Scene-behavior combined automatic driving vehicle acceleration test method

Abstract

The invention discloses a scene-behavior combined automatic driving vehicle acceleration test method, which comprises the following steps of: generating probability explicit acceleration test distribution by establishing scene and behavior combined distribution in a natural driving environment and introducing a risk indication function, and realizing combined sampling of a high-risk scene and an adversarial behavior. The initial scene is subjected to marginalization processing and combined with risk weighted screening to generate a high-risk scene, background vehicle behaviors are subjected to acceptance-rejection sampling to obtain an antagonism sequence, and the antagonism sequence is combined with the high-risk scene to construct a complete test scene. In a simulation test, an accident event is recorded, reweighting is carried out through importance sampling, unbiased accident rate estimation in a natural driving environment is obtained, and test convergence is judged according to the relative half width of the accident rate estimation. According to the method, the dynamic association between the scene and the behavior is effectively captured, the exposure rate of the high-risk event is improved, rapid, accurate and unbiased accident rate evaluation is realized, and the safety verification efficiency of the automatic driving vehicle is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of autonomous vehicle safety testing technology, specifically relating to a scenario-behavior joint acceleration testing method for autonomous vehicles. Background Technology

[0002] Highly Automated Vehicles (HAVs) have attracted widespread attention and research due to their enormous potential in driving safety and maneuverability. To ensure the safe operation of HAVs in complex traffic flow environments, their evaluation methods are crucial. Rapid estimation of the accident rate is one of the core testing indicators, directly reflecting the vehicle's safety performance boundaries and serving as a key presentation form of the evaluation results, providing core quantitative evidence for HAV safety verification. However, the complexity of traffic environments and the rarity of accidents (such as collisions and injuries) mean that traditional testing requires covering billions of miles to effectively estimate the accident rate, severely limiting evaluation efficiency.

[0003] Accelerated testing is an advanced method for testing autonomous driving systems. It achieves rapid and unbiased estimation of accident rates by increasing the exposure rate of high-risk, rare events and reweighting probability measures. Existing accelerated testing methods fall into two categories: one is based on initial cross-sectional scenarios, which adjusts the initial scenario distribution through methods such as cross-entropy and deep generative models to increase the proportion of high-risk scenarios, but ignores the dynamic impact of driving behavior. For example, CN118446012A discloses a virtual accelerated testing method for autonomous vehicles based on an optimal distribution model. This method extracts the probability distribution of key variables from a natural driving dataset, changes the original distribution characteristics of key variables through importance sampling, increases the sampling probability of high-risk events, and obtains the optimal distribution for accelerated testing. However, this method only adjusts the initial scenario distribution, fixing the driving behavior as a natural driving distribution, and cannot capture the synergistic effect of high-risk scenarios and adversarial behaviors. Its acceleration efficiency is limited by the strategy of only adjusting the initial scenario.

[0004] Another type is adversarial acceleration testing based on driving behavior. This method generates adversarial driving behaviors through reinforcement learning, but it uses a fixed initial scenario with a natural driving distribution, making it impossible to capture the synergistic effect of "high-risk scenario + adversarial behavior". However, accidents often depend on the combination of both, and adjusting the scenario or behavior alone can limit acceleration efficiency and make it difficult to capture the complex relationship between the two, further affecting the reliability of the evaluation results.

[0005] Furthermore, the high dimensionality, continuity, and dependency of scenarios and behaviors also pose challenges to joint modeling: driving scenarios are high-dimensional variables involving multiple vehicles and time sequences; scenarios and behaviors have dynamic dependencies; and each scenario in natural driving data typically contains only one behavior record, making it difficult to directly obtain the distribution of behavior conditions. Therefore, there is an urgent need for an accelerated testing method that can jointly model scenarios and behaviors while balancing efficiency and unbiasedness. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a scenario-behavior joint acceleration testing method for autonomous vehicles.

[0007] The objective of this invention can be achieved through the following technical solutions: This invention provides a scenario-behavior joint acceleration testing method for autonomous vehicles, comprising the following steps: Step S1: Model the natural driving environment, construct scene variables and behavior variables, and establish a natural driving distribution that includes the joint distribution of scene and behavior as well as the edge distribution of scene. Step S2: Introduce a risk indication mechanism, model the acceleration test environment based on the natural driving distribution, and construct the acceleration test distribution; Step S3: Obtain natural driving data and extract the probabilistic explicit acceleration test distribution from the natural driving data using a normalized flow; Step S4: Based on the probabilistic explicit accelerated test distribution, sample the initial scenario to generate a high-risk initial scenario; Step S5: Use the acceptance-rejection sampling method to sample the temporal behavior variables of the background vehicle, generate the adversarial behavior variable sequence of the background vehicle, and combine it with the high-risk initial scenario to form a complete test scenario; Step S6: Perform autonomous vehicle simulation testing based on the complete test scenario, obtain test results, and reweight the test results according to the distribution relationship between the natural driving environment and the acceleration test environment to obtain an accident rate estimate. Step S7: Determine whether the estimated accident rate meets the preset reliability requirements. If it does, output the test results and the estimated accident rate. If it does not, return to step S4 to continue sampling and testing.

