Automatic driving edge scene generation method and device, electronic equipment and storage medium

Abstract

The present application belongs to the field of automatic driving technology test and verification, and particularly relates to an automatic driving edge scene generation method and device, electronic equipment and storage medium. The method comprises the following steps: taking a driving scene element parameter space as a basis, combining a fidelity model and a multi-dimensional value function to determine an initial observation data set; constructing a Gaussian process model based on the initial observation data set; obtaining a final observation data set based on the driving scene element parameter space, the Gaussian process model, an edge perception acquisition function, a global optimization algorithm, the fidelity model and the multi-dimensional value function; and obtaining an automatic driving edge scene based on the final observation data set and the multi-dimensional value function, thereby solving the technical problem of high search and calculation cost and incomplete coverage of automatic driving edge scenes.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology testing and verification, and in particular to autonomous driving edge scene generation methods, devices, electronic devices and storage media. Background Technology

[0002] As autonomous driving technology continues to evolve towards Level 3 and above, the demand for testing and verification of autonomous driving systems is growing exponentially. Simulation environments have become the core support platform for generating qualified autonomous driving test scenarios. According to relevant autonomous driving testing specifications, the generated autonomous driving test scenarios must simultaneously meet the requirements of spatial completeness, temporal completeness, and interactive completeness. Various autonomous driving test scenarios in the simulation environment can be uniformly represented through a high-dimensional parameter space. This space specifically refers to the Cartesian product space composed of multi-dimensional continuous variables such as ambient lighting, road friction coefficient, traffic flow density, obstacle trajectory, and vehicle motion state. Within this high-dimensional parameter space, the vast majority of regions correspond to conventional safety-type autonomous driving scenarios, while only a very small portion corresponds to autonomous driving edge scenarios. This extremely sparse distribution of autonomous driving edge scenarios is defined as the "curse of dimensionality," which has become a core bottleneck restricting the safety verification process of autonomous driving technology. Autonomous driving edge scenarios specifically refer to critical scenarios that can trigger autonomous driving system failure or are on the verge of system failure.

[0003] To overcome the aforementioned bottlenecks, existing conventional technologies mainly employ four methods to generate autonomous driving test scenarios: grid search-based high-dimensional parameter space scenario generation methods, random sampling-based high-dimensional parameter space scenario generation methods, rule-based scenario generation methods, and data-driven scenario generation methods. All four methods suffer from technical drawbacks: the sampling complexity of grid search-based methods increases exponentially with the parameter dimension, leading to a sharp increase in scenario generation costs. For example, in an 8-dimensional parameter space, if 10 sampling points are selected in a single dimension, the grid search needs to perform 10... 8 The computational complexity of this simulation evaluation is completely impractical in engineering practice; while random sampling-based methods are unbiased, they are limited by the limited simulation resource budget and cannot accurately target sparsely distributed autonomous driving edge scenarios; rule-based methods rely entirely on the experience of domain experts to design test cases, making it difficult to cover the potential risks caused by nonlinear parameter coupling; data-driven methods do not establish a pre-quantification mechanism for scenario hazard levels, the generation process is blind, and they cannot effectively distinguish between autonomous driving edge scenarios and regular dangerous scenarios, resulting in a waste of test resources.

[0004] None of the four conventional methods mentioned above have established a mathematical quantification standard for "edge degree", making it impossible to actively search for critical scenarios with extremely high testing value. At the same time, they all use a uniform high-fidelity simulation mode for all sampled samples, failing to utilize the kinematic simplification low-fidelity model with extremely low computational cost to quickly screen the parameter space in the early stages of optimization, ultimately resulting in a continuous limitation on the efficiency of mining edge scenarios for autonomous driving.

[0005] With technological iteration, the industry has gradually introduced two types of improved scene generation technologies: one is a key safety scene generation system for autonomous driving based on diffusion models, and the other is a method for generating key test scenes for autonomous driving that integrates visual language models and diffusion models. However, both of these improved technologies still have technical shortcomings: the diffusion model-based method can generate driving trajectories that fit real road conditions, but it lacks directional optimization capabilities for scene search targeting specific safety goals. This not only consumes a lot of computing resources but is also prone to getting stuck in local optima, making it difficult to efficiently mine edge scenes for autonomous driving. The method that integrates visual language models and diffusion models can guide scene generation through high-level semantics, but the iterative process relies on a large number of simulation evaluations, resulting in low sample utilization efficiency. In addition, when dealing with high-dimensional continuous parameter spaces composed of multi-dimensional coupled variables such as weather illumination, traffic flow density, and traffic participant behavior parameters, it is impossible to achieve a balance between "exploring unknown dangerous areas" and "utilizing known dangerous patterns," further exacerbating the problem of insufficient efficiency in mining edge scenes for autonomous driving.

[0006] In summary, the field of autonomous driving testing urgently needs a scene generation technology that has high sample utilization efficiency, value orientation, and edge perception capabilities to solve the technical problems of high computational cost and incomplete coverage in autonomous driving edge scene search under high-dimensional parameter space. Summary of the Invention

[0007] In order to solve the above-mentioned technical problems, or at least partially solve the above-mentioned technical problems, this application provides an autonomous driving edge scene generation method, apparatus, electronic device and storage medium.

[0008] Firstly, this application provides a method for generating edge scenes for autonomous driving, including: Based on the parameter space of driving scene elements, and combined with the fidelity model and multidimensional value function, the initial observation dataset is determined; Based on the initial observation dataset, a Gaussian process model is constructed; Based on the driving scene element parameter space, Gaussian process model, edge perception acquisition function, global optimization algorithm, fidelity model, and multidimensional value function, the final observation dataset is obtained. Based on the final observation dataset and the multidimensional value function, the edge scenario of autonomous driving is obtained.

[0009] A further optional implementation involves obtaining the final observation dataset based on the driving scene element parameter space, Gaussian process model, edge-aware acquisition function, global optimization algorithm, fidelity model, and multidimensional value function, including: Based on the driving scene element parameter space and the initial observation dataset, combined with the Gaussian process model, edge perception acquisition function, global optimization algorithm, fidelity model and multidimensional value function, an intermediate observation dataset is obtained; Based on the intermediate observation dataset, the Gaussian process model is optimized until the Gaussian process model converges, resulting in a converged Gaussian process model. Based on the driving scene element parameter space, combined with the convergent Gaussian process model, edge-aware acquisition function, global optimization algorithm, fidelity model, and multidimensional value function, the final observation dataset is obtained.

[0010] In a further optional implementation, the fidelity model includes a low-fidelity model; Based on the parameter space of driving scene elements, and combined with a fidelity model and a multidimensional value function, the initial observation dataset is determined, including: A space-filling sampling strategy is adopted to uniformly collect multiple sample points in the driving scene element parameter space; The low-fidelity model is used to process each sample point to obtain the physical feature set corresponding to each sample point; The physical feature set of each sample point is processed using the multidimensional value function to obtain the observation value corresponding to each sample point; The initial observation dataset is composed of all the sample points and their corresponding observation values.

[0011] A further optional implementation involves constructing a Gaussian process model based on the initial observation dataset, including: The initial observation dataset is processed using a Gaussian process surrogate model to obtain a Gaussian process model.

[0012] A further optional implementation involves obtaining an intermediate observation dataset based on the driving scene element parameter space and the initial observation dataset, combined with the Gaussian process model, edge-aware acquisition function, global optimization algorithm, fidelity model, and multidimensional value function. This intermediate observation dataset includes: Using the Gaussian process model, edge-aware acquisition function, and global optimization algorithm, candidate points are selected in the parameter space of the driving scene elements; Based on the candidate points, and combined with the fidelity model and the multidimensional value function, the candidate observations corresponding to the candidate points are obtained. The observation dataset is updated with the candidate points and their corresponding candidate observations, and the Gaussian process model is updated based on the updated initial observation dataset. Based on the updated Gaussian process model, the operations of selecting candidate points in the driving scene element parameter space, updating the initial observation dataset, and updating the Gaussian process model are repeatedly performed until a preset number of repetitions are reached to obtain the intermediate observation dataset.

