scene evaluation;Beneficial effect: the hierarchical classification of scene elements is realized, thereby establishing the hierarchical model of scene elements. Select the scene element with larger importance weight value as the decision variable, thereby realizing the dimension optimization problem of scene element. A new ideal scene space construction method is established. The occurrence probability of scene in low-dimensional ideal scene space is solved. The search efficiency is greatly improved.

Description

Technical Field

[0001] This invention relates to a general method for generating dynamic scenes for autonomous driving, and particularly to a general method for generating adaptive dynamic scenes for autonomous driving oriented towards high- and low-dimensional evaluation scenarios. Background Technology

[0002] Currently, autonomous driving has become one of the mainstream automotive technology development directions, and autonomous vehicles represent the future trend. The prerequisite for autonomous vehicles to be allowed on public roads is that their driving safety has been fully verified and meets relevant standards. To fully test the safety of autonomous vehicles and the reliability of their intelligent algorithms, simulation testing, field testing, and road testing are necessary sequentially. Although these three testing methods differ in their stages, processes, and implementation, they all require the design of specific test scenarios to execute the testing process. Therefore, generating representative, simple adaptive autonomous driving dynamic scenarios for low-dimensional evaluation scenarios and complex adaptive autonomous driving dynamic scenarios for high-dimensional evaluation scenarios is a crucial issue. This will greatly improve the testing efficiency and reliability of autonomous vehicles, accelerating their deployment.

[0003] Current methods for generating dynamic scenarios for adaptive autonomous driving have two main shortcomings. First, they only consider the danger of the scenario as the sole criterion, focusing solely on generating scenarios that are dangerous for autonomous vehicles. While this principle can test the safety boundaries of autonomous vehicles, it fails to consider real-world road traffic conditions. Many highly dangerous scenarios have an extremely low probability of occurring on real roads, making such scenarios largely meaningless for testing autonomous vehicles. Second, current research on adaptive autonomous driving dynamic scenario generation is entirely geared towards low-dimensional evaluation scenarios. There is no comprehensive method for generating adaptive autonomous driving dynamic scenarios for high-dimensional evaluation scenarios. Research on this is essential because autonomous vehicles, for truly viable road use, cannot be limited to handling simple low-dimensional scenarios but must also be able to handle complex high-dimensional scenarios.

[0004] Chinese patent CN202210941004.4 discloses a method, device, equipment, and storage medium for generating autonomous driving test scenarios, which can generate autonomous driving test scenarios with high risk based on target traffic participant data; Chinese patent CN202210804060.3 discloses a method, device, vehicle, and storage medium for generating autonomous driving test scenarios, which can generate simulation test scenarios by loading the key parameter range of the simulation test scenario into the scenario design document and using a preset script; Chinese patent CN202210741420.X discloses an automated simulation test system and related equipment for intelligent driving, whose scenario generation module can create test scenarios based on the driving parameters of the main vehicle and generalize the test scenarios to generate one or more test scenarios for testing intelligent driving algorithms. However, none of the above three patents consider real road traffic conditions when generating scenarios, and they do not address the problem of adaptive dynamic scenario generation for high-dimensional evaluation scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a general method for generating adaptive dynamic scenarios for autonomous driving, which can accelerate the testing of autonomous vehicles, improve the testing efficiency and reliability of autonomous vehicles, and accelerate the deployment process of autonomous vehicles.

[0006] The present invention provides a general method for generating adaptive autonomous driving dynamic scenes for high- and low-dimensional evaluation scenarios, the method comprising the following steps:

[0007] The first step is to optimize the dimensions of the evaluation scenario elements based on the analytic hierarchy process. The specific process is as follows:

[0008] Step 1: Construct a scene element hierarchy model based on scene parameterization. Describe the scene as a collection of scene elements. According to the basic attributes of the scene elements and their geographical location conditions, classify the scene elements level by level to establish a scene element hierarchy model.

[0009] A scene is a general dynamic description of an autonomous vehicle and its driving environment over a period of time. Based on the basic attributes of scene elements, scene elements are divided into static elements and dynamic elements. According to the different geographical locations of scene elements, static elements and dynamic elements are further divided into off-road elements and on-road elements.

[0010] Static elements outside the road mainly include static roadside objects, such as buildings, trees and green belts, as well as traffic signs. Static elements inside the road mainly include lane-related elements and environment-related elements. Lane-related elements include lane type, lane alignment, number of lanes, lane width, lane slope, lane curvature, lane line type, lane line width, pavement type, and pavement adhesion coefficient. Environment-related elements include weather and light conditions.

[0011] The off-road elements of dynamic elements mainly include traffic participants, such as traffic participant type and number. The on-road elements of dynamic elements mainly include elements related to traffic participants, elements related to the initial state, and elements related to the driving state sequence. Elements related to traffic participants include traffic participant type and number; elements related to the initial state include initial position, initial lane, and initial speed; elements related to the driving state sequence include trigger mode, relative distance sequence, relative speed sequence, and relative acceleration sequence.

[0012] Step 2: Optimization of Scene Element Dimensions Based on Analytic Hierarchy Process (AHP). Based on the scene element hierarchy model, the influence propagation model is used to calculate the number of times scene elements at the same level propagate influence across different levels of the autonomous driving system. The difference in the number of times scene elements propagate influence within the same level is calculated. Based on the correspondence between the difference in the number of times influence propagation occurs and the 1-9 scaling method, the difference in the number of times influence propagation occurs is converted into a scaling value, thereby establishing a judgment matrix for scene elements at the same level. The largest eigenvalue of the judgment matrix and its corresponding eigenvector are solved. The eigenvector is normalized to obtain the importance weight value of each scene element at the same level relative to its current level. The rationality of the importance weight value is verified through a consistency check. Based on the scene element hierarchy model and the importance weight value of each scene element at each level relative to its current level, the importance weight value of each scene element among all scene elements is calculated. Based on the type of the target scene and the dimensional requirements of the evaluation scene to be established, scene elements with larger importance weight values ​​are selected as decision variables.

[0013] Using the scene element hierarchy model from step one as input, the influence propagation model is used to solve for the number of times the influence of scene elements at the same level is propagated across different levels of the autonomous driving system. The influence propagation model has three assumptions, as follows:

[0014] 1) The impact of scene elements on the autonomous driving system will be passed down level by level with the autonomous driving system, and this impact will not decrease with the passage of levels.

[0015] 2) Different types of scene elements have the same impact on the autonomous driving system;

[0016] 3) The magnitude of the influence of scene elements on the autonomous driving system is represented by the number of times the influence of scene elements is transmitted between the levels of the autonomous driving system. The more times the influence is transmitted, the greater the influence of the scene element, and the influence is linearly related to the number of transmissions.

[0017] The number of times the influence of scene elements in the influence propagation model propagates through each level of the autonomous driving system can be calculated using the following formula:

[0018] (1)

[0019] In the formula, P(n) represents the number of times the influence of a scene element is transmitted through each level of the autonomous driving system, n represents the number of element attributes of the scene element, and E i This represents the number of times the influence of the i-th element attribute in this scene is propagated across all levels of the autonomous driving system.

[0020] After obtaining the number of influence transmissions of scene elements at each level in the scene element hierarchy model based on the influence transmission model, the difference in the number of influence transmissions of scene elements at the same level is calculated. The difference in the number of influence transmissions is converted into a scale in the 1-9 scaling method. After converting the difference in the number of influence transmissions into a scale in the 1-9 scaling method, a judgment matrix for scene elements at the same level is established. The largest eigenvalue of the judgment matrix and its corresponding eigenvector are solved. The eigenvector is normalized to obtain the importance weight value of scene elements at the same level.

[0021] To verify the rationality of the importance weight values, a consistency check is performed on the judgment matrix. The formula for the consistency check is:

[0022] (2)

[0023] In the formula, CR represents the consistency index, and RI is the standard value of the hierarchical overall ranking average random consistency index, which takes different values ​​depending on the order of the judgment matrix.

[0024] CI stands for Overall Hierarchical Ranking Consistency Index. The formula for calculating CI is:

[0025] (3)

[0026] In the formula, λ max CR represents the largest eigenvalue of the judgment matrix, and n is the order of the judgment matrix. When CR < 0.1, it indicates that the judgment matrix has good consistency and meets the consistency requirements.

[0027] Finally, based on the hierarchical model of scene elements and the importance weight value of each scene element relative to its own level, the importance weight value of each scene element in all scene elements is calculated. Then, the type of the target scene to be studied is determined. Based on the type of the target scene and the dimensional requirements of the evaluation scene to be established, some scene elements with larger importance weight values ​​are selected as decision variables, thus completing the dimensional optimization of scene elements.

[0028] The second step, constructing the evaluation scenario space and key long-tail functions, is as follows:

[0029] Step 1: Constructing the Ideal Scenario Space Based on Scenario Element Dimension Optimization. The ideal test scenario space is obtained by discretizing the decision variables obtained from the scenario element dimension optimization in Step 1. Each decision variable is discretized into different values ​​according to its own element attributes. The range and dispersion of the decision variables are determined based on the importance weight value of the decision variable, while fully considering the constraints of real road conditions. Decision variables with higher importance weight values ​​are given a larger range and a smaller dispersion value to ensure that as many scenarios as possible can be included in the ideal test scenario space. For decision variables with lower importance weight values, the selection of their range and dispersion value follows the principle that key scenarios can be included in the ideal test scenario space.

[0030] Step 2: Constructing a real-world scene space based on natural driving data. Scene data is collected by vehicles driving on real roads. The collected sensor data is preprocessed, and target scene data is selected from the preprocessed scene data. Then, the data corresponding to the decision variables are selected from the target scene data to construct a real-world scene space.

[0031] Scene data collection should also follow these two principles:

[0032] 1) The scene data collection area should include cities and regions with different road and traffic characteristics, and the total collection mileage should be large enough;

[0033] 2) The scene data collection environment should include different weather types and lighting types;

[0034] The specific contents of sensor data preprocessing include: time alignment and spatial alignment of each sensor data; verification of the validity of sensor data; generation of vehicle bus alignment signal, vehicle status alignment signal and multimodal environment sensor alignment signal;

[0035] Based on the type of target scenario selected in the first step, the scene data of all target scenarios is extracted from all preprocessed data by manually watching the video. Each complete segment of target scenario data is called a scenario condition, which represents a complete target scenario event that occurs on a real road. The data corresponding to the decision variables in each scenario condition are filtered out to form the real scenario space.

[0036] Step 3: Constructing a critical long-tail function based on scene occurrence probability and scene hazard level. Following the adaptive dynamic scene generation principle of "challenging autonomous vehicles and having a certain probability of occurrence on real roads," a critical long-tail function is designed based on scene occurrence probability and scene hazard level as the basis for selecting critical adaptive dynamic scenes. The calculation formula for the critical long-tail function is as follows:

[0037] (4)

[0038] In the formula, I(x) represents the critical long-tail function value of the scenario, P(x) represents the probability of the scenario occurring, and V(x) represents the risk level of the scenario.

[0039] The third step is adaptive dynamic scene generation for low-dimensional evaluation scenarios, and the specific process is as follows:

[0040] Step 1: Solve the probability of scene occurrence based on the convex combination algorithm. The convex combination algorithm is used to transform the scene in the low-dimensional real scene space from the natural driving data into the scene in the low-dimensional ideal scene space, and then solve the probability of scene occurrence in the low-dimensional ideal scene space.

[0041] Suppose there exist vectors {x1, x2, x3...x} n If there is a real number λ i ≥ 0, i=1,2,3...n, and λ1+λ2+...+λ n =1, then λ1x1+λ2x2+...λ n x n Let the vector {x1, x2, x3...x} be... n A convex combination of};

[0042] The following assumptions are made when using the convex combination algorithm to solve for the probability of scenario occurrence:

[0043] 1) Within a certain distance range of the ideal scene space, the probability of the scene occurring will not change abruptly, but will change continuously and follow a linear change law;

[0044] 2) Within a certain distance range of the ideal scene space, the linear change law of the probability of scene occurrence can be represented by the Euclidean distance between scenes in the ideal scene space.