[0008] Furthermore, the natural driving data includes vehicle operation data and environmental information data. The vehicle operation data includes vehicle position, speed, acceleration, and driving trajectory information. The environmental information data includes road structure information, traffic participant information, and traffic environment status information. The traffic participant information includes the status information of surrounding vehicles and other traffic entities. The traffic environment status information includes weather conditions and road adhesion conditions.

[0009] Furthermore, the scenario variable is , indicating the first The traffic scene state variables at the nth time step are used to characterize the nth time step. The environmental status of all vehicles within each time step, including vehicle position, speed, acceleration, road structure information, and traffic environment status information; The behavioral variables are , for the first The car in The driving behavior variables at the nth time step are used to characterize the corresponding vehicle at the nth time step. Decision-making behavior at each time step, including acceleration, steering angle and other driving operations.

[0010] Furthermore, the natural driving distribution is constructed as a chain probability model based on scene variables and behavioral variables, representing the entire driving process as a sequence driven by the initial scene and walking at each time step; the probability expression of the natural driving distribution is expressed as: in, for A set of scenario variables for each time step, used to describe the state evolution of all vehicles during the entire test period; This represents the total number of time steps. As the initial scene, For the first Scene variables at each time step, For the first The car in Behavioral variables at each time step The total number of vehicles. The scenario-behavior joint distribution represents the distribution of a given scenario variable. Next, the Vehicle driving behavior variables The probability of; The scene edge distribution represents the first... Time step scene variables The probability of occurrence; the decisions of different vehicles at the same time step are independent of each other; It represents the joint probability distribution of the entire driving process in a natural driving environment, used to quantify the probability of different scenarios and behavioral combinations occurring; in, , , The distribution is calculated using the Monte Carlo method based on natural driving data; the behavioral variables of different vehicles at the same time step are independent of each other, and the natural driving distribution is used to characterize the statistical characteristics of scenes and behaviors in the natural driving environment.

[0011] Furthermore, step S2 specifically includes: Introducing a risk indication function in the natural driving environment This is used to characterize the risk level of different scenarios and behavioral combinations, where the risk indicator function... , is represented as: in, For traffic safety index functions, it is used to quantify the risk level of scenario-behavior combinations, represented by the time-to-collision distance (TTC) between vehicles in highway scenarios or the estimated time to collision (PET) after intrusion in intersection scenarios; The risk indication function By combining the probability expression of the natural driving distribution with the weighted processing, the acceleration test environment is modeled and the acceleration test distribution is constructed.

[0012] Furthermore, the accelerated test distribution is expressed by the following probability expression: in, This represents the total number of time steps. The total number of vehicles; for A set of scene variables at each time step; This is the initial scene; For the first Scene variables at each time step; For the first The car in Driving behavior variables at each time step; For the joint probability distribution in the overall accelerated testing environment; To accelerate the joint distribution of scenarios and behaviors in the testing environment; To accelerate scene edge distribution in the testing environment; For scene-behavior joint distribution; It is distributed at the edge of the scene.

[0013] Furthermore, step S3 specifically includes: A sample sequence constructed from scene variables and behavioral variables extracted from natural driving data. Input a standardized flow model; the standardized flow model is a deep reversible generative model built on the reversible and differentiable affine coupling layer RealNVP, used to transform the distribution of simple prior latent variables. Mapped to a complex joint distribution of scene and behavior; A risk indicator function is introduced during model training. We weight the natural driving data and construct a weighted KL divergence loss function, which is expressed as: in, For risk indication function; For the standardized flow model Layer affine coupling transformation, Let be the Jacobian matrix of the transformation; For the prior distribution of latent variables; The distribution of accelerated testing environments is obtained by standardizing the flow model learning. Distribution by margin and constitute; n This represents the total number of affine coupling layers; The weighted KL divergence loss function; The normalized flow model is trained by minimizing the weighted KL divergence loss function, thereby making the distribution of the model output... It can characterize the probabilistic features of high-risk scenarios and behavior combinations in an accelerated testing environment, and obtain the probabilistic explicit accelerated testing distribution. .

[0014] Furthermore, step S4 specifically includes: Scene edge distribution in the probabilistic explicit accelerated testing distribution obtained based on the normalized flow model learning. ,exist The probability distribution of the initial scene variables is obtained directly at any time. ; Based on the probability distribution of the initial scene variables Perform random sampling to obtain initial scene samples ; Among them, the scene edge distribution in the probabilistic explicit accelerated test distribution Scenario variables from natural driving data are processed through a risk indicator function. Weighted samples for each scenario Corresponding to its source time sequence scenario and inherit the corresponding risk indicator function value; The initial scene sample obtained from sampling As a high-risk initial scenario.