[0013] A further optional implementation involves using the Gaussian process model, edge-aware acquisition function, and global optimization algorithm to select candidate points in the driving scene element parameter space, including... The predicted mean coefficient and the exploration weight coefficient in the Gaussian process model are used to update the predicted mean coefficient and the exploration weight coefficient in the edge sensing acquisition function to obtain the updated edge sensing acquisition function. Candidate points are selected in the driving scene element parameter space using the global optimization algorithm and the updated edge perception acquisition function.

[0014] In a further optional implementation, the fidelity model includes a high-fidelity model and a low-fidelity model; Based on the candidate points, combined with the fidelity model and multidimensional value function, candidate observations are obtained. The observation dataset is updated with the convergence points and their corresponding candidate observations, including: The updated edge sensing acquisition function is used to process the candidate points to obtain the sensing values ​​of the candidate points; The perceived value of the candidate point is compared with a preset perception threshold; if the perceived value of the candidate point is not less than the preset perception threshold, the candidate point is processed using the low-fidelity model to obtain the physical feature set of the candidate point; if the perceived value of the candidate point is less than the preset perception threshold, the candidate point is processed using the high-fidelity model to obtain the physical feature set of the candidate point. The physical feature set of the candidate points is processed using the multidimensional value function to obtain the candidate observation values ​​corresponding to the candidate points; The initial observation dataset is updated with the candidate points and their corresponding candidate observations, and each data point in the updated initial observation dataset is labeled with its corresponding fidelity data type.

[0015] A further optional implementation involves updating the Gaussian process model based on the updated initial observation dataset, including: The updated initial observation dataset is processed using the Gaussian process model to update the predicted mean coefficient and exploration weight coefficient of the Gaussian process model, thus completing the update of the Gaussian process model.

[0016] A further optional implementation includes, based on the updated Gaussian process model, repeatedly performing the operations of selecting candidate points in the driving scene element parameter space, updating the initial observation dataset, and updating the Gaussian process model until a preset number of repetitions is reached to obtain the intermediate observation dataset, the process comprising: For each repeated operation, the weight values ​​in the multidimensional value function are optimized using a smooth transition algorithm.

[0017] In a further optional implementation, the fidelity data types include low-fidelity data types and high-fidelity data types; Based on the intermediate observation dataset, the Gaussian process model is optimized until it converges, resulting in a converged Gaussian process model, including: Data labeled as high-fidelity data type are extracted from the intermediate observation dataset to form a high-fidelity dataset; The Gaussian process model is optimized using the high-fidelity dataset until it reaches a convergent state, thus obtaining the converged Gaussian process model.

[0018] A further optional implementation involves obtaining the final observation dataset based on the driving scene element parameter space, combined with the convergent Gaussian process model, edge-aware acquisition function, global optimization algorithm, fidelity model, and multidimensional value function, including: Using the convergent Gaussian process model, edge-aware acquisition function, and global optimization algorithm, a convergence point is selected in the parameter space of the driving scene elements; based on the convergence point, combined with the fidelity model and multidimensional value function, candidate observations corresponding to the convergence point are obtained; The initial observation dataset is updated with the convergence point and its corresponding candidate observations, and the convergent Gaussian process model is updated based on the updated initial observation dataset. Based on the updated convergent Gaussian process model, the operations of selecting convergence points in the driving scene element parameter space, updating the initial observation dataset, and updating the convergent Gaussian process model are repeated until the amount of data of the specified type in the updated initial observation dataset is not less than the data amount threshold, thus obtaining the final observation dataset.

[0019] A further optional implementation involves obtaining the autonomous driving edge scenario based on the final observation dataset and the multidimensional value function, including: A high-fidelity data type was selected from the final observation dataset to form a screening dataset; The high-fidelity model is used to process the sample points corresponding to each data point in the filtered dataset to obtain the physical feature set of the sample points corresponding to each data point in the filtered dataset; The physical feature set of each data point in the selected dataset is processed using the multidimensional value function to obtain the physical collision risk value and edge metric value of each data point in the selected dataset. In the selected dataset, data with edge metric values ​​greater than a preset edge metric threshold and physical collision risk values ​​between the minimum and maximum physical collision risk thresholds are selected as autonomous driving edge scenarios.

[0020] Secondly, this application provides an autonomous driving edge scene generation device, comprising: The initial observation dataset determination module is used to determine the initial observation dataset based on the parameter space of driving scene elements, combined with a fidelity model and a multidimensional value function. A Gaussian process model building module is used to build a Gaussian process model based on the initial observation dataset; The final observation dataset determination module is used to obtain the final observation dataset based on the driving scene element parameter space, Gaussian process model, edge perception acquisition function, global optimization algorithm, fidelity model and multidimensional value function. The autonomous driving edge scene determination module is used to obtain the autonomous driving edge scene based on the final observation dataset and the multidimensional value function.

[0021] Thirdly, this application provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements the above-mentioned method for generating autonomous driving edge scenes.

[0022] Fourthly, this application provides a computer-readable storage medium storing a program for an autonomous driving edge scene generation method, wherein the program for the autonomous driving edge scene generation method, when executed by a processor, implements the steps of the aforementioned autonomous driving edge scene generation method.

[0023] The beneficial effects of this invention are: This invention achieves targeted optimization for specific safety objectives through a multidimensional value function, effectively avoiding blind searches and local optima, and improving the targeting and efficiency of edge scene mining. By fitting the observation dataset with a Gaussian process model, it replaces the extensive repetitive simulation evaluations in existing technologies, improving sample utilization efficiency and reducing computational resource consumption. Through the combination of an edge-aware acquisition function and a global optimization algorithm, a dynamic balance is achieved between "exploring unknown dangerous areas" and "utilizing known dangerous patterns" in the high-dimensional continuous parameter space of the driving scene element parameter space, solving the problem of low edge scene search efficiency in high-dimensional parameter spaces. Simultaneously, through the reasonable application of a fidelity model and multiple rounds of iterative optimization, it ensures that the final observation dataset comprehensively covers high-value edge scene areas. Combined with the precise screening of the multidimensional value function, it further guarantees the comprehensiveness and accuracy of autonomous driving edge scene coverage. Ultimately, within a limited simulation budget, it efficiently and accurately generates autonomous driving edge scenes with high testing value, meeting the needs of the autonomous driving testing field for scene generation technology with high sample utilization efficiency, value orientation, and edge awareness capabilities. This effectively solves the technical problems of high computational cost and incomplete coverage in autonomous driving edge scene search in high-dimensional parameter spaces. Attached Figure Description

[0024] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A flowchart illustrating an autonomous driving edge scene generation method provided in this application embodiment; Figure 2 This is a schematic diagram illustrating the convergence Gaussian process in the generation of autonomous driving edge scenes according to an embodiment of this application. Figure 3 This is a schematic diagram of the structure of the autonomous driving edge scene generation device according to an embodiment of this application; Figure 4 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] Example 1 Embodiment 1 of the present invention provides a method for generating edge scenes for autonomous driving, such as... Figure 1 As shown, it includes the following steps: Step S101: Based on the parameter space of driving scene elements, and combined with the fidelity model and multidimensional value function, determine the initial observation dataset.

[0029] Step S102: Construct a Gaussian process model based on the initial observation dataset.

[0030] Step S103: Based on the driving scene element parameter space, Gaussian process model, edge perception acquisition function, global optimization algorithm, fidelity model, and multidimensional value function, the final observation dataset is obtained.

[0031] Step S104: Based on the final observation dataset and the multidimensional value function, obtain the autonomous driving edge scenario.

