An automatic driving test method and system based on pre-boundary scene guidance

Abstract

The application discloses an automatic driving vehicle test method and system based on pre-boundary scene guidance, and relates to the technical field of intelligent driving test, comprising the following steps: identifying the real-time operation scene of a target vehicle based on a pre-boundary scene identification method, and determining whether the target vehicle enters a pre-boundary scene; if the target vehicle enters the pre-boundary scene, switching to an intervention guidance driving state; switching to an automatic driving state, and autonomously deciding and testing based on the current vehicle state and current vehicle environment of the target vehicle by using an automatic driving model until the test is terminated, wherein the current vehicle state comprises multi-dimensional space-time features such as the critical vehicle speed of the target vehicle entering the boundary scene. The method provided by the application directionally guides the target vehicle from the pre-boundary scene to the boundary scene, overcomes the problem that the prior art relies on natural evolution to cause low boundary scene state generation efficiency and long waiting period, and realizes efficient and automatic generation of the boundary test scene state.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving testing technology, and in particular to an autonomous driving testing method and system based on pre-boundary scenario guidance. Background Technology

[0002] With the rapid development of autonomous driving technology, the importance of its intelligent testing phase is becoming increasingly prominent. Based on the invention patent CN120951839B, which describes a method and system for generating vehicle test boundary scenarios based on pre-boundary scenarios, a sufficient amount of diverse boundary scenario data has been acquired. This boundary scenario data not only covers typical risk evolution processes but also reflects key dynamic characteristics under different traffic behaviors and interaction conditions, providing data support for further system-level testing. Building upon this, the proposed further development goal of this application is to construct an accelerated testing system for autonomous driving safety, using pre-boundary scenarios and boundary scenarios as core inputs. This system drives the testing process with risk scenarios, unifying scenario identification, scenario guidance, behavioral response evaluation, and result feedback, thereby achieving systematization and standardization of the testing process. However, in existing technologies, testing autonomous driving models is only based on the natural evolution of pre-boundary scenarios into boundary scenarios. Due to the sparse data in boundary scenarios, the testing efficiency of autonomous driving models is low.

[0003] Therefore, it is necessary to provide a pre-boundary scenario-guided autonomous driving testing method, which involves manually intervening in the pre-boundary scenario to the boundary scenario and then switching to the autonomous driving model for testing, in order to address the technical issue of testing efficiency of the autonomous driving model. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide an autonomous driving test method based on pre-boundary scenario guidance, which aims to solve the technical problem of limited number of boundary vehicle test scenarios generated by the natural evolution of pre-boundary scenarios into boundary scenarios in the prior art, resulting in the test efficiency of autonomous driving models.

[0005] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: An autonomous driving testing method based on pre-boundary scenario guidance includes the following steps: S10, Based on the pre-boundary scene recognition method, the scene where the target vehicle is located is recognized. If the target vehicle is in the pre-boundary scene, proceed to step S20. If the target vehicle is in the safe scene, the target vehicle maintains the autonomous driving operation state and continues to perform scene recognition. S20, switch to intervention-guided driving state, obtain the speed increment of the target vehicle from the pre-boundary scene to the boundary scene based on the current driving characteristics of the target vehicle, determine the guidance acceleration of the target vehicle based on the speed increment, and use the guidance acceleration to apply a disturbance to guide the target vehicle until the target vehicle enters the boundary scene; S30, switch to automatic driving mode, use the autonomous driving model to make autonomous decision-making tests based on the current vehicle state and current vehicle environment of the target vehicle until the test ends. The current vehicle state includes the critical speed at which the target vehicle enters the boundary scene.

[0006] Furthermore, in step S10, Obtain the real-time scene feature matrix of the target vehicle, and use the pre-boundary scene discrimination model based on the real-time scene feature matrix to identify whether the scene in which the target vehicle is located is a pre-boundary scene or a safe scene. Among them, the pre-boundary scene discrimination model is one of the following: random forest model, logistic regression model, support vector machine and single decision tree model.

[0007] Furthermore, the pre-boundary scene discrimination model is a random forest model. When the random forest model is trained using the training and test sets, the training input label is the real-time scene feature matrix, and the training output label is the judgment label with the pre-boundary scene or the safe scene as binary variables. In step S10, the target vehicle autonomously drives within a designated test area; the sensor module continuously collects the vehicle's operating status and the operating status of surrounding vehicles, and calculates the MTTC risk threshold based on the vehicle's operating status and the operating status of surrounding vehicles; when the MTTC risk threshold is less than the first threshold, the pre-boundary scene discrimination model is activated to determine whether the target vehicle is in a pre-boundary scene.

[0008] Furthermore, the test will terminate if the total test duration of the autonomous decision-making test reaches the preset duration.

[0009] Furthermore, in step S20, a throttle-acceleration mapping table is established based on the statistical law of throttle speed change in the real pre-boundary samples; The valve opening ratio of the acceleration control valve is determined based on the throttle-acceleration mapping table using the guided acceleration. Based on the valve opening ratio, a guiding disturbance is applied to the target vehicle until the target vehicle enters the boundary scenario.

[0010] Furthermore, it also includes the steps of: obtaining the high-risk exposure rate, overall accident rate, accident rate under boundary scenario conditions, average high-risk duration, and risk accumulation level corresponding to the total test duration. The accelerated testing method guided by pre-boundary scenarios is evaluated based on high-risk exposure rate, overall accident rate, accident rate under boundary scenario conditions, average duration of high risk, and degree of risk accumulation.

[0011] Furthermore, let the total test duration be... The sampling interval is The total number of sampling points is ,time Risk indicators are The number of collision events that occurred during the entire test process within the total test duration was The number of boundary scene events identified was , No. The duration of continuous events in sub-boundary scenario events that satisfy the MTTC risk threshold ≤ 0.9 is: ; Using formula Perform calculations. high risk exposure rate , This is an indicator function that takes the value 1 when the condition is true and 0 otherwise. This high-risk exposure rate is used to represent the proportion of time spent in the boundary scenario to the total test duration during the entire test process. Using formula Calculations were performed to obtain the overall accident rate. The overall accident rate is expressed as the frequency of collision events occurring over time during the test. Using formula Calculations are performed to obtain the accident rate under boundary scenario conditions. The accident rate under boundary scenario conditions is used to reflect the degree of danger that triggers the boundary scenario; Using formula Calculations were performed to obtain the average duration of high risk. The average duration of high risk is used to reflect the sustained level of danger in the test scenario; Using formula Calculations are performed to obtain the degree of risk accumulation. The risk accumulation level is used to reflect the overall severity of risk exposure during the testing process.

