Automatic driving safety test system and method

Abstract

The invention provides an automatic driving safety test system and method, and relates to the technical field of automatic driving, and the system comprises a conventional traffic scene generation subsystem which is used for generating a conventional driving scene data set in a simulation platform; the safety key scene generation subsystem is used for generating a safety key scene by utilizing an automatic driving scene generation network SAGA based on adversarial disturbance according to the conventional driving scene data set; and the safety key scene clustering analysis subsystem is used for dynamically discovering a new category and optimizing a clustering structure by introducing a progressive unsupervised clustering method according to the generated safety key scene to obtain a scene analysis result. According to the method, the problems of insufficient test scene coverage, weak resistance test capability, low key scene analysis efficiency and the like in the safety verification of the existing automatic driving system are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of autonomous driving, and in particular relates to an autonomous driving safety test system and method. Background Art

[0002] With the rapid development of autonomous driving technology, it has shown great potential in improving traffic efficiency and reducing the accident rate. However, there are still many challenges in the robustness and safety of autonomous driving systems in the actual complex traffic environment, and safety issues have become an important factor restricting their large-scale application. Therefore, conducting comprehensive and efficient safety testing and verification of autonomous driving systems has become a key task for the research and industrial communities. Currently, the testing methods of autonomous driving systems mainly include real-road testing and simulation platform testing. The former has problems such as high cost, long cycle, and insufficient safety, while the latter depends on the simulation platform and its scenario generation ability. Although there have been relevant studies on the architecture of autonomous driving systems, key scenario generation, and test analysis methods, there are still many deficiencies in aspects such as automated scenario generation, diversity testing, and dynamic risk analysis. The relevant technical background is elaborated as follows.

[0003] (1) Autonomous Driving System Architecture and Its Existing Challenges

[0004] Currently, most mainstream autonomous driving systems adopt a modular design strategy. The modular system divides functions such as perception, prediction, and planning into independent modules for development and integration, with good interpretability and the ability to optimize modules independently. However, due to the weak coupling between modules, the overall system optimization is insufficient, and problems such as information redundancy and decreased decision-making efficiency may occur during the module interaction process. To make up for the deficiencies of the modular architecture, end-to-end autonomous driving systems have gradually emerged in recent years. The end-to-end system directly maps sensor inputs to control outputs through a single neural network, avoiding information loss between modules and improving the overall system optimization ability. End-to-end autonomous driving systems mainly adopt two methods: imitation learning and reinforcement learning. IL realizes the training of efficient and stable autonomous driving strategies by learning expert behavior demonstrations, but its generalization ability is limited, and it is prone to system failure due to input perturbations, especially in safety-critical scenarios. RL has the ability to actively explore the environment and optimize complex strategies, but the training process requires a large amount of data and usually relies on a simulation environment for high-risk decision-making training. The RL method is sensitive to the design of the reward function, and existing systems often use simple objective functions, which are prone to risky behaviors. In addition, RL is difficult to optimize in high-dimensional environments, with problems such as insufficient gradients and sparse feedback.

[0005] Generally speaking, the robustness and stability of existing autonomous driving systems in safety-critical scenarios still need to be improved. Therefore, designing testing and analysis methods for autonomous driving systems facing complex scenarios has become an urgent problem to be solved.

[0006] (2) Safety-critical scenario generation technology

[0007] Existing safety-critical scenario generation methods can be roughly divided into three categories: Data-driven methods perform scenario sampling or distribution learning based on large-scale traffic datasets to generate high-risk traffic scenarios. Although such methods improve scenario diversity, they are limited by the scarcity of data sources and are difficult to cover long-tail scenarios. In addition, data-driven generation lacks flexibility and is difficult to dynamically adapt to different test requirements. Knowledge-based scenario generation methods utilize traffic rules and expert knowledge to generate safety-critical test scenarios through rule reasoning or model constraints. Although they can ensure the logical rationality and safety of the generated scenarios, they rely too much on manual design, resulting in insufficient innovation in the generated scenarios and difficulty in capturing potential risk scenarios of the system. Adversarial generation methods generate challenging test scenarios by simulating adversarial behaviors. Although adversarial generation methods improve scenario complexity and challenge, they have the following problems: high computational complexity for large-scale optimization, which is prone to the curse of dimensionality; difficulty in guaranteeing the authenticity of the generated scenarios, relying on strong rule constraints to improve rationality; and relying on adversarial model design, with limited adaptability and generalization ability. Existing generation methods have not yet formed an integrated solution for automated, dynamic, and diverse scenario generation and system robustness evaluation.

[0008] (3) Current situation of safety-critical scenario clustering and analysis technology

[0009] The generation of large-scale autonomous driving test scenarios has given rise to a vast amount of test data, and effective scenario clustering and analysis technologies are urgently needed to achieve risk pattern mining and system vulnerability location. Currently, commonly used methods include:

[0010] Traditional clustering algorithms: Unsupervised clustering algorithms such as K-means and DBSCAN have been applied to the classification and analysis of autonomous driving test data. These methods rely on preset parameters (such as the number of categories and density thresholds), and in high-dimensional complex data and dynamic environments, there are problems such as unstable clustering results and fuzzy category boundaries.

[0011] Spectral clustering and anchor graph clustering algorithms: Spectral clustering constructs a graph structure through a similarity matrix and performs eigen-decomposition to achieve clustering analysis, which is suitable for the classification of complex scenario data. However, this method has a high computational complexity, especially in a large-scale data environment, where the Laplacian matrix decomposition consumes a large amount of computing resources. The anchor graph clustering method simplifies the calculation of data similarity by selecting representative anchor points, reducing the computational complexity of spectral clustering, but faces challenges such as uneven anchor point selection and insufficient stability of the clustering structure.

[0012] Progressive Unsupervised Clustering Method: The progressive clustering method combines anchor point optimization and label propagation algorithm to achieve dynamic category discovery and adaptive clustering. Through gradual optimization and category boundary adjustment, this method has the ability to handle dynamically growing data and discover new categories, improving the adaptability and scalability of the clustering algorithm. However, existing progressive clustering methods still have problems such as insufficient computational efficiency and weak robustness in large-scale autonomous driving test environments, making it difficult to achieve real-time and efficient dynamic risk analysis.

