Online generation method of key edge test scenarios for automatic driving acceleration test

Abstract

This invention relates to an online method for generating critical boundary test scenarios for autonomous driving acceleration testing, comprising: generalizing and generating natural driving test scenarios; constructing a scenario complexity model, including static scenario complexity and dynamic scenario complexity; evaluating and testing autonomous vehicles based on the generated natural driving test scenarios and the constructed scenario complexity model; and performing scenario adaptive adjustments based on the test results, including static scenario adaptive adjustments and dynamic scenario adaptive adjustments, to output critical boundary test scenarios. Compared with existing technologies, this invention can generate online critical boundary test scenarios that satisfy the distribution and diversity of natural traffic behavior and have multiple interactions, in order to test the performance boundaries of autonomous vehicles.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic driving, and in particular to a key edge test scene online generation method for automatic driving acceleration test. BACKGROUND

[0002] With the development of automatic driving technology, in recent years, automatic driving test methods have also been paid more and more attention, and a large number of researchers have made great efforts to accelerate the test of automatic driving, so as to promote the practical application deployment of automatic driving vehicles. The scene-based test method has been widely concerned by researchers, and the key step of intelligent driving test is to find the test scene that meets the natural distribution and is key. Only by comprehensively covering various scenes, can the safety and intelligence of automatic driving be effectively solved. However, from the point of view of calculation, it is not feasible to search all possible traffic scenes to find key test scenes, because the natural traffic environment contains thousands of parameter combinations. How to improve the generation efficiency and diversity of key test scenes has become an important problem. The driving scene in the real world is very complex and diverse, which contains various structural types of static road environment and various types of dynamic traffic participants. Because there are infinite possible scenes in the real world, the scene generation technology must provide extensive changes to solve the infinite scenes. The key of the test scene is to effectively reproduce the complexity and variability of the real world scene to ensure the efficiency and effectiveness of intelligent driving test.

[0003] The existing test scenes of automatic driving vehicles mainly include two types of typical scenes and key scenes, as shown in Table 1. The former contains a large number of low-risk scenes, mainly testing the implementation of various conventional functions of the vehicle. The latter mainly expands the proportion of high-risk scenes in the test scene set, realizes the supplement of typical scenes, and further improves the coverage rate of test scenes, which is usually used for safety test of automatic driving vehicles, to find the boundary scene between the occurrence of accidents and the non-occurrence of accidents of automatic driving vehicles, and then guide the optimization of robustness and safety of automatic driving vehicles.

[0004] Table 1 Comparison of typical scenes and key scenes

[0005]

[0006]

[0007] In the prior art automatic driving test scene generation method, attention is focused on the generation of high-risk collision scenes, such as using adversarial methods to cause the test automatic driving vehicle to collide, but these generated test scenes can only test the safety of the automatic driving vehicle, and cannot comprehensively test the comprehensive performance of the automatic driving vehicle. In addition, the existing method mainly focuses on the generation of uniformly sampled scenes and dynamic traffic scenes, and rarely considers the generation of static key test scenes and dynamic key test scenes, which will inevitably affect the test efficiency of the automatic driving vehicle and the diversity of the test scene. And the existing key scene generation method rarely considers the complexity of the scene, and it is difficult to generate test scenes with different complexity levels. The above defects will affect the acceleration test efficiency of the automatic driving vehicle to a certain extent, and cannot test the performance boundary of the automatic driving vehicle (referring to the performance limit of the automatic driving vehicle without collision). SUMMARY

[0008] The purpose of the present application is to overcome the defects of the prior art and provide a key edge test scene online generation method for automatic driving acceleration test, which can generate key boundary test scenes that meet the natural traffic flow behavior distribution, diversity and multiple interactions online to test the performance boundary of the automatic driving vehicle.

[0009] The purpose of the present application can be achieved by the following technical solution: a key edge test scene online generation method for automatic driving acceleration test, comprising the following steps:

[0010] S1, generalizing natural driving test scenes;

[0011] S2, constructing a scene complexity model, including static scene complexity and dynamic scene complexity;

[0012] S3, based on the generated natural driving test scenes and the constructed scene complexity model, evaluating and testing the automatic driving vehicle;

[0013] S4, according to the test results obtained in step S3, performing scene adaptive adjustment, including static scene adaptive adjustment and dynamic scene adaptive adjustment, and outputting the key boundary test scene.

[0014] Further, the specific process of step S1 is:

[0015] S11, identifying the driving environment of the real world, and converting the static traffic environment and weather environment into a static traffic scene;

[0016] S12, identifying the driving environment of the real world, and obtaining a natural driving behavior distribution model of dynamic traffic participants;

[0017] S13. Generalize the natural traffic scenario on the simulation platform and convert it into a natural driving test scenario.