[0032] This embodiment achieves targeted optimization for specific safety objectives through a multidimensional value function, effectively avoiding blind searches and local optima, and improving the targeting and efficiency of edge scene mining. By fitting the observation dataset with a Gaussian process model, it replaces the extensive repetitive simulation evaluations in existing technologies, improving sample utilization efficiency and reducing computational resource consumption. Through the combination of an edge-aware acquisition function and a global optimization algorithm, a dynamic balance between "exploring unknown hazardous areas" and "utilizing known hazardous patterns" is achieved in the high-dimensional continuous parameter space of the driving scene element parameter space, solving the problem of low edge scene search efficiency in high-dimensional parameter spaces. Simultaneously, through the reasonable application of a fidelity model and multiple rounds of iterative optimization, it ensures that the final observation dataset comprehensively covers high-value edge scene areas. Combined with the precise screening of the multidimensional value function, it further guarantees the comprehensiveness and accuracy of autonomous driving edge scene coverage. Ultimately, within a limited simulation budget, it efficiently and accurately generates autonomous driving edge scenes with high testing value, meeting the needs of the autonomous driving testing field for scene generation technologies with high sample utilization efficiency, value orientation, and edge awareness capabilities. This effectively solves the technical problems of high computational cost and incomplete coverage in autonomous driving edge scene search in high-dimensional parameter spaces.

[0033] In a further optional embodiment, in step S101, the fidelity model includes a low-fidelity model and a high-fidelity model; the process of determining the initial observation dataset based on the driving scene element parameter space, combined with the fidelity model and the multidimensional value function, involves: using a space-filling sampling strategy to uniformly collect multiple sample points in the driving scene element parameter space; processing each sample point using the low-fidelity model to obtain the physical feature set corresponding to each sample point; processing the physical feature set of each sample point using the multidimensional value function to obtain the observation value corresponding to each sample point; and assembling all the sample points and their corresponding observation values ​​into the initial observation dataset. This process rapidly processes sample points using the low-fidelity model, reducing the computational cost in the observation dataset determination stage. Simultaneously, the space-filling sampling strategy ensures a uniform distribution of samples in the driving scene element parameter space, laying a reliable data foundation for the accurate construction of the subsequent Gaussian process model. This approach balances comprehensive data coverage with computational efficiency, effectively avoiding the problem of insufficient model fitting accuracy caused by initial sampling bias.

[0034] In this embodiment, the space-filling sampling strategy is a sampling method used in the generation of autonomous driving edge scenes. It targets the high-dimensional driving scene element parameter space (including continuous parameter dimensions such as ambient light, road friction coefficient, vehicle speed, and target vehicle distance), aiming for uniform coverage, unbiased distribution, and high representativeness. The goal is to cover the entire parameter range with a limited number of samples, avoiding local sample clustering or blank areas, and providing high-quality initial samples for subsequent model training. The core specific method is Latin hypercube sampling (LHS): each dimension of the parameter space is uniformly divided into several equally wide intervals. One sample is randomly selected from each interval, and samples from different dimensions are randomly combined to achieve full coverage of a single dimension and uniform distribution across dimensions. Auxiliary methods include uniform grid sampling (taking grid intersections at equal steps in low-dimensional scenes), Sobol low-discrepancy sequence sampling (generating quasi-random samples with more uniform distribution and faster convergence), and Halton sequence sampling (layered uniform distribution, adaptable to continuous parameters). This invention preferably uses Latin hypercube sampling, which equally divides the 8-dimensional driving parameter space for sampling, generating uniformly distributed initial samples, effectively balancing sampling efficiency and spatial coverage.

[0035] In this embodiment, the low-fidelity model is a lightweight and simplified simulation model in the field of autonomous driving simulation testing and edge scene generation, and is a key component of the high-fidelity model system. Its core design idea is to discard non-critical physical, geometric and rendering details, retain core features that are strongly related to vehicle motion and scene risks, simplify secondary elements such as suspension deformation, fine tire friction, complex textures, high-precision lighting and weather optical effects, and adopt a simplified kinematic / basic dynamics model, low-polygon scene modeling and a lightweight simulation engine. Its core function is to quickly output physical feature sets such as vehicle position, speed, collision time, and relative distance. It has the characteristics of high efficiency, low cost and adaptability to large-scale sample screening. In the autonomous driving edge scene generation of this invention, it is mainly used for rapid screening of spatial filling sampling samples and iterative candidate points, efficiently generating physical features to support multi-dimensional value function calculation, and balancing simulation efficiency and scene effectiveness.

[0036] In this embodiment, the high-fidelity model is a professional simulation model in the field of autonomous driving simulation testing and edge scene generation, characterized by high precision, high fidelity, and full-element representation. It belongs to the high-fidelity model system, and its core is to completely replicate the physical characteristics, geometric details, environmental effects, and vehicle dynamics of real driving scenarios. At the physical modeling level, it fully reproduces complex dynamic details such as vehicle suspension deformation, fine tire friction, air resistance, collision mechanics, and the powertrain system. At the scene geometry level, it uses a high-precision 3D model, preserving all structures such as road textures, building surfaces, obstacle details, and road markings. At the rendering and environment level, it accurately reproduces weather optical effects such as lighting, shadows, reflections, rain, snow, and fog, supporting high-resolution, high-frame-rate simulation rendering. At the simulation engine level, it is equipped with a full-featured high-precision physics engine, including fine collision detection, fluid calculation, and complex interaction modules. The core function is to output high-precision vehicle motion data, collision risk data, and scene interaction data for accurate verification of high-value candidate samples and final confirmation of edge scenes, ensuring the authenticity and reliability of autonomous driving tests. In the generation of autonomous driving edge scenes, high-precision simulation is mainly performed on candidate points with low perception values ​​and high value to verify the authenticity of edge scenes and support the final edge scene selection.

[0037] In this embodiment, the Gaussian process surrogate model is a nonparametric Bayesian machine learning model used as a low-cost alternative model in autonomous driving edge scene generation. It fits the complex nonlinear mapping relationship between high-dimensional driving scene element parameters and multidimensional value functions. Its core is to characterize the statistical properties of stochastic processes through the mean function and covariance function. It requires no pre-defined fixed model structure and only relies on a small amount of observation data (sample points and corresponding observation values) to complete modeling, while simultaneously outputting two key pieces of information: the predicted mean and the uncertainty variance. It possesses four core characteristics: small sample adaptation, nonlinear fitting, probabilistic prediction, and uncertainty quantification. This effectively solves the problems of scarce samples and high simulation costs in high-dimensional parameter spaces. In this invention, the model is built based on an initial observation dataset and continuously updated and optimized during iteration. It is used to predict the value function value of unsampled parameter regions and quantify the prediction uncertainty, providing a decision-making basis for the edge perception acquisition function, balancing the exploration of unknown dangerous areas with the utilization of known dangerous patterns, and efficiently supporting the mining of high-value edge scenes.

[0038] In this embodiment, the Gaussian process model is a nonparametric probabilistic model based on Bayesian theory. It does not require a pre-defined fixed function form; the mean function and covariance function jointly define all the statistical characteristics of the stochastic process, making it naturally suitable for high-dimensional, nonlinear, and small-sample data modeling scenarios. In the generation of autonomous driving edge scenarios, this model is trained on an observation dataset containing driving scene parameters and corresponding value observations. Through Bayesian inference, it learns the complex nonlinear mapping relationship between the high-dimensional scene parameter space and the multidimensional value function. It can output the value prediction mean at any position in the parameter space and quantify the prediction uncertainty (variance). It possesses core advantages such as small-sample adaptation, nonlinear fitting, probabilistic output, and uncertainty quantification. It can replace high-cost simulation evaluation, provide decision-making basis for edge perception acquisition functions, balance the exploration of unknown dangerous areas with the utilization of known dangerous patterns, and support the accurate mining and iterative optimization of high-value edge scenarios.