[0045] 3) Within a certain distance range of the ideal scene space, scenes that are closer to high-probability scenes in Euclidean distance have a higher probability of occurrence, while scenes that are farther from high-probability scenes in Euclidean distance have a lower probability of occurrence.

[0046] Taking a scene space within a walking distance as an example, A1(R) A1 ,∆v A1 ), A2(R A2 , ∆v A2 ), A3(R A3 , ∆v A3 ), A4(R A4 , ∆v A4 B(R) represents four uniformly discrete scenes in a two-dimensional ideal scene space. B , ∆v B ( ) is a scene in a real-world scene space obtained from natural driving data, and B is inside the rectangle formed by A1, A2, A3, and A4, L 11 L is the Euclidean distance from B to the lines containing A2 and A4. 12 L is the Euclidean distance from B to the lines containing A1 and A3. 21 L is the Euclidean distance from B to the lines containing A3 and A4. 22 Let B be the Euclidean distance from the lines containing A1 and A4. Then, scene B in the real scene space can be transformed into scenes A1, A2, A3, and A4 in the two-dimensional ideal scene space through a convex combination algorithm.

[0047] (5)

[0048] In the formula, ω1, ω2, ω3, and ω4 are the weighting coefficients of A1, A2, A3, and A4, respectively, and their calculation formula is as follows:

[0049] (6)

[0050] (7)

[0051] (8)

[0052] (9)

[0053] The above method is used to transform all scenes in the real scene space into uniformly discrete scenes in the ideal scene space. The number of uniformly discrete scenes is counted by combining their weight coefficients, which gives the probability of the scene occurring in the two-dimensional ideal scene space.

[0054] Taking a scene space within a walking distance as an example, A1(R) A1 , ∆v A1 , ∆aA1 ), A2(R A2 , ∆v A2 , ∆a A2 ), A3(R A3 , ∆v A3 , ∆a A3 ), A4(R A4 , ∆v A4 , ∆a A4 ), A5(R A5 , ∆v A5 , ∆a A5 ), A6(R A6 , ∆v A6 , ∆a A6 ), A7(R A7 , ∆v A7 , ∆a A7 ), A8(R A8 , ∆v A8 , ∆a A8 B(R) represents eight uniformly discrete scenes in a three-dimensional scene space. B , ∆v B , ∆a B ( ) is a scene in a real-world scene space obtained from natural driving data, and B is inside the cube formed by A1, A2, A3, A4, A5, A6, A7, and A8, L 11 L is the Euclidean distance from B to the plane containing A2, A4, A6, and A8. 12 L is the Euclidean distance from B to the plane containing A1, A3, A5, and A7. 21 L is the Euclidean distance from B to the plane containing A3, A4, A7, and A8. 22 L is the Euclidean distance from B to the plane containing A1, A2, A5, and A6. 31 L is the Euclidean distance from B to the plane containing A5, A6, A7, and A8. 32 If the distance from B to the plane containing A1, A2, A3, A4 is Euclidean distance, then the scene B in the real scene space can be transformed into scenes A1, A2, A3, A4, A5, A6, A7, A8 in the three-dimensional ideal scene space through the convex combination algorithm.

[0055] (10)

[0056] In the formula, ω1, ω2, ω3, ω4, ω5, ω6, ω7, and ω8 are the weight coefficients of A1, A2, A3, A4, A5, A6, A7, and A8, respectively, and their calculation formula is as follows:

[0057] (11)

[0058] (12)

[0059] (13)

[0060] (14)

[0061] (15)

[0062] (16)

[0063] (17)

[0064] (18)

[0065] The above method is used to transform all scenes in the real scene space into uniformly discrete scenes in the ideal scene space. The number of uniformly discrete scenes is counted by combining their weight coefficients, which gives the probability of scene occurrence in the three-dimensional ideal scene space.

[0066] Step 2: Solving the scene hazard boundary based on the support vector regression algorithm. The scene hazard boundary in the real scene space is pre-divided according to the collision time. A joint distribution of collision time and relative acceleration is constructed. Boundary points belonging to the same boundary line in the joint distribution are used as a training sample set and input into the support vector regression algorithm separately. Each training sample set corresponds to a scene hazard boundary. All scene hazard boundary lines are summarized to obtain the complete scene hazard boundary. Finally, the scene hazard boundary in the real scene space is mapped to the ideal scene space, and hazard coefficients are set for scenes with different hazard levels.

[0067] TTC represents the time required for a collision to occur between the two vehicles while maintaining their current states of motion. A smaller TTC value indicates a higher level of danger. TTC is calculated using the following formula:

[0068] (19)

[0069] In the formula, R represents the relative distance between the two vehicles, and ∆v represents the relative speed between the vehicle and the target vehicle;

[0070] The pre-defined criteria for defining the hazard boundary of a scenario are as follows: when TTC∈[0s,1s], the hazard level of the scenario is a collision scenario; when TTC∈(1s,3s], the hazard level of the scenario is an emergency scenario; when TTC∈(3s,5s], the hazard level of the scenario is a conflict scenario; when TTC∈(5s,+∞) or (-∞,0s), the hazard level of the scenario is a safe scenario.

[0071] A joint distribution of TTC and ∆a in the real scene space is constructed. This joint distribution is equivalent to the existence of four scene hazard boundaries defined by TTC=0, TTC=1, TTC=3, and TTC=5. The boundary points on both sides of the joint distribution are used as separate training sample sets and input into a support vector regression algorithm for learning. Since the boundaries of this joint distribution map are close to linear, a linear kernel function is chosen to solve for them, obtaining the scene hazard boundaries on both sides of the joint distribution. The scene hazard boundaries pre-defined based on TTC are merged with the scene hazard boundaries based on support vector regression to obtain the hazard boundaries in the real scene space:

[0072] (20)

[0073] In the formula, l l With l r These represent the scene hazard boundaries on the left and right sides of the joint distribution map obtained through support vector regression, Δa and Δa, respectively. l and Δa r These represent the distances between scene B and the scene hazard boundary, l m1 l m2 l m3 and l m4 These are the scene hazard boundaries obtained from the TTC pre-classification;

[0074] Map the above scenario hazard boundaries to the ideal scenario space, and set the scenario hazard level corresponding to the collision scenario to 1, the scenario hazard level corresponding to the emergency scenario to 0.7, the scenario hazard level corresponding to the conflict scenario to 0.3, and the scenario hazard level corresponding to the safe scenario to 0.

[0075] Step 3: Generate key adaptive dynamic scenes based on multi-starting point optimization algorithm and seed filling algorithm. Calculate the key long-tail function value in the ideal scene space by using the scene occurrence probability and scene hazard level. Input the key long-tail function, ideal scene space, and key threshold into the multi-starting point optimization algorithm to solve for local key adaptive dynamic scenes. Input the local key scenes, key long-tail function, ideal scene space, and key threshold into the seed filling algorithm to solve for all key adaptive dynamic scenes.

[0076] The critical long-tail function of the scene in the ideal scene space can be obtained by combining the scene occurrence probability obtained in step one with the scene danger level obtained in step two. The critical long-tail function, the ideal scene space, and the critical threshold γ are input into the multi-start point optimization algorithm. Some points are sampled in the ideal scene space by manual setting or random sampling as the starting point of the algorithm. The maximum value of the critical long-tail function in the attraction domain of each starting point is solved. The scene corresponding to the maximum value of all critical long-tail functions greater than γ is output, thus obtaining the local critical adaptive dynamic scene for low-dimensional evaluation scenarios.

[0077] The local key scenes, key long-tail functions, ideal scene space, and key threshold γ output by the multi-start point optimization algorithm are input into the seed filling algorithm. Starting from each local key scene, the algorithm searches for scenes in the ideal scene space whose key long-tail function values ​​are greater than γ in the neighborhood around the starting point. These scenes are then used as new starting points to continue searching their neighborhoods. The above steps are repeated until the key long-tail function values ​​of all neighborhood scenes around the starting point are less than γ. At this point, the algorithm ends and outputs all the scenes marked as starting points, thus generating all the key adaptive dynamic scenes for low-dimensional evaluation scenarios.

[0078] Step 4: Adaptive dynamic scene generation for high-dimensional evaluation scenarios, the specific process is as follows:

[0079] Step 1: Constructing a scene hazard identification model based on the Hammerstein identification process and solving the scene occurrence probability based on the convex combination algorithm. The Hammerstein identification process consists of a series of static nonlinear and dynamic linear components. The relative state variables of the vehicle and the target vehicle, as well as the acceleration of the target vehicle, are used as inputs to the scene hazard identification model, while the acceleration of the vehicle is used as the output. After training with a large amount of scene data, the model parameters of the scene hazard identification model are used as key parameters characterizing the intrinsic attributes of scene hazard. Principal component analysis is used to decouple and reduce the dimensionality of the key parameters. Ant colony clustering algorithm is used to cluster the scene hazards. Representative "state-action" pairs are selected from each scene hazard level and input into the scene hazard identification model. Based on the cluster category of the parameters, a mapping relationship between the scene hazard level and the clustering results is established. The scene occurrence probability solution based on the convex combination algorithm can directly solve the occurrence probability of the "state-action" pairs according to the method for solving the scene occurrence probability of low-dimensional evaluation scenarios.

[0080] Viewing the high-dimensional assessment scenario as a Markov decision process, the relative state variables of the vehicle and the target vehicle are considered as states, and the acceleration of the target vehicle is considered as an action. The state and action at the same time step are considered as a "state-action" pair. The relative state variables of the vehicle and the target vehicle at the same moment, as well as the acceleration of the target vehicle, are used as inputs to the scenario hazard identification model, and the acceleration of the vehicle is used as the output of the model. Therefore, the scenario hazard identification model is a multi-input single-output system. The Hammerstein identification process consists of a static nonlinear element and a dynamic linear element connected in series. The static nonlinear element uses a dead-zone function, a sigmoid function, or a saturation function. The z-transform of the dynamic linear element is shown in the following equation:

[0081] (twenty one)

[0082] In the formula, O p (k) represents the set of accelerations of the main vehicle, N(k) represents the set of outputs of the static nonlinear element, and d

[0083] A(z) represents the order of the input delay and is defined as an integer multiple of the sampling time. -1 ) and B(z) -1 It can be calculated using the following formula:

[0084] (twenty two)

[0085] In the formula, (a1,…a q ) and (b1,…b n ) are all coefficients of the dynamic linear element, and q and n are the orders of the dynamic linear element;

[0086] After training with a large amount of scenario data from natural driving data, the model parameters contained in the static nonlinear and dynamic linear links are key data that characterize the intrinsic attributes of scenario hazard. Therefore, they are used as data samples for scenario hazard assessment. In order to reduce the spatial dimension of the data samples as much as possible to improve computational efficiency while expressing the same model features, principal component analysis is used to decouple and reduce the dimensionality of the key parameters in the scenario hazard identification model.