[0018] Furthermore, step S13 specifically involves using natural driving data from the natural driving environment to generalize parameters and generate simulation test scenarios in OpenDrive and OpenScenario formats.

[0019] Furthermore, the scene complexity model in step S2 is specifically as follows:

[0020] C = α S C S +α D C D

[0021] Where, α S and α D The static scene complexity is C. S and dynamic scene complexity C D The weighting coefficients.

[0022] Furthermore, the quantification dimensions of the static scene complexity in step S2 include the area of ​​the drivable area, weather visibility, and road friction coefficient;

[0023] Key variables for quantifying the complexity of dynamic scenarios include the encounter angle, relative distance, and relative speed between the tested autonomous vehicle and other traffic participants.

[0024] Furthermore, the complexity of the static scene is specifically as follows:

[0025]

[0026] Among them, A max A represents the maximum area of ​​the drivable zone. min W is the minimum area of ​​the drivable zone. max W is the visibility distance at which lighting and clarity are optimal. min F represents the visibility distance at which illumination and transparency are at their worst. max F is the coefficient of road friction under the driest conditions. min γ is the coefficient of road friction under the wettest road conditions. d γ v γ f These correspond to the drivable area A. d Weather visibility W v and road friction coefficient F r The weighting coefficients.

[0027] Furthermore, the calculation process for the dynamic scene complexity is as follows:

[0028] First, define the effective range of potentially complex traffic participants around the autonomous vehicle.

[0029] Then the complexity of the dynamic interaction pair, which consists of the autonomous vehicle under test and surrounding traffic participants, is calculated.

[0030] Finally, the complexity of all interaction pairs is taken into account to calculate the complexity of the dynamic traffic scenario.

[0031] The effective range of potential traffic participants around the autonomous vehicle is described by a semicircle centered on the rear axle center of the vehicle, with a radius of:

[0032] r range =max{r safe ,r l}

[0033]

[0034] Where, r range r is the radius of the area affected by the autonomous vehicle. safe r is the radius of the safe zone for the autonomous vehicle under test. l Let v0 be the length of the intersection area, v0 be the initial speed of the autonomous vehicle, and τ be the response lag time of the autonomous vehicle. and These represent the maximum acceleration and minimum deceleration of an autonomous vehicle, respectively.

[0035] The complexity of the dynamic interaction pair is:

[0036] C i,0 =Γ(θ) i,0 ,d i,0 ,v i,0 )=f1(θ i,0 )×f2(d i,0 )×f3(v i,0 )

[0037]

[0038]

[0039]

[0040]

[0041]

[0042] Among them, C i,0 For the complexity among dynamic traffic participants, θ i,0d is the meeting angle between dynamic traffic participants. i,0 v represents the relative distance between dynamic traffic participants. i,0 For the relative velocities between dynamic traffic participants, f1(θ) i,0 ) represents the meeting angle θ i,0 With dynamic interaction complexity C i,0 The relationship between f2(d) i,0 ) represents the relative distance d i,0 With dynamic interaction complexity C i,0 The relationship between f3(v) i,0 (v) represents the relative velocity. i,0 With dynamic interaction complexity C i,0 The relationship between D′ i,0 d represents the standardized relative distance. i,0 d represents the relative distance between different traffic participants. min The relative distance in the least complex case is equal to the radius r of the autonomous vehicle's influence area. range d max The relative distance in the most complex case is zero, indicating that the two cars are infinitely close, v′ i,0 v is the standardized relative velocity. i,01 v represents the relative speed between different traffic participants. min The relative velocity in the least complex case is equal to the divergent relative velocity of 3.5 m / s, v. max The relative velocity under the most complex condition is equal to the convergent relative velocity of 3.5 m / s;

[0043] The complexity of the dynamic traffic scenario is:

[0044]

[0045]

[0046] Among them, C D For dynamic scene complexity, C t The dynamic scene complexity of the integration before smoothing at time step t, where N is the number of vehicle pairs within the affected region, and λ i For the impact of traffic participant i, C i,0 For the set of interactive pairs E N+1 The complexity is given by k, where k represents the length of the sliding window.

[0047] Furthermore, step S4 specifically includes the following steps:

[0048] S41. Based on the test results obtained in step S3, determine whether the current test scenario belongs to a safety-critical test scenario, a critical edge test scenario, or a test scenario that requires further adaptive optimization.