[0012] The present invention also provides an autonomous driving test system based on pre-boundary scenario guidance, comprising: The pre-boundary scene recognition module identifies the scene in which the target vehicle is located based on the pre-boundary scene recognition method. If the target vehicle is in a pre-boundary scene, it enters the intervention and guidance driving state. If the target vehicle is in a safe scene, the target vehicle maintains the autonomous driving operation state and continues to perform scene recognition. The guided driving control module is used to obtain the speed increment of the target vehicle as it evolves from the pre-boundary scene to the boundary scene based on the current driving characteristics of the target vehicle, determine the guiding acceleration of the target vehicle based on the speed increment, and use the guiding acceleration to apply a disturbance to guide the target vehicle until the target vehicle enters the boundary scene. The automatic driving test module is used to switch to automatic driving mode when the target vehicle enters the boundary scene. It uses the autonomous driving model to make autonomous decision-making tests based on the current vehicle state and current vehicle environment of the target vehicle until the test is terminated. The current vehicle state includes the critical speed at which the target vehicle enters the boundary scene. The test evaluation module is used to obtain five indicators: high-risk exposure rate, overall accident rate, accident rate under boundary scenario conditions, average high-risk duration, and risk accumulation level. It performs quantitative analysis on the performance of the test method and obtains the risk avoidance success rate, highest risk index, risk fluctuation range, risk convergence time, speed fluctuation range, and maximum deceleration to evaluate the decision-making and control capabilities of the autonomous driving model in high-risk scenarios.

[0013] Compared with the prior art, the advantages of the present invention are as follows: The autonomous driving testing method based on pre-boundary scenario guidance provided by this invention, after identifying that the target vehicle has entered the pre-boundary scenario, involves manual intervention to accelerate the target vehicle and guide it to the boundary scenario. In the boundary scenario, the vehicle enters an autonomous driving state, utilizing an autonomous driving model to make autonomous decisions based on the target vehicle's current state and environment until the test terminates. By actively applying guidance acceleration determined based on the statistical laws of real pre-boundary samples, the target vehicle is directionally guided from the pre-boundary scenario to the boundary scenario. This overcomes the problems of low efficiency and long waiting periods in boundary scenario state generation caused by reliance on natural evolution in existing technologies, achieving efficient and proactive generation of boundary test scenario states. Furthermore, guidance based on guidance acceleration ensures that the guidance process conforms to the objective evolution laws of real traffic scenarios, avoiding scenario distortion caused by forceful human intervention. Using a clear critical speed as the starting point for model testing improves the accuracy and repeatability of the autonomous driving model's decision-making performance evaluation in boundary scenario states, and significantly enhances the testing efficiency compared to existing autonomous driving models. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the autonomous driving testing method based on pre-boundary scenario guidance according to the present invention. Detailed Implementation

[0015] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please refer to Figure 1 An embodiment of the present invention provides an autonomous driving test method based on pre-boundary scenario guidance, comprising the following steps: The scene of the target vehicle is identified based on the pre-boundary scene recognition method. If the target vehicle is in the pre-boundary scene, the process proceeds to step S20. If the target vehicle is in the safe scene, the target vehicle maintains its autonomous driving state and continues to perform scene recognition. S20, switch to intervention-guided driving state, obtain the speed increment of the target vehicle from the pre-boundary scene to the boundary scene based on the current driving characteristics of the target vehicle, determine the guidance acceleration of the target vehicle based on the speed increment, and use the guidance acceleration to apply a disturbance to guide the target vehicle until the target vehicle enters the boundary scene; S30, switch to automatic driving mode, use the autonomous driving model to make autonomous decision-making tests based on the current vehicle state and current vehicle environment of the target vehicle until the test ends. The current vehicle state includes multi-dimensional spatiotemporal features such as the critical speed of the target vehicle when entering the boundary scene.

[0018] The autonomous driving testing method based on pre-boundary scenario guidance provided by this invention, after identifying that the target vehicle has entered the pre-boundary scenario, involves manual intervention to accelerate the target vehicle and guide it to the boundary scenario. In the boundary scenario, the vehicle enters an autonomous driving state, utilizing an autonomous driving model to make autonomous decisions based on the target vehicle's current state and environment until the test terminates. By actively applying guidance acceleration determined based on the statistical laws of real pre-boundary samples, the target vehicle is directionally guided from the pre-boundary scenario to the boundary scenario. This overcomes the problems of low efficiency and long waiting periods in boundary scenario state generation caused by reliance on natural evolution in existing technologies, achieving efficient and proactive generation of boundary test scenario states. Furthermore, guidance based on guidance acceleration ensures that the guidance process conforms to the objective evolution laws of real traffic scenarios, avoiding scenario distortion caused by forceful human intervention. Using a clear critical speed as the starting point for model testing improves the accuracy and repeatability of the autonomous driving model's decision-making performance evaluation in boundary scenario states, and significantly enhances the testing efficiency compared to existing autonomous driving models.

[0019] Optionally, the invention patent with patent number ZL202511494758.X, which discloses a method and system for generating vehicle test boundary scenarios based on pre-boundary scenarios, discloses an offline backtracking method for determining pre-boundary scenarios based on boundary scenarios, used for scene data generation and model training. In the solution of this invention, the minimum threshold and maximum threshold of MTTC level 2 risk can be determined based on the MTTC risk threshold corresponding to the boundary scenario and a preset leniency value, thereby clarifying the scene frame corresponding to the MTTC level 2 risk range as a pre-boundary scenario, and thus realizing online real-time judgment of pre-boundary scenarios based on preset leniency values. Alternatively, the real-time scene feature matrix of the target vehicle can be obtained during the dynamic test in the automatic driving state, and the real-time scene feature matrix can be input into the pre-boundary scenario discrimination model. After the pre-boundary scenario discrimination model determines whether the scene in which the vehicle is currently located is a pre-boundary scenario or a safe scenario, the dynamic identification of the pre-boundary scenario can be performed.

[0020] Understandably, the autonomous driving model can be used to drive autonomously before the target vehicle enters the pre-boundary scenario; after entering the pre-boundary scenario, the autonomous driving model can be switched to a human intervention and guidance driving state, and a disturbance can be applied to the target vehicle based on the guidance acceleration until the target vehicle enters the boundary scenario; after entering the boundary scenario, the autonomous driving model can be switched back to drive, thereby testing the main decision-making logic of the autonomous driving model.

[0021] Understandably, the autonomous driving testing method based on pre-boundary scenario guidance provided by this invention is implemented based on the CaRLa-SUMO test scenario. Scenario construction is a fundamental step in the accelerated testing of autonomous driving. Only when the test scenario simultaneously possesses real-world environmental characteristics and a certain degree of controllability can subsequent algorithm verification, test implementation, and result analysis proceed smoothly. Based on this, this chapter focuses on the testing requirements of autonomous driving systems, researching scenario construction methods, and selecting the Town04 map from the CaRLa simulation platform as the primary experimental environment. It should be noted that although the Town04 map includes both urban roads and highways, this solution only uses the highway-related portion for verification during the actual testing process. The Town04 map is one of the standard test maps provided by the CaRLa platform. Its overall design is based on a highway traffic environment. This map includes multi-lane main roads, curved road sections, and ramp connection structures. The roads have strong continuity, and the traffic flow organization is relatively complete. Under these environmental conditions, vehicles can more easily form stable following and lane-changing behaviors. Therefore, this map is suitable for studying the longitudinal and lateral interaction processes of vehicles in a highway environment. In addition, the Town04 map has a relatively complete road geometry structure, including straight road sections, curved road sections, and on-ramps and off-ramps, which can provide good support for high-speed traffic flow simulation. Based on this environment, a variety of typical traffic test scenarios can be constructed, such as car-following scenarios, lane-changing scenarios, and lane-merging scenarios. These scenarios are of great help in analyzing the decision-making and control capabilities of autonomous driving systems under high-speed conditions.