[0015] Furthermore, step S5 specifically includes: At each time step For each background vehicle behavioral variables Sampling is performed to generate a sequence of adversarial behavior variables, and a complete test scenario is constructed. The specific sampling steps are as follows: Step S501: Calculate the walking distance of each background vehicle at each time point using the acceptance-rejection sampling method. The probability of acceptance is expressed as: in, Indicates the probability of acceptance; To accelerate the joint distribution of scenarios and behaviors in the testing environment, To accelerate scene edge distribution in the testing environment; For easy sampling, auxiliary distributions include uniform or Gaussian distributions; Let be the upper bound of the probability, satisfying ; Step S502: From uniform distribution Sampling, if Then accept Otherwise, resample; Repeat steps S501-S502 until the behavioral variables of all background vehicles at each time step are determined, forming a complete sequence of adversarial behavioral variables; combine the complete sequence of behavioral variables with the high-risk initial scenario to construct a complete test scenario.

[0016] Furthermore, the estimated accident rate is calculated using the following formula: in, This is an estimate of the accident rate. This represents the total number of samples in the complete test scenario. For the first A complete set of sequence of scenario variables and behavioral variables for a test scenario; For the natural driving distribution The joint probability of a sequence of scenes; To accelerate the test distribution under the first The joint probability of a sequence of scenes; For indicator functions, when the first A complete test scenario is assigned a value of 1 if an accident occurs during simulation testing, and 0 otherwise.

[0017] Compared with the prior art, the present invention has the following advantages: (1) In the prior art, accidents involving high-level autonomous vehicles are extremely rare. Traditional testing requires covering billions of miles to obtain effective accident statistics, resulting in extremely low testing efficiency and making it difficult to complete safety verification within a reasonable timeframe. This invention constructs an accelerated testing method based on a scenario-behavior joint distribution, utilizes a standardized flow model to generate a probabilistic explicit accelerated testing distribution, and reweights the sampling results to map the accelerated testing scenario back to the natural driving environment, thereby achieving a rapid and unbiased estimation of the accident rate. Through this technology, this invention significantly improves the efficiency of accident rate estimation, greatly shortens the testing time compared to natural driving environment testing, and maintains the statistical reliability of the results.

[0018] (2) Existing technologies in accelerated testing can only adjust the initial scenario or driving behavior individually, failing to capture the synergistic effect between high-risk scenarios and adversarial behaviors, thus limiting acceleration efficiency, especially in complex traffic flow and multi-vehicle interaction scenarios where it is difficult to generate high-risk events. This invention performs high-risk screening on the initial scenario and uses accept-rejection sampling to generate a sequence of adversarial behaviors of background vehicles at each time step, jointly modeling the dynamic relationship between scenarios and behaviors, thereby forming a high-risk scenario-behavior combination. This technical feature solves the problem that existing methods cannot capture the synergistic effect between the two, improving the efficiency of accelerated testing by tens to thousands of times compared to traditional methods, while ensuring the diversity and representativeness of the generated scenarios.

[0019] (3) Existing technologies struggle to handle high-dimensional, continuous, and time-dependent scenario and behavior data when generating accelerated test scenarios. Traditional methods cannot effectively represent the joint scenario-behavior distribution, leading to simulation results that fail to reflect the risk characteristics of the real driving environment. This invention employs a Normalized Flow (RealNVP) model based on a reversible, differentiable affine coupling layer to map a simple prior distribution into a complex joint scenario-behavior distribution. Furthermore, it weights the training data using a risk indicator function to obtain a probabilistically explicit high-risk accelerated test distribution. This technology addresses the difficulty of modeling high-dimensional continuous variables, enabling the generated accelerated test distribution to accurately reflect the probabilistic characteristics of high-risk areas, ensuring the scientific rigor and unbiasedness of the simulation test, while simultaneously improving simulation accuracy and safety event capture rate.

[0020] (4) Existing technologies suffer from high computational costs and low sampling efficiency when generating adversarial behavior sequences of background vehicles, especially in high-dimensional time-series environments, which can easily lead to complex engineering implementation and limit the scale and practicality of testing. In the acceptance-rejection sampling process, this invention uses a truncated Gaussian mixture or uniform distribution as an easy-to-sample auxiliary distribution and optimizes the upper bound of sampling, significantly reducing the computational cost of behavior sequence generation and sampling, improving the acceptance rate, and making the generation of large-scale test scenarios feasible. This technical feature not only improves engineering applicability but also ensures the statistical representativeness of the generated adversarial behaviors, further improving the reliability of accident rate estimation and testing efficiency.