[0039] In this embodiment, the global optimization algorithm is an optimization algorithm aimed at the global optimal solution and avoiding local optimal traps in a high-dimensional, nonlinear, and non-convex driving scene element parameter space. Its core function is to efficiently traverse and search for high-value candidate points / convergence points in the entire parameter space based on the prediction of the Gaussian process model and the edge perception acquisition function. The global optimization algorithm covers the entire parameter space through a global search strategy (such as random sampling, evolutionary iteration, and probabilistic exploration). It combines the predicted mean and uncertainty variance output by the Gaussian process model with the guidance of the edge perception acquisition function to dynamically balance "exploring unknown high-risk areas" and "utilizing known dangerous modes", adapting to the search requirements of high-dimensional parameter spaces in the generation of autonomous driving edge scenes. Commonly used types include Bayesian optimization, genetic algorithm, particle swarm optimization, and simulated annealing. This invention preferably uses a Bayesian-type global optimization algorithm that is compatible with the Gaussian process model to accurately locate the optimal sampling points that conform to the characteristics of the edge scene, ensuring the comprehensiveness, accuracy, and efficiency of the search.

[0040] In this embodiment, the smooth transition algorithm is an adaptive weight optimization algorithm in the iterative process of generating edge scenes for autonomous driving. Its core function is to dynamically adjust the weights of each risk term in the multidimensional value function, achieving a smooth transition in the search strategy from "extensive exploration of high-risk areas in the initial stage" to "refined mining of edge scenes in the later stage." This avoids drastic fluctuations in the search strategy, getting stuck in local optima, or missing edge scenes due to sudden weight changes. The algorithm uses the number of iterations as a variable to dynamically adjust the weights of the physical collision risk term, the edge metric indicator term, and the decision stability risk term: initially, the physical collision risk term is given a high initial weight, which gradually decreases with iteration; the weight of the edge metric indicator term increases with iteration; and the weight of the decision stability risk term dynamically balances the former two. A preset decay rate controls the rate of weight change, achieving a smooth and gradual weight transition. During multiple rounds of iterative optimization of the multidimensional value function, the algorithm continuously outputs adaptive weights, ensuring a smooth transition in the search strategy and improving the accuracy and stability of edge scene mining.

[0041] In this embodiment, for example, a real driving environment consists of numerous continuous driving scenario element parameters; for example, a real driving environment includes environmental conditions, road features, and the states and behaviors of traffic participants. These continuous driving scenario element parameters are encoded into a driving scenario element parameter space. x n =[ i 1 ,i 2, i D ] T The driving scene element parameters in the driving scene element parameter space include, but are not limited to: ambient light intensity. i 1. Road surface friction coefficient i 2∈[0.3,1.0], initial speed of the main vehicle i 3∈[0,120]km / h, longitudinal distance of the target vehicle i 4∈[5,100]m, target vehicle intrusion rate i 5∈[-10,10]m / s, driver reaction delay i 6∈[0.1,2.0]s, road curvature i 7∈[-0.1,0.1]m -1 Crosswind speed i 8∈[0,20]m / s. This solves the problem of digitizing and standardizing the parameters of driving scene elements.

[0042] To address the differences in physical characteristics and dimensions across different dimensions within the driving scene element parameter space, Min-Max normalization is applied to map the driving scene element parameters to the [0,1] interval. For example, periodic variables such as wind direction angle are mapped to sine / cosine components to eliminate numerical calculation biases and improve the optimizer's convergence efficiency. The driving scene element parameter space lays the mathematical foundation for subsequent optimization searches.

[0043] In this embodiment, a multidimensional value function is constructed based on physical laws, higher-order derivatives of control variables, edge quantification algorithms, and noise. This multidimensional value function provides scene search guidance for generating autonomous driving edge scenes and determines the search direction of the optimization algorithm, thereby overcoming the technical deficiency of traditional single risk indicators in effectively capturing autonomous driving edge scenes. The addition of noise makes the multidimensional value function more closely resemble the real situation.

[0044] The multidimensional value function is shown in the following formula: ; in, w 1. w 2. w 3 represents the weight. R collision Represents the physical collision risk item. R instability This represents the risk factor for decision-making stability. C edge Represents the edge metric, ⊙ represents the Hadamard product. The noise level is represented by the physical risk term, which directly reflects the physical risk of a collision in the scenario. The decision stability risk term identifies abnormal decision scenarios that may lead to vehicle instability or control oscillations. If the control quantity exhibits high-frequency oscillations or amplitude saturation, this term will increase significantly, indicating that the scenario has triggered the stability boundary of the controller. The edge quantification term quantifies the degree to which a scenario is located at the boundary between safe and dangerous decisions. The larger the value, the closer the scenario is to the "cliff edge" of system performance, and the higher the test value.

[0045] The process of optimizing the weight values ​​in the multidimensional value function using a smooth transition algorithm involves: in the initial optimization phase, assigning a larger initial weight to the physical collision risk term, with its decay law set as follows: (Where λ is the decay rate, set to 0.01), to guide the algorithm to quickly locate physical risk areas; as the iteration progresses, w The weight of 1(n) gradually decreases, thus preventing the algorithm from getting stuck in a local optimum where collisions are inevitable in the later stages of optimization. Conversely, the weight of the marginal metric term increases with iteration, set to... In the later stages of optimization, the ability to refine the safety-hazard decision boundary is significantly improved. The weight of the decision stability risk term is also adjusted. w 2(n) is used as a smoothing adjustment term, and a normalization constraint strategy is preferred. w 2(n) = 1 - w 1(n)- w 3(n). Through the above mechanism, the system automatically realizes the shift in search strategy from "extensive search for high-risk scenarios in the early stage" to "refined mining of edge scenarios in the later stage". w 1 (0) is set to 0.7. w 2(0) is set to 0.2. w 3(0) is set to 0.1.

[0046] In this embodiment, the edge-aware acquisition function is shown in the following formula: ; in, Represents the edge-aware acquisition function. m n Represents the predicted mean coefficient. s n Represents the exploration weighting coefficient. β n Represents the adaptive exploration weight coefficient. Represents the gradient term; The project encourages exploration of regions with high forecasting uncertainty. The newly added... The algorithm is specifically guided to search for regions where model uncertainty changes drastically. These regions are often the boundary between safety and danger—that is, the edge cases. β n It will adaptively decay with iteration, realizing the transformation from "extensive exploration" to "key utilization".

[0047] In a further optional embodiment, in step S102, the process of constructing a Gaussian process model based on the initial observation dataset involves using a Gaussian process surrogate model. The initial observation dataset is processed to obtain a Gaussian process model; this process is performed using a Gaussian process proxy model. An efficient fitting method was used to quickly construct a Gaussian process model that can characterize the risk distribution of driving scenario element parameter space by fitting the initial observation dataset. This provides reliable model support for subsequent acquisition of intermediate observation datasets and optimization of the Gaussian process model, effectively improving the accuracy and efficiency of subsequent edge scene search and reducing the search complexity in high-dimensional parameter space. Among these methods, the Gaussian process surrogate model... It can learn and predict multidimensional value functions across the entire parameter space using limited simulation observation data. R ( X The distribution of ) is the posterior distribution. p ( f ( x n )∣D n ).

[0048] In a further optional embodiment, step S103 involves obtaining the final observation dataset based on the driving scene element parameter space, Gaussian process model, edge-aware acquisition function, global optimization algorithm, fidelity model, and multidimensional value function: Based on the driving scene element parameter space and the initial observation dataset, and combining the Gaussian process model, edge-aware acquisition function, global optimization algorithm, fidelity model, and multidimensional value function, an intermediate observation dataset is obtained; based on the intermediate observation dataset, the Gaussian process model is optimized until it converges, resulting in a converged Gaussian process model; based on the driving scene element parameter space, and combining the converged Gaussian process model, edge-aware acquisition function, global optimization algorithm, fidelity model, and multidimensional value function, the final observation dataset is obtained.