[0022] It should be noted that the Town04 map also includes urban road areas of a certain scale, featuring intersections, curves, and connections between different roads. Theoretically, researchers can also construct low-speed traffic scenarios (such as vehicle following, yielding, or lane changing) in this area to analyze the perception and decision-making capabilities of autonomous driving systems in more complex traffic environments. Preferably, the present invention is primarily aimed at highway testing environments; background traffic flow is generated based on SUMO. Overall, the Town04 map provides a relatively clear road structure, complete traffic elements, and good environmental expansion capabilities, while supporting stable and repeatable testing processes. Based on these characteristics, this solution constructs subsequent traffic scenarios on the Town04 map and conducts autonomous driving tests and related data collection within this environment.

[0023] Understandably, the main purpose of this invention is to test the autonomous driving control (autonomous decision-making) algorithm corresponding to the autonomous driving model after the target vehicle enters the boundary scenario. The autonomous driving control algorithm plays a crucial role in the safety and operational stability of the autonomous driving system. It directly outputs vehicle control commands and directly impacts the vehicle's final driving decision. Based on the aforementioned Town04 test environment, this solution selects three control algorithms—TCP, Ekostinoudis, and Justin900429—as test objects. These three algorithms are used only for the control of the tested master vehicle and do not replace the original behavior model of the background traffic vehicles.

[0024] The three control methods selected in this scheme represent different technical implementation paths. The TCP method mainly represents an end-to-end learning framework, characterized by its ability to directly generate vehicle control outputs based on perceived inputs. The Ekostinoudis method belongs to the type that combines imitation learning and deep learning, and can further improve decision-making performance on the basis of data-driven approaches. The Justin900429 method uses a diffusion model for control modeling, which has a certain ability to characterize uncertainties in complex dynamic processes. Among them, TCP (Trajectory-guided Control Prediction for End-to-end autonomous driving) is a type of end-to-end autonomous driving control method. Unlike the modular structure of "perception-decision-control" in traditional autonomous driving systems, TCP does not emphasize the obvious separation of these parts. Instead, it directly uses environmental perception information as input and outputs vehicle control commands, thereby establishing a direct mapping relationship between the external environment and control behavior. This method can handle real-time control tasks under relatively complex traffic conditions. Its main advantage is that it has a low dependence on manual rules. After the model is trained with a large amount of scene data, it can learn the driving operation rules under different traffic conditions. For example, in a highway environment, it can respond to situations such as the vehicle in front slowing down and the adjacent vehicle merging; while in an urban road environment, this method can also complete operations such as following, steering, and avoidance. The Ekostinoudis method is an autonomous driving control algorithm that integrates imitation learning and deep learning. Its basic idea is to train the control model using real human driving data to form an autonomous driving strategy with certain human-like driving characteristics. This method retains reasonable experiential features from human driving behavior while leveraging deep learning models to enhance the system's adaptability to complex traffic environments. Its technical process can be broadly divided into two stages: the first stage is data acquisition and preprocessing, which involves collecting human driving data under different road conditions and cleaning, filtering, and normalizing key variables such as vehicle speed, steering angle, vehicle spacing, and relative position; the second stage is model training and optimization, which involves establishing a supervised learning model based on a deep learning network and using human driving behavior as a supervisory signal to teach the model the correspondence between "environmental input and control output." To improve the model's generalization performance, regularization and other optimization strategies are typically added during training. The Justin900429 method, on the other hand, is an autonomous driving control algorithm based on a diffusion model. Its main characteristic is its strong modeling ability for complex probability distributions and highly random traffic scenarios.Because it can generate corresponding control strategies for various traffic conditions, it is well-suited for robustness testing of autonomous driving systems. The basic principle of the diffusion model is to learn the distribution of real data through two processes: forward diffusion and reverse diffusion. In the forward process, the model gradually adds noise to the original data; while in the reverse process, the model gradually recovers the effective data structure from the noise. Through this process, the model can learn the probabilistic relationship between traffic scene states and control behaviors, and further generate control outputs that adapt to different operating conditions.

[0025] Further, in step S10, the real-time scene feature matrix of the target vehicle is obtained, and the scene in which the target vehicle is located is identified as a pre-boundary scene or a safe scene based on the real-time scene feature matrix using a pre-boundary scene discrimination model; wherein, the pre-boundary scene discrimination model is one of the following: random forest model, logistic regression model, support vector machine and single decision tree model.

[0026] Furthermore, the pre-boundary scene discrimination model is a random forest model. When the random forest model is trained using a training and test set, the training input label is the real-time scene feature matrix, and the training output label is a judgment label with the pre-boundary scene or the safe scene as binary variables. In step S10, the target vehicle autonomously drives within a designated test area. The sensor module continuously collects the vehicle's operating status and the operating status of surrounding vehicles, and calculates the MTTC risk threshold based on the vehicle's operating status and the operating status of surrounding vehicles. When the MTTC risk threshold is less than a first threshold, the pre-boundary scene discrimination model is activated to determine whether the target vehicle is in a pre-boundary scene. The first threshold can be a value such as 0.8s, 0.9s, 1.0s, or 1.1s, preferably 0.9s.

[0027] Optionally, in an optional embodiment of the present invention, pre-boundary scene recognition is achieved by designing a pre-boundary scene discrimination model. Specifically, at the data layer, this application constructs a training dataset composed of real pre-boundary scenes and safe scenes. The real pre-boundary scenes are the preceding scenes before the boundary scenes with MTTC values ​​less than or equal to 0.9 s. At the same time, in order to form a comparison, this application also selects regular driving data with MTTC values ​​greater than 4 s and no obvious risk accumulation characteristics as safe scenes. This type of data usually has a high safety margin, the vehicle's operating state is relatively stable, and there are no signs of continuous risk development. Considering the sample organization needs in model training, this paper finally selects 200 sets of real pre-boundary scenes and 800 sets of real safe scenes. Both types of samples contain the basic behavioral features in the following or lane changing process, and retain m frames of continuous temporal information so that the subsequent model can learn the changing patterns of the scene in the time dimension. Secondly, in terms of the model layer, this application uses a classification model to learn the risk feature patterns. The input of the model is the temporal feature matrix of the vehicle, and the output is the preliminary judgment label of the pre-boundary scenario or the safety scenario, of which 20% of the scenarios are used for the model test set.

[0028] In the solution of this invention, the core definition of the real pre-boundary scenario is: in the vehicle trajectory data sequence, within a continuous m-frame time window before the boundary scenario determination frame, a trajectory data set containing full-dimensional features of the vehicle and environment can supplement the historical context information lacking in the boundary scenario, reveal the behavioral adjustment pattern of the target vehicle before the risk reaches a critical state (such as speed changes when following, the incubation process of lane-changing decisions, etc.), providing key support for subsequent data-driven model learning of risk evolution patterns and improving the accuracy of scenario generation and risk prediction. If the time frame corresponding to the boundary scenario is t (i.e., the moment when the risk first falls below the threshold θ), then the time range of the real pre-boundary scenario is defined as [Tm-1, T-1]. The setting of this interval ensures the temporal continuity of historical data and boundary scenarios.