[0013] Main Problems in the Existing Technology: Although existing research has made certain progress in the architecture of autonomous driving systems, scenario generation, and test data analysis, there are still the following deficiencies:

[0014] The automation degree of scenario generation methods is not high, the cost is high, it is difficult to cover complex and extreme safety-critical scenarios, and the generation efficiency and authenticity are insufficient; the computational complexity of test data analysis methods is high, the ability of dynamic category recognition and risk analysis is insufficient, and it is difficult to achieve real-time analysis and risk warning of large-scale test data; there is a lack of a unified autonomous driving safety test platform, and the closed-loop integration of scenario generation, system vulnerability mining, and risk analysis cannot be achieved, affecting the automation and efficiency of the test process. Summary of the Invention

[0015] In view of the above deficiencies in the existing technology, an autonomous driving safety test system and method provided by the present invention solve the problems such as insufficient coverage of test scenarios, weak adversarial testing ability, and low efficiency of key scenario analysis in the safety verification of existing autonomous driving systems.

[0016] To achieve the above objectives, the technical solution adopted by the present invention is: An autonomous driving safety test system, including:

[0017] A conventional traffic scenario generation subsystem for generating a conventional driving scenario dataset in a simulation platform;

[0018] A safety-critical scenario generation subsystem for generating safety-critical scenarios according to the conventional driving scenario dataset by using the autonomous driving scenario generation network SAGA based on adversarial perturbations;

[0019] A safety-critical scenario clustering and analysis subsystem for dynamically discovering new categories and optimizing the clustering structure by using the introduced progressive unsupervised clustering method according to the generated safety-critical scenarios, obtaining scenario analysis results, and completing the autonomous driving safety test.

[0020] Further, the conventional traffic scenario generation subsystem includes:

[0021] A road network topology modeling module for extracting road topology information by analyzing the road network in the simulation platform and modeling the road network topology based on the road topology information;

[0022] A path planning and optimization module, which is used to generate a smooth driving path from the starting point to the target point based on the constructed road network topology by using a path planning algorithm, and optimize the smooth driving path by using the interpolation method;

[0023] A multi-modal data acquisition module, which is used to synchronously collect various types of scenario data by using multi-modal sensors based on the optimized smooth driving path, and construct a conventional driving scenario dataset.

[0024] Furthermore, the safety-critical scenario generation subsystem includes:

[0025] A conventional training module, which is used to train an autonomous driving agent by using the conventional driving scenario dataset to form a benchmark model with driving decision-making and environment perception capabilities;

[0026] An adversarial scenario generation module, which is used to generate an adversarial scenario for the victim model in the initial scenario based on the training results by using a gradient-based adversarial perturbation strategy and the behavior trajectory of the attacker. Among them, the optimal perturbation sequence is found by minimizing the objective function to generate a safety-critical scenario;

[0027] A victim model module, which is used to provide real-time feedback during the perturbation generation process to optimize the adversarial perturbation strategy. Among them, the victim model is an autonomous driving decision-making system that serves as the attack target during the adversarial scenario generation process.

[0028] Furthermore, the training of the autonomous driving agent by using the conventional driving scenario dataset is specifically as follows:

[0029] According to the conventional driving scenario dataset, the observation data O t is combined with the target trajectory point G to train the autonomous driving policy network π θ , forming a benchmark model with driving decision-making and environment perception capabilities.

[0030] Furthermore, the expression of the optimal perturbation sequence is as follows:

[0031]

[0032] Among them, θ* represents the optimal perturbation sequence, θ represents the set of all adversarial agent action sequences, J(S) represents the minimized objective function, θ i' represents the driving strategy of the i'-th adversarial agent action sequence. i = 0 represents the autonomous driving agent associated with the victim model, i ∈ {1,..., N} represents the traffic agents in the scenario except i = 0, and N represents the number of adversarial agents in the adversarial sequence. represents the action of the i-th autonomous driving agent at time t, T represents the number of simulation steps, and t represents the time. represents the collision distance cost of the autonomous driving agent, where α, β, and γ all represent adjustable hyperparameters, represents the speed difference cost of the autonomous driving agent, represents the safety distance cost between adversarial agents, represents the out-of-bounds constraint cost, and D() represents the Euclidean distance operation, represents the state corresponding to the victim model, represents the state of the i-th autonomous driving agent at time t, v bb represents the speed difference, and τ represents the minimum safety threshold, represents the state of the j-th autonomous driving agent at time t, L represents the number of corner points included in each vehicle bounding box, m represents the index of the currently processed corner point, which is used to traverse all the corner point positions of the vehicle bounding box at each time step, and f() represents the Gaussian potential function around the corner point, represents the bounding box of the adversarial agent i at time t, p m represents the corner point of the bounding box, represents the convolution operation with the non-traversable area in the map.

[0033] Furthermore, based on the generated safety-critical scenarios, by introducing a progressive unsupervised clustering method to dynamically discover new categories and optimize the clustering structure, a scenario analysis result is obtained, specifically as follows:

[0034] Perform hierarchical partitioning on the generated safety-critical scenarios. Among them, in each partitioning process, select representative samples to construct an anchor point set;

[0035] Generate pseudo-labels based on the anchor point set;

[0036] Based on the pseudo-labels, construct an anchor point graph based on data similarity, and use representative samples to describe the distribution characteristics of the data;

[0037] Based on the distribution characteristics, spread high-confidence information through the label propagation mechanism to make the clustering boundary reach a clear state;

[0038] When a new sample different from the existing categories is detected, create a new category and dynamically update the clustering result until the clustering result meets the requirements, obtaining the scenario analysis result.

[0039] The present invention also provides an autonomous driving safety testing method. The autonomous driving safety testing method is used to execute the autonomous driving safety testing system, including the following steps:

[0040] Generate a conventional driving scenario dataset in the simulation platform;

[0041] According to the conventional driving scenario dataset, the safety-critical scenarios are generated by using the self-driving scenario generation network SAGA based on adversarial perturbations.

[0042] According to the generated safety-critical scenarios, the progressive unsupervised clustering method is introduced to dynamically discover new categories and optimize the clustering structure, and the scenario analysis results are obtained to complete the safety test of autonomous driving.