[0021] (5) Existing technologies struggle to guarantee the unbiasedness of the final accident rate estimate in accelerated testing, especially when generating rare events through accelerated scenarios. This makes it difficult to effectively map back to the natural driving probability distribution, potentially leading to biased safety assessment results. This invention introduces importance sampling weights into the simulation test results and reweights them using the joint probability ratio of the natural driving distribution and the accelerated test distribution, achieving unbiased estimation of the accident rate. This technical feature solves the bias problem of existing methods, enabling accelerated test results to not only quickly generate high-risk events but also ensure statistical accuracy, providing a reliable quantitative basis for HAV safety assessment.

[0022] (6) Existing technologies lack flexibility and scalability in handling scenario-behavior joint distributions, failing to adapt to the testing needs of different types of scenarios or complex scenarios involving multiple vehicle interactions. This invention describes temporal scenario-behavior sequences using a chain-like probability model, and combines a risk indicator function with a standardized flow model, enabling flexible sampling and generation for different scenario types, vehicle numbers, and behavioral strategies. This technology enhances the scalability and applicability of the method, efficiently generating high-risk test scenarios in diverse traffic environments, improving HAV test coverage and scenario richness, while maximizing the synergistic advantages of joint modeling. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the acceleration testing method for autonomous vehicles according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the natural driving environment and acceleration test environment in steps S1 and S2 of the present invention; Figure 3 This is a structural diagram of the standardized flow model in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] Example 1: This implementation case presents an accelerated testing method for autonomous vehicles based on explicit probabilistic reasoning, specifically a scenario-behavior joint testing approach (hereinafter referred to as the accelerated testing method). The core idea is as follows: extract the joint scenario-behavior distribution and scenario edge distribution from natural driving data; simplify high-dimensional driving environment modeling based on Markov decision processes; shift the distribution towards high-risk areas through normalizing flow (NF) to generate an accelerated driving environment; combine importance sampling (IS) to ensure the evaluation results are unbiased towards the natural driving environment; and finally, generate high-risk scenarios and adversarial behaviors through acceptance-rejection sampling to complete the safety performance evaluation of the autonomous vehicle. The overall technical route is as follows: Figure 1 As shown, the experimental verification was carried out in two typical scenarios: a three-lane highway and an unprotected left-turn intersection.

[0026] Specifically, the following steps are included: Step S1: Model the natural driving environment, construct scene variables and behavior variables, and establish a natural driving distribution including the joint distribution of scene and behavior, as well as the edge distribution of the scene, such as... Figure 2 As shown, it includes: Creating a natural driving environment: Probabilistic features of scenes and behaviors are extracted from the natural driving dataset to construct the scene distribution of the natural driving environment, providing a foundation for subsequent acceleration distribution generation. The specific process is as follows: (1) Extract probabilistic features of scenarios and behaviors from the natural driving dataset, and obtain the state information of each vehicle at different time steps through data statistics and processing. Definition T Each time step and K The set of scene variables for each vehicle is The state of each vehicle at each time step includes parameters such as position, velocity, acceleration, and road conditions, which can be represented as: in, For the first The car is The state variable at any given time.

[0027] (2) Scene probability decomposition: In order to reduce sampling complexity and capture dynamic correlations, the following steps are taken: Following the chain decomposition of the Markov decision process, a structure of "initial scenario + sequential behavior" is formed. Define the first... t The time frame is: The driving behavior variable is defined as Based on vehicle dynamics and traffic physics constraints, the evolution of the scene over time is represented by a state transition equation. ,in, This represents a combined function of vehicle kinematics and road constraints, used to ensure the physical feasibility of the generated scene. This decomposition method maps high-dimensional temporal scenes to an initial scene and a sequence of behaviors at each time step, so that only one behavior sequence needs to be generated for each sampling, while preserving the dynamic correlation between vehicles and improving modeling efficiency.

[0028] Furthermore, by combining the conditional probability formula, it can be decomposed into a joint distribution of scene and behavior. Distribution with scene edges The quotient form reduces the dimensionality of a single sampling in the scene: in, For the first The car is Actions at any moment Describe the probability of a scenario and an action occurring simultaneously. Describes the probability of a scenario occurring alone. The scenario variable is... , indicating the first The traffic scene state variables at the nth time step are used to characterize the nth time step. The environmental status of all vehicles within each time step, including vehicle position, speed, acceleration, road structure information, and traffic environment status information; behavioral variables are , for the first The car in The driving behavior variables at the nth time step are used to characterize the corresponding vehicle at the nth time step. Decision-making behavior at each time step, including acceleration, steering angle and other driving operations.

[0029] (3) Quantification of safety-critical events. Introducing indicator functions. Marking an incident: when the scene When an accident occurs Otherwise, it is 0; the accident rate is calculated using the Monte Carlo method. The expression is: in, For the number of samples, From The first sampled A scenario.