[0029] Optionally, Table 1 compares the performance of the random forest model, logistic regression model, support vector machine, and single decision tree model. Analysis shows that the performance of these models in the pre-boundary scene recognition task was compared, with evaluation metrics including accuracy, precision, recall, F1 score, and inference speed. Overall, the random forest model demonstrates the best overall classification performance and achieves a good balance between recognition effectiveness and computational cost. While the single-sample inference time of the random forest is 9 ms, slightly higher than logistic regression, support vector machine, and single decision tree, it remains relatively low and meets the real-time requirements of subsequent scene recognition tasks. Preferably, the pre-boundary scene discrimination model is the random forest model, which is ultimately chosen as the core model for the pre-boundary scene recognition stage, providing fundamental support for subsequent risk evolution analysis, boundary scene generation, and accelerated testing research.

[0030] Table 1 Furthermore, the test terminates if the autonomous decision-making test reaches the preset duration. This scheme uses a fixed duration as the primary condition for test termination; that is, each round of testing automatically ends after reaching the preset duration. In addition, to ensure test safety and data validity, when a collision occurs during the test, the target vehicle is promptly moved to another location to continue the test. After the test, the system extracts and saves key performance data such as avoidance success rate, changes in risk indicators, and operational stability for subsequent statistical analysis and comparative studies between different autonomous driving algorithms.

[0031] Furthermore, in step S20, a throttle-acceleration mapping table is established based on the statistical law of throttle speed change in the real pre-boundary samples; The valve opening ratio of the acceleration control valve is determined based on the throttle-acceleration mapping table using the guided acceleration. Based on the valve opening ratio, a guiding disturbance is applied to the target vehicle until the target vehicle enters the boundary scenario.

[0032] Optionally, Table 2 is a throttle-acceleration mapping relationship table established based on the statistical law of throttle speed change in the real pre-boundary sample. Preferably, the accurate valve opening ratio is obtained by interpolation processing using the throttle-acceleration mapping relationship table based on the guided acceleration.

[0033] Table 2 Furthermore, in order to evaluate the effectiveness of the proposed accelerated testing method, an accelerated testing performance evaluation system is constructed based on high-risk exposure rate, overall accident rate, accident rate under boundary scenario conditions, average high-risk duration, and risk accumulation level. Specifically, the first-level risk threshold MTTC is used as a unified risk characterization index, and MTTC ≤ 0.9s is used as an important criterion for determining whether a scenario enters the boundary scenario. Based on this, five indicators are selected to quantitatively analyze the performance of the testing method, including: Let the total test duration be The sampling interval is The total number of sampling points is ,time Risk indicators are The number of collision events that occurred during the entire test process within the total test duration was The number of boundary scene events identified was , No. The duration of continuous events in sub-boundary scenario events that satisfy the MTTC risk threshold ≤ 0.9 is: ; Using formula Perform calculations. high risk exposure rate , This is an indicator function, which takes the value 1 when the condition is met and 0 otherwise. This high-risk exposure rate is used to represent the proportion of time spent in the boundary scenario during the entire testing process to the total testing time. The larger the value, the easier it is for the testing method to push the scenario into the high-risk area, and it also indicates that the testing process has a stronger risk exposure capability. Using formula Calculations were performed to obtain the overall accident rate. The unit is times per hour (times / h), and the overall accident rate is expressed as the frequency of collision events occurring in the time dimension during the test; Using formula Calculations are performed to obtain the accident rate under boundary scenario conditions. The accident rate under boundary scenario conditions is used to reflect the degree of danger that triggers the boundary scenario; Using formula Calculations were performed to obtain the average duration of high risk. The average duration of high risk is used to reflect the sustained level of danger in the test scenario. Indicates the first In sub-boundary scenario events, satisfying The continuous duration is ≤0.9, in seconds (s). The average high-risk duration is used to describe the average time spent in a scenario under high-risk conditions. Using formula Calculations are performed to obtain the degree of risk accumulation. The unit is seconds (s), and the risk accumulation level is used to reflect the overall severity of risk exposure during the testing process.

[0034] Optionally, in specific implementation, the high-risk exposure rate is used to measure the frequency with which the scenario enters a high-risk state during the testing process. The high-risk exposure rate represents the proportion of time spent in boundary scenarios during the entire testing process. A higher high-risk exposure rate indicates that the acceleration method is more likely to push the scenario into a high-risk area, and also indicates that the testing process has a stronger risk exposure capability. The overall accident rate is defined as the frequency of collision events occurring within a unit of testing time, measured in times per hour (h). It reflects the frequency of collision events occurring over time during the testing process. A higher overall accident rate indicates that the acceleration method is more likely to trigger the failure of the control algorithm within a limited testing time. Therefore, this metric can be used to evaluate the testing method's ability to expose the consequences of accidents. To enhance the readability of the results, this metric can be further converted into the number of accidents per hour, etc. The accident rate under boundary scenario conditions measures the likelihood that a boundary scenario will further evolve into a collision accident. This metric represents the proportion of all identified boundary scene events that ultimately develop into collision accidents. It reflects the degree of danger of the triggered boundary scene. A higher accident rate under boundary scene conditions indicates that the acceleration method not only triggers boundary scenes but also that the triggered boundary scenes have a stronger tendency to evolve into accidents. Regarding the average high-risk duration, this describes the average time a scene remains in a high-risk state, measured in seconds (s). ,in, Indicates the first In sub-boundary scenario events, satisfying A continuous duration of ≤0.9 indicates a higher average high-risk duration, meaning the scenario remains in the high-risk area for a longer period. This implies the control algorithm needs to continuously address the corresponding risks for an extended time. Therefore, the average high-risk duration serves as an indicator of the sustained danger level of the test scenario. To comprehensively characterize the duration and depth of high-risk states, a risk accumulation index, measured in seconds (s), is introduced to assess risk accumulation. This indicator only applies to The portion ≤0.9 is cumulatively calculated. The greater the MTTC falls below the threshold, or the longer the high-risk duration, the higher the value of the risk accumulation index. Therefore, this index can more comprehensively reflect the overall severity of risk exposure during testing. This invention uses the above five indicators to construct the test performance evaluation system of this scheme. The high-risk exposure rate measures the ability of the testing method to push the scenario into a high-risk area; the overall accident rate and the accident rate under boundary scenario conditions measure the testing method's ability to trigger accident results; the average high-risk duration and risk accumulation level measure the persistence and severity of the high-risk state. Through these indicators, not only can we evaluate whether the testing method can effectively trigger boundary scenarios, but we can also further analyze the operational performance and failure characteristics of different control algorithms under high-risk conditions. This evaluation system provides a unified quantitative basis for subsequent test result analysis.

[0035] Furthermore, the autonomous driving model is evaluated based on at least one of the following: risk avoidance success rate, highest risk index, risk fluctuation range, risk convergence time, speed fluctuation range, and maximum deceleration.