[0049] A further optional embodiment involves obtaining an intermediate observation dataset based on the driving scene element parameter space, the initial observation dataset, and the Gaussian process model, edge-aware acquisition function, global optimization algorithm, fidelity model, and multidimensional value function: Candidate points are selected in the driving scene element parameter space using the Gaussian process model, edge-aware acquisition function, and global optimization algorithm; candidate observation values ​​corresponding to the candidate points are obtained based on the candidate points and the fidelity model and multidimensional value function; the observation dataset is updated with the candidate points and their corresponding candidate observation values, and the Gaussian process model is updated based on the updated initial observation dataset; the updated Gaussian process model... Based on this, the process of selecting candidate points in the driving scene element parameter space, updating the initial observation dataset, and updating the Gaussian process model is repeated until a preset number of repetitions is reached to obtain the intermediate observation dataset. This process updates the initial observation dataset and the Gaussian process model through multiple rounds of iteration. By using the edge-aware acquisition function and the global optimization algorithm, a balance is achieved between exploring unknown dangerous areas and utilizing known dangerous patterns in the driving scene element parameter space. The accuracy of candidate observations is ensured by relying on the fidelity model and the multidimensional value function, thereby obtaining a high-quality intermediate observation dataset. This provides solid data support for subsequent optimization of the Gaussian process model based on this dataset and obtaining a convergent Gaussian process model.

[0050] A further optional embodiment involves selecting candidate points in the driving scene element parameter space using the Gaussian process model, the edge-aware acquisition function, and the global optimization algorithm: The process involves updating the predicted mean coefficient and exploration weight coefficient of the edge-aware acquisition function using the predicted mean coefficient and exploration weight coefficient of the Gaussian process model to obtain an updated edge-aware acquisition function; then, candidate points are selected in the driving scene element parameter space using the global optimization algorithm and the updated edge-aware acquisition function. This process updates the edge-aware acquisition function using the predicted mean coefficient and exploration weight coefficient of the Gaussian process model, ensuring that the acquisition function always closely matches the risk distribution characteristics of the driving scene element parameter space. Combined with the global optimization algorithm, it accurately selects high-potential candidate points, effectively improving the targeting and efficiency of candidate point selection, and providing reliable support for the accurate acquisition of subsequent candidate observations and the high-quality generation of intermediate observation datasets.

[0051] A further optional embodiment involves obtaining candidate observations based on the candidate points, combined with a low-fidelity model and a multidimensional value function. The process of updating the observation dataset with the candidate points and their corresponding candidate observations includes: processing the candidate points using the updated edge-aware acquisition function to obtain the perceived values ​​of the candidate points; comparing the perceived values ​​of the candidate points with a preset perception threshold; if the perceived value of the candidate points is not less than the preset perception threshold, processing the candidate points using the low-fidelity model to obtain the physical feature set of the candidate points; if the perceived value of the candidate points is less than the preset perception threshold, processing the candidate points using the high-fidelity model to obtain... The physical feature set of the candidate points; the multidimensional value function is used to process the physical feature set of the candidate points to obtain the candidate observation values ​​corresponding to the candidate points; the initial observation dataset is updated with the candidate points and their corresponding candidate observation values, and each data in the updated initial observation dataset is labeled with the corresponding fidelity data type; this process dynamically switches the fidelity model through the perceptual value threshold, taking into account both the accuracy of the candidate observation values ​​and the efficient use of simulation resources. At the same time, labeling the fidelity data type provides standardized and reliable data support for the subsequent update of the Gaussian process model and the high-quality generation of intermediate observation datasets, further improving the efficiency and accuracy of edge scene mining.

[0052] A further optional embodiment involves updating the Gaussian process model based on the updated initial observation dataset: The updated initial observation dataset is processed using the Gaussian process model to update the predicted mean coefficient and exploration weight coefficient of the Gaussian process model, thus completing the update of the Gaussian process model. This process, by updating the predicted mean coefficient and exploration weight coefficient of the Gaussian process model, ensures that the model always conforms to the risk distribution characteristics of the driving scenario element parameter space, improves the model's prediction accuracy, and provides reliable model support for subsequent updates to the edge perception acquisition function, accurate selection of candidate points, and high-quality generation of intermediate observation datasets.

[0053] In a further optional embodiment, based on the updated Gaussian process model, the operations of selecting candidate points in the driving scene element parameter space, updating the initial observation dataset, and updating the Gaussian process model are repeatedly performed until a preset number of repetitions are reached to obtain the intermediate observation dataset. For each repeated operation, a smooth transition algorithm is used to optimize the weight values ​​in the multidimensional value function. This process dynamically optimizes the weight values ​​of the multidimensional value function through the smooth transition algorithm, so that the guidance of the multidimensional value function is adaptively adjusted with the iteration process, further improving the accuracy of candidate point selection and the quality of the observation dataset, and providing a strong guarantee for the efficient generation of the intermediate observation dataset and the subsequent convergence optimization of the Gaussian process model.

[0054] In this embodiment, the preset number of repetitions is the optimal value determined by the inventors through statistical analysis and repeated verification of a large amount of experimental data. This preset number of repetitions ensures that the quantity and quality of the intermediate observation dataset meet the optimization requirements of the Gaussian process model, providing reliable data support for its convergence, while also avoiding excessive and unnecessary iterations. This effectively balances the convergence effect of the Gaussian process model with the computational efficiency of the entire autonomous driving edge scene generation process, taking into account both the reliability of the technical effect and the efficiency of engineering applications.

[0055] A further optional embodiment involves optimizing a Gaussian process model based on an intermediate observation dataset until the Gaussian process model converges, resulting in a converged Gaussian process model. This process involves extracting data labeled as high-fidelity data types from the intermediate observation dataset to form a high-fidelity dataset; optimizing the Gaussian process model using the high-fidelity dataset until it reaches a convergent state, thus obtaining the converged Gaussian process model. This process, by optimizing the Gaussian process model using a high-fidelity dataset and ensuring its convergence, significantly improves the prediction accuracy and reliability of the converged Gaussian process model, ensuring that the model can accurately represent the risk distribution in the parameter space of driving scene elements. This provides high-quality model support for subsequently obtaining the final observation dataset based on the converged Gaussian process model and accurately mining edge scenarios of autonomous driving.

[0056] In this embodiment, the convergence condition of the Gaussian process model is as follows: the logarithmic marginal likelihood value of the Gaussian process model at the current iteration number is compared with the logarithmic marginal likelihood value corresponding to the previous iteration; if the difference between the two is less than or equal to a preset convergence threshold, the Gaussian process model is determined to have converged.

[0057] A further optional embodiment involves obtaining the final observation dataset based on the driving scene element parameter space, combined with the convergent Gaussian process model, edge-aware acquisition function, global optimization algorithm, fidelity model, and multidimensional value function: Using the convergent Gaussian process model, edge-aware acquisition function, and global optimization algorithm, a convergence point is selected in the driving scene element parameter space; based on the convergence point, combined with the fidelity model and multidimensional value function, candidate observation values ​​corresponding to the convergence point are obtained; the initial observation dataset is updated with the convergence point and its corresponding candidate observation values, and the convergent Gaussian process model is updated based on the updated initial observation dataset; based on the updated convergent Gaussian process model, the final observation dataset is re-selected. The process repeats the steps of selecting a convergence point in the parameter space of the driving scene elements, updating the initial observation dataset, and updating the convergent Gaussian process model until the amount of data of a specified type in the updated initial observation dataset is not less than the data amount threshold, thus obtaining the final observation dataset. This process relies on the high prediction accuracy of the convergent Gaussian process model, combined with the edge perception acquisition function and the global optimization algorithm to select the convergence point in a targeted manner, and uses a fidelity model and a multidimensional value function to ensure the accuracy of candidate observations. Through multiple rounds of iterative updates, the final observation dataset that meets the threshold requirements is obtained, providing comprehensive and high-quality data support for the subsequent accurate acquisition of autonomous driving edge scenes by combining the multidimensional value function, and further improving the accuracy and efficiency of edge scene mining.