[0030] Step S2: Introduce a risk indication mechanism, model the acceleration test environment based on the natural driving distribution, and construct the acceleration test distribution; To increase accident exposure, a risk indicator function and importance sampling are introduced, and the natural driving distribution is reweighted into an acceleration test distribution while ensuring unbiased evaluation. The specific steps are as follows: (1) Definition of risk indicator function. A risk indicator function is used. The distribution is being guided to shift towards high-risk areas, among which These are key indicators for traffic safety. For highway scenarios, Time-to-Collision (TTC, range 0-20s) is used; for intersection scenarios, Time After Intrusion (PET, range 0-10s) is used. A smaller value indicates a higher scenario risk, thus affecting the distribution adjustment. Scenarios that are high-risk will receive higher sampling weights, allowing for focused attention on high-risk areas. This design can increase the probability of incidents occurring during testing without altering the physical characteristics and behavioral rationality of the scenario, thereby significantly accelerating the convergence speed of incident rate estimation.

[0031] (2) Accelerating test distribution generation. Based on importance sampling theory, an accelerated test distribution is constructed. Its structure and Symmetry, only through Adjust the distribution weights: in, , ; Indicates the first k The car is Behavioral variables at any given time For the joint probability of scene and behavior under the natural driving distribution, The sampling is distributed along the edge of the scene. By explicitly amplifying the originally low-probability high-risk scenes, the sampling is more concentrated in potentially dangerous areas, which can effectively increase the sampling frequency of high-risk events, reduce the number of samples required for testing, improve testing efficiency, and provide rich high-risk data for subsequent simulation analysis.

[0032] (3) Unbiased weighted accident rate. The assessment results of the accelerated testing environment are mapped back to the natural driving environment through importance sampling to ensure... The unbiased estimate is expressed as: in, From The first sampled One scenario, This is the probability reweighting factor.

[0033] By guiding sampling through a risk indicator function, the sampling frequency of high-risk events is significantly increased, reducing the sample size required for accident rate estimation. Probability reweighting is performed through importance sampling to achieve unbiased estimation of the accident rate. The distribution structure is symmetrical with the natural driving distribution, ensuring physical rationality and statistical consistency. Combined with scenario-behavior joint modeling, the dynamic correlation characteristics of multi-vehicle and temporal behavior are preserved, making the acceleration test results reliable and interpretable.

[0034] Step S3: Acquire natural driving data and extract the probabilistic explicit acceleration test distribution from the natural driving data using a normalized flow; The NF model is used to learn the distribution characteristics of natural driving and acceleration testing environments. NF maps simple latent distributions to complex target distributions through invertible and differentiable transformations, and can provide explicit probability densities, meeting the requirements of unbiased evaluation. The model structure is as follows: Figure 3 As shown, the specific implementation is as follows: (1) NF model structure selection. A RealNVP (Real-valued Non-Volume Preserving) structure is chosen, achieving reversible transformation through multiple affine coupling layers. For D-dimensional input... and Affine coupling layer output satisfy: in and These are the scale and offset parameters learned through a neural network.

[0035] (2) Distribution learning and loss function. When learning the natural driving distribution, KL divergence is used. The loss function minimizes the data distribution. Learning distribution with the model The difference in driving patterns leads to the following loss function when learning the natural driving distribution: When learning the accelerated test distribution, the loss function is adjusted to... ,pass The weighted guiding model shifts towards high-risk areas.

[0036] in, , The distribution learned by the model, As a latent variable, Represents affine coupling layer transformation. Let be the Jacobian matrix of the transformation.

[0037] (3) Edge distribution reconstruction. An edge-truncated Gaussian mixture distribution is designed. By using the scene-behavior joint distribution learned by NF, edge distributions of arbitrary dimensions (such as scene edge distribution and single vehicle behavior distribution) can be directly generated without training a model independently for each vehicle or scene, thus reducing the difficulty of engineering implementation.

[0038] in represent dimensional variables, To truncate the normal distribution, and Let be the mean and covariance of a normal distribution. This is to truncate the interval.

[0039] The NF model allows for the simultaneous learning of natural driving distributions and accelerated testing distributions, enabling explicit probabilistic modeling and meeting the requirements for unbiased accident rate assessment. Weighted KL divergence guides the model to focus on high-risk areas, increasing the sampling frequency of high-risk events and significantly improving accident rate estimation efficiency. Reversible transformations maintain data interpretability and density computability, providing a reliable probabilistic basis for subsequent accept-rejection sampling to generate complete test scenarios, enhancing the scientific rigor and engineering feasibility of accelerated testing. Advantages include strong high-dimensional scene-behavior sequence modeling capabilities, high sampling efficiency, and statistically reliable evaluation results.