[0043] Advantages of the present invention:

[0044] (1) Compared with the prior art, the present invention proposes a closed-loop autonomous driving algorithm safety test platform framework, which has significant innovation and superiority. Existing testing methods for autonomous driving systems mostly rely on actual road collection or rule-based preset scenarios, which not only have high testing costs but also are difficult to comprehensively cover the long-tail and extreme situations in complex traffic environments, resulting in insufficient verification of system safety. To solve the above problems, based on the CARLA simulation platform, the present invention designs and implements functional modules such as automatic generation of conventional traffic scenarios, generation of adversarial safety-critical scenarios, and dynamic clustering analysis of large-scale test data, thereby constructing a unified autonomous driving algorithm safety test platform and realizing the automation and closed-loop integration of the test process.

[0045] (2) Through the innovative scenario generation method, adversarial testing mechanism, and dynamic data analysis strategy, the present invention significantly improves the safety verification ability of autonomous driving systems and the optimization efficiency of algorithm robustness, and has significant advantages such as high automation, strong system integration, rich scenario diversity, and accurate risk mining. It breaks through the technical bottleneck of existing testing methods and provides a solid technical guarantee for the safety improvement and industrial application of autonomous driving systems.

[0046] (3) Through road topology modeling and the JPS path planning algorithm, combined with B-spline interpolation to optimize path smoothness, the present invention automatically generates conventional traffic scenarios, significantly reducing the cost of test data collection and improving the data collection efficiency and traffic scenario diversity. At the same time, the privileged expert agent model is used to drive multi-modal data collection to ensure the high precision and stability of data collection, providing rich and reliable basic data for the training and evaluation of autonomous driving algorithms. Compared with traditional real vehicle testing methods, the present invention significantly improves the automation of data generation and the realism of simulation scenarios, greatly expanding the coverage of the test environment.

[0047] (4) In terms of generating safety-critical scenarios, the present invention proposes an adversarial scenario generation method based on the generator-victim model architecture, which automatically generates test scenarios with high challenges and strong risks through the policy gradient optimization algorithm. Through attacker trajectory perturbation and multi-dimensional objective function optimization, precise mining of potential decision-making vulnerabilities of the victim model is achieved, significantly improving the comprehensiveness and effectiveness of the robustness test of the autonomous driving system. In addition, traffic rules and physical constraints are introduced during the generation process to ensure that the generated scenarios not only have challenges but also meet the rationality and feasibility of the real traffic environment, solving problems such as scenario distortion and policy failure in existing adversarial generation methods.

[0048] (5) For the need of test data analysis, the present invention adopts a progressive unsupervised clustering analysis method, which realizes preliminary hierarchical clustering based on the BKHK algorithm. Through anchor graph construction and label propagation mechanism, new categories are dynamically discovered and the clustering structure is optimized. This method significantly improves the processing efficiency of large-scale high-dimensional test data, enhances the automatic recognition and clustering ability of high-risk scenarios, and overcomes the problem that existing static clustering methods cannot adapt to the dynamic evolution of data. Through continuous iterative optimization, the system can efficiently refine typical high-risk scenario patterns, providing reliable data support for the dynamic risk warning and safety optimization of the autonomous driving system.

[0049] (6) The present invention builds a unified closed-loop safety test platform for autonomous driving algorithms, integrating functions such as conventional scenario generation, adversarial scenario testing, and risk analysis, forming a full-process automated processing mechanism of data generation - vulnerability mining - dynamic analysis. Compared with the existing decentralized test system, the test process of the present invention is more systematic and automated, greatly improving the test efficiency and the algorithm optimization iteration speed, shortening the development cycle of the autonomous driving system, and enhancing the overall safety and robustness of the system. The platform is designed based on the open-source CARLA simulation environment, has good scalability and compatibility, supports the access of multiple types of autonomous driving algorithms, can adapt to different sensor configurations and test requirements, and is convenient for integration and practical application promotion. Description of the Drawings

[0050] Figure 1 It is a schematic diagram of the system structure of the present invention.

[0051] Figure 2 It is a framework diagram of the platform of the present invention.

[0052] Figure 3 It is a schematic diagram of the generation process of the conventional traffic scenario dataset in the present invention.

[0053] Figure 4 It is a schematic diagram of the overall network architecture and the interaction relationship between each module in the present invention.

[0054] Figure 5 It is a flowchart of the method of the present invention. Detailed implementation manners

[0055] The following describes the detailed implementation manners of the present invention to facilitate those skilled in the art of this technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the detailed implementation manners. For those of ordinary skill in the art of this technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

[0056] Example 1

[0057] Aiming at the problems existing in the safety verification of existing autonomous driving systems, such as insufficient coverage of test scenarios, weak adversarial testing capabilities, and low efficiency of critical scenario analysis, the present invention provides a framework for an autonomous driving algorithm test platform. Through the testing and analysis of the autonomous driving system in a simulation environment, this framework realizes functions such as automatic generation of safety-critical scenarios, vulnerability mining of system robustness, and dynamic clustering analysis, forming a closed-loop autonomous driving safety verification solution.

[0058] The present invention focuses on solving the following technical problems:

[0059] (1) Insufficient efficiency and diversity in generating safety-critical scenarios: Existing data-driven and knowledge-rule scenario generation methods have problems such as scarce samples, limited coverage, and insufficient challenge, making it difficult to generate diverse and challenging long-tail extreme test scenarios. The present invention proposes an automated scenario generation method based on a generative adversarial mechanism, which automatically mines system vulnerabilities through an adversarial optimization strategy to realize the automatic generation of complex extreme traffic scenarios, significantly improving the test coverage rate and challenge.

[0060] (2) Insufficient authenticity and rationality in generating safety-critical scenarios: When current adversarial generation methods improve the challenge of scenarios, they often ignore the rationality and interpretability of traffic scenarios, resulting in generated results lacking the physical constraints and behavior norms of the real traffic environment. The present invention designs a fusion mechanism based on expert knowledge and traffic rules, combining traffic behavior norms with the optimization process to ensure that the generated scenarios are both challenging and meet the rationality and feasibility in the real traffic environment.

[0061] (3) Low efficiency in analyzing large-scale test data and insufficient ability to discover categories: With the continuous growth of the number and complexity of test scenarios, existing test result analysis methods are difficult to effectively handle large-scale high-dimensional data and are difficult to dynamically discover new high-risk scenario patterns. The present invention proposes a progressive unsupervised clustering analysis method, which realizes the dynamic classification and risk analysis of test data through an anchor point optimization and label propagation mechanism, supports the automatic discovery of new categories, and significantly improves the analysis efficiency and accuracy of the test results of the autonomous driving system.