[0058] A further optional embodiment involves selecting convergence points in the driving scene element parameter space using the convergent Gaussian process model, the edge-aware acquisition function, and the global optimization algorithm: The process involves updating the predicted mean coefficient and exploration weight coefficient of the edge-aware acquisition function using the predicted mean coefficient and exploration weight coefficient of the convergent Gaussian process model to obtain an edge-aware acquisition function adapted to the convergence stage; and then using the global optimization algorithm and the adapted edge-aware acquisition function to directionally select convergence points in the driving scene element parameter space. This process updates the edge-aware acquisition function with the parameters of the convergent Gaussian process model, ensuring that the acquisition function fits the risk distribution of the driving scene element parameter space during the convergence stage. Combined with the global optimization algorithm, it accurately identifies high-value convergence points, effectively improving the targeting of convergence point selection and providing reliable support for the accurate acquisition of subsequent candidate observations and the high-quality generation of the final observation dataset.

[0059] A further optional embodiment involves obtaining candidate observations corresponding to the convergence point based on the convergence point, combined with the fidelity model and the multidimensional value function: The convergence point is processed using an adapted edge-aware acquisition function to obtain its perceived value; the perceived value is compared with a preset perception threshold; if the perceived value is not less than the preset perception threshold, the convergence point is processed using the low-fidelity model to obtain its physical feature set; if the perceived value is less than the preset perception threshold, the convergence point is processed using the high-fidelity model to obtain its physical feature set; the physical feature set is processed using the multidimensional value function to obtain candidate observations corresponding to the convergence point. This process dynamically switches the fidelity model through the perceived value threshold, balancing the accuracy of the candidate observations at the convergence point with the efficient use of simulation resources, providing accurate data support for subsequent updates of the observation dataset and iterative optimization of the convergent Gaussian process model.

[0060] A further optional embodiment involves updating the initial observation dataset with the convergence point and its corresponding candidate observations, and updating the convergent Gaussian process model based on the updated initial observation dataset: The convergence point and its corresponding candidate observations are added to the initial observation dataset to complete the update of the observation dataset; the updated initial observation dataset is used to process the convergent Gaussian process model, updating the predicted mean coefficient and exploration weight coefficient of the convergent Gaussian process model to complete the iterative update of the convergent Gaussian process model; this process, by updating the parameters of the observation dataset and the convergent Gaussian process model, ensures that the model always fits the risk distribution characteristics of the driving scenario element parameter space, continuously improves the model's prediction accuracy, and provides high-quality model support for the selection of the next convergence point and the iterative generation of the final observation dataset.

[0061] A further optional embodiment involves determining whether the amount of data of a specified type in the updated initial observation dataset is not less than a data volume threshold: the specified data type refers to high-fidelity data type, and the total amount of data marked as the specified type in the updated initial observation dataset is counted; the total amount of data is compared with the data volume threshold, and if the total amount of data is not less than the preset data volume threshold, the iteration stops and the final observation dataset is output; if the total amount of data is less than the data volume threshold, the process returns to repeat the operations of selecting a convergence point, updating the initial observation dataset, and updating the convergent Gaussian process model; this process, through a clear threshold judgment mechanism, ensures that the final observation dataset can meet the needs of subsequent edge scene screening, avoids invalid iterations, and balances the quality and generation efficiency of the final observation dataset.

[0062] In a further optional embodiment, in step S104, based on the final observation dataset and combined with the multidimensional value function, the process of obtaining the autonomous driving edge scene is as follows: High-fidelity data types are selected from the final observation dataset to form a screening dataset; the high-fidelity model is used to process the sample points corresponding to each data point in the screening dataset to obtain the physical feature set of each data point in the screening dataset; the multidimensional value function is used to process the physical feature set of each data point in the screening dataset to obtain the physical collision risk value and edge metric of each data point in the screening dataset. Value; In the selected dataset, data with edge metric values ​​greater than a preset edge metric threshold and physical collision risk values ​​between the minimum and maximum physical collision risk thresholds are selected as autonomous driving edge scenarios; This process selects high-fidelity data by targeted screening, relies on a high-fidelity model to ensure the accuracy of the physical feature set, and combines a multi-dimensional value function to achieve dual quantitative screening of physical collision risk values ​​and edge metric values, accurately locking autonomous driving edge scenarios at the safety-danger threshold, providing high-value and high-precision scenario samples for autonomous driving testing, and effectively solving the problems of inaccurate and incomplete edge scenario screening in existing technologies.

[0063] In this embodiment, it should be noted that for each data point in the selected dataset, the physical collision risk value and marginal metric value of the corresponding sample point are obtained by processing the corresponding sub-items of the multidimensional value function. Specifically, the physical collision risk term in the multidimensional value function is used to quantify the physical feature set of each data point in the selected dataset to obtain the physical collision risk value of the sample point; at the same time, the marginal metric term in the multidimensional value function is used to quantify the physical feature set of each data point in the selected dataset to obtain the marginal metric value of the sample point.

[0064] In this embodiment, it should be noted that both the high-fidelity model and the low-fidelity model are based on existing technologies in the field and can be implemented by those skilled in the art.

[0065] In this embodiment, the convergence Gaussian process exploration process in the autonomous driving edge scene generation process is specifically as follows: Figure 2 As shown.

[0066] (a) shows the Gaussian process surrogate model and uncertainty distribution diagram. Part (a) specifically demonstrates how the Gaussian process surrogate model, built based on limited observation data, achieves full-domain fitting and uncertainty quantification of the true risk value function in the high-dimensional driving scenario parameter space, clarifying the model's representation logic of the risk distribution in the parameter space. The solid black line represents the posterior mean μ(x) of the surrogate model, used to predict and estimate the true risk value function f(x); the dashed black line represents the true risk value function, and the deviation between the two intuitively reflects the model's prediction error. The shaded area represents the model's 95% confidence interval. The width of this interval quantifies the uncertainty of the model's prediction; a wider interval indicates scarcer observation data in the corresponding area and lower reliability of the model's prediction. The black dots represent observed scenario points. This type of data is the basis for the construction and iterative update of the Gaussian process surrogate model. As the number of observed scenario points increases, the model's 95% confidence interval gradually narrows, and the model's prediction accuracy improves accordingly. The core function of this Gaussian process proxy model is to learn the risk distribution characteristics within the parameter space of driving scene elements in a low-cost manner, replacing the large number of repetitive and costly simulation evaluation operations in existing technologies, thereby significantly improving sample utilization efficiency and reducing the consumption of computing resources in the edge scene generation process.

[0067] (b) is a schematic diagram of the edge-aware acquisition function optimization. Part (b) clearly demonstrates the specific process by which the edge-aware acquisition function guides the optimization algorithm, selecting the next high-value evaluation point, thus clarifying the guiding role of the acquisition function in edge scene directional search. In part (b), the black solid line represents the edge-aware acquisition function α(x) proposed in this invention. This acquisition function is an improved version obtained by adding an edge-aware gradient term to the traditional confidence bound (UCB, shown as a dashed line in the figure). Compared with the traditional acquisition function, it has a more accurate edge scene localization capability. The peak point of the acquisition function (marked with an asterisk as x) is shown. n+1 This is the next evaluation point selected by the algorithm. This evaluation point is located in the "search preference region," characterized by high model prediction uncertainty and drastic changes in uncertainty. It belongs to a typical safety-danger critical region and has extremely high testing value. The specific expression of the edge-aware acquisition function used in this invention is as follows: , where μ n (x n The ) item is used to leverage known danger patterns to improve search targeting; β n σ n (x n The first item is used to encourage exploration of unknown areas with high predictive uncertainty, avoiding search blind spots; the second item is... This specialized algorithm guides the search towards the boundary regions of the driving scene element parameter space and areas with drastic changes in model uncertainty, achieving precise localization of edge scenes. The core mechanism of this acquisition function achieves a dynamic balance between "exploring unknown dangerous areas" and "utilizing known dangerous patterns," effectively solving the technical shortcomings of traditional algorithms that are prone to getting trapped in local optima and cannot efficiently mine edge scenes. This provides core support for the efficient advancement of the convergent Gaussian process exploration process.