[0036] Optionally, the accelerated testing method based on pre-boundary scenario guidance is evaluated based on high-risk exposure rate, overall accident rate, accident rate under boundary scenario conditions, average duration of high risk, and degree of risk accumulation. Based on the evaluation results of the testing method, the ability of the testing method to induce key dangerous scenarios is determined. For example, if the high-risk exposure rate of this testing method is high within the same testing time, it means that this testing method has a strong ability to induce high-risk scenarios, and can conduct high-risk scenario stress tests on the autonomous driving system, which means that the efficiency of autonomous driving testing is accelerated.

[0037] In practical implementation, to evaluate the performance of the autonomous driving control (autonomous decision-making) algorithm corresponding to the autonomous driving model in boundary scenarios, this study constructs an autonomous driving performance evaluation system (core test index system) from two aspects: risk control capability and vehicle operation stability. The core test index system includes six indicators, namely, risk avoidance success rate, highest risk index, risk fluctuation range, risk convergence time, speed fluctuation range, and maximum deceleration. Each indicator can reflect the decision-making and control capabilities of the control algorithm in high-risk scenarios from different perspectives. In order to ensure the comparability between different test results, this study defines the calculation method of each indicator in a unified manner.

[0038] Assume there is a total Group Boundary Scene Test Samples, Number The total duration of the test was During the test, the vehicle was at a constant time The risk value is denoted as This study uses the MTTC value (Level 1 Risk Threshold MTTC) as the risk quantification standard, with vehicle speed denoted as... The longitudinal acceleration of the vehicle is denoted as The safety risk threshold is denoted as , set as When the risk value falls below this threshold and remains stable over a subsequent period, the vehicle can be considered to have escaped the dangerous state and entered a safe state.

[0039] Regarding the collision avoidance success rate, this metric is used to statistically analyze the proportion of collisions successfully avoided by the control algorithm across all boundary scenario tests. This metric directly reflects the basic safety performance of the control algorithm. If no collision occurs during the entire boundary scenario test, the test is considered successful; if a collision occurs during the test, it is considered a failure. Let the nth... The risk avoidance result of this test is denoted as a binary variable. : Then the success rate of risk avoidance The calculation formula is: The higher the value of this indicator, the stronger the collision avoidance capability of the control algorithm in boundary scenarios. If the collision avoidance success rate reaches a high level, it indicates that the algorithm has a good safety guarantee capability.

[0040] The highest risk index (MTTC) characterizes the extreme risk level encountered during a single test, measured in seconds (s). This index reflects whether the control algorithm enters an excessively dangerous state during operation and also represents the maximum risk pressure exerted on the algorithm by the scenario. The smaller the MTTC value (Level 1 Risk Threshold), the higher the risk. The highest risk index of this test Defined as: If it is necessary to perform overall statistics on all test samples, the average highest risk index can be further calculated: The higher the value of this indicator, the higher the maximum risk the algorithm experiences during the test. If the maximum risk index corresponding to a certain control algorithm is low, it means that the algorithm can take effective control measures before the risk escalates further.

[0041] Regarding the risk fluctuation amplitude, it measures the range of risk value changes during testing and reflects the smoothness of the risk evolution process. The unit is seconds (s). A larger amplitude indicates more significant risk fluctuations during vehicle status adjustment, suggesting potentially strong fluctuations in the control process; conversely, a smaller amplitude indicates relatively stable risk changes. For the first... This test has a high degree of risk volatility. Defined as Similarly, the average risk volatility of all test samples can be expressed as: This indicator mainly reflects the overall degree of change in risk from accumulation and peak to decline. The smaller the value of the indicator for risk fluctuation, the more stable the risk adjustment of the control algorithm is in handling high risks.

[0042] Regarding risk convergence time, it describes the time required for a risk value to decrease from its peak to a safe range and remain stable, measured in seconds (s). This metric reflects the response speed and recovery capability of the control algorithm to risk. If the algorithm can control the risk to a safe range in a short time, it indicates that its risk management efficiency is high. First, define the... The peak risk time for this test is: Then define the risk convergence time. For the risk value to first fall below the safety threshold and remain below it over subsequent consecutive time intervals Internally always satisfied The earliest moment, of which, To stabilize the time window for judgment and avoid misjudging short-term fluctuations as risk convergence, this study sets a safety risk threshold. It is 4s, therefore the first Risk convergence time of the second test Defined as: The average risk convergence time for all test samples is: The smaller this indicator, the shorter the time required for the control algorithm to recover from a high-risk state to a safe state, and the stronger its risk response capability.

[0043] Regarding speed fluctuation amplitude, it measures the range of vehicle speed changes during the test, and the unit is meters per second (m / s). This indicator reflects the smoothness of longitudinal control. Large speed fluctuations suggest the algorithm may frequently adjust for acceleration and deceleration, resulting in poor vehicle comfort. Smaller speed fluctuations indicate smoother speed changes. For the... The speed fluctuation range in the second test Defined as: The average speed fluctuation of all test samples is: The smaller this indicator, the smoother the control algorithm can adjust the vehicle speed when dealing with boundary scenarios. This result usually means that the algorithm can ensure safety while taking into account both operational stability and passenger comfort.

[0044] Regarding the maximum deceleration, this metric records the maximum braking intensity encountered during the test. It reflects the algorithm's control strategy in emergency situations and is expressed in meters per second squared. If the maximum deceleration value is large, it indicates that the algorithm relies heavily on forced braking to avoid danger; while this method can prevent collisions, it reduces ride comfort and may induce new driving risks. Since the longitudinal acceleration of a vehicle is usually negative during braking, the first... The maximum deceleration in this test It can be defined as: The average maximum deceleration of all test samples can be expressed as: The higher this indicator, the more aggressive the braking action taken by the control algorithm in extreme situations. If an algorithm has a high success rate in avoiding hazards but its maximum deceleration is significantly large, it indicates that its safety mainly relies on a strong emergency braking strategy.

[0045] In summary, the six indicators described above describe the performance of autonomous driving control algorithms in boundary scenarios from different perspectives. The success rate of risk avoidance and the highest risk index primarily reflect the algorithm's safety level; the risk fluctuation amplitude and risk convergence time primarily reflect the algorithm's ability to adapt to changes in risk; and the speed fluctuation amplitude and maximum deceleration primarily reflect the algorithm's control smoothness and comfort performance. By jointly analyzing these indicators, a more comprehensive comparison of the performance differences of different control algorithms in boundary scenarios can be made. Based on the evaluation system of these six indicators, the testing process can conduct multi-dimensional evaluations of the behavior of different control algorithms in boundary scenarios. This process, through unified indicator definitions, unified calculation methods, and unified testing conditions, improves the standardization, repeatability, and comparability of test results, and provides a clearly structured data foundation for subsequent experimental result analysis.