[0062] (4) Lack of a unified closed-loop safety testing platform for autonomous driving algorithms: The current processes of scenario generation, system testing, and data analysis are scattered, making it difficult to form a systematic and automated safety testing platform. Through integrating key technologies such as simulation testing, adversarial scenario generation, and dynamic clustering analysis, the present invention constructs a unified testing platform for autonomous driving algorithms, realizing a closed-loop automated processing flow of data generation - vulnerability mining - dynamic analysis.

[0063] As Figure 1 and Figure 2 shown, the present invention provides an autonomous driving safety testing system, including:

[0064] A conventional traffic scenario generation subsystem, configured to generate a conventional driving scenario dataset in a simulation platform, including:

[0065] A road network topology modeling module, configured to extract road topology information by analyzing the road network in the simulation platform, and model the road network topology based on the road topology information;

[0066] A path planning and optimization module, configured to generate a smooth driving path from a starting point to a target point by using a path planning algorithm based on the constructed road network topology, and optimize the smooth driving path by using an interpolation method;

[0067] A multi-modal data acquisition module, configured to synchronously acquire various types of scenario data by using multi-modal sensors based on the optimized smooth driving path, and construct a conventional driving scenario dataset;

[0068] A safety-critical scenario generation subsystem, configured to generate safety-critical scenarios according to the conventional driving scenario dataset by using an autonomous driving scenario generation network SAGA based on adversarial perturbations, including:

[0069] A conventional training module, configured to train an autonomous driving agent by using the conventional driving scenario dataset to form a benchmark model with driving decision-making and environment perception capabilities, specifically:

[0070] According to the conventional driving scenario dataset, combine the observation data O t with the target trajectory point G to train the autonomous driving policy network π θ , and form a benchmark model with driving decision-making and environment perception capabilities;

[0071] An adversarial scenario generation module, configured to generate adversarial scenarios for the victim model in the initial scenario according to the training results by using a gradient-based adversarial perturbation strategy, where the optimal perturbation sequence is found by minimizing the objective function to generate safety-critical scenarios;

[0072] The victim model module is used to provide real-time feedback during the perturbation generation process to optimize the adversarial perturbation strategy. Here, the victim model is the autonomous driving decision-making system that serves as the attack target during the generation process of adversarial scenarios;

[0073] The safety-critical scenario clustering analysis subsystem is used to, based on the generated safety-critical scenarios, dynamically discover new categories and optimize the clustering structure by introducing a progressive unsupervised clustering method to obtain the scenario analysis results and complete the safety test for autonomous driving. Specifically:

[0074] Perform hierarchical partitioning on the generated safety-critical scenarios. Among them, representative samples are selected to construct an anchor point set during each partitioning process;

[0075] Generate pseudo-labels based on the anchor point set;

[0076] Based on the pseudo-labels, construct an anchor graph based on data similarity and use representative samples to characterize the distribution characteristics of the data;

[0077] Based on the distribution characteristics, spread high-confidence information through the label propagation mechanism to make the clustering boundary reach a clear state;

[0078] When a new sample different from the existing categories is detected, create a new category and dynamically update the clustering results until the clustering results meet the requirements to obtain the scenario analysis results.

[0079] In this embodiment, the present invention proposes a closed-loop autonomous driving algorithm safety test platform framework. Based on the CARLA simulation platform, it integrates functions such as conventional traffic scenario generation, autonomous driving system behavior training and evaluation, adversarial scenario generation and testing, and risk clustering analysis to form a full-process automated test system from data generation to system testing, and then to risk mining and analysis. The platform includes the following three core technology subsystems: the conventional traffic scenario generation subsystem, the safety-critical scenario generation subsystem, and the clustering analysis subsystem. The platform framework diagram is as Figure 2 shown.

[0080] In this embodiment, the conventional traffic scenario generation subsystem: In the actual road environment, data for conventional driving scenarios usually needs to be collected through real vehicle tests, recording the operating data of the vehicle under different traffic conditions to support the development and verification of autonomous driving technologies. However, actual road tests are costly, time-consuming, and it is difficult to effectively cover a rich variety of traffic situations. To solve the above problems, the present invention selects to generate a conventional driving scenario dataset in the CARLA simulation platform to break through the limitations of traditional data collection methods. First, by analyzing the road network of the simulation town in the CARLA platform, detailed road topology information is extracted, including road punctuation marks, road segments, and their corresponding geometric and semantic attributes, and then a directed graph model is constructed to accurately describe the road structure. On this basis, a path planning algorithm is applied to generate a smooth driving path from the starting point to the target point, and the interpolation method is used to further optimize the continuity and feasibility of the path to ensure that the generated path meets the vehicle dynamics constraints and actual road operating conditions. Subsequently, a privileged expert agent model is deployed to synchronously collect various types of scenario data such as RGB images, depth images, semantic segmentation images, point cloud data, and vehicle dynamic state information using multi-modal sensors, constructing a conventional traffic scenario dataset with both authenticity and diversity. This conventional traffic scenario dataset not only provides reliable and rich data support for the development and performance evaluation of autonomous driving algorithms, but also provides a solid data foundation for the generation of subsequent safety-critical scenarios. The specific design flowchart is as shown in Figure 3 shown.

[0081] In this embodiment, a path planning algorithm based on the Jump Point Search (JPS) algorithm and B-spline interpolation is designed and implemented. The specific implementation process of the JPS algorithm is as follows:

[0082] (1) Node initialization: Initialize the open set and the closed set. Add the starting point to the open set and set the cost gScore = 0, and set the costs of the remaining nodes to infinity.

[0083] (2) Jump expansion: Only perform jumps in the current search direction until an obstacle is encountered or a forced neighbor is met, and mark it as a key node.

[0084] (3) Cost update: Update the cost value of the key node after the jump and add it to the open set to ensure the optimality of the path.

[0085] (4) Path backtracking: When the target node is added to the closed set, use the backtracking method to backtrack from the target node to the starting point in turn to generate the final path.

[0086] In this embodiment, the B-spline interpolation algorithm optimizes the path generated by JPS, and the specific steps are as follows:

[0087] (1) Path point filtering: Remove dense and redundant points in the path to reduce the complexity of subsequent interpolation calculations.