[0040] Step S4: Based on the probabilistic explicit accelerated test distribution, sample the initial scenario to generate a high-risk initial scenario, specifically including: Scene edge distribution in the probabilistic explicit accelerated testing distribution obtained based on the normalized flow model learning. ,exist The probability distribution of the initial scene variables is obtained directly at any time. ; Based on the probability distribution of the initial scene variables Perform random sampling to obtain initial scene samples ; Among them, the scene edge distribution in the probabilistic explicit accelerated test distribution Scenario variables from natural driving data are processed through a risk indicator function. Weighted samples for each scenario Corresponding to its source time sequence scenario and inherit the corresponding risk indicator function value; The initial scene sample obtained from sampling As a high-risk initial scenario.

[0041] By marginalizing the joint distribution, the dimensionality of the initial scene sampling is reduced, improving computational efficiency. A risk indicator function is used to weight and filter high-risk initial scenes, achieving efficient coverage of rare incidents. Probability-proportional filtering ensures priority sampling of high-risk events, while combining importance sampling theory maintains statistical unbiasedness in incident rate estimation. Advantages include significantly improved sampling efficiency for high-risk scenes, preservation of probabilistic information integrity, and a reliable foundation for generating complete test scenarios.

[0042] Step S5: Use the acceptance-rejection sampling method to sample the temporal behavior variables of the background vehicle, generate the adversarial behavior variable sequence of the background vehicle, and combine it with the high-risk initial scenario to form a complete test scenario; At each time step For each background vehicle behavioral variables Sampling is performed to generate a sequence of adversarial behavior variables, and a complete test scenario is constructed. The specific sampling steps are as follows: Step S501: Calculate the walking distance of each background vehicle at each time point using the acceptance-rejection sampling method. The probability of acceptance is expressed as: in, Indicates the probability of acceptance; To accelerate the joint distribution of scenarios and behaviors in the testing environment, To accelerate scene edge distribution in the testing environment; For easy sampling, auxiliary distributions include uniform or Gaussian distributions; Let be the upper bound of the probability, satisfying ; Step S502: From uniform distribution Sampling, if Then accept Otherwise, resample; Repeat steps S501-S502 until the behavioral variables of all background vehicles at each time step are determined, forming a complete sequence of adversarial behavioral variables; combine the complete sequence of behavioral variables with the high-risk initial scenario to construct a complete test scenario.

[0043] Accelerating test distributions through acceptance-rejection sampling Mapping to a sampleable auxiliary distribution Above, retain the joint probability features of high-risk scenarios and behaviors; acceptance probability The settings ensure that the sampling results match the target distribution. Maintain consistency and avoid bias. Effectively generate high-risk adversarial behavior sequences for background vehicles to increase the exposure rate of accident events in the simulation, while ensuring statistical unbiasedness; rationally design auxiliary distributions and upper bounds. M It can balance sampling efficiency and accuracy, and reduce computational costs; through serialized sampling, it ensures the synergistic effect of complete temporal behavior and initial high-risk scenarios, thereby improving the realism and reliability of accelerated testing.

[0044] Step S6: Perform autonomous vehicle simulation testing based on the complete test scenario, obtain test results, and reweight the test results according to the distribution relationship between the natural driving environment and the acceleration test environment to obtain an accident rate estimate, specifically including: In the complete test scenario The following is the process of generating the behavior of a simulated autonomous vehicle: First, the initial scenario is input into the autonomous vehicle (HAV) decision-making module to obtain its status in the current scenario. Behavioral decision-making Subsequently, the background vehicle behavior will be... With autonomous vehicle behavior Execute synchronously to obtain the scene at the next moment. Record scene evolution and incident markers .

[0045] This process occurs at time step The test runs in a loop until an accident occurs or test resources are exhausted. The entire test process includes the following steps: 1. Sampling high-risk initial scenarios ; 2. Sampling Background: Vehicle Adversarial Behavior ; 3. Obtaining the behavior of autonomous vehicles ; 4. Execute all vehicle actions and record the scene at the next moment. ; 5. Repeat steps 2-4 until an accident occurs or all time steps are completed; 6. When relative half width Stop testing when necessary to ensure the confidence level and accuracy of the accident rate estimate, whereby... Take 1.96 (the quantile corresponding to the 95% confidence level). for The standard deviation of the estimated value.

[0046] Stop test relative half-width The calculation formula is in These are the quantiles corresponding to the confidence level. This represents the standard deviation of the accident rate estimate.

[0047] Importance sampling maps high-risk scenarios in the accelerated testing environment back to the natural driving distribution, achieving unbiased estimation of accident rates. Temporal interactions between autonomous vehicles and background vehicles maintain the realistic characteristics of scenario-behavior joint dynamic evolution. Introducing relative half-width judgment automatically determines the adequacy of testing resources, improving testing efficiency and ensuring statistical confidence. This significantly improves the exposure rate of accident events, achieving fast, accurate, and reliable accident rate estimation while controlling simulation resource consumption.