[0087] Let H represent the parameter dimension of the scene hazard identification model, and E represent the amount of training data. Then the model parameter dataset X can be represented as:

[0088] (twenty three)

[0089] In the formula, x i The intrinsic parameter vector representing the scene hazard identification model;

[0090] Taking X as input, the principal component analysis algorithm defines the percentage of the sum of the eigenvalues ​​of the first m principal components to the sum of all eigenvalues ​​as the principal component contribution rate. The cumulative principal component contribution rate M is then calculated. m Calculated using the following formula:

[0091] (twenty four)

[0092] In the formula, λ i Represents the eigenvector;

[0093] To ensure the dimensionality reduction effect, we take M. m The value of m corresponding to ≥85% is used as the dimension of the independent parameters of the model calculated by the algorithm, and finally an m×E matrix L is obtained:

[0094] (25)

[0095] Ant colony clustering is used to cluster scene hazards. Taking L as the input to the ant colony clustering algorithm, scene hazard classification based on ant colony clustering is to find the partitioning method in L that minimizes the sum of distances from each data sample to the cluster centers with a known number of clusters. Referring to the scene hazard grading method for low-dimensional assessment scenes in step two of step three, the scene hazard is divided into 4 levels, resulting in 4 clusters. The ant colony clustering algorithm is expressed by the following formula:

[0096] (26)

[0097] In the formula, J represents the sum of the distances from each data sample to the four cluster centers, l ip c represents the p-th model parameter feature of the i-th data sample. jp The p-th model parameter feature for the j-th class center is calculated using the following formula:

[0098] (27)

[0099] In the formula, E j Let ω be the observed variable corresponding to the j-th class in observed variable E. ij The subordinate relationship indicator between the observed variable and the category is calculated using the following formula:

[0100] (28)

[0101] The classification quality of ant colony clustering can be improved through iterative update equations, as shown in the following equation:

[0102] (29)

[0103] In the formula, P ijLet τ be the probability of cross-class transformation of a data sample. ij The standardized pheromone between data sample i and its class j is calculated using the following formula:

[0104] (30)

[0105] In the formula, ρ represents the pheromone volatility, and t represents the time step;

[0106] By iteratively updating the equation, the clustering results of the "state-action" pairs are obtained. Since the clustering results do not yet have physical meaning, it is necessary to establish a mapping relationship between the clustering results and the scene hazard level. Select some representative "state-action" pairs from each scene hazard level and input them into the scene hazard identification model. Based on the clustering category of their parameters, establish a mapping relationship between the scene hazard level and the clustering results.

[0107] The probability of scene occurrence based on the convex combination algorithm is to directly solve the probability of occurrence of the "state-action" pair by following the method for solving the probability of scene occurrence in low-dimensional evaluation scenarios.

[0108] Step 2: Reconstruct the key long-tail function based on the Markov decision process. By discretizing the decision variables, the ideal scenario space of the high-dimensional evaluation scenario is obtained. By viewing the high-dimensional evaluation scenario as a Markov decision process, the key long-tail function is reconstructed in the form of "state-action" pairs.

[0109] The initial velocity of the target vehicle, the initial relative distance between the current vehicle and the target vehicle, the initial relative velocity between the current vehicle and the target vehicle, and the acceleration sequence of the target vehicle are used as decision variables x:

[0110] (31)

[0111] In the formula, v o R represents the initial velocity of the target vehicle. o ∆v represents the initial relative distance between this vehicle and the target vehicle. o a represents the initial relative velocity between this vehicle and the target vehicle. 0k This represents the acceleration of the target vehicle at the k-th time step;

[0112] When setting v o The value range is [20m / s, 40m / s], and the distance from the walk is 2m / s; R o The value range is (0m, 90m], and the distance from the walk is 2m; ∆v o The value range is [-20m / s, 20m / s], and the distance from the walk is 2m / s; a 0k The value range is [-4m / s 2 2m / s 2The walking distance is 0.2m / s. 2 When k is 10s, the number of ideal scene spaces is 21×45×21×31. 10 To reduce the dimensionality of the ideal scenario space and simplify the computational complexity of the key long-tail function, the high-dimensional evaluation scenario is treated as a Markov decision process. The relative distance and relative speed between the vehicle and the target vehicle are considered as states, and the acceleration of the target vehicle is considered as an action. The acceleration of the target vehicle at a given time step depends only on the relative states of the vehicle and the target vehicle at that time step. Therefore, the states and actions at the same time step are considered as a whole, i.e., a "state-action" pair. The number of "state-action" pairs in this ideal scenario space is 21 × 45 × 21 × 31 = 615195, which is significantly reduced compared to the number of scenarios in the previous ideal scenario space. The key long-tail function is reconstructed as follows:

[0113] (32)

[0114] In the formula, s i a represents the state at time step i. i G(s) represents the action at time step i. i ,a i It can be calculated using the following formula:

[0115] (33)

[0116] In the formula, V(s) i ,a i ) represents s i With a i The corresponding danger level of the "state-action" pair, P(s) i ,a i ) represents s i With a i The probability of occurrence of the corresponding "state-action" pair;

[0117] Step 3: Generate key adaptive dynamic scenarios based on Q-learning algorithm. Construct Bellman equations, solve the objective function of Markov decision process through Bellman equations, and solve the optimal action sequence corresponding to each initial state by updating expected returns, which is the key adaptive dynamic scenario.

[0118] Let Q(s,a) be the expected reward of the agent taking action a in state s, and r be the reward given by the environment when the agent takes action a. The Bellman equation is used to solve for the optimal policy in a Markov decision process:

[0119] (34)

[0120] In the formula, π represents the policy, S represents the state set, R represents the reward set, ξ represents the discount factor, t represents the time step, s represents the state at time t, and V π (s) represents the cost function, and Q can be updated according to the following formula:

[0121] (35)

[0122] In the formula, A represents the action set, a represents the action at time t, and α represents the learning rate;

[0123] Thus, the optimal action sequence corresponding to each initial state can be obtained, which is the key adaptive dynamic scene;

[0124] Step 5: Adaptive dynamic scenario evaluation for high- and low-dimensional assessment scenarios. The specific process is as follows:

[0125] Step 1: Adaptive dynamic scene sampling based on the ϵ-greededy sampling strategy. Using the same sampling method, scenes are sampled from the adaptive dynamic scenes for both high-dimensional and low-dimensional evaluation scenarios. Taking the low-dimensional adaptive dynamic scene as an example, a small probability value ϵ is set. Scenes are randomly sampled from the low-dimensional adaptive dynamic scene with a probability of 1-ϵ. Scenes are also randomly sampled from the scenes outside the low-dimensional adaptive dynamic scene in the low-dimensional ideal scene space with a probability of ϵ. This forms a test scene library, named Test Group 1. The test scene library obtained in the high-dimensional case is named Test Group 2 using the same sampling method.

[0126] Step 2: Adaptive dynamic scene evaluation based on accident rate and number of tests. Test group 1 and test group 2 test autonomous vehicles respectively, setting a certain confidence level, such as 80%, until the accident rate converges. Record the number of tests and the accident rate at which the accident rate converges for both test groups at this confidence level. At the same time, set up a control experiment, randomly sampling scenes from the low-dimensional ideal scene space and the high-dimensional ideal scene space respectively, denoted as control group 1 and control group 2, and test autonomous vehicles with the same confidence level until the accident rate converges. Record the accident rate for the two control groups at this confidence level. The number of tests and the accident rate at convergence are compared between test group 1 and control group 1, and between test group 2 and control group 2. If the accident rate of test group 1 is much higher than that of control group 1, and the number of tests at convergence is much lower than that of control group 1, then the adaptive dynamic scenario evaluation for low-dimensional test scenarios is effective. If the accident rate of test group 2 is much higher than that of control group 2, and the number of tests at convergence is much lower than that of control group 2, then the adaptive dynamic scenario evaluation for high-dimensional test scenarios is effective.

[0127] The beneficial effects of this invention are:

[0128] This invention provides a general adaptive dynamic scene generation method for autonomous driving, applicable to both high- and low-dimensional evaluation scenarios. This method generates key adaptive dynamic scenes that accelerate the testing of autonomous vehicles, significantly improving testing efficiency and reliability, and speeding up deployment. Specific benefits are as follows:

[0129] 1) This invention provides a method for constructing a scene element hierarchy model based on scene parameterization. The scene is described as a collection of scene elements. Based on the basic attributes of the scene elements and their geographical location, the scene elements are classified hierarchically, thereby establishing a scene element hierarchy model.

[0130] 2) This invention provides a method for optimizing the dimensionality of scene elements based on the Analytic Hierarchy Process (AHP). By using AHP, the importance of scene elements, which is difficult to quantify, is quantified, and the rationality of the importance weight values ​​of scene elements is verified. Therefore, based on the type of the target scene and the dimensional requirements of the evaluation scene to be established, scene elements with larger importance weight values ​​can be selected as decision variables, thereby achieving the optimization of scene element dimensions.

[0131] 3) This invention provides a method for constructing an ideal scene space based on scene element dimension optimization. The ideal test scene space is obtained by discretizing decision variables. A larger value range and a smaller deviation length are set for decision variables with higher importance weights, so that as many scenes as possible are included in the ideal test scene space. For decision variables with lower importance weights, the selection principle for the value range and deviation length is "key scenes can be included in the ideal test scene space", thereby establishing a new method for constructing an ideal scene space.

[0132] 4) This invention provides a method for constructing a realistic scene space based on natural driving data. Scene data is collected by a vehicle driving on real roads. The collected sensor data is preprocessed, target scene data is selected from the preprocessed scene data, and data corresponding to decision variables are selected from the target scene data to construct a realistic scene space.

[0133] 5) This invention provides a method for constructing critical long-tail functions based on the probability of scene occurrence and the level of scene danger. A new adaptive dynamic scene generation principle is designed, which considers scenes that are challenging for autonomous vehicles and have a certain probability of occurrence on real roads. Critical long-tail functions are designed based on the probability of scene occurrence and the level of scene danger as the basis for selecting critical adaptive dynamic scenes.

[0134] 6) This invention provides a method for solving the probability of scene occurrence based on a convex combination algorithm. The convex combination algorithm transforms scenes in a low-dimensional real-world scene space derived from natural driving data into scenes in a low-dimensional ideal scene space, thereby solving for the probability of scene occurrence in the low-dimensional ideal scene space.

[0135] 7) This invention provides a method for solving scene hazard boundaries based on support vector regression algorithm and a method for constructing scene hazard identification model based on Hammerstein identification process, which realizes the division of scene hazard boundaries and the solution of scene hazard in low-dimensional ideal scene space and high-dimensional ideal scene space.

[0136] 8) This invention provides a low-dimensional key adaptive dynamic scene screening method based on multi-starting point optimization algorithm and seed filling algorithm, and a high-dimensional key adaptive dynamic scene screening method based on Q-learning algorithm, thereby realizing fast search for key adaptive dynamic scenes in the entire ideal scene space and greatly improving search efficiency.

[0137] 9) This invention provides a method for reconstructing critical long-tail functions based on Markov decision processes. The high-dimensional evaluation scenario is viewed as a Markov decision process, the relative state variables of the vehicle and the target vehicle are considered states, the acceleration of the target vehicle is considered an action, and the state and action at the same time step are considered a "state-action" pair. This allows for the reconstruction of critical long-tail functions through "state-action" pairs.

[0138] 10) This invention provides an adaptive dynamic scene evaluation method for high- and low-dimensional evaluation scenarios. It achieves sampling of low- and high-dimensional adaptive dynamic scenes through an ε-greedy sampling strategy. By setting a control group and comparing the convergence failure rate with its corresponding number of tests, it realizes the evaluation of adaptive dynamic scenes for high- and low-dimensional evaluation scenarios. Attached Figure Description

[0139] Figure 1 This is a schematic diagram illustrating the overall steps of the general dynamic scene generation method described in this invention.

[0140] Figure 2 This is a block diagram of the method architecture for the general dynamic scene generation method described in this invention.

[0141] Figure 3 This is a schematic diagram of an exemplary embodiment of step one of the first steps of the present invention.

[0142] Figure 4 This is a schematic diagram of an exemplary embodiment of the influence transfer model described in this invention.

[0143] Figure 5This is a schematic diagram of an exemplary embodiment of step two of the second step of the present invention.