[0088] (2) B-spline curve generation: Use the filtered path points as control points to generate a B-spline curve to make the path smoother and more continuous.

[0089] (3) Kinematic constraint adjustment: Appropriately adjust the spacing of the control points to ensure that the turning radius of the path meets the constraint conditions of vehicle kinematics and avoid excessive steering angle changes.

[0090] In this embodiment, the safety-critical scenario generation subsystem: The present invention proposes an autonomous driving scenario generation network SAGA based on adversarial perturbations, aiming to solve the problem of insufficient testing of high-risk scenarios in autonomous driving systems. By designing a perturbation strategy for the attacker's trajectory, this network specifically exploits the decision-making vulnerabilities of the victim model to generate extreme and diverse safety-critical scenarios, thereby expanding the distribution range and diversity of training data. Specifically, the autonomous driving scenario generation network SAGA is based on a conventional scenario dataset, extracting the geographical structure, traffic participant behavior, and environmental conditions therein as the initial input information to support the evolution of adversarial scenarios. The autonomous driving scenario generation network SAGA consists of a conventional training module, an adversarial scenario generation module, a victim model module (Victim Models), and an objective function optimization module. These modules work together to gradually complicate the scenario and approach extreme situations. The conventional training module initially trains the autonomous driving policy model based on the conventional scenario dataset Dreg to ensure its basic driving ability and decision-making ability, thus forming a baseline model under attack. This module mainly trains the victim model and the adversarial agent sequence. The adversarial scenario generation module is the core part of the framework. Through a gradient-based adversarial perturbation strategy, it generates adversarial scenarios for the victim model in the initial scenario using the attacker's behavior trajectory. The perturbation generation process guides the scenario to evolve to a high-risk state, simulating an extreme driving environment. The victim model module consists of specific autonomous driving models (AIM-BEV and TF++), which provide real-time feedback during the generation process as the attack targets to assist in optimizing the adversarial perturbation strategy. Finally, the objective function optimization module quantifies the extremity and diversity of the scenario through a multi-dimensional cost function (such as collision distance, speed difference, out-of-bounds penalty, etc.) to ensure that the generated scenarios are more challenging and representative. The generation process of the autonomous driving scenario generation network SAGA starts from the initial scenario, optimizes the attacker's behavior trajectory through the perturbation strategy, and gradually adjusts and generates new adversarial scenarios. In each iteration, the framework evaluates the perturbation effect through the objective function and optimizes the perturbation strategy according to the feedback until the generated adversarial scenario causes a collision or other high-risk behaviors between the attack target and the victim model, completing the optimization process. As Figure 4 shown, Figure 4 it shows the overall network architecture of the algorithm framework and the interaction relationships between various modules.

[0091] In this embodiment, the generation of adversarial scenarios is as follows:

[0092] The adversarial scenario generation module aims to generate high-risk adversarial scenarios that can induce the failure of the autonomous driving system by guiding the attacker to perform interference behaviors through a carefully designed perturbation strategy based on existing conventional driving scenarios. The specific process is as follows: First, the adversarial scenario generation module selects a conventional scenario as the initial environment and deploys the victim model (such as AIM-BEV, TF++) and the attacker agent in it. The attacker acts as a perturber in the environment, and its behavioral parameters (such as position, speed, acceleration, etc.) will be continuously adjusted as optimization variables. Then, the module initializes the attacker's action trajectory and defines the perturbation range. In each frame of the simulation process, the attacker executes actions according to the current perturbation parameters, the environmental state is updated accordingly, and the victim model responds to the environmental changes based on its strategy. Subsequently, the system evaluates the perturbation effect according to the risk situation caused by the perturbation and optimizes and adjusts the perturbation trajectory according to the feedback results. The evaluation process depends on a set of preset risk indicators, such as: the minimum distance between two vehicles, speed difference, whether a collision occurs, whether there is a boundary crossing, etc., to comprehensively measure the extremity of the current scenario. The optimization process is carried out iteratively, and the attacker's trajectory is adjusted according to the victim's behavior feedback after each perturbation attempt, gradually guiding the scenario to evolve towards a higher-risk state. Finally, when a collision or significant abnormal behavior is successfully induced in the scenario, this round of perturbation process terminates, and the generated adversarial scenario is output as a test sample. This module emphasizes the goal-orientation of perturbation and the diversity of scenarios in its design, can effectively discover potential decision-making loopholes in the victim model, and then expand the coverage of extreme cases in the scenario dataset, improving the robustness testing ability of the autonomous driving system in high-risk environments.

[0093] In this embodiment, the conventional training module is the first step in the entire safety-critical scenario generation framework. Its core goal is to preliminarily train the autonomous driving agent through an imitation learning strategy to improve its driving ability and robustness in a conventional traffic environment. This module uses the conventional traffic scenario dataset Dreg to perform supervised learning training on the autonomous driving policy model to form a benchmark model with basic driving decision-making and environmental perception capabilities. During the training process, the conventional training module combines the input observation data O t , including in-vehicle camera images, lidar point clouds, and vehicle state information, with the target trajectory points G to train the autonomous driving policy network π θ .

[0094] In this embodiment, the core goal of the adversarial scenario generation module is to generate extreme driving scenarios by iteratively optimizing the perturbation sequence to test and challenge the driving strategy of the target victim model. The design of this module includes the following parts: the definition of states and actions, the adversarial perturbation control sequence and scenario generation, and the corresponding cost function design.

[0095] (1) States and Actions in Traffic Simulation

[0096] First, define the state variables of each agent in traffic simulation: planar position heading speed magnitude Among them, the index i = 0 represents the autonomous driving agent associated with the victim model, and the remaining i ∈ {1,..., N} represent other traffic agents in the scenario (including adversarial agents). At time t, the states of all agents can be represented by the set The state corresponding to the victim model is denoted as The state set of the entire traffic scenario at the time sequence t = 0, 1, 2,..., T is then represented as:

[0097]

[0098] Among them, T is the fixed number of simulation steps or simulation duration. At initialization, X will be set as a regular and non - colliding traffic scenario as the starting point. In each step of the simulation, if the state x t at time t and the actions of all agents (including the victim model and adversarial agents) are known, then the state x t+1 at the next time is given by the following formula:

[0099] x t+1 = K(x t , a t )

[0100] Among them, K represents a differentiable bicycle kinematic model. Since this model can well depict the physical motion characteristics of non - holonomic vehicles and is differentiable, it is conducive to subsequent gradient propagation and adversarial perturbation optimization.