[0048] Step S7: Determine whether the pre-set reliability requirements are met based on the accident rate estimate. If they are met, output the test results and the accident rate estimate. If they are not met, return to step S4 to continue sampling and testing.

[0049] Example 2: This embodiment demonstrates experimental verification; in high-speed scenarios, the acceleration test only requires... A single sampling is sufficient to meet the accuracy requirements, which is more than what is needed in a natural driving environment. In a single sampling, the speedup ratio reached 1181 times; in intersection scenarios, the acceleration test only required... Secondary sampling, compared to natural driving environments The acceleration ratio reached 78 times in the second sampling. The weighted accident rate distribution in the acceleration test was not significantly different from that in the natural driving environment. The relative error of the accident rate estimation was 5.2% in the high-speed scenario and 3.1% in the intersection scenario. In the ablation experiment, the efficiency of the scenario-behavior joint acceleration was significantly better than that of the single scenario acceleration and the single behavior acceleration, which proved the value of the synergistic effect of the two. The results are shown in Table 1.

[0050] Table 1. Results of accelerated testing In summary, this embodiment achieves "high efficiency and unbiasedness" in the safety assessment of autonomous vehicles through joint scene-behavior modeling, standardized flow distribution generation, and importance sampling. It demonstrates excellent performance in both highway and intersection scenarios, providing key technical support for safety verification before the large-scale deployment of autonomous vehicles.

[0051] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0052] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A scenario-behavior joint acceleration testing method for autonomous vehicles, characterized in that, Includes the following steps: Step S1: Model the natural driving environment, construct scene variables and behavior variables, and establish a natural driving distribution that includes the joint distribution of scene and behavior as well as the edge distribution of scene. Step S2: Introduce a risk indication mechanism, model the acceleration test environment based on the natural driving distribution, and construct the acceleration test distribution; Step S3: Obtain natural driving data and extract the probabilistic explicit acceleration test distribution from the natural driving data using a normalized flow; Step S4: Based on the probabilistic explicit accelerated test distribution, sample the initial scenario to generate a high-risk initial scenario; Step S5: Use the acceptance-rejection sampling method to sample the temporal behavior variables of the background vehicle, generate the adversarial behavior variable sequence of the background vehicle, and combine it with the high-risk initial scenario to form a complete test scenario; Step S6: Perform autonomous vehicle simulation testing based on the complete test scenario, obtain test results, and reweight the test results according to the distribution relationship between the natural driving environment and the acceleration test environment to obtain an accident rate estimate. Step S7: Determine whether the estimated accident rate meets the preset reliability requirements. If it does, output the test results and the estimated accident rate. If it does not, return to step S4 to continue sampling and testing.

2. The scenario-behavior joint acceleration testing method for autonomous vehicles according to claim 1, characterized in that, The natural driving data includes vehicle operation data and environmental information data. The vehicle operation data includes vehicle position, speed, acceleration, and driving trajectory information. The environmental information data includes road structure information, traffic participant information, and traffic environment status information. The traffic participant information includes the status information of surrounding vehicles and other traffic entities. The traffic environment status information includes weather conditions and road adhesion conditions.

3. The scenario-behavior joint acceleration testing method for autonomous vehicles according to claim 1, characterized in that, The scenario variables are , indicating the first The traffic scene state variables at the nth time step are used to characterize the nth time step. The environmental status of all vehicles within each time step, including vehicle position, speed, acceleration, road structure information, and traffic environment status information; The behavioral variables are , for the first The car in The driving behavior variables at the nth time step are used to characterize the corresponding vehicle at the nth time step. Decision-making behavior at each time step, including acceleration, steering angle and other driving operations.

4. The scenario-behavior joint acceleration testing method for autonomous vehicles according to claim 1, characterized in that, The natural driving distribution is constructed as a chain probability model based on scene variables and behavioral variables, representing the entire driving process as a sequence driven by the initial scene and actions at each time step; the probability expression of the natural driving distribution is as follows: in, for A set of scenario variables for each time step, used to describe the state evolution of all vehicles during the entire test period; This represents the total number of time steps. As the initial scene, For the first Scene variables at each time step, For the first The car in Behavioral variables at each time step The total number of vehicles. The scenario-behavior joint distribution represents the distribution of a given scenario variable. Next, the Vehicle driving behavior variables The probability of; The scene edge distribution represents the first... Time step scene variables The probability of occurrence; the decisions of different vehicles at the same time step are independent of each other; It represents the joint probability distribution of the entire driving process in a natural driving environment, used to quantify the probability of different scenarios and behavioral combinations occurring; in, , , The distribution is calculated using the Monte Carlo method based on natural driving data; the behavioral variables of different vehicles at the same time step are independent of each other, and the natural driving distribution is used to characterize the statistical characteristics of scenes and behaviors in the natural driving environment.