[0144] Figure 6 This is a schematic diagram of an exemplary embodiment when the ideal scene space described in this invention is a two-dimensional space.

[0145] Figure 7 This is a schematic diagram of an exemplary embodiment of the ideal scene space described in this invention, which is a three-dimensional space.

[0146] Figure 8 This is a schematic diagram illustrating an exemplary implementation result of step two of the third step of the present invention.

[0147] Figure 9 This is a schematic diagram of an exemplary embodiment of the entry scenario described in this invention.

[0148] Figure 10 This is a schematic diagram of an exemplary implementation of the multi-starting point optimization algorithm described in this invention.

[0149] Figure 11 This is a schematic diagram of an exemplary embodiment of the seed filling algorithm described in this invention.

[0150] Figure 12 This is a schematic diagram of an exemplary embodiment of the following scenario described in this invention. Detailed Implementation

[0151] Please see Figures 1 to 12 As shown:

[0152] The present invention provides a general method for generating adaptive autonomous driving dynamic scenes for high- and low-dimensional evaluation scenarios, the method of which is as follows:

[0153] Step 1: Optimize the dimensions of evaluation scenario elements based on analytic hierarchy process (AHP);

[0154] The second step is to construct the evaluation scenario space and key long-tail functions.

[0155] The third step is adaptive dynamic scene generation for low-dimensional evaluation scenarios.

[0156] Step 4: Adaptive dynamic scene generation for high-dimensional evaluation scenarios;

[0157] Step 5: Adaptive dynamic scenario evaluation for high- and low-dimensional evaluation scenarios.

[0158] The process of optimizing the dimensions of the evaluation scenario elements based on the analytic hierarchy process in the first step is as follows:

[0159] Step 1: Constructing a scene element hierarchy model based on scene parameterization. The scene is described as a collection of scene elements. Based on the basic attributes of the scene elements and their geographical location, the scene elements are classified hierarchically to establish a scene element hierarchy model.

[0160] exist Figure 3 The diagram illustrates an exemplary implementation of step one. A scene is a general dynamic description of an autonomous vehicle and its driving environment over a period of time. Based on the basic attributes of scene elements, scene elements can be divided into static elements and dynamic elements. Furthermore, according to the different geographical locations of scene elements, both static and dynamic elements can be divided into off-road elements and on-road elements.

[0161] Static elements outside the road mainly include static roadside objects such as buildings, trees, green belts, and traffic signs. Static elements inside the road mainly include lane-related elements and environment-related elements. Lane-related elements include lane type (highway, urban road, rural road), lane alignment (straight, curved, intersection), number of lanes, lane width, lane slope, lane curvature, lane line type, lane line width, pavement type (asphalt pavement, cement concrete pavement), and pavement adhesion coefficient. Environment-related elements include weather (sunny, cloudy, rainy, snowy, foggy) and light (sufficient, dim, changing).

[0162] The off-road elements of dynamic elements mainly include traffic participants, such as traffic participant types (pedestrians, bicycles, other living beings) and the number of traffic participants. The on-road elements of dynamic elements mainly include elements related to traffic participants, elements related to the initial state, and elements related to the driving state sequence. Elements related to traffic participants include traffic participant types (cars, trucks, motorcycles, tricycles, buses) and the number of traffic participants; elements related to the initial state include initial position, initial lane, and initial speed; elements related to the driving state sequence include triggering patterns (distance triggering, time triggering), relative distance sequences, relative speed sequences, and relative acceleration sequences.

[0163] Step 2: Scene Element Dimension Optimization Based on Analytic Hierarchy Process (AHP). Based on the scene element hierarchy model, the influence propagation model is used to calculate the number of times scene elements at the same level propagate influence across different levels of the autonomous driving system. The difference in the number of times scene elements propagate influence within the same level is calculated. Based on the correspondence between the difference in the number of times influence propagation occurs and the 1-9 scaling method, the difference in the number of times influence propagation occurs is converted into a scaling value, thereby establishing a judgment matrix for scene elements at the same level. The largest eigenvalue of the judgment matrix and its corresponding eigenvector are calculated. The eigenvector is normalized to obtain the importance weight value of each scene element at the same level relative to its current level. The rationality of the importance weight value is verified through a consistency check. Based on the scene element hierarchy model and the importance weight values ​​of each scene element at each level relative to its current level, the importance weight value of each scene element among all scene elements is calculated. According to the type of the target scene and the dimensional requirements of the evaluation scene to be established, scene elements with larger importance weight values ​​are selected as decision variables.

[0164] Using the scene element hierarchy model from step one as input, the influence propagation model is used to solve for the number of influence propagations of scene elements at the same level in each level of the autonomous driving system. Figure 4 An exemplary implementation of the influence transmission model is shown. The influence transmission model has three assumptions:

[0165] 1) The impact of scene elements on the autonomous driving system will be passed down level by level with the autonomous driving system, and this impact will not decrease with the passage of levels.

[0166] 2) Different types of scene elements have the same impact on the autonomous driving system;

[0167] 3) The magnitude of the influence of scene elements on the autonomous driving system can be represented by the number of times the influence of scene elements is transmitted between the levels of the autonomous driving system. The more times the influence is transmitted, the greater the influence of the scene element, and the influence is linearly related to the number of transmissions.

[0168] The number of times the influence of scene elements in the influence propagation model propagates through each level of the autonomous driving system can be calculated using the following formula:

[0169] (1)

[0170] In the formula, P(n) represents the number of times the influence of a scene element is transmitted through each level of the autonomous driving system, n represents the number of element attributes of the scene element, and E i This represents the number of times the influence of the i-th element attribute in this scene is propagated across all levels of the autonomous driving system.

[0171] After obtaining the number of influence transmissions of scene elements at each level in the scene element hierarchy model based on the influence transmission model, the difference in the number of influence transmissions of scene elements at the same level is calculated, and the difference in the number of influence transmissions is converted into a scale in the 1-9 scale method according to Table 1.

[0172] Table 1: Correspondence between the difference in the number of transmissions and the 1-9 scale method

[0173] Scale 1 2 3 Difference in the number of transmissions [7.5,10) [10,12.5) [12.5,15) Scale 4 5 6 Difference in the number of transmissions [15,17.5) [17.5,20) [20,23) Scale 7 8 9

[0174] The meanings of each scale in the 1-9 scale method are shown in Table 2:

[0175] Table 2: Meaning of the 1-9 Scale Method

[0176] 1 Both scene elements are equally important 3 One scene element is slightly more important than another scene element. 5 One scene element is significantly more important than another scene element. 7 One scene element is more strongly important than another scene element. 9 One scene element is absolutely more important than another scene element. 2，4，6，8 The case where it is in the middle of its two adjacent scales

[0177] After converting the difference in the number of influence transmissions into the scale in the 1-9 scaling method, a judgment matrix for scene elements at the same level can be established. The largest eigenvalue of the judgment matrix and its corresponding eigenvector can be solved. The normalization of the eigenvector is the importance weight value of scene elements at the same level.

[0178] To verify the rationality of the importance weight values, a consistency check must be performed on the judgment matrix. The formula for the consistency check is:

[0179] (2)

[0180] In the formula, CR represents the consistency index, and RI is the standard value of the hierarchical overall ranking average random consistency index, which takes different values ​​according to the order of the judgment matrix, as shown in Table 3:

[0181] Table 3: Average Random Consistency Index of Overall Hierarchical Ranking

[0182] RI 0 0 0.58 0.90 1.12 1.24 1.32 1.41 1.45

[0183] CI stands for Overall Hierarchical Ranking Consistency Index. The formula for calculating CI is:

[0184] (3)

[0185] In the formula, λ max CR represents the largest eigenvalue of the judgment matrix, and n is the order of the judgment matrix. When CR < 0.1, it indicates that the judgment matrix has good consistency and meets the consistency requirements.

[0186] Finally, based on the hierarchical model of scene elements and the importance weight value of each scene element relative to its own level, the importance weight value of each scene element in all scene elements is calculated. Then, the type of the target scene to be studied is determined. Based on the type of the target scene and the dimensional requirements of the evaluation scene to be established, some scene elements with larger importance weight values ​​are selected as decision variables, thus completing the dimensional optimization of scene elements.

[0187] The process of constructing the evaluation scenario space and key long-tail functions in the second step is as follows:

[0188] Step 1: Constructing the Ideal Test Scenario Space Based on Scene Element Dimension Optimization. The ideal test scenario space is obtained by discretizing the decision variables obtained from the scene element dimension optimization in Step 1. Each decision variable can be discretized into different values ​​according to its own element attributes. The range and dispersion of the decision variable should be determined based on its importance weight value, while fully considering the constraints of real-world road conditions. Decision variables with higher importance weight values ​​should be assigned a larger range and a smaller dispersion value to ensure that as many scenarios as possible can be included in the ideal test scenario space. For decision variables with lower importance weight values, the selection of their range and dispersion value should follow the principle that key scenarios should be included in the ideal test scenario space.

[0189] Step 2: Constructing a Real-World Scene Space Based on Natural Driving Data. Scene data is collected by the vehicle driving on real roads. The collected sensor data is preprocessed, and target scene data is selected from the preprocessed scene data. From the target scene data, the data corresponding to the decision variables are selected to construct the real-world scene space.

[0190] exist Figure 5 The diagram illustrates an exemplary implementation of step two in the second step. The scene data acquisition vehicle is equipped with sensors such as LiDAR, millimeter-wave radar, GPS high-precision inertial navigation, high-definition cameras, an onboard CAN bus, lane line sensors, rain sensors, and light sensors. Data from all sensors is collected according to a fixed acquisition cycle. The collected sensor data includes: spatial 3D point clouds in frames generated by LiDAR, obstacle status lists in frames generated by millimeter-wave radar, positioning and attitude data in time series generated by GPS high-precision inertial navigation, color images in frames generated by the high-definition camera and lane line sensors, vehicle handling and motion state data in time series generated by the onboard CAN bus, and voltage data in time series generated by the rain and light sensors. Furthermore, scene data acquisition should also adhere to the following two principles:

[0191] 1) The scene data collection area should include cities and regions with different road and traffic characteristics, and the total collection mileage should be large enough;

[0192] 2) The scene data collection environment should include different weather types and lighting types.

[0193] The specific contents of sensor data preprocessing include: time alignment and spatial alignment of each sensor data; verification of the validity of sensor data; generation of vehicle bus alignment signal, vehicle status alignment signal and multimodal environment sensor alignment signal.

[0194] Based on the type of target scenario selected in the first step, scene data of all target scenarios are extracted from all preprocessed data by manually watching videos. Each complete segment of target scenario data is called a scenario condition, representing a complete target scenario event that occurs on a real road. The data corresponding to the decision variables in each scenario condition are filtered out to form the real scenario space.

[0195] Step 3: Constructing a critical long-tail function based on scenario occurrence probability and scenario hazard level. Based on the adaptive dynamic scenario generation principle of "challenging autonomous vehicles and having a certain probability of occurrence on real roads," a critical long-tail function is designed based on scenario occurrence probability and scenario hazard level as the basis for selecting critical adaptive dynamic scenarios. The calculation formula for the critical long-tail function is:

[0196] (4)

[0197] In the formula, I(x) represents the critical long-tail function value of the scenario, P(x) represents the probability of the scenario occurring, and V(x) represents the risk level of the scenario.

[0198] The process of adaptive dynamic scene generation for low-dimensional evaluation scenarios in the third step is as follows:

[0199] Step 1: Solving the scenario occurrence probability based on the convex combination algorithm. The convex combination algorithm is used to transform the low-dimensional real-world scenario space from natural driving data into a low-dimensional ideal scenario space, and then the occurrence probability of the scenario in the ideal scenario space is calculated.