[0101] During the simulation process, the victim model will determine its own actions t according to the observation information O provided by the environment. These observation data are given by the rendering function, as shown in the following formula:

[0102] Q t = F(x t , M)

[0103] Among them, M represents the simulation environment map (such as road information, drivable area, etc.). The rendering function R is used to convert the simulation state and map information into input data for the attacked model to make decisions. Depending on the type of attacked model, the rendering method may be different. The AIM-BEV model (BEV perspective autonomous driving model based on imitation) uses a differentiable rasterization method to convert vehicle and road information into a bird's-eye view (BEV) format to ensure that the data processing process is differentiable. The TF++ model (improved TransFuser fusion model) uses the non-differentiable graphics engine of the CARLA simulator to generate LIDAR point clouds, and uses PointNet technology to convert the point clouds into differentiable forms, thereby supporting subsequent cost functions and gradient calculations.

[0104] (2) Action sequence and scene generation of adversarial agents

[0105] In order to explore and simulate safety-critical traffic scenarios, the present invention focuses on optimizing the action sequence of the adversarial agent (i.e., the agent with i≠0) By continuously updating the action sequence of the adversarial agent, extreme scenarios can be gradually approached during the simulation process to conduct more rigorous tests on the victim model.

[0106] Let the parameter set θ represent the set of all adversarial agent action sequences, that is, Since the action of each agent at each moment may be composed of two control variables (such as steering angle and acceleration), the dimension of θ is N×T×2. We find the optimal perturbation sequence θ* by minimizing the objective function J(S), as follows:

[0107]

[0108] The objective function J(S) comprehensively considers factors such as the scenario's safety, feasibility, and attack target, where S represents the set of spatiotemporal evolution sequences of the entire scenario. During adversarial scenario generation, the adversarial agent observes and leverages the victim model's control actions and environmental reactions from the previous step, gradually adjusting its own actions to produce unfavorable scenarios such as collisions, sudden braking, and dangerous lane changes. After each iteration, θ is updated via gradient descent, allowing the adversarial scenario to gradually converge to extreme conditions that pose a significant threat to the autonomous driving system.

[0109] (3) Cost function design

[0110] The design of the objective function J(S) is crucial. This paper comprehensively considers key factors such as collision induction, safe spacing between opposing vehicles, and cross-boundary penalties, and sets the overall objective function as the following weighted sum:

[0111]

[0112] Among them, α, β, and γ all represent adjustable hyperparameters used to control the weights of each loss term in the overall objective. The definitions of specific loss terms are as follows:

[0113] 1) Collision distance cost of the autonomous driving agent

[0114] To quantify the collision risk between the autonomous driving agent i = 0 and the adversarial agent i > 0, this term is defined as the average value of the shortest distance between the bounding boxes of this agent and the nearest adversarial agent over the simulation steps T, as shown in the following formula:

[0115]

[0116] By minimizing this distance, the adversarial agent is guided to choose actions that are likely to cause collisions or very close interactions.

[0117] 2) Speed difference cost of the autonomous driving agent

[0118] Equally important as the impact distance is the influence of the speed difference between vehicles on the collision risk. The greater the speed difference, the higher the potential collision risk and destructiveness tend to be, as shown in the following formula:

[0119]

[0120] 3) Safety distance cost between adversarial agents

[0121] To avoid unreasonable mutual collisions within the adversarial agents, the present invention sets a minimum safety threshold τ. If the distance between any two adversarial agents is less than τ, a corresponding penalty is generated. This constraint helps to maintain the physical rationality of the adversarial scenario and avoid overly extreme (or stacked) results, as shown in the following formula:

[0122]

[0123] 4) Out-of-bounds constraint cost

[0124] In addition to the relative collisions between adversarial agents, the present invention also needs to consider the road boundary constraints. If an adversarial vehicle crosses the boundary and enters an impassable area, it means that the vehicle behavior lacks authenticity or rationality. Therefore, an out-of-bounds penalty is defined, as shown in the following formula:

[0125]

[0126] Among them, θ* represents the optimal perturbation sequence, θ represents the set of action sequences of all adversarial agents, J(S) represents the minimized objective function, θ i'Denotes the driving strategy of the i'-th adversarial agent action sequence, where i = 0 represents the autonomous driving agent associated with the victim model, and i ∈ {1,..., N} represents traffic agents in the scenario other than i = 0. Traffic agents include adversarial agent agents and victim models (the autonomous driving model under test), and N represents the number of adversarial agents in the adversarial sequence. Denotes the action of the i-th autonomous driving agent at time t, T represents the number of simulation steps, and t represents the time. Denotes the collision distance cost of the autonomous driving agent, where α, β, and γ all represent adjustable hyperparameters. Denotes the speed difference cost of the autonomous driving agent. Denotes the safety distance cost between adversarial agents. Denotes the out-of-bounds constraint cost, and D() represents the Euclidean distance operation. Denotes the state corresponding to the victim model. Denotes the state of the i-th autonomous driving agent at time t, v bb Denotes the speed difference, and τ represents the minimum safety threshold. Denotes the state of the j-th autonomous driving agent at time t, L represents the number of corner points included in each vehicle bounding box. Usually, a rectangular bounding box has 4 corner points, the bounding box is rectangular, L = 3 (because from m = 0 to L =, there are a total of 4 corner points), and m represents the index of the currently processed corner point, which is used to traverse all the corner point positions p of the vehicle bounding box at each time step. m , and f() represents the Gaussian potential function around the corner point, which is used to measure the distance from the map boundary or non-traversable area. Denotes the bounding box of the adversarial agent (vehicle) i at time t, p m Denotes the corner points of the bounding box. Denotes the convolution operation with the non-traversable area in the map (such as sidewalks, green belts, etc.). If the adversarial vehicle goes out of bounds, this term will increase significantly, corresponding to a high penalty or directly causing the simulation to terminate.

[0127] In this embodiment, traffic agents include adversarial agents and victim models, and each adversarial agent in the adversarial sequence has two contents: state and action.