[0046] The performance of the above-mentioned technical solutions was tested and verified. To verify the effectiveness of the accelerated testing method proposed in this solution, the method was analyzed from two aspects: testing efficiency and accident triggering results. Ordinary natural simulation testing was used as a control group. Ordinary natural simulation testing did not apply a boundary scene guidance mechanism (no intervention guidance), and all other settings were consistent with the proposed method. The test vehicle algorithm was the built-in traffic management method based on rules and phased planning control after the CaRLa autonomous driving module was activated. The testing efficiency analysis mainly focused on the ability of the testing method to trigger boundary scenes under the same testing resource constraints. The accident triggering result analysis mainly focused on the danger level of the high-risk scenarios triggered by the testing method and its tendency to evolve into collision consequences. Tables 3 and 4 show the comparison results between ordinary natural simulation and the proposed method, where Table 3 is a comparison analysis table of the efficiency of different testing methods, and Table 4 is a comparison analysis table of the accident triggering results of different testing methods.

[0047] Table 3 Table 4 Based on Table 3, a comparison was made between the two testing methods under the same total test duration. Ordinary natural simulation only obtained 73.5 s of boundary scene data, while the proposed method obtained 285 s, approximately 3.88 times longer. This demonstrates that the proposed method can more efficiently expose high-risk states under limited test resources. Furthermore, ordinary natural simulation requires an average of 1030.75 frames to obtain one frame of boundary scene data, while the proposed method requires only 265.82 frames, indicating higher boundary scene triggering efficiency. Regarding the high-risk exposure rate, ordinary natural simulation is 0.097%, while the proposed method is 0.376%, showing a significant improvement. In terms of the overall accident rate, ordinary natural simulation is 1.09 incidents / hour, while the proposed method reaches 4.89 incidents / hour, approximately 4.49 times higher. This indicates that the proposed method can more easily expose high-risk states and trigger control failure consequences more quickly within a limited test time. The proposed method outperforms ordinary natural simulation in boundary scene triggering, high-risk exposure, and accident triggering, making it more suitable for efficient driving tests of autonomous driving systems.

[0048] Based on Table 4, the differences in accident triggering and high-risk exposure of different testing methods were further compared, and the number of collisions, average duration of high risk, risk accumulation level, and accident rate under boundary scenario conditions were analyzed. The results show that, under the same total test duration of 75760 s, the proposed method generated 103 collisions, while the ordinary natural simulation only generated 23, approximately 4.48 times more, indicating that the proposed method is more likely to expose potential system failure behaviors. Regarding the average duration of high risk, the ordinary natural simulation was 0.90 s, while the proposed method was 0.96 s. Although the difference is not significant, it indicates that the high-risk state triggered by the proposed method is slightly more persistent, not just a short-term fluctuation. In terms of risk accumulation level, the proposed method reached 133.12, significantly higher than the ordinary natural simulation's 26.84, indicating that its triggering scenarios not only enter high-risk situations more frequently but also have a more concentrated risk level. Furthermore, regarding the accident rate under boundary scenario conditions, the ordinary natural simulation was 28.05%, while the proposed method increased it to 36.14%. The results show that the boundary scenarios triggered by the proposed method are more likely to evolve into collision accidents. Overall, the proposed method outperforms ordinary natural simulation in terms of the intensity of dangerous scenario triggering and accident exposure capability, and is therefore more suitable for high-risk testing of autonomous driving systems. It should be noted that the proposed method does not increase the accident rate by arbitrarily enhancing the disturbance, but rather guides the scenario based on the behavioral statistical patterns in real pre-boundary samples, thereby improving the exposure efficiency of high-risk states while maintaining a certain level of traffic rationality. Therefore, the increase in the number of boundary scenarios and the number of accident triggers mainly reflects the test method's higher accessibility to the system's capability boundaries.

[0049] Furthermore, analyzing the performance test results of autonomous driving vehicles can facilitate the selection of suitable autonomous driving control algorithms. Among these, the autonomous driving control algorithms include three algorithms: TCP, Ekostinoudis, and Justin900429.

[0050] Specifically, to verify the effectiveness of the constructed accelerated testing process and to compare and analyze the performance of different autonomous driving control algorithms in boundary scenarios, a statistical study was conducted on the test results of three algorithms: TCP, Ekostinoudis, and Justin900429. The study focuses on the differences in risk control capabilities and operational stability of the algorithms.

[0051] The boundary scenario triggered by the highway area in the Town04 environment was selected as the analysis object. All indicators were derived from real-time recorded data during the test to ensure the objectivity and traceability of the results. The evaluation indicators include risk avoidance success rate, maximum risk index, risk fluctuation range, risk convergence time, speed fluctuation range, and maximum deceleration. These indicators characterize the control performance of the algorithm under high-risk conditions from different perspectives, providing a basis for subsequent performance analysis.

[0052] Risk control capabilities of the three algorithms were analyzed separately. To further compare the risk response characteristics of different algorithms in boundary scenarios, risk-related indicators of the three algorithms were statistically analyzed separately, as shown in Table 5. Table 5 shows the statistical results of risk indicators for different autonomous driving control algorithms: Table 5 According to the results in the table, the TCP, Ekostinoudis, and Justin900429 algorithms all possess a certain degree of risk control capability, but there are significant differences in their control effectiveness. Justin900429 has the highest risk avoidance success rate at 93.0%, and its average risk convergence time is the shortest at only 3.37 seconds, indicating that this algorithm can restore the system to a stable state more quickly after a risk occurs. Furthermore, this algorithm has the lowest average risk index, highest risk index, and risk fluctuation amplitude among the three, indicating that it can not only effectively reduce the overall risk level but also suppress risk peaks and reduce risk fluctuations. Therefore, this invention provides a preferred solution: if the autonomous driving mode is determined with comprehensive risk control capability as the objective, then the Justin900429 algorithm is selected as the algorithm for the autonomous driving model after entering the boundary scenario.

[0053] Besides risk control capabilities, the vehicle's operational stability in boundary scenarios is also an important aspect of evaluating autonomous driving control algorithms. Further statistical analysis of vehicle operational status indicators was conducted, as shown in Table 6. Table 6 presents the vehicle operational stability indicators for different autonomous driving control algorithms: Table 6 As shown in Table 6, the three algorithms exhibit little difference in average vehicle speed, all falling within the reasonable operating range for highways. This indicates that all algorithms can maintain stable driving speeds under normal cruise conditions. However, after boundary scenarios are triggered, the speed adjustment methods of the different algorithms show significant divergence. TCP's speed fluctuation amplitude is 5.21 m / s, and its maximum deceleration is 4.86 m / s², both the highest among the three. This suggests that TCP tends to employ stronger braking and speed adjustments to avoid risks under high-risk conditions. This strategy can reduce risk in a short time but causes significant speed changes, thus affecting driving stability. Ekostinoudis's speed fluctuation amplitude is 4.48 m / s, and its maximum deceleration is 4.23 m / s². This result indicates that this algorithm maintains relatively smooth speed changes during risk handling. This method has human-like driving characteristics, thus providing more continuity in longitudinal control and balancing safety and comfort. The Justin900429 algorithm exhibits a speed fluctuation of 4.11 m / s and a maximum deceleration of 4.01 m / s², both at the lowest levels. This indicates that the algorithm can not only achieve hazard avoidance in boundary scenarios but also maintain good operational smoothness. The method considers both risk suppression and control continuity when generating control strategies, thus demonstrating superior overall stability. Therefore, this invention provides a preferred solution: if the autonomous driving mode is determined based on vehicle operating state indicators, the Justin900429 algorithm is selected as the autonomous driving model algorithm after entering a boundary scenario.