[0200] Suppose there exist vectors {x1, x2, x3...x} n If there is a real number λ i ≥ 0, i=1,2,3...n, and λ1+λ2+...+λ n =1, then λ1x1+λ2x2+...λ n x n Let the vector {x1, x2, x3...x} be... n A convex combination of}.

[0201] The following assumptions are made when using the convex combination algorithm to solve for the probability of scenario occurrence:

[0202] 1) Within a certain distance range of the ideal scene space, the probability of the scene occurring will not change abruptly, but will change continuously and follow a linear change law;

[0203] 2) Within a certain distance range of the ideal scene space, the linear change law of the probability of scene occurrence can be represented by the Euclidean distance between scenes in the ideal scene space;

[0204] 3) Within a certain distance range of the ideal scene space, the probability of a scene occurring is high if it is close to the Euclidean distance of a high-probability scene, and low if it is far from the Euclidean distance of a high-probability scene.

[0205] exist Figure 6 This illustrates an exemplary embodiment when the ideal scene space is a two-dimensional space. When the ideal scene space is two-dimensional, i.e., formed by discretizing two decision variables, taking relative distance R and relative velocity ∆v as decision variables as an example, this paper introduces a method for solving the probability of scene occurrence in a two-dimensional ideal scene space. Taking a scene space within a discretized step length as an example, A1(R... A1 , ∆v A1 ), A2(R A2 , ∆v A2 ), A3(R A3 , ∆v A3 ), A4(R A4 , ∆v A4 B(R) represents four uniformly discrete scenes in a two-dimensional ideal scene space. B , ∆v B (A) is a scene in a real-world scene space derived from natural driving data, and B is located within the rectangle formed by A1, A2, A3, and A4. 11 L is the Euclidean distance from B to the lines containing A2 and A4. 12 L is the Euclidean distance from B to the lines containing A1 and A3. 21 L is the Euclidean distance from B to the lines containing A3 and A4. 22 Let be the Euclidean distance from B to the lines containing A1 and A4. Then, scene B in the real-world scene space can be transformed into scenes A1, A2, A3, and A4 in a two-dimensional ideal scene space using a convex combination algorithm.

[0206] (5)

[0207] In the formula, ω1, ω2, ω3, and ω4 are the weighting coefficients of A1, A2, A3, and A4, respectively, and their calculation formula is as follows:

[0208] (6)

[0209] (7)

[0210] (8)

[0211] (9)

[0212] Using the above method, all scenes in the real scene space can be transformed into uniformly discrete scenes in the ideal scene space. By combining their weight coefficients to count the number of uniformly discrete scenes, the probability of a scene occurring in the two-dimensional ideal scene space can be obtained.

[0213] exist Figure 7 This document illustrates an exemplary embodiment when the ideal scene space is three-dimensional. When the ideal scene space is three-dimensional, i.e., formed by discretizing three decision variables, using relative distance R, relative velocity ∆v, and relative acceleration ∆a as examples, it introduces a method for solving the probability of scene occurrence in a three-dimensional ideal scene space. Taking a scene space within a discretized step size as an example, A1(R... A1 , ∆v A1 , ∆a A1 ), A2(R A2 , ∆v A2 , ∆a A2 ), A3(R A3 , ∆v A3 , ∆a A3 ),A4(R A4 , ∆v A4 , ∆a A4 ), A5(R A5 , ∆v A5 , ∆a A5 ), A6(R A6 , ∆v A6 , ∆a A6 ), A7(R A7 , ∆v A7 , ∆a A7 ), A8(R A8 , ∆v A8 , ∆a A8 B(R) represents eight uniformly discrete scenes in a three-dimensional scene space. B , ∆v B , ∆a B ( ) is a scene in the real-world scene space obtained from natural driving data, and B is inside the cube formed by A1, A2, A3, A4, A5, A6, A7, and A8. L 11 L is the Euclidean distance from B to the plane containing A2, A4, A6, and A8.12 L is the Euclidean distance from B to the plane containing A1, A3, A5, and A7. 21 L is the Euclidean distance from B to the plane containing A3, A4, A7, and A8. 22 L is the Euclidean distance from B to the plane containing A1, A2, A5, and A6. 31 L is the Euclidean distance from B to the plane containing A5, A6, A7, and A8. 32 This is the Euclidean distance from B to the planes containing A1, A2, A3, and A4. Therefore, scene B in the real-world scene space can be transformed into scenes A1, A2, A3, A4, A5, A6, A7, and A8 in the ideal 3D scene space using a convex combination algorithm.

[0214] (10)

[0215] In the formula, ω1, ω2, ω3, ω4, ω5, ω6, ω7, and ω8 are the weight coefficients of A1, A2, A3, A4, A5, A6, A7, and A8, respectively, and their calculation formula is as follows:

[0216] (11)

[0217] (12)

[0218] (13)

[0219] (14)

[0220] (15)

[0221] (16)

[0222] (17)

[0223] (18)

[0224] Using the above method, all scenes in the real scene space can be transformed into uniformly discrete scenes in the ideal scene space. By combining their weight coefficients to count the number of uniformly discrete scenes, the probability of a scene occurring in the three-dimensional ideal scene space can be obtained.

[0225] Step 2: Solving the scene hazard boundary based on the support vector regression algorithm. The scene hazard boundary in the real scene space is pre-divided according to the collision time, constructing a joint distribution of collision time and relative acceleration. Boundary points belonging to the same boundary line in the joint distribution are used as a training sample set and input separately into the support vector regression algorithm. Each training sample set corresponds to one scene hazard boundary. All scene hazard boundary lines are summarized to obtain the complete scene hazard boundary. Finally, the scene hazard boundaries in the real scene space are mapped to the ideal scene space, and hazard coefficients are set for scenes with different hazard levels.

[0226] exist Figure 8 The image shows an exemplary implementation result of step two in the third step, taking the entry scenario as the target scenario as an example. Figure 9 The diagram illustrates an exemplary embodiment of the scenario. At the moment when the target vehicle's center coincides with the lane line, the relative distance R, relative velocity ∆v, and relative acceleration ∆a between the vehicle and the target vehicle are selected as decision variables to construct an ideal scenario space and a real scenario space based on natural driving data. The time to collision (TTC) in the real scenario space is then calculated. TTC represents the time required for a collision to occur from the current moment while both vehicles maintain their current states of motion. The smaller the TTC value, the higher the risk of the scenario. TTC can be calculated using the following formula:

[0227] (19)

[0228] In the formula, R represents the relative distance between the two vehicles, and ∆v represents the relative speed between the vehicle and the target vehicle.

[0229] The pre-defined criteria for defining the hazard boundary of a scenario are as follows: When TTC∈[0s,1s], the hazard level of the scenario is a crash scenario; when TTC∈(1s,3s], the hazard level of the scenario is an emergency scenario; when TTC∈(3s,5s], the hazard level of the scenario is a conflict scenario; when TTC∈(5s,+∞) or (-∞,0s), the hazard level of the scenario is a safe scenario.

[0230] A joint distribution of TTC and ∆a in the real scene space is constructed. This joint distribution is equivalent to the existence of four scene hazard boundaries defined by TTC=0, TTC=1, TTC=3, and TTC=5. The boundary points on both sides of the joint distribution are used as separate training sample sets and input into a support vector regression algorithm for learning. Since the boundaries of this joint distribution are nearly linear, a linear kernel function is chosen to solve for the scene hazard boundaries on both sides of the joint distribution. The scene hazard boundaries pre-defined based on TTC are then merged with the scene hazard boundaries based on support vector regression to obtain the hazard boundaries in the real scene space:

[0231] (20)

[0232] In the formula, l l With l r These represent the scene hazard boundaries on the left and right sides of the joint distribution map obtained through support vector regression, Δa and Δa, respectively. l and Δa r These represent the distances between scene B and the scene hazard boundary, l m1 l m2 l m3 and l m4 These are the scene hazard boundaries obtained from the TTC pre-division.

[0233] The above-mentioned scenario hazard boundaries are mapped to the ideal scenario space. The scenario hazard level corresponding to the collision scenario is set to 1, the scenario hazard level corresponding to the emergency scenario is set to 0.7, the scenario hazard level corresponding to the conflict scenario is set to 0.3, and the scenario hazard level corresponding to the safe scenario is set to 0.

[0234] Step 3: Key Adaptive Dynamic Scene Generation Based on Multi-Startpoint Optimization and Seed Filling Algorithms. The key long-tail function values ​​in the ideal scene space are calculated using the scene occurrence probability and scene hazard level. The key long-tail function, the ideal scene space, and the key threshold are input into the multi-startpoint optimization algorithm to obtain local key adaptive dynamic scenes. The local key scenes, key long-tail functions, the ideal scene space, and the key thresholds are then input into the seed filling algorithm to obtain all key adaptive dynamic scenes.

[0235] exist Figure 10An exemplary implementation of the multi-starting-point optimization algorithm is shown. The critical long-tail function of the scene in the ideal scene space can be obtained by combining the scene occurrence probability obtained in step one with the scene danger level obtained in step two. The critical long-tail function, the ideal scene space, and the critical threshold γ are input into the multi-starting-point optimization algorithm. Points are sampled in the ideal scene space, either manually or randomly, as the starting points of the algorithm. The maxima of the critical long-tail function in the attraction domain of each starting point are solved. The scenes corresponding to the maxima of all critical long-tail functions greater than γ are output, thus obtaining the locally critical adaptive dynamic scene for low-dimensional evaluation scenarios.

[0236] exist Figure 11 The diagram illustrates an exemplary implementation of the seed filling algorithm. The local key scenes, key long-tail functions, ideal scene space, and key threshold γ output by the multi-starting-point optimization algorithm are input into the seed filling algorithm. Starting with each local key scene, the algorithm searches the ideal scene space for scenes in the neighborhood around the starting point whose key long-tail function values ​​are greater than γ. These scenes are then used as new starting points, and the search continues in their surrounding neighborhoods. This process is repeated until the key long-tail function values ​​of all scenes in the neighborhood around the starting point are less than γ. The algorithm then terminates, outputting all scenes marked as starting points, thus generating all key adaptive dynamic scenes for low-dimensional evaluation scenarios.

[0237] The process of adaptive dynamic scene generation for high-dimensional evaluation scenarios in the fourth step is as follows:

[0238] Step 1: Construction of a Scene Hazard Identification Model Based on the Hammerstein Identification Process and Solution of Scene Occurrence Probability Based on the Convex Combination Algorithm. The Hammerstein identification process consists of a series of static nonlinear and dynamic linear components. The relative state variables of the vehicle and the target vehicle, as well as the acceleration of the target vehicle, are used as inputs to the scene hazard identification model, while the acceleration of the vehicle is used as the model output. After training with a large amount of scene data, the model parameters of the scene hazard identification model are used as key parameters characterizing the intrinsic attributes of scene hazard. Principal component analysis is used to decouple and reduce the dimensionality of the key parameters. Ant colony clustering is used to cluster scene hazards. Representative "state-action" pairs are selected from each scene hazard level and input into the scene hazard identification model. Based on the cluster category of the parameters, a mapping relationship between the scene hazard level and the clustering results is established. The scene occurrence probability solution based on the convex combination algorithm can directly solve for the occurrence probability of "state-action" pairs using the method for solving scene occurrence probability in low-dimensional assessment scenarios.