[0049] If it belongs to a safety-critical test scenario, the generated scenario will be added to the safety-critical test scenario library;

[0050] If it belongs to a critical edge test scenario, the generated scenario will be placed in the critical edge test scenario library;

[0051] If the requirements of the desired scenario are not met, proceed to step S42;

[0052] S42. The current test scenario is further adjusted and optimized online through the critical edge test scenario adaptive generation algorithm. The critical edge test scenario adaptive generation algorithm performs adaptive optimization of the scenario from both static and dynamic perspectives simultaneously.

[0053] The adaptive optimization of static scenes is carried out from three aspects: drivable area, weather visibility, and road friction coefficient;

[0054] Adaptive optimization of dynamic traffic scenarios is achieved by selecting other major adversarial traffic participants and optimizing their natural and adversarial behaviors.

[0055] In the process of adaptive optimization of static and dynamic traffic scenarios, the scenario complexity model is used to quantify the complexity of the optimized scenario. Based on the scenario complexity improvement coefficient, the complexity and danger of the test scenario are adaptively improved, and finally a key test scenario library with different complexity levels, used for performance testing in different dimensions, and covering various scenarios is generated.

[0056] Furthermore, the specific process of static scene adaptive optimization in step S42 is as follows:

[0057] First, based on the geographical coordinates G of the actual traffic environment c and distance threshold d th Extract road network data from OpenStreetMap (OSM);

[0058] After data extraction, a static traffic scene is generated and post-processed. Then, the parameter vector R is used to... it The static traffic scene is parameterized, where the parameter vector is defined as R. it =[P s V s W v ,F r ] T P s and V sLet R0 represent the position and size of the static obstacle, respectively. An initial state vector R0 is used to initialize the static traffic scene. Then, the autonomous vehicle is tested based on the generated scene, and trajectory data T is obtained. s and the corresponding test results;

[0059] Finally, the area of ​​the drivable region is calculated using the reachable state set:

[0060]

[0061] x(R it Let u(t) represent the motion state of the tested autonomous vehicle at time t, u(t) represent the action of the vehicle at time t, and proj(x) project the vehicle's state x onto the two-dimensional position domain. lane \O(R it O(R) indicates that it does not contain static obstacles. it Two-dimensional space;

[0062] To improve the criticality of static scenes, the parameter vector R is optimized. it Obtain the critical static traffic scenario S static (it+1) is used for the (it+1)th intelligent driving test:

[0063]

[0064] stA d >0

[0065] Using the static scene parameters R calculated in the it-th intelligent driving test it And scalar β∈[0,1] to quantize and improve the complexity of static scenes: C S,crit (it+1)=(1+β)·C s (it).

[0066] Furthermore, the specific process of dynamic scene adaptive optimization in step S42 is as follows:

[0067] First, collect natural driving data D n ;

[0068] Then, the behaviors of different traffic participants in the natural driving data are identified and analyzed, and the motion states s and actions u of the traffic participants are obtained. In order to generate a dynamic traffic scene that meets the requirements, the state frequency P(s) and action frequency P(u|s) of different traffic participants in the corresponding states are calculated. The dynamic scene parameters (s(t), u(t), P(s), P(u|s)) are initialized and compared with the static traffic scene S. static Combined to generate the initial scene;

[0069] Subsequently, the autonomous vehicle is tested according to the generated test scenarios. In order to generate critical boundary test scenarios efficiently and reasonably, it is necessary to identify the main other traffic participants who have a significant impact on the autonomous vehicle under test from the background traffic participants. The behavior of the main other traffic participants is dynamically adjusted using dynamic scenario generation and adaptive optimization algorithms, and the behavior of other background traffic participants is controlled using a model based on natural driving environment data, so that the generated critical boundary test scenarios conform to the distribution of natural driving data as much as possible.

[0070] To identify the major other traffic participants who pose the greatest threat to the tested autonomous vehicle, the concept of the autonomous vehicle's influence zone is used to traverse and select traffic participants within that zone. First, natural traffic flow data in the test environment is analyzed, and the state frequency P(s) and action frequency P(u|s) of different traffic participants under different states are calculated. At time steps t = 0, ..., T, the states and actions are represented as follows:

[0071] s(t)=[s0(t),s1(t),…,s M+N+L (t)]

[0072] u(t) = [u0(t), u1(t), ..., u M+N+L (t)]

[0073] Where s0(t) and u0(t) represent the state and action of the autonomous vehicle being tested at time t, respectively. i (t)(i=1,2,…,M+N+L) represents the state of the i-th background traffic participant at time t, where M, N, and L represent the number of background vehicles, background pedestrians, and background bicycles within the affected area, respectively. i (t) represents the action of the i-th background traffic participant at time t.