[0068] A specific example of this embodiment is as follows: 1. Basic parameter settings: 1.1 Driving scenario element parameter space (8 dimensions).

[0069] Define the 8-dimensional normalized parameter vector to be searched, with the initial physical parameter range as follows: Ambient light intensity θ1: 6000 lux (cloudy daytime); Road surface friction coefficient θ2: 0.4 (wet asphalt road); Initial speed of the main vehicle θ3: 50km / h (speed limit on urban roads); Longitudinal distance to the target vehicle θ4: 15m (close-range vehicle intrusion); Target vehicle intrusion rate θ5: -3m / s (lateral intrusion to the left); Driver reaction delay θ6: 0.8s; Road curvature θ7:0m - ¹(straight road); Lateral wind speed θ8: 5 m / s.

[0070] 1.2 Core threshold setting.

[0071] Perception threshold: 0.5; Edge quantification threshold: 0.8; Physical collision risk threshold range: [0.6, 0.9]; Data volume threshold: 50 sets of high-fidelity data.

[0072] 1.3 Initial weights of the multidimensional value function.

[0073] w1(0)=0.7 (physical collision risk), w2(0)=0.2 (decision stability risk), w3(0)=0.1 (marginality), and decay rate λ=0.01.

[0074] 2. Step-by-step execution process.

[0075] 2.1 Generate the initial observation dataset.

[0076] 2.11 Space Fill Sampling.

[0077] Latin hypercube sampling was used to uniformly collect 30 sample points in an 8-dimensional parameter space, including the basic parameter points set above.

[0078] 2.12 Low-fidelity model processing.

[0079] The kinematic simplified low-fidelity model is invoked to calculate the physical feature set (vehicle deceleration, relative distance, collision time TTC) for each sample point.

[0080] 2.13 Calculate the observed values.

[0081] Substitute into the multidimensional value function: ; Example sample point observation: 0.72 (medium to high risk).

[0082] 2.14 Form the initial dataset.

[0083] The initial observation dataset D0 is formed by 30 sample points and their corresponding observation values.

[0084] 2.2 Constructing a Gaussian process model.

[0085] The initial observation dataset D0 is fitted using a Gaussian process surrogate model to learn the posterior distribution p(f(x) of the multidimensional value function). n (∣D0), to obtain the initial Gaussian process model and complete the preliminary fitting of the risk distribution in the 8-dimensional parameter space.

[0086] 2.3 Iteratively generate the final observation dataset.

[0087] 2.31 Generate intermediate observation dataset.

[0088] Using an edge-aware acquisition function and a global optimization algorithm, 20 candidate points are selected in the parameter space; Candidate point perception value = 0.6 (≥0.5), call low-fidelity model to calculate physical features; In each iteration, the weights are updated using a smooth transition algorithm: w1 decays and w3 increases. After 15 iterations, an intermediate observation dataset containing 200 sample points was obtained.

[0089] 2.32 The convergent Gaussian process model was obtained through optimization.

[0090] Extract 60 sets of high-fidelity data from the intermediate dataset, optimize the Gaussian process model until the logarithmic marginal likelihood value change is ≤0.001, then determine convergence and obtain the converged Gaussian process model.

[0091] 2.33 Generate the final observation dataset.

[0092] The convergence point is selected using the convergence model, the perceptual value is 0.3 (<0.5), and the high-fidelity model is called for accurate calculation; Iteratively update the initial dataset until the amount of high-fidelity data reaches 55 sets (≥50), then stop the iteration and obtain the final observation dataset D_final.

[0093] 2.4 Screening yields autonomous driving edge scenarios.

[0094] 2.41 Constructing the filtered dataset.

[0095] Extract 55 sets of high-fidelity data from D_final to form a filtered dataset.

[0096] 2.42 Calculate risk and marginality.

[0097] Substituting into the multidimensional value function, calculate the example target sample points: Physical collision risk value: 0.75 (within the range of 0.6 to 0.9); Edge metric: 0.85 (>0.8 threshold).

[0098] 2.43 Determine the edge scene.

[0099] The sample point met the screening criteria, and finally generated an autonomous driving edge scenario with a wet urban road surface, a main vehicle traveling straight at 50km / h, and a target vehicle laterally intruding at 3m / s 15m away. This provides a high-value test scenario for the adversarial training of the autonomous driving planner.

[0100] In another embodiment of this application, an autonomous driving edge scene generation device is also provided, such as... Figure 3 As shown, it includes: an initial observation dataset determination module 301, a Gaussian process model construction module 302, a final observation dataset determination module 303, and an autonomous driving edge scene determination module 304; The initial observation dataset determination module 301 is used to determine the initial observation dataset based on the parameter space of driving scene elements, combined with a fidelity model and a multidimensional value function. Gaussian process model construction module 302 is used to construct a Gaussian process model based on the initial observation dataset; The intermediate observation data determination module 303 is used to obtain the final observation dataset based on the driving scene element parameter space, Gaussian process model, edge perception acquisition function, global optimization algorithm, fidelity model and multidimensional value function. The autonomous driving edge scene determination module 304 is used to obtain the autonomous driving edge scene based on the final observation dataset and the multidimensional value function.

[0101] In another embodiment of this application, an electronic device is also provided, characterized in that it includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements the autonomous driving edge scene generation method described in any of the foregoing method embodiments.

[0102] The electronic device provided in this invention generates autonomous driving edge scenes by executing programs stored in memory. This ensures targeted optimization for specific safety objectives through a multi-dimensional value function, effectively avoiding blind searches and local optima, and improving the targeting and efficiency of edge scene mining. By fitting the initial observation dataset using a Gaussian process model, it replaces the extensive repetitive simulation evaluations in existing technologies, improving sample utilization efficiency and reducing computational resource consumption. Furthermore, by combining an edge perception acquisition function with a global optimization algorithm, it achieves both "exploring unknown danger zones" and "utilizing known danger patterns" within the high-dimensional continuous parameter space of the driving scene element parameter space. The dynamic balance solves the problem of low efficiency in edge scene search under high-dimensional parameter space. At the same time, through the reasonable application of the fidelity model and multiple rounds of iterative optimization, it ensures that the final observation dataset can fully cover high-value edge scene areas. Combined with the accurate screening of multi-dimensional value functions, it further guarantees the comprehensiveness and accuracy of autonomous driving edge scene coverage. Finally, within the limited simulation budget, it efficiently and accurately generates autonomous driving edge scenes with high testing value, meeting the needs of the autonomous driving testing field for scene generation technology with high sample utilization efficiency, value orientation and edge perception capabilities. It effectively solves the technical problems of high computational cost and incomplete coverage in autonomous driving edge scene search under high-dimensional parameter space.

[0103] The communication bus 1140 mentioned in the above electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus 1140 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 4 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0104] The communication interface 1120 is used for communication between the above-mentioned electronic device and other devices.

[0105] The memory 1130 may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0106] The processor 1110 mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0107] In another embodiment of this application, a computer-readable storage medium is provided, on which a program for an autonomous driving edge scene generation method is stored. When the program for the autonomous driving edge scene generation method is executed by a processor, it implements the steps of the autonomous driving edge scene generation method described in any of the foregoing method embodiments.