[0239] Viewing the high-dimensional assessment scenario as a Markov decision process, the relative state variables of the vehicle and the target vehicle are considered as states, and the acceleration of the target vehicle is considered as an action. The state and action at the same time step are considered a "state-action" pair. The relative state variables of the vehicle and the target vehicle at the same moment, along with the acceleration of the target vehicle, are used as inputs to the scenario hazard identification model, while the acceleration of the vehicle is used as the model's output. Therefore, the scenario hazard identification model is a multi-input single-output system. The Hammerstein identification process consists of a static nonlinear element and a dynamic linear element connected in series. The static nonlinear element can use various functions such as dead-zone functions, sigmoid functions, or saturation functions. The z-transform of the dynamic linear element is shown in the following equation:

[0240] (twenty one)

[0241] In the formula, O p (k) represents the set of accelerations of the main vehicle, N(k) represents the set of outputs of the static nonlinear element, and d

[0242] A(z) represents the order of the input delay and is defined as an integer multiple of the sampling time. -1 ) and B(z) -1 It can be calculated using the following formula:

[0243] (twenty two)

[0244] In the formula, (a1,…a q ) and (b1,…b n All of these are coefficients of the dynamic linear element, and q and n are the orders of the dynamic linear element.

[0245] After training with a large amount of scenario data from natural driving data, the model parameters contained in the static nonlinear and dynamic linear components are key data characterizing the intrinsic attributes of scenario hazard, and therefore are used as data samples for scenario hazard assessment. To ensure that the spatial dimension of the data samples is reduced as much as possible to improve computational efficiency while expressing the same model features, principal component analysis is used to decouple and reduce the dimensionality of key parameters in the scenario hazard identification model.

[0246] Let H represent the parameter dimension of the scene hazard identification model, and E represent the amount of training data. Then the model parameter dataset X can be represented as:

[0247] (twenty three)

[0248] In the formula, x i The intrinsic parameter vector represents the scene hazard identification model.

[0249] X is input to the principal component analysis algorithm. The percentage of the sum of the eigenvalues ​​of the first m principal components to the sum of all eigenvalues ​​is defined as the principal component contribution rate. The cumulative principal component contribution rate M is then calculated. m It can be calculated using the following formula:

[0250] (twenty four)

[0251] In the formula, λ i This represents the eigenvector.

[0252] To ensure the dimensionality reduction effect, we take M. m The value of m corresponding to ≥85% is used as the dimension of the independent parameters of the model calculated by the algorithm, and finally an m×E matrix L is obtained:

[0253] (25)

[0254] Ant colony clustering is used to cluster scene hazards. Taking L as the input to the ant colony clustering algorithm, scene hazard classification based on ant colony clustering is to find the partitioning method in L that minimizes the sum of distances from each data sample to the cluster centers with a known number of clusters. Referring to the scene hazard grading method for low-dimensional assessment scenes in step two of step three, the scene hazard is divided into 4 levels, resulting in 4 clusters. The ant colony clustering algorithm can be expressed by the following formula:

[0255] (26)

[0256] In the formula, J represents the sum of the distances from each data sample to the four cluster centers, l ip c represents the p-th model parameter feature of the i-th data sample. jp The p-th model parameter feature for the j-th class center can be calculated using the following formula:

[0257] (27)

[0258] In the formula, E j Let ω be the observed variable corresponding to the j-th class in observed variable E. ij The subordinate relationship indicator between the observed variable and the category can be calculated using the following formula:

[0259] (28)

[0260] The classification quality of ant colony clustering can be improved through iterative update equations, as shown in the following equation:

[0261] (29)

[0262] In the formula, P ijLet τ be the probability of cross-class transformation of a data sample. ij The standardized pheromone between data sample i and its class j can be calculated using the following formula:

[0263] (30)

[0264] In the formula, ρ represents the pheromone volatility, and t represents the time step.

[0265] By iteratively updating the equations, clustering results for "state-action" pairs can be obtained. Since the clustering results do not yet have physical meaning, it is necessary to establish a mapping relationship between the clustering results and the scene hazard level. Representative "state-action" pairs are selected from each scene hazard level and input into the scene hazard identification model. Based on the cluster category of their parameters, a mapping relationship between the scene hazard level and the clustering results is established.

[0266] The probability of scene occurrence based on the convex combination algorithm can be directly solved by the method for solving the probability of scene occurrence in low-dimensional evaluation scenarios, which is used to solve the probability of occurrence of "state-action" pairs.

[0267] Step 2: Reconstruction of the key long-tail function based on the Markov decision process. By discretizing the decision variables, the ideal scenario space of the high-dimensional evaluation scenario is obtained. By treating the high-dimensional evaluation scenario as a Markov decision process, the key long-tail function is reconstructed in the form of "state-action" pairs.

[0268] exist Figure 12 An exemplary embodiment of a following scenario is illustrated. Taking a following scenario as the target scenario as an example, since the following scenario can no longer represent the entire scenario condition by key moments, but rather studies the dynamic changes of the scenario over a period of time, the number of scenarios in the ideal scenario space will increase exponentially with the increase of the scenario dimension when constructing the ideal scenario space from decision variables. The initial speed of the target vehicle, the initial relative distance between the current vehicle and the target vehicle, the initial relative speed between the current vehicle and the target vehicle, and the acceleration sequence of the target vehicle are taken as decision variables x:

[0269] (31)

[0270] In the formula, v o R represents the initial velocity of the target vehicle. o ∆v represents the initial relative distance between this vehicle and the target vehicle. o a represents the initial relative velocity between this vehicle and the target vehicle. 0k This represents the acceleration of the target vehicle at the k-th time step.

[0271] When setting v o The value range is [20m / s, 40m / s], and the distance from the walk is 2m / s; Ro The value range is (0m, 90m], and the distance from the walk is 2m; ∆v o The value range is [-20m / s, 20m / s], and the distance from the walk is 2m / s; a 0k The value range is [-4m / s 2 2m / s 2 The walking distance is 0.2m / s. 2 When k is 10s, the number of ideal scene spaces is 21×45×21×31. 10 To reduce the dimensionality of the ideal scenario space and simplify the computational complexity of the critical long-tail function, the high-dimensional evaluation scenario can be viewed as a Markov decision process. The relative distance and speed between the vehicle and the target vehicle can be considered as states, and the acceleration of the target vehicle as an action. The acceleration of the target vehicle at a given time step depends only on the relative states of the vehicle and the target vehicle at that time step. Therefore, the states and actions at the same time step can be considered as a whole, i.e., a "state-action" pair. The number of "state-action" pairs in this ideal scenario space is 21 × 45 × 21 × 31 = 615195, significantly reduced compared to the number of scenarios in the previous ideal scenario space. The critical long-tail function is reconstructed as follows:

[0272] (32)

[0273] In the formula, s i a represents the state at time step i. i G(s) represents the action at time step i. i ,a i It can be calculated using the following formula:

[0274] (33)

[0275] In the formula, V(s) i ,a i ) represents s i With a i The corresponding danger level of the "state-action" pair, P(s) i ,a i ) represents s i With a i The probability of occurrence of the corresponding "state-action" pair.

[0276] Step 3: Key Adaptive Dynamic Scene Generation Based on Q-learning Algorithm. Construct the Bellman equation, solve for the objective function of the Markov decision process using the Bellman equation, and solve for the optimal action sequence corresponding to each initial state by updating the expected return; this is the key adaptive dynamic scene.

[0277] Let Q(s,a) be the expected reward of the agent taking action a in state s, and r be the reward given by the environment when the agent takes action a. The Bellman equation is used to solve for the optimal policy in a Markov decision process:

[0278] (34)

[0279] In the formula, π represents the policy, S represents the state set, R represents the reward set, ξ represents the discount factor, t represents the time step, s represents the state at time t, and V π (s) represents the cost function. Q can be updated according to the following formula:

[0280] (35)

[0281] In the formula, A represents the action set, a represents the action at time t, and α represents the learning rate.

[0282] This allows us to obtain the optimal action sequence for each initial state, which is the key adaptive dynamic scene.

[0283] The adaptive dynamic scenario evaluation process for high- and low-dimensional evaluation scenarios in step five is as follows:

[0284] Step 1: Adaptive Dynamic Scene Sampling Based on the ϵ-greededy Sampling Strategy. Using the same sampling method, scenes are sampled from both the adaptive dynamic scenes for high- and low-dimensional evaluation scenarios. Taking the low-dimensional adaptive dynamic scene as an example, a small probability value ϵ is set. Scenes are randomly sampled from the low-dimensional adaptive dynamic scene with a probability of 1-ϵ. Scenes are also randomly sampled from scenes outside the low-dimensional adaptive dynamic scene in the low-dimensional ideal scene space with a probability of ϵ. This forms a test scene library, named Test Group 1. The same sampling method is used to name the test scene library obtained in the high-dimensional case Test Group 2.

[0285] Step 2: Adaptive Dynamic Scene Evaluation Based on Accident Rate and Number of Tests. Test Group 1 and Test Group 2 are used to test autonomous vehicles, with a certain confidence level (e.g., 80%), until the accident rate converges. The number of tests and the accident rate at which the accident rate converges are recorded for both test groups at this confidence level. Simultaneously, a control experiment is set up, randomly sampling scenes from the low-dimensional ideal scene space and the high-dimensional ideal scene space, respectively designated as Control Group 1 and Control Group 2. Autonomous vehicles are tested in these control groups with the same confidence level until the accident rate converges. The number of tests and the accident rate at which the accident rate converges for both control groups are recorded for each control group at this confidence level. By comparing the accident rates and the number of tests required for accident rate convergence between test group 1 and control group 1, and between test group 2 and control group 2, if the accident rate of test group 1 is much higher than that of control group 1, and the number of tests required for accident rate convergence is much lower than that of control group 1, then the adaptive dynamic scenario evaluation for low-dimensional test scenarios is effective. If the accident rate of test group 2 is much higher than that of control group 2, and the number of tests required for accident rate convergence is much lower than that of control group 2, then the adaptive dynamic scenario evaluation for high-dimensional test scenarios is effective.