[0054] The autonomous driving testing method based on pre-boundary scenario guidance proposed in this invention can comprehensively examine the actual performance of algorithms in risk perception, decision response, control execution, and stability recovery. It can more easily reveal the essential differences between different autonomous driving algorithms under high-risk conditions. Finally, the Justin900429 algorithm can be reasonably selected as the algorithm for the autonomous driving model. Alternatively, based on the test results of the corresponding algorithm, the poorly performing indicators can be incorporated into the model training objective or reward function design. At the same time, data augmentation, adversarial training, and safety constraint optimization can be carried out for the weak scenarios exposed in the test, thereby forming a closed-loop iterative path of testing, problem discovery, algorithm improvement, and retesting, providing a basis for optimizing autonomous driving control strategies.

[0055] In summary, based on research on pre-boundary scenarios and boundary scenarios, the present invention further introduces risk scenarios into the accelerated testing process, establishing an accelerated testing framework that includes risk identification, active guidance, algorithm response, and result recording. This method enables scenarios to gradually evolve from pre-boundary scenarios to boundary scenarios under continuous interaction conditions. Compared with testing methods that rely directly on the natural evolution of traffic flow, the present invention can more effectively target testing at key risk stages. Furthermore, the test results show that, under the condition of maintaining a consistent total test duration, the proposed method is more likely to trigger boundary scenarios compared to ordinary natural simulation, and also significantly improves the ability to expose high risks and trigger accidents. Specifically, the proposed method shows better results in terms of boundary scenario duration, high-risk exposure rate, overall accident rate, and risk accumulation degree. This indicates that the proposed method can concentrate limited test resources more on high-risk critical segments, thereby improving the exposure efficiency of potential defects in autonomous driving systems. In other words, the proposed method not only improves the triggering efficiency of boundary scenarios, but also enhances the test's ability to identify dangerous states and their failure consequences. In addition, the performance of different autonomous driving control algorithms under boundary scenario conditions was compared based on the evaluation system, which can provide methodological support for subsequent autonomous driving system driving tests and performance evaluations, and provide a decision-making basis for determining and optimizing autonomous driving models.

[0056] This invention provides a specific method for testing autonomous driving based on pre-boundary scenario guidance. After identifying a pre-boundary state in a continuously running scenario, instead of directly placing the vehicle into the boundary scenario, it guides the target vehicle based on the statistical patterns of speed changes in real pre-boundary samples, allowing the scenario to continue developing along the path of "pre-boundary scenario – boundary scenario". In this way, while preserving the gradual accumulation of risk, it can more effectively trigger boundary scenarios and further evaluate the response capability of autonomous driving algorithms under high-risk conditions, including: Step one: Prepare for testing. Load the completed Town04 test scenario into the CaRLa platform and check the road geometry, lane attributes, and initial states of traffic participants to ensure the test environment meets the preset requirements. Based on this, deploy the tested autonomous driving control algorithms, including TCP, Ekostinoudis, and Justin900429 methods, onto the target vehicle and complete the connection between the algorithm module and the vehicle control interface. Simultaneously, configure the system with LiDAR, cameras, and positioning modules, and complete the relevant sensor count to ensure continuous operation of each functional unit during subsequent testing.

[0057] Step 2, Scene Operation and Pre-Boundary Scene Recognition: After initialization, the target vehicle begins autonomous driving within the designated test area. The sensor module continuously collects the target vehicle's operating status, surrounding vehicle behavior information, and road environment information, and transmits the relevant data to the risk recognition module in real time. Based on the pre-boundary scene recognition method, the current traffic status is dynamically evaluated, and it is determined in real time whether the scene has entered the pre-boundary state. When the system determines that the current segment does not belong to the pre-boundary scene, the vehicle continues to operate normally, and the test process continues to execute status monitoring and data recording.

[0058] Step 3: Pre-boundary Scene Identification and Guidance Triggering; Once the system identifies that the current scene has entered the pre-boundary state, the test process enters the guidance triggering phase. In this phase, the system does not randomly apply disturbances, nor does it directly construct the terminal boundary state. Instead, it guides the target vehicle based on the statistical patterns of speed changes in real pre-boundary samples. Specifically, the SOFTS-based boundary scene prediction model predicts the speed increment after the current scene transforms into a boundary scene, applies acceleration control to the target vehicle, and ensures that the current traffic state continues to evolve along the existing risk evolution direction. In one specific implementation, taking statistical results as an example, when the target vehicle enters the pre-boundary scenario, the predicted speed increment of the boundary scenario is 1.46 km / h. To achieve this guidance process, the system selects the corresponding control quantity from the throttle and acceleration mapping relationship shown in Table 2 based on the target vehicle's current speed, and inputs the control quantity into the vehicle's execution end. For example, if the current speed is 70 km / h, the throttle opening should be 0.6. Through this guidance method based on the statistical laws of real samples, the test scenario can evolve from the pre-boundary state to the boundary state in a more natural way, thereby increasing the probability of boundary scenario triggering and enhancing the pertinence of the test process.

[0059] Step four, boundary scenario response; as the pre-boundary scenario further enters the boundary state under guidance, the guidance control exits, and vehicle control is reassigned to the autonomous driving algorithm on the target vehicle. At this point, the system continues to record the target vehicle's longitudinal and lateral control behavior, changes in the surrounding environment, and the evolution of risk indicators. The focus is on analyzing the algorithm's braking, steering, avoidance, and recovery capabilities under high-risk conditions. This stage primarily evaluates the actual performance of the autonomous driving algorithm in boundary scenarios. When the test reaches the termination condition, the system automatically terminates the current round and saves the relevant test data. This paper uses a fixed duration as the main condition for test termination; each round of testing automatically ends after reaching the preset test duration. In addition, to ensure test safety and data validity, when a collision occurs during the test, the vehicle is promptly moved to another location to continue the test. After the test, the system extracts and saves key performance data such as avoidance success rate, changes in risk indicators, and operational stability for subsequent statistical analysis and comparative studies between different autonomous driving algorithms. Understandably, high-risk exposure rate, overall accident rate, accident rate under boundary scenario conditions, average duration of high risk, and degree of risk accumulation are used to evaluate the performance of accelerated testing methods; risk avoidance success rate, highest risk index, risk fluctuation range, risk convergence time, speed fluctuation range, and maximum deceleration are used to evaluate the performance of each control algorithm.