5. The scenario-behavior joint acceleration testing method for autonomous vehicles according to claim 1, characterized in that, Step S2 specifically includes: Introducing a risk indication function in the natural driving environment This is used to characterize the risk level of different scenarios and behavioral combinations, where the risk indicator function... , is represented as: in, For traffic safety index functions, it is used to quantify the risk level of scenario-behavior combinations, represented by the time-to-collision distance (TTC) between vehicles in highway scenarios or the estimated time to collision (PET) after intrusion in intersection scenarios; The risk indication function By combining the probability expression of the natural driving distribution with the weighted processing, the acceleration test environment is modeled and the acceleration test distribution is constructed.

6. The scenario-behavior joint acceleration testing method for autonomous vehicles according to claim 5, characterized in that, The accelerated test distribution is expressed by the following probability expression: in, This represents the total number of time steps. The total number of vehicles; for A set of scene variables at each time step; This is the initial scene; For the first Scene variables at each time step; For the first The car in Driving behavior variables at each time step; For the joint probability distribution in the overall accelerated testing environment; To accelerate the joint distribution of scenarios and behaviors in the testing environment; To accelerate scene edge distribution in the testing environment; For scene-behavior joint distribution; It is distributed at the edge of the scene.

7. The scenario-behavior joint acceleration testing method for autonomous vehicles according to claim 1, characterized in that, Step S3 specifically includes: A sample sequence constructed from scene variables and behavioral variables extracted from natural driving data. Input a standardized flow model; the standardized flow model is a deep reversible generative model built on the reversible and differentiable affine coupling layer RealNVP, used to transform the distribution of simple prior latent variables. Mapped to a complex joint distribution of scene and behavior; A risk indicator function is introduced during model training. We weight the natural driving data and construct a weighted KL divergence loss function, which is expressed as: in, For risk indication function; For the standardized flow model Layer affine coupling transformation, Let be the Jacobian matrix of the transformation; For the prior distribution of latent variables; The distribution of accelerated testing environments is obtained by standardizing the flow model learning. Distribution by margin and constitute; n This represents the total number of affine coupling layers; The weighted KL divergence loss function; The normalized flow model is trained by minimizing the weighted KL divergence loss function, thereby making the distribution of the model output... It can characterize the probabilistic features of high-risk scenarios and behavior combinations in an accelerated testing environment, and obtain the probabilistic explicit accelerated testing distribution. .

8. The scenario-behavior joint acceleration testing method for autonomous vehicles according to claim 1, characterized in that, Step S4 specifically includes: Scene edge distribution in the probabilistic explicit accelerated testing distribution obtained based on the normalized flow model learning. ,exist The probability distribution of the initial scene variables is obtained directly at any time. ; Based on the probability distribution of the initial scene variables Perform random sampling to obtain initial scene samples ; Among them, the scene edge distribution in the probabilistic explicit accelerated test distribution Scenario variables from natural driving data are processed through a risk indicator function. Weighted samples for each scenario Corresponding to its source time sequence scenario and inherit the corresponding risk indicator function value; The initial scene sample obtained from sampling As a high-risk initial scenario.

9. The scenario-behavior joint acceleration testing method for autonomous vehicles according to claim 1, characterized in that, Step S5 specifically includes: At each time step For each background vehicle behavioral variables Sampling is performed to generate a sequence of adversarial behavior variables, and a complete test scenario is constructed. The specific sampling steps are as follows: Step S501: Calculate the walking distance of each background vehicle at each time point using the acceptance-rejection sampling method. The probability of acceptance is expressed as: in, Indicates the probability of acceptance; To accelerate the joint distribution of scenarios and behaviors in the testing environment, To accelerate scene edge distribution in the testing environment; For easy sampling, auxiliary distributions include uniform or Gaussian distributions; Let be the upper bound of the probability, satisfying ; Step S502: From uniform distribution Sampling, if Then accept Otherwise, resample; Repeat steps S501-S502 until the behavioral variables of all background vehicles at each time step are determined, forming a complete sequence of adversarial behavioral variables; combine the complete sequence of behavioral variables with the high-risk initial scenario to construct a complete test scenario.

10. The scenario-behavior joint acceleration testing method for autonomous vehicles according to claim 1, characterized in that, The formula for the accident rate estimate is as follows: in, This is an estimate of the accident rate. This represents the total number of samples in the complete test scenario. For the first A complete set of sequence of scenario variables and behavioral variables for a test scenario; For the natural driving distribution The joint probability of a sequence of scenes; To accelerate the test distribution under the first The joint probability of a sequence of scenes; For indicator functions, when the first A complete test scenario is assigned a value of 1 if an accident occurs during simulation testing, and 0 otherwise.

Images