[0128] The above cost function comprehensively considers multiple constraints, generates high-risk scenarios while ensuring the rationality and physical authenticity of the scenarios, so as to achieve more effective generation of safety-critical scenarios.

[0129] In this embodiment, the victim model refers to the autonomous driving decision-making system that is the target of attack during the generation of adversarial scenarios. The victim model in the present invention adopts an autonomous driving agent trained by an end-to-end imitation learning method, aiming to comprehensively test and verify the safety and robustness of the autonomous driving system in a simulation environment. The overall structure of this model includes three core functional modules: perception, decision-making, and control. For different autonomous driving models, the structures in these three parts are different. The perception module collects environmental data in real time through sensors (such as cameras and lidar), identifies roads, obstacles, and traffic rules, and provides the "eyes" for the system; the decision-making module plans paths, predicts risks, and formulates driving strategies (such as lane changes and braking) based on the perception information, acting as the "brain" to complete intelligent judgments; the control module converts the decisions into vehicle execution instructions (such as steering, throttle, and braking), precisely controls the hardware to achieve actions, just like the "hands and feet" to complete the closed-loop operation.

[0130] In this embodiment, the safety-critical scenario clustering analysis subsystem: In view of the diversity, dynamics, and complexity of data in safety-critical scenarios, the present invention introduces a progressive unsupervised clustering method to achieve adaptive clustering update and new category discovery. This method iteratively optimizes the clustering structure, makes full use of the pseudo-labels generated by the preliminary clustering results, and deeply explores the similarity relationships between data, continuously optimizing the category boundaries, thereby improving the accuracy and adaptability of the clustering effect. Specifically, first, the BKHK algorithm (binary K-Means) is used to perform a preliminary hierarchical partitioning of the data. BKHK adaptively hierarchizes the data based on a binary tree structure. In each partitioning process, representative samples are selected to construct an anchor point set, and preliminary pseudo-labels are generated based on the anchor point set; subsequently, an anchor point graph is constructed based on data similarity, and the distribution characteristics of the data are characterized using representative samples; then, the high-confidence information is diffused through the label propagation mechanism to gradually clarify the clustering boundaries; when new samples that are significantly different from the existing categories are detected, the algorithm can adaptively identify and create new categories, and at the same time dynamically update the clustering structure until the clustering result tends to be stable.

[0131] The adversarial scenario generation method based on the generator-victim model architecture proposed by the present invention uses a policy gradient optimization algorithm to generate adversarial scenarios with high risk and challenge, so as to automatically discover potential vulnerabilities in the autonomous driving system. This method effectively breaks through the bottleneck of the insufficient coverage of extremely complex scenarios by existing testing methods, and ensures the robustness and safety of the autonomous driving system. This technical point is one of the important protected contents of the present invention. For the generated large-scale test scenario data, the present invention designs a progressive unsupervised clustering analysis method, combines the anchor graph construction and label propagation mechanism to realize the dynamic discovery of new categories and the automatic clustering of high-risk scenarios. By this method, the efficiency of test data analysis is improved, the weak scenario patterns of the system are accurately identified, and the dynamic risk warning ability is enhanced. This technical point is the core method of test data analysis and is an important protected content of the present invention. The present invention uses the CARLA simulation platform, combines road topology modeling, JPS path planning and B-spline interpolation to realize the automatic generation of conventional traffic scenarios. Through the privileged expert agent model, multi-modal data synchronization acquisition is completed, providing high-quality data for the training and evaluation of the autonomous driving system. This module effectively improves the data generation efficiency and the authenticity of the simulation scenario, and is the basic technical content of the present invention. The protection scope covers the complete data generation and acquisition process. The present invention constructs a unified safety test platform for autonomous driving algorithms by integrating modules such as conventional scenario generation, adversarial testing and dynamic risk analysis, and realizes a closed-loop automated process of data generation-system testing-risk discovery-dynamic analysis. This platform has high integration, strong scalability and automation, and is the architecture foundation of the overall technical solution of the present invention. The protection points cover the platform architecture design, module cooperation method and system integration mechanism.

[0132] In summary, the present invention takes the safety-critical scenario generation method as the core, combines dynamic clustering analysis and conventional scenario automatic generation technologies to form a closed-loop safety test platform for autonomous driving systems. Each technical solution complements each other, comprehensively improving the test coverage and risk discovery ability of the autonomous driving system. The related technical methods and system architectures are the main protected contents of the present invention. Through innovative scenario generation methods, adversarial testing mechanisms and dynamic data analysis strategies, the present invention significantly improves the safety verification ability of autonomous driving systems and the optimization efficiency of algorithm robustness. It has significant advantages such as high automation, strong system integration, rich scenario diversity, and accurate risk discovery. It breaks through the technical bottleneck of existing testing methods and provides a solid technical guarantee for the safety improvement and industrial application of autonomous driving systems.

[0133] Embodiment 2

[0134] As Figure 5 shown, the present invention provides an autonomous driving safety test method, and the autonomous driving safety test method is used to execute the autonomous driving safety test system described in Embodiment 1, including the following steps:

[0135] S1. Generate a conventional driving scenario dataset in the simulation platform;

[0136] S2. According to the conventional driving scenario dataset, use the self-driving scenario generation network SAGA based on adversarial perturbations to generate safety-critical scenarios;

[0137] S3. According to the generated safety-critical scenarios, use the method of introducing progressive unsupervised clustering to dynamically discover new categories and optimize the clustering structure, obtain the scenario analysis results, and complete the safety test of autonomous driving.

[0138] The purpose of the present invention is to design and implement a test platform framework for autonomous driving algorithms with efficient scenario generation, vulnerability mining, and risk analysis capabilities in view of the deficiencies in the existing safety test means for autonomous driving algorithms, specifically including the following aspects:

[0139] 1) Provide an automated safety-critical scenario generation method: Based on the adversarial optimization algorithm, adopt the generator-victim model architecture, and use the policy gradient optimization method to achieve the automated and efficient generation of safety-critical scenarios, covering different types of traffic participant behaviors and diverse extreme environmental conditions.

[0140] 2) Provide an adversarial generation mechanism with traffic rationality constraints: By introducing traffic rules, expert experience, and physical constraints, optimize the adversarial scenario generation process to ensure that the generated test scenarios conform to the actual traffic logic and behavior norms, and improve the credibility and interpretability of the test data.