Claims

1. A general method for generating adaptive autonomous driving dynamic scenes for high- and low-dimensional evaluation scenarios, characterized in that: The method includes the following steps: The first step is to optimize the dimensions of the evaluation scenario elements based on the analytic hierarchy process. The specific process is as follows: Step 1: Construct a scene element hierarchy model based on scene parameterization. Describe the scene as a collection of scene elements. According to the basic attributes of the scene elements and their geographical location conditions, classify the scene elements level by level to establish a scene element hierarchy model. A scene is a general dynamic description of an autonomous vehicle and its driving environment over a period of time. Based on the basic attributes of scene elements, scene elements are divided into static elements and dynamic elements. According to the different geographical locations of scene elements, static elements and dynamic elements are further divided into off-road elements and on-road elements. Static elements outside the road mainly include static roadside objects, while static elements inside the road mainly include lane-related elements and environment-related elements. The off-road elements of dynamic elements mainly include traffic participants, while the on-road elements of dynamic elements mainly include elements related to traffic participants, elements related to the initial state, and elements related to the driving state sequence. Step 2: Optimization of Scene Element Dimensions Based on Analytic Hierarchy Process (AHP). Based on the scene element hierarchy model, the influence propagation model is used to calculate the number of times scene elements at the same level propagate influence across different levels of the autonomous driving system. The difference in the number of times scene elements propagate influence within the same level is calculated. Based on the correspondence between the difference in the number of times influence propagation occurs and the 1-9 scaling method, the difference in the number of times influence propagation occurs is converted into a scaling value, thereby establishing a judgment matrix for scene elements at the same level. The largest eigenvalue of the judgment matrix and its corresponding eigenvector are solved. The eigenvector is normalized to obtain the importance weight value of each scene element at the same level relative to its current level. The rationality of the importance weight value is verified through a consistency check. Based on the scene element hierarchy model and the importance weight value of each scene element at each level relative to its current level, the importance weight value of each scene element among all scene elements is calculated. Based on the type of the target scene and the dimensional requirements of the evaluation scene to be established, scene elements with larger importance weight values ​​are selected as decision variables. Using the scene element hierarchy model from step one as input, the influence propagation model is used to solve for the number of times the influence of scene elements at the same level is propagated across different levels of the autonomous driving system. The influence propagation model has three assumptions, as follows: 1) The impact of scene elements on the autonomous driving system will be passed down level by level with the autonomous driving system, and this impact will not decrease with the passage of levels. 2) Different types of scene elements have the same impact on the autonomous driving system; 3) The magnitude of the influence of scene elements on the autonomous driving system is represented by the number of times the influence of scene elements is transmitted between the levels of the autonomous driving system. The more times the influence is transmitted, the greater the influence of the scene element, and the influence is linearly related to the number of transmissions. The number of times the influence of scene elements in the influence propagation model propagates through each level of the autonomous driving system can be calculated using the following formula: (1)； In the formula, P(n) represents the number of times the influence of a scene element is transmitted through each level of the autonomous driving system, n represents the number of element attributes of the scene element, and E i This represents the number of times the influence of the i-th element attribute in this scene is propagated across all levels of the autonomous driving system. After obtaining the number of influence transmissions of scene elements at each level in the scene element hierarchy model based on the influence transmission model, the difference in the number of influence transmissions of scene elements at the same level is calculated. The difference in the number of influence transmissions is converted into a scale in the 1-9 scaling method. After converting the difference in the number of influence transmissions into a scale in the 1-9 scaling method, a judgment matrix for scene elements at the same level is established. The largest eigenvalue of the judgment matrix and its corresponding eigenvector are solved. The eigenvector is normalized to obtain the importance weight value of scene elements at the same level. To verify the rationality of the importance weight values, a consistency check is performed on the judgment matrix. The formula for the consistency check is: (2)； In the formula, CR represents the consistency index, and RI is the standard value of the hierarchical overall ranking average random consistency index, which takes different values ​​depending on the order of the judgment matrix. CI stands for Overall Hierarchical Ranking Consistency Index. The formula for calculating CI is: (3)； In the formula, λ max CR represents the largest eigenvalue of the judgment matrix, and n is the order of the judgment matrix. When CR < 0.1, it indicates that the judgment matrix has good consistency and meets the consistency requirements. Finally, based on the hierarchical model of scene elements and the importance weight value of each scene element relative to its own level, the importance weight value of each scene element in all scene elements is calculated. Then, the type of the target scene to be studied is determined. Based on the type of the target scene and the dimensional requirements of the evaluation scene to be established, some scene elements with larger importance weight values ​​are selected as decision variables, thus completing the dimensional optimization of scene elements. The second step, constructing the evaluation scenario space and key long-tail functions, is as follows: Step 1: Construction of the ideal test scenario space based on scenario element dimension optimization. The ideal test scenario space is obtained by discretizing the decision variables obtained from the scenario element dimension optimization in Step 1. Each decision variable is discretized into different values ​​according to its own element attributes. The range of values ​​and the deviation length of the decision variable are determined according to the importance weight value of the decision variable, taking into full account the constraints of real road conditions. Decision variables with higher importance weight values ​​are set with a larger range of values ​​and a smaller deviation length to ensure that as many scenarios as possible can be included in the ideal test scenario space. For decision variables with lower importance weights, the selection of their value range and dispersion size follows the principle that the key scenarios can be included in the ideal test scenario space; Step 2: Constructing a real-world scene space based on natural driving data. Scene data is collected by vehicles driving on real roads. The collected sensor data is preprocessed, and target scene data is selected from the preprocessed scene data. Then, the data corresponding to the decision variables are selected from the target scene data to construct a real-world scene space. Scene data collection should also follow these two principles: 1) The scene data collection area should include cities and regions with different road and traffic characteristics, and the total collection mileage should be large enough; 2) The scene data collection environment should include different weather types and lighting types; The specific contents of sensor data preprocessing include: time alignment and spatial alignment of each sensor data; verification of the validity of sensor data; generation of vehicle bus alignment signal, vehicle status alignment signal and multimodal environment sensor alignment signal; Based on the type of target scenario selected in the first step, the scene data of all target scenarios is extracted from all preprocessed data by manually watching the video. Each complete segment of target scenario data is called a scenario condition, which represents a complete target scenario event that occurs on a real road. The data corresponding to the decision variables in each scenario condition are filtered out to form the real scenario space. Step 3: Constructing a critical long-tail function based on scene occurrence probability and scene hazard level. Following the adaptive dynamic scene generation principle of "challenging autonomous vehicles and having a certain probability of occurrence on real roads," a critical long-tail function is designed based on scene occurrence probability and scene hazard level as the basis for selecting critical adaptive dynamic scenes. The calculation formula for the critical long-tail function is as follows: (4)； In the formula, I(x) represents the critical long-tail function value of the scenario, P(x) represents the probability of the scenario occurring, and V(x) represents the risk level of the scenario. The third step is adaptive dynamic scene generation for low-dimensional evaluation scenarios, and the specific process is as follows: Step 1: Solve the probability of scene occurrence based on the convex combination algorithm. The convex combination algorithm is used to transform the scene in the low-dimensional real scene space from the natural driving data into the scene in the low-dimensional ideal scene space, and then solve the probability of scene occurrence in the low-dimensional ideal scene space. Suppose there exist vectors {x1, x2, x3...x} n If there is a real number λ i ≥ 0, i=1,2,3...n, and λ1+λ2+...+λ n =1, then λ1x1+λ2x2+...λ n x n Let the vector {x1, x2, x3...x} be... n A convex combination of}; The following assumptions are made when using the convex combination algorithm to solve for the probability of scenario occurrence: 1) Within a certain distance range of the ideal scene space, the probability of the scene occurring will not change abruptly, but will change continuously and follow a linear change law; 2) Within a certain distance range of the ideal scene space, the linear change law of the probability of scene occurrence can be represented by the Euclidean distance between scenes in the ideal scene space. 3) Within a certain distance range of the ideal scene space, scenes that are closer to high-probability scenes in Euclidean distance have a higher probability of occurrence, while scenes that are farther from high-probability scenes in Euclidean distance have a lower probability of occurrence. Taking a scene space within a walking distance as an example, A1(R) A1 ,∆v A1 ), A2(R A2 , ∆v A2 ), A3(R A3 , ∆v A3 ),A4(R A4 , ∆v A4 B(R) represents four uniformly discrete scenes in a two-dimensional ideal scene space. B , ∆v B ( ) is a scene in a real-world scene space obtained from natural driving data, and B is inside the rectangle formed by A1, A2, A3, and A4, L 11 L is the Euclidean distance from B to the lines containing A2 and A4. 12 L is the Euclidean distance from B to the lines containing A1 and A3. 21 L is the Euclidean distance from B to the lines containing A3 and A4. 22 Let B be the Euclidean distance from the lines containing A1 and A4. Then, scene B in the real scene space can be transformed into scenes A1, A2, A3, and A4 in the two-dimensional ideal scene space through a convex combination algorithm. (5)； In the formula, ω1, ω2, ω3, and ω4 are the weighting coefficients of A1, A2, A3, and A4, respectively, and their calculation formula is as follows: (6)； (7)； (8)； (9)； The above method is used to transform all scenes in the real scene space into uniformly discrete scenes in the ideal scene space. The number of uniformly discrete scenes is counted by combining their weight coefficients, which gives the probability of the scene occurring in the two-dimensional ideal scene space. Taking a scene space within a walking distance as an example, A1(R) A1 , ∆v A1 , ∆a A1 ), A2(R A2 , ∆v A2 , ∆a A2 ),A3(R A3 , ∆v A3 , ∆a A3 ), A4(R A4 , ∆v A4 , ∆a A4 ), A5(R A5 , ∆v A5 , ∆a A5 ), A6(R A6 , ∆v A6 , ∆a A6 ), A7(R A7 , ∆v A7 , ∆a A7 ), A8(R A8 , ∆v A8 , ∆a A8 B(R) represents eight uniformly discrete scenes in a three-dimensional scene space. B , ∆v B , ∆a B ( ) is a scene in a real-world scene space obtained from natural driving data, and B is inside the cube formed by A1, A2, A3, A4, A5, A6, A7, and A8, L 11 L is the Euclidean distance from B to the plane containing A2, A4, A6, and A8. 12 L is the Euclidean distance from B to the plane containing A1, A3, A5, and A7. 21 L is the Euclidean distance from B to the plane containing A3, A4, A7, and A8. 22 L is the Euclidean distance from B to the plane containing A1, A2, A5, and A6. 31 L is the Euclidean distance from B to the plane containing A5, A6, A7, and A8. 32 If the distance from B to the plane containing A1, A2, A3, A4 is Euclidean distance, then the scene B in the real scene space can be transformed into scenes A1, A2, A3, A4, A5, A6, A7, A8 in the three-dimensional ideal scene space through the convex combination algorithm. (10)； In the formula, ω1, ω2, ω3, ω4, ω5, ω6, ω7, and ω8 are the weight coefficients of A1, A2, A3, A4, A5, A6, A7, and A8, respectively, and their calculation formula is as follows: (11)； (12)； (13)； (14)； (15)； (16)； (17)； (18)； The above method is used to transform all scenes in the real scene space into uniformly discrete scenes in the ideal scene space. The number of uniformly discrete scenes is counted by combining their weight coefficients, which gives the probability of scene occurrence in the three-dimensional ideal scene space. Step 2: Solving the scene hazard boundary based on the support vector regression algorithm. The scene hazard boundary in the real scene space is pre-divided according to the collision time. A joint distribution of collision time and relative acceleration is constructed. Boundary points belonging to the same boundary line in the joint distribution are used as a training sample set and input into the support vector regression algorithm separately. Each training sample set corresponds to a scene hazard boundary. All scene hazard boundary lines are summarized to obtain the complete scene hazard boundary. Finally, the scene hazard boundary in the real scene space is mapped to the ideal scene space, and hazard coefficients are set for scenes with different hazard levels. TTC represents the time required for a collision to occur between the two vehicles while maintaining their current states of motion. A smaller TTC value indicates a higher level of danger. TTC is calculated using the following formula: (19)； In the formula, R represents the relative distance between the two vehicles, and ∆v represents the relative speed between the vehicle and the target vehicle; The pre-defined criteria for defining the hazard boundary of a scenario are as follows: when TTC∈[0s,1s], the hazard level of the scenario is a collision scenario; when TTC∈(1s,3s], the hazard level of the scenario is an emergency scenario; when TTC∈(3s,5s], the hazard level of the scenario is a conflict scenario; when TTC∈(5s,+∞) or (-∞,0s), the hazard level of the scenario is a safe scenario. A joint distribution of TTC and ∆v in the real scene space is constructed. This joint distribution is equivalent to four existing scene hazard boundaries defined by TTC=0, TTC=1, TTC=3, and TTC=5. The boundary points on both sides of the joint distribution are used as separate training sample sets and input into a support vector regression algorithm for learning. Since the boundaries of this joint distribution are close to linear, a linear kernel