[0108] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0109] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for generating edge scenes for autonomous driving, characterized in that, include: A space-filling sampling strategy is adopted to uniformly collect multiple sample points in the parameter space of the driving scene elements. Each sample point is processed using a low-fidelity model to obtain the physical feature set corresponding to each sample point; The physical feature set of each sample point is processed using a multidimensional value function to obtain the observation value corresponding to each sample point; all sample points and their corresponding observation values ​​are combined to form an initial observation dataset; Based on the initial observation dataset, a Gaussian process model is constructed; The predicted mean coefficient and exploration weight coefficient in the Gaussian process model are used to update the predicted mean coefficient and exploration weight coefficient in the edge perception acquisition function to obtain the updated edge perception acquisition function; candidate points are selected in the driving scene element parameter space using the global optimization algorithm and the updated edge perception acquisition function. The edge-aware acquisition function is shown below: , Represents the edge-aware acquisition function. μ n Represents the predicted mean coefficient. σ n Represents the exploration weighting coefficient. β n Represents the adaptive exploration weight coefficient. Represents the gradient term; The candidate points are processed using an updated edge-aware acquisition function to obtain their perceived values. These perceived values ​​are then compared to a preset perception threshold. If the perceived value is not less than the preset threshold, a low-fidelity model is used to process the candidate points, resulting in a physical feature set. If the perceived value is less than the preset threshold, a high-fidelity model is used to process the candidate points, again obtaining a physical feature set. Finally, a multidimensional value function is used to process the physical feature set of the candidate points, yielding candidate observation values ​​corresponding to each candidate point. The initial observation dataset is updated with the candidate points and their corresponding candidate observations. Each data point in the updated initial observation dataset is labeled with its corresponding fidelity data type. The updated initial observation dataset is then processed using the Gaussian process model to update the prediction mean coefficient and exploration weight coefficient of the Gaussian process model, thus completing the update of the Gaussian process model. Based on the updated Gaussian process model, the operations of selecting candidate points in the driving scene element parameter space, updating the initial observation dataset, and updating the Gaussian process model are repeatedly performed until the preset number of repetitions is reached to obtain the intermediate observation dataset. Based on the intermediate observation dataset, the Gaussian process model is optimized until it converges, resulting in a converged Gaussian process model. Based on the driving scene element parameter space, combined with the converged Gaussian process model, edge perception acquisition function, global optimization algorithm, high-fidelity model, low-fidelity model, and multidimensional value function, the final observation dataset is obtained. Based on the final observation dataset and the multidimensional value function, the edge scenario of autonomous driving is obtained.

2. The autonomous driving edge scene generation method as described in claim 1, characterized in that, Based on the initial observation dataset, a Gaussian process model is constructed, including: The initial observation dataset is processed using a Gaussian process surrogate model to obtain a Gaussian process model.

3. The autonomous driving edge scene generation method as described in claim 1, characterized in that, Based on the updated Gaussian process model, the process of repeatedly selecting candidate points in the driving scene element parameter space, updating the initial observation dataset, and updating the Gaussian process model until a preset number of repetitions is reached to obtain the intermediate observation dataset includes: For each repeated operation, the weight values ​​in the multidimensional value function are optimized using a smooth transition algorithm.

4. The autonomous driving edge scene generation method as described in claim 1, characterized in that, Fidelity data types include low-fidelity data types and high-fidelity data types; Based on the intermediate observation dataset, the Gaussian process model is optimized until it converges, resulting in a converged Gaussian process model, including: Data labeled as high-fidelity data type are extracted from the intermediate observation dataset to form a high-fidelity dataset; The Gaussian process model is optimized using the high-fidelity dataset until it reaches a convergent state, thus obtaining the converged Gaussian process model.

5. The autonomous driving edge scene generation method as described in any one of claims 4, characterized in that, Based on the driving scene element parameter space, combined with the convergent Gaussian process model, edge-aware acquisition function, global optimization algorithm, high-fidelity model, low-fidelity model, and multi-dimensional value function, the final observation dataset is obtained, including: Using the convergent Gaussian process model, edge-aware acquisition function, and global optimization algorithm, a convergence point is selected in the parameter space of the driving scene elements; based on the convergence point, combined with the high-fidelity model, the low-fidelity model, and the multidimensional value function, candidate observations corresponding to the convergence point are obtained; The initial observation dataset is updated with the convergence point and its corresponding candidate observations, and the convergent Gaussian process model is updated based on the updated initial observation dataset. Based on the updated convergent Gaussian process model, the operations of selecting convergence points in the driving scene element parameter space, updating the initial observation dataset, and updating the convergent Gaussian process model are repeated until the amount of data of the specified type in the updated initial observation dataset is not less than the data amount threshold, thus obtaining the final observation dataset.

6. The autonomous driving edge scene generation method as described in claim 5, characterized in that, Based on the final observation dataset and the multidimensional value function, the autonomous driving edge scenario is obtained, including: A high-fidelity data type was selected from the final observation dataset to form a screening dataset; The high-fidelity model is used to process the sample points corresponding to each data point in the filtered dataset to obtain the physical feature set of the sample points corresponding to each data point in the filtered dataset; The physical feature set of each data point in the selected dataset is processed using the multidimensional value function to obtain the physical collision risk value and edge metric value of each data point in the selected dataset. In the selected dataset, data with edge metric values ​​greater than a preset edge metric threshold and physical collision risk values ​​between the minimum and maximum physical collision risk thresholds are selected as autonomous driving edge scenarios.

7. An autonomous driving edge scene generation device, characterized in that, include: The initial observation dataset determination module is used to uniformly collect multiple sample points in the driving scene element parameter space using a space-filling sampling strategy. Each sample point is processed using a low-fidelity model to obtain the physical feature set corresponding to each sample point; The physical feature set of each sample point is processed using a multidimensional value function to obtain the observation value corresponding to each sample point; all sample points and their corresponding observation values ​​are combined to form an initial observation dataset; A Gaussian process model building module is used to build a Gaussian process model based on the initial observation dataset; The final observation dataset determination module is used to update the predicted mean coefficient and exploration weight coefficient in the edge perception acquisition function using the predicted mean coefficient and exploration weight coefficient in the Gaussian process model, so as to obtain the updated edge perception acquisition function; and to select candidate points in the driving scene element parameter space using the global optimization algorithm and the updated edge perception acquisition function. The edge-aware acquisition function is shown below: , Represents the edge-aware acquisition function. μ n Represents the predicted mean coefficient. σ n Represents the exploration weighting coefficient. β n Represents the adaptive exploration weight coefficient. Represents the gradient term; The candidate points are processed using an updated edge-aware acquisition function to obtain their perceived values. These perceived values ​​are then compared to a preset perception threshold. If the perceived value is not less than the threshold, a low-fidelity model is used to process the candidate points, resulting in a physical feature set. If the perceived value is less than the threshold, a high-fidelity model is used to process the candidate points, again obtaining a physical feature set. A multidimensional value function is then used to process the physical feature set to obtain candidate observation values. The initial observation dataset is updated with the candidate points and their corresponding candidate observation values, and each data point in the updated dataset is labeled with its corresponding fidelity data type. The updated initial observation dataset is processed using the Gaussian process model to update the predicted mean coefficient and exploration weight coefficient of the Gaussian process model, thus completing the update of the Gaussian process model. Based on the updated Gaussian process model, the operations of selecting candidate points in the driving scene element parameter space, updating the initial observation dataset, and updating the Gaussian process model are repeatedly performed until a preset number of repetitions are reached to obtain an intermediate observation dataset. Based on the intermediate observation dataset, the Gaussian process model is optimized until it converges, resulting in a converged Gaussian process model. Based on the driving scene element parameter space, combined with the converged Gaussian process model, edge perception acquisition function, global optimization algorithm, high-fidelity model, low-fidelity model, and multidimensional value function, the final observation dataset is obtained. The autonomous driving edge scene determination module is used to obtain the autonomous driving edge scene based on the final observation dataset and the multidimensional value function.

8. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements the autonomous driving edge scene generation method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for an autonomous driving edge scene generation method, which, when executed by a processor, implements the steps of the autonomous driving edge scene generation method according to any one of claims 1-6.

Images