[0074] Then use dynamic scene interaction to reduce complexity C i,0 To measure the threat level posed by surrounding traffic participants to the tested autonomous vehicle, state-action pair (s) i ,u i The interaction complexity under ) is C i,0 (s i ,u i )for:

[0075]

[0076] Where P(u0|s0) represents the probability that the vehicle will take action u0 in state s0;

[0077] Based on the obtained action frequency P(u) of background traffic participants i |si ) and interaction complexity C i,0 (s i ,u i ), and comprehensively calculate the critical value q(u) of the corresponding background traffic participants. i |s i ):

[0078] q(u i |s i )=P(u i |s i )×C i,0 (s i ,u i )

[0079] Then, the threshold value for each background traffic participant is calculated as the sum of the action threshold values ​​for all actions of the corresponding background traffic participant:

[0080]

[0081] Based on the calculated threshold value for each background traffic participant, the background vehicles, pedestrians, and bicycles with the largest threshold values ​​are selected as the main other traffic participants. The identification of the main other vehicles, pedestrians, and bicycles is as follows:

[0082]

[0083]

[0084]

[0085] Where, q v q p and q b Let M, N, and L represent the numbers of the main other vehicles, pedestrians, and bicycles, respectively. Let Q represent the sets of background vehicles, pedestrians, and bicycles within the affected area. v Q P and Q b These represent predetermined critical thresholds for identifying major other vehicles, pedestrians, and bicycles, specifically adversarial vehicles, adversarial pedestrians, adversarial bicycles, or any combination thereof, to maximize the efficiency of generating critical boundary test scenarios.

[0086] After identifying the major other traffic participants, the complexity of the dynamic scenario is gradually increased by optimizing the behavior of the critical boundary test scenario. The behavior of the major other traffic participants has a crucial impact on the complexity of the dynamic scenario. Changes in scenario complexity can accelerate the evaluation of the performance boundaries of autonomous vehicles and identify critical boundary test scenarios. Scenario complexity is optimized during the adaptive adjustment of scenario parameters. If the autonomous vehicle passes the current test, the complexity threshold of the dynamic traffic scenario is further increased, and its calculation is as follows:

[0087] C crit = (1+β)C D (it)

[0088] Where β∈(0,1] is the scene complexity enhancement coefficient;

[0089] After determining the complexity threshold of the scenario, the action parameters u of the main other traffic participants are... q Optimization is performed to generate critical boundary test scenarios that meet complexity requirements and minimize deviation from the distribution of natural driving data. The behavioral trajectories of the main other traffic participants obtained from natural driving traffic flow are represented as T. nat (t,u q,nat ), and represent the optimized behavioral trajectories of the main other traffic participants as T crit (t,u q The corresponding objective function and constraints are shown below:

[0090]

[0091] st D (it+1)≥C crit ,

[0092]

[0093] Where t0 represents the start time of behavior optimization by other major traffic participants, t f TTC(t) represents the duration during which major other traffic participants influence the behavior of the tested autonomous vehicle. TTC(t) represents the minimum collision time between the major other traffic participants and the nearest obstacle or other traffic participant at time t. min This represents the minimum safe collision time, ensuring, through the aforementioned constraints, that no other road users in the vicinity actively collide with the tested autonomous vehicle.

[0094] Compared with existing technologies, this invention proposes an online generation method for critical edge test scenarios in autonomous driving acceleration testing. By constructing an autonomous driving test scenario complexity model to quantify the complexity of different test scenarios, and comprehensively considering the complex interactive behaviors of autonomous vehicles, human-driven vehicles, bicycles, and pedestrians, as well as their impact on the tested autonomous vehicles, the method systematically describes, generates, selects, optimizes, and uses test scenarios. This combines virtual scenarios with real traffic flow data, effectively improving the flexibility, complexity, and diversity of critical edge test scenarios.

[0095] This invention establishes a driving scenario complexity model, which mainly consists of two parts: a static scenario complexity model and a dynamic scenario complexity model. In the static scenario complexity model, scenario complexity is quantified from three dimensions: the area of ​​the driving area, weather visibility, and the road friction coefficient. In the dynamic scenario complexity model, complexity is quantified based on three key variables: the encounter angle, relative distance, and relative speed between the tested autonomous vehicle and other surrounding traffic participants. Then, based on potential field theory and other methods, the scenario complexity and hazard assessment of the entire test scenario are performed. This allows for better quantification of the complexity and hazard of the test scenario and the identification of key test scenarios to accelerate testing.