[0141] 3) Provide a progressive unsupervised clustering analysis method: Adopt the anchor graph construction and label propagation mechanism to support the dynamic classification of autonomous driving test data, category boundary optimization, and new category discovery, achieve the efficient analysis of large-scale test data and the identification of key risk scenarios, and improve the system risk assessment ability.

[0142] 4) Provide a closed-loop autonomous driving algorithm safety test platform framework: The present invention is based on the CARLA simulation platform, integrates functions such as automatic generation of conventional traffic scenarios, behavior training and evaluation of autonomous driving systems, adversarial scenario generation, automated testing, and risk clustering analysis, realizes the full-process safety verification of autonomous driving systems, and provides a reliable basis for algorithm optimization and robustness improvement.

[0143] In summary, the present invention realizes the automation, systematization, and intelligence of autonomous driving algorithm testing, significantly improves the test scenario generation efficiency, system vulnerability discovery ability, and the intelligence level of test data analysis, and provides a solid technical guarantee for the safety improvement and wide application of autonomous driving systems.

[0144] Those skilled in the art will realize that the embodiments described herein are provided to assist the reader in understanding the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific deformations and combinations without departing from the essence of the present invention based on the technical revelations disclosed in the present invention, and these deformations and combinations are still within the scope of protection of the present invention.

Claims

1. An autonomous driving safety test system, characterized in that, Including: A conventional traffic scenario generation subsystem for generating a conventional driving scenario dataset in a simulation platform; A safety-critical scenario generation subsystem for generating safety-critical scenarios according to the conventional driving scenario dataset by using the self-driving scenario generation network SAGA based on adversarial perturbations; A safety-critical scenario clustering and analysis subsystem for obtaining a scenario analysis result and completing the safety test for autonomous driving by using the generated safety-critical scenarios and introducing a progressive unsupervised clustering method to dynamically discover new categories and optimize the clustering structure; 2. The autonomous driving safety test system according to claim 1, wherein The conventional traffic scenario generation subsystem includes: A road network topology modeling module for extracting road topology information by analyzing the road network in the simulation platform and modeling the road network topology based on the road topology information; A path planning and optimization module for generating a smooth driving path from the starting point to the target point by using a path planning algorithm based on the constructed road network topology and optimizing the smooth driving path by using the interpolation method; A multi-modal data acquisition module for synchronously acquiring various types of scenario data by using multi-modal sensors based on the optimized smooth driving path to construct a conventional driving scenario dataset; 3. The autonomous driving safety test system according to claim 1, characterized in that, The safety-critical scenario generation subsystem includes: A conventional training module for training an autonomous driving agent by using the conventional driving scenario dataset to form a benchmark model with driving decision-making and environment perception capabilities; An adversarial scenario generation module for generating adversarial scenarios for the victim model in the initial scenario according to the training result by using a gradient-based adversarial perturbation strategy, where the optimal perturbation sequence is found by minimizing the objective function to generate safety-critical scenarios; A victim model module for providing real-time feedback during the perturbation generation process to optimize the adversarial perturbation strategy, where the victim model is an autonomous driving decision-making system that is used as the attack target during the adversarial scenario generation process; 4. The autonomous driving safety test system according to claim 3, wherein The training of the autonomous driving agent by using the conventional driving scenario dataset is specifically: Based on the conventional driving scenario dataset, the observation data O is combined with the target trajectory points G to train the autonomous driving policy network π θ , forming a baseline model with driving decision-making and environmental perception capabilities.

5. The autonomous driving safety test system according to claim 3, characterized in that, The expression of the optimal perturbation sequence is as follows: Among them, θ* represents the optimal perturbation sequence, θ represents the set of all adversarial agent action sequences, J(S) represents the minimized objective function, and θ i' represents the driving strategy of the i'-th adversarial agent action sequence. i = 0 represents the autonomous driving agent associated with the victim model, i ∈ {1,..., N} represents the traffic agents in the scenario except i = 0, and N represents the number of adversarial agents in the adversarial sequence. represents the action of the i-th autonomous driving agent at time t. T represents the number of simulation steps, and t represents the time. represents the collision distance cost of the autonomous driving agent. α, β, and γ all represent adjustable hyperparameters. represents the speed difference cost of the autonomous driving agent. represents the safety distance cost between adversarial agents. represents the out-of-bounds constraint cost. D() represents the Euclidean distance operation. represents the state corresponding to the victim model. represents the state of the i-th autonomous driving agent at time t, and v bb represents the speed difference, and τ represents the minimum safety threshold. represents the state of the j-th autonomous driving agent at time t. L represents the number of corner points included in each vehicle bounding box, m represents the index of the currently processed corner point, which is used to traverse all the corner point positions of the vehicle bounding box at each time step, and f() represents the Gaussian potential function around the corner point. represents the bounding box of the adversarial agent i at time t, and p m represents the corner points of the bounding box. represents the convolution operation with the non-traversable area in the map.

6. The autonomous driving safety test system according to claim 1, characterized in that The obtaining of the scenario analysis result by using the generated safety-critical scenarios and introducing a progressive unsupervised clustering method to dynamically discover new categories and optimize the clustering structure is specifically: Performing hierarchical partitioning on the generated safety-critical scenarios, where a set of anchor points is constructed by selecting representative samples during each partitioning process; Generating pseudo-labels based on the set of anchor points; Constructing an anchor graph based on the pseudo-labels and the data similarity and characterizing the distribution characteristics of the data by using the representative samples; Diffusing high-confidence information through a label propagation mechanism based on the distribution characteristics to make the clustering boundary reach a clear state; When new samples different from the existing categories are detected, new categories are created and the clustering result is dynamically updated until the clustering result meets the requirements to obtain the scenario analysis result; 7. A method for autonomous driving safety testing, characterized in that, The autonomous driving safety test method is used to execute the autonomous driving safety test system according to any one of claims 1-6, and includes the following steps: Generating a conventional driving scenario dataset in a simulation platform; Generating safety-critical scenarios according to the conventional driving scenario dataset by using the self-driving scenario generation network SAGA based on adversarial perturbations; According to the generated safety-critical scenarios, a progressive unsupervised clustering method is introduced to dynamically discover new categories and optimize the clustering structure, obtaining the scenario analysis results and completing the safety testing for autonomous driving.

Images