[0060] The present invention also provides an autonomous driving test system based on pre-boundary scenario guidance, comprising: The pre-boundary scene recognition module identifies the scene in which the target vehicle is located based on the pre-boundary scene recognition method. If the target vehicle is in a pre-boundary scene, it enters the intervention and guidance driving state. If the target vehicle is in a safe scene, the target vehicle maintains the autonomous driving operation state and continues to perform scene recognition. The guided driving control module is used to obtain the speed increment of the target vehicle as it evolves from the pre-boundary scene to the boundary scene based on the current driving characteristics of the target vehicle, determine the guiding acceleration of the target vehicle based on the speed increment, and use the guiding acceleration to apply a disturbance to guide the target vehicle until the target vehicle enters the boundary scene. The automatic driving test module is used to switch to automatic driving mode when the target vehicle enters the boundary scene. It uses the autonomous driving model to make autonomous decision-making tests based on the current vehicle state and current vehicle environment of the target vehicle until the test is terminated. The current vehicle state includes the critical speed at which the target vehicle enters the boundary scene. The test evaluation module is used to obtain five indicators: high-risk exposure rate, overall accident rate, accident rate under boundary scenario conditions, average high-risk duration, and risk accumulation level. It performs quantitative analysis on the performance of the test method and obtains the risk avoidance success rate, highest risk index, risk fluctuation range, risk convergence time, speed fluctuation range, and maximum deceleration to evaluate the decision-making and control capabilities of the autonomous driving model in high-risk scenarios.

[0061] The above are preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments; all technical solutions falling within the scope of the present invention's concept are within its protection. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within its protection scope. Therefore, the patent protection scope of the present invention should be determined by the appended claims.

Claims

1. A method for testing autonomous driving based on pre-boundary scenario guidance, characterized in that, Including the following steps: S10, Based on the pre-boundary scene recognition method, the scene where the target vehicle is located is recognized. If the target vehicle is in the pre-boundary scene, proceed to step S20. If the target vehicle is in the safe scene, the target vehicle maintains the autonomous driving operation state and continues to perform scene recognition. S20, switch to intervention-guided driving state, obtain the speed increment of the target vehicle from the pre-boundary scene to the boundary scene based on the current driving characteristics of the target vehicle, determine the guidance acceleration of the target vehicle based on the speed increment, and apply a disturbance to the target vehicle using the guidance acceleration until the target vehicle enters the boundary scene; S30, switch to automatic driving mode, and use the automatic driving model to make autonomous decision-making tests based on the current vehicle state and current vehicle environment of the target vehicle until the test is terminated. The current vehicle state includes the critical speed at which the target vehicle enters the boundary scene.

2. The autonomous driving test method based on pre-boundary scenario guidance according to claim 1, characterized in that, In step S10, Obtain the real-time scene feature matrix of the target vehicle, and use the pre-boundary scene discrimination model based on the real-time scene feature matrix to identify and determine whether the scene in which the target vehicle is located is the pre-boundary scene or the safety scene; The pre-boundary scene discrimination model is one of the following: random forest model, logistic regression model, support vector machine, and single decision tree model.

3. The autonomous driving test method based on pre-boundary scenario guidance according to claim 2, characterized in that, The pre-boundary scene discrimination model is a random forest model. When the random forest model is trained using a training and test set, the training input label is the real-time scene feature matrix, and the training output label is the judgment label with the pre-boundary scene or the safe scene as binary variables. In step S10, the target vehicle autonomously drives within the designated test area; The sensor module continuously collects the vehicle operating status of the target vehicle and the operating status of surrounding vehicles, and calculates the MTTC risk threshold based on the vehicle operating status and the operating status of surrounding vehicles; when the MTTC risk threshold is less than the first threshold, the pre-boundary scene discrimination model is activated to determine whether the target vehicle is in the pre-boundary scene.

4. The autonomous driving test method based on pre-boundary scenario guidance according to claim 1, characterized in that, The test will terminate if the total test duration of the autonomous decision-making test reaches the preset duration.

5. The autonomous driving test method based on pre-boundary scenario guidance according to any one of claims 1 to 4, characterized in that, In step S20, a throttle-acceleration mapping table is established based on the statistical law of throttle speed change in the real pre-boundary samples; Based on the guided acceleration, the valve opening ratio of the acceleration control valve is determined using the throttle-acceleration mapping table; Based on the valve opening ratio, a guiding disturbance is applied to the target vehicle until the target vehicle enters the boundary scene.

6. The autonomous driving test method based on pre-boundary scenario guidance according to claim 5, characterized in that, It also includes the following steps: obtaining the high-risk exposure rate, overall incident rate, incident rate under boundary scenario conditions, average high-risk duration, and risk accumulation level corresponding to the total test duration. The accelerated testing method guided by pre-boundary scenarios is evaluated based on the high-risk exposure rate, overall accident rate, accident rate under boundary scenario conditions, average duration of high risk, and degree of risk accumulation.

7. The autonomous driving test method based on pre-boundary scenario guidance according to claim 5, characterized in that, Let the total test duration be The sampling interval is The total number of sampling points is , ,time Risk indicators are The number of collision events that occurred during the entire test process within the total test duration was The number of boundary scene events identified was , No. The duration of continuous events in sub-boundary scenario events that satisfy the MTTC risk threshold ≤ 0.9 is: ; Using formula Calculations were performed to obtain the high-risk exposure rate. , This is an indicator function that takes the value 1 when the condition is true and 0 otherwise. This high-risk exposure rate is used to represent the proportion of time spent in the boundary scenario to the total test duration during the entire test process. Using formula Calculations were performed to obtain the overall accident rate. The overall accident rate is expressed as the frequency of collision events occurring over time during the test. Using formula Calculations are performed to obtain the accident rate under boundary scenario conditions. The accident rate under the boundary scenario conditions is used to reflect the degree of danger in triggering the boundary scenario; the formula is used. Calculations were performed to obtain the average duration of high risk. The average duration of high risk is used to reflect the sustained level of danger in the test scenario; Using formula Calculations are performed to obtain the degree of risk accumulation. The risk accumulation level is used to reflect the overall severity of risk exposure during the testing process.

8. An autonomous driving test system based on pre-boundary scenario guidance, characterized in that, The method for implementing the pre-boundary scenario-guided autonomous driving test method as described in any one of claims 1 to 7 includes: The pre-boundary scene recognition module identifies the scene in which the target vehicle is located based on the pre-boundary scene recognition method. If the target vehicle is in a pre-boundary scene, it enters the intervention and guidance driving state. If the target vehicle is in a safe scene, the target vehicle maintains the autonomous driving operation state and continues to perform scene recognition. The guided driving control module is used to obtain the speed increment of the target vehicle as it evolves from the pre-boundary scene to the boundary scene based on the current driving characteristics of the target vehicle, determine the guiding acceleration of the target vehicle based on the speed increment, and apply a disturbance to the target vehicle using the guiding acceleration to guide the target vehicle until the target vehicle enters the boundary scene. The automatic driving test module is used to switch to automatic driving state when the target vehicle enters the boundary scene, and use the automatic driving model to make autonomous decision-making tests based on the current vehicle state and current vehicle environment of the target vehicle until the test is terminated. The current vehicle state includes the critical speed at which the target vehicle enters the boundary scene. The test evaluation module is used to obtain five indicators: high-risk exposure rate, overall accident rate, accident rate under boundary scenario conditions, average high-risk duration, and risk accumulation level. It performs quantitative analysis on the performance of the test method and obtains the risk avoidance success rate, highest risk index, risk fluctuation range, risk convergence time, speed fluctuation range, and maximum deceleration to evaluate the decision-making and control capabilities of the autonomous driving model in high-risk scenarios.

Images