function is chosen to solve for them, obtaining the scene hazard boundaries on both sides of the joint distribution. The scene hazard boundaries pre-defined based on TTC are merged with the scene hazard boundaries based on support vector regression to obtain the hazard boundaries in the real scene space: (20)； In the formula, l l With l r These represent the scene hazard boundaries on the left and right sides of the joint distribution map obtained through support vector regression, Δa and Δa, respectively. l and Δa r These represent the distances between scene B and the scene hazard boundary, l m1 l m2 l m3 and l m4 These are the scene hazard boundaries obtained from the TTC pre-classification; Map the above scenario hazard boundaries to the ideal scenario space, and set the scenario hazard level corresponding to the collision scenario to 1, the scenario hazard level corresponding to the emergency scenario to 0.7, the scenario hazard level corresponding to the conflict scenario to 0.3, and the scenario hazard level corresponding to the safe scenario to 0. Step 3: Generate key adaptive dynamic scenes based on multi-starting point optimization algorithm and seed filling algorithm. Calculate the key long-tail function value in the ideal scene space by using the scene occurrence probability and scene hazard level. Input the key long-tail function, ideal scene space, and key threshold into the multi-starting point optimization algorithm to solve for local key adaptive dynamic scenes. Input the local key scenes, key long-tail function, ideal scene space, and key threshold into the seed filling algorithm to solve for all key adaptive dynamic scenes. The critical long-tail function of the scene in the ideal scene space can be obtained by combining the scene occurrence probability obtained in step one with the scene danger level obtained in step two. The critical long-tail function, the ideal scene space, and the critical threshold γ are input into the multi-start point optimization algorithm. Some points are sampled in the ideal scene space by manual setting or random sampling as the starting point of the algorithm. The maximum value of the critical long-tail function in the attraction domain of each starting point is solved. The scene corresponding to the maximum value of all critical long-tail functions greater than γ is output, thus obtaining the local critical adaptive dynamic scene for low-dimensional evaluation scenarios. The local key scenes, key long-tail functions, ideal scene space, and key threshold γ output by the multi-start point optimization algorithm are input into the seed filling algorithm. Starting from each local key scene, the algorithm searches for scenes in the ideal scene space whose key long-tail function values ​​are greater than γ in the neighborhood around the starting point. These scenes are then used as new starting points to continue searching their neighborhoods. The above steps are repeated until the key long-tail function values ​​of all neighborhood scenes around the starting point are less than γ. At this point, the algorithm ends and outputs all the scenes marked as starting points, thus generating all the key adaptive dynamic scenes for low-dimensional evaluation scenarios. Step 4: Adaptive dynamic scene generation for high-dimensional evaluation scenarios, the specific process is as follows: Step 1: Constructing a scene hazard identification model based on the Hammerstein identification process and solving the scene occurrence probability based on the convex combination algorithm. The Hammerstein identification process consists of a series of static nonlinear and dynamic linear components. The relative state variables of the vehicle and the target vehicle, as well as the acceleration of the target vehicle, are used as inputs to the scene hazard identification model, while the acceleration of the vehicle is used as the output. After training with a large amount of scene data, the model parameters of the scene hazard identification model are used as key parameters characterizing the intrinsic attributes of scene hazard. Principal component analysis is used to decouple and reduce the dimensionality of the key parameters. Ant colony clustering algorithm is used to cluster the scene hazards. Representative "state-action" pairs are selected from each scene hazard level and input into the scene hazard identification model. Based on the cluster category of the parameters, a mapping relationship between the scene hazard level and the clustering results is established. The scene occurrence probability solution based on the convex combination algorithm can directly solve the occurrence probability of the "state-action" pairs according to the method for solving the scene occurrence probability of low-dimensional evaluation scenarios. Viewing the high-dimensional assessment scenario as a Markov decision process, the relative state variables of the vehicle and the target vehicle are considered as states, and the acceleration of the target vehicle is considered as an action. The state and action at the same time step are considered as a "state-action" pair. The relative state variables of the vehicle and the target vehicle at the same moment, as well as the acceleration of the target vehicle, are used as inputs to the scenario hazard identification model, and the acceleration of the vehicle is used as the output of the model. Therefore, the scenario hazard identification model is a multi-input single-output system. The Hammerstein identification process consists of a static nonlinear element and a dynamic linear element connected in series. The static nonlinear element uses a dead-zone function, a sigmoid function, or a saturation function. The z-transform of the dynamic linear element is shown in the following equation: (21)； In the formula, O p (k) represents the set of accelerations of the main vehicle, N(k) represents the set of outputs of the static nonlinear element, and d A(z) represents the order of the input delay and is defined as an integer multiple of the sampling time. -1 ) and B(z) -1 It can be calculated using the following formula: (22)； In the formula, (a1,…a q ) and (b1,…b n ) are all coefficients of the dynamic linear element, and q and n are the orders of the dynamic linear element; After training with a large amount of scenario data from natural driving data, the model parameters contained in the static nonlinear and dynamic linear links are key data that characterize the intrinsic attributes of scenario hazard. Therefore, they are used as data samples for scenario hazard assessment. In order to reduce the spatial dimension of the data samples as much as possible to improve computational efficiency while expressing the same model features, principal component analysis is used to decouple and reduce the dimensionality of the key parameters in the scenario hazard identification model. Let H represent the parameter dimension of the scene hazard identification model, and E represent the amount of training data. Then the model parameter dataset X can be represented as: (23)； In the formula, x i The intrinsic parameter vector representing the scene hazard identification model; Taking X as input, the principal component analysis algorithm defines the percentage of the sum of the eigenvalues ​​of the first m principal components to the sum of all eigenvalues ​​as the principal component contribution rate. The cumulative principal component contribution rate M is then calculated. m Calculated using the following formula: (24)； In the formula, λ i Represents the eigenvector; To ensure the dimensionality reduction effect, we take M. m The value of m corresponding to ≥85% is used as the dimension of the independent parameters of the model calculated by the algorithm, and finally an m×E matrix L is obtained: (25)； Ant colony clustering is used to cluster scene hazards. Taking L as the input to the ant colony clustering algorithm, scene hazard classification based on ant colony clustering is to find the partitioning method in L that minimizes the sum of distances from each data sample to the cluster centers with a known number of clusters. Referring to the scene hazard grading method for low-dimensional assessment scenes in step two of step three, the scene hazard is divided into 4 levels, resulting in 4 clusters. The ant colony clustering algorithm is expressed by the following formula: (26)； In the formula, J represents the sum of the distances from each data sample to the four cluster centers, l ip c represents the p-th model parameter feature of the i-th data sample. jp The p-th model parameter feature for the j-th class center is calculated using the following formula: (27)； In the formula, E j Let ω be the observed variable corresponding to the j-th class in observed variable E. ij The subordinate relationship indicator between the observed variable and the category is calculated using the following formula: (28)； The classification quality of ant colony clustering can be improved through iterative update equations, as shown in the following equation: (29)； In the formula, P ij Let τ be the probability of a data sample transitioning across classes. ij The standardized pheromone between data sample i and its class j is calculated using the following formula: (30)； In the formula, ρ represents the pheromone volatility, and t represents the time step; By iteratively updating the equation, the clustering results of the "state-action" pairs are obtained. Since the clustering results do not yet have physical meaning, it is necessary to establish a mapping relationship between the clustering results and the scene hazard level. Select some representative "state-action" pairs from each scene hazard level and input them into the scene hazard identification model. Based on the clustering category of their parameters, establish a mapping relationship between the scene hazard level and the clustering results. The probability of scene occurrence based on the convex combination algorithm is to directly solve the probability of occurrence of the "state-action" pair by following the method for solving the probability of scene occurrence in low-dimensional evaluation scenarios. Step 2: Reconstruct the key long-tail function based on the Markov decision process. By discretizing the decision variables, the ideal scenario space of the high-dimensional evaluation scenario is obtained. By viewing the high-dimensional evaluation scenario as a Markov decision process, the key long-tail function is reconstructed in the form of "state-action" pairs. The initial velocity of the target vehicle, the initial relative distance between the current vehicle and the target vehicle, the initial relative velocity between the current vehicle and the target vehicle, and the acceleration sequence of the target vehicle are used as decision variables x: (31)； In the formula, v o R represents the initial velocity of the target vehicle. o △v represents the initial relative distance between this vehicle and the target vehicle. o a represents the initial relative speed between this vehicle and the target vehicle. 0k This represents the acceleration of the target vehicle at the k-th time step; When setting v o The value range is [20m / s, 40m / s], and the distance from the walk is 2m / s; R o The value range is (0m, 90m], and the distance from the walking distance is 2m; △v o The value range is [-20m / s, 20m / s], and the distance from the walk is 2m / s; a 0k The value range is [-4m / s 2 2m / s 2 The walking distance is 0.2m / s. 2 When k is 10s, the number of ideal scene spaces is 21×45×21×31. 10 To reduce the dimensionality of the ideal scenario space and simplify the computational complexity of the key long-tail function, the high-dimensional evaluation scenario is treated as a Markov decision process. The relative distance and relative speed between the vehicle and the target vehicle are considered as states, and the acceleration of the target vehicle is considered as an action. The acceleration of the target vehicle at a given time step depends only on the relative states of the vehicle and the target vehicle at that time step. Therefore, the states and actions at the same time step are considered as a whole, i.e., a "state-action" pair. The number of "state-action" pairs in this ideal scenario space is 21 × 45 × 21 × 31 = 615195, which is significantly reduced compared to the number of scenarios in the previous ideal scenario space. The key long-tail function is reconstructed as follows: (32)； In the formula, s i a represents the state at time step i. i G(s) represents the action at time step i. i ,a i It can be calculated using the following formula: (33)； In the formula, V(s) i ,a i ) represents s i With a i The corresponding danger level of the "state-action" pair, P(s) i ,a i ) represents s i With a i The probability of occurrence of the corresponding "state-action" pair; Step 3: Generate key adaptive dynamic scenarios based on Q-learning algorithm. Construct Bellman equations, solve the objective function of Markov decision process through Bellman equations, and solve the optimal action sequence corresponding to each initial state by updating expected returns, which is the key adaptive dynamic scenario. Let Q(s,a) be the expected reward of the agent taking action a in state s, and r be the reward given by the environment when the agent takes action a. The Bellman equation is used to solve for the optimal policy in a Markov decision process: (34)； In the formula, π represents the policy, S represents the state set, R represents the reward set, ξ represents the discount factor, t represents the time step, s represents the state at time t, and V π (s) represents the cost function, and Q can be updated according to the following formula: (35)； In the formula, A represents the action set, a represents the action at time t, and α represents the learning rate; Thus, the optimal action sequence corresponding to each initial state can be obtained, which is the key adaptive dynamic scene; Step 5: Adaptive dynamic scenario evaluation for high- and low-dimensional assessment scenarios. The specific process is as follows: Step 1: Adaptive dynamic scene sampling based on the ϵ-greededy sampling strategy. Using the same sampling method, scenes are sampled from the adaptive dynamic scenes for both high-dimensional and low-dimensional evaluation scenarios. Taking the low-dimensional adaptive dynamic scene as an example, a small probability value ϵ is set. Scenes are randomly sampled from the low-dimensional adaptive dynamic scene with a probability of 1-ϵ. Scenes are also randomly sampled from the scenes outside the low-dimensional adaptive dynamic scene in the low-dimensional ideal scene space with a probability of ϵ. This forms a test scene library, named Test Group 1. The test scene library obtained in the high-dimensional case is named Test Group 2 using the same sampling method. Step 2: Adaptive dynamic scene evaluation based on accident rate and number of tests. Test group 1 and test group 2 test autonomous vehicles respectively, setting a certain confidence level until the accident rate converges. Record the number of tests and the accident rate when the accident rate converges for both test groups at this confidence level. Simultaneously, set up a control experiment, randomly sampling scenes from the low-dimensional ideal scene space and the high-dimensional ideal scene space respectively, denoted as control group 1 and control group 2, and test autonomous vehicles with the same confidence level until the accident rate converges. Record the number of tests and the accident rate convergence for both control groups at this confidence level. The number of tests and the accident rate were compared between test group 1 and control group 1, and between test group 2 and control group 2. The number of tests at which the accident rate converged was significantly higher than that of control group 1, and the number of tests at which the accident rate converged was significantly lower than that of control group 1. This indicates that the adaptive dynamic scenario evaluation for low-dimensional test scenarios is effective. If the accident rate of test group 2 is significantly higher than that of control group 2, and the number of tests at which the accident rate converged is significantly lower than that of control group 2, this indicates that the adaptive dynamic scenario evaluation for high-dimensional test scenarios is effective.

Images