[0096] This invention comprehensively considers the complex interactions of mixed traffic flows consisting of autonomous vehicles, human-driven vehicles, bicycles, and pedestrians. First, it defines the effective range of potential traffic participants with varying complexity around the autonomous vehicle. Then, it calculates the complexity of dynamic interaction pairs, which consist of the tested autonomous vehicle and surrounding traffic participants. Finally, it comprehensively considers the complexity of all interaction pairs to calculate the complexity of the dynamic traffic scenario. This allows for a more comprehensive assessment of the complexity of dynamic traffic participants around the autonomous vehicle.

[0097] This invention conducts comprehensive intelligent driving tests on autonomous driving systems based on generated natural driving simulation test scenarios and scenario complexity models. The test results are analyzed to determine whether critical edge test scenario adaptive generation algorithms are needed for test scenario optimization. These algorithms simultaneously optimize scenarios from both static and dynamic perspectives. Static scenario adaptive optimization considers three aspects: drivable area, weather visibility, and road friction coefficient. Dynamic traffic scenario adaptive optimization is achieved by selecting key adversarial traffic participants and optimizing their natural and adversarial behaviors. During the adaptive optimization of static and dynamic traffic scenarios, a scenario complexity model is used to quantify the optimized scenario complexity. Based on the scenario complexity enhancement coefficient, the complexity and danger of the test scenarios are adaptively increased. Ultimately, this generates a critical test scenario library with different complexity levels, covering various scenarios for performance testing across different dimensions. Attached Figure Description

[0098] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0099] Figure 2 This is the adaptive generation architecture for simulating key autonomous driving scenarios in the embodiment;

[0100] Figure 3 This is a schematic diagram of the drivable area and weather visibility for a static scene complexity model;

[0101] Figure 4 A schematic diagram illustrating the virtual formation method and the calculation of relative distances between different traffic participants;

[0102] Figure 5 Flowcharts that are adaptively generated and optimized for critical edge scenes;

[0103] Figures 6a to 6f This is a schematic diagram showing the simulation test comparison of key edge scenes in the embodiments. Detailed Implementation

[0104] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0105] Example

[0106] like Figure 1 As shown, an online method for generating critical edge test scenarios for autonomous driving acceleration testing includes the following steps:

[0107] S1. Generalize and generate natural driving test scenarios;

[0108] S2. Construct a scene complexity model, including static scene complexity and dynamic scene complexity;

[0109] S3. Based on the generated natural driving test scenarios and the constructed scenario complexity model, evaluate and test autonomous vehicles.

[0110] S4. Based on the test results obtained in step S3, perform scene adaptive adjustment, including static scene adaptive adjustment and dynamic scene adaptive adjustment, and output the critical boundary test scene.

[0111] This embodiment applies the above technical solution, mainly including:

[0112] The core of autonomous driving key scenario simulation generation and accelerated testing technology is to generate test scenarios that satisfy natural traffic flow distribution, are adversarial, and have high complexity. Furthermore, these test scenarios need to cover as many corner cases as possible that autonomous vehicles might encounter in real-world driving. These key test scenarios are truly meaningful for autonomous driving testing and are crucial for improving the performance of autonomous driving systems.

[0113] In order to effectively generate key test scenarios that meet the distribution of natural traffic scenarios, this embodiment builds an online key test scenario adaptive generation system to accelerate the testing process of autonomous vehicles from multiple dimensions. Figure 2 The diagram illustrates the architecture of the online key test scenario adaptive generation system in this embodiment. The entire system consists of four parts: a natural driving scenario generalization generation module, a scenario complexity evaluation model, an automated testing platform for intelligent driving algorithms, and a key test scenario adaptive generation algorithm. To achieve the generation and automated testing of simulation test scenarios that satisfy natural traffic flow distribution, natural driving data from the natural driving environment is used for parameter generalization to generate simulation test scenarios in OpenDrive and OpenScenario formats. Furthermore, virtual simulation testing software such as VTD is used for real-world rendering to make the generated key test scenarios closer to real traffic scenarios. During the generation of key test scenarios, the impact of autonomous and human-driven vehicles around the test vehicle on the tested autonomous vehicle is considered, as well as the combined movement and behavior of pedestrians and bicycles in the surrounding mixed traffic flow.

[0114] I. Generalized Generation of Natural Driving Scenarios

[0115] In the natural driving scenario generalization generation module, environmental identification is performed based on the real-world driving environment, converting static traffic and weather environments into static traffic scenarios in OpenX format. Dynamic traffic participant behavior identification and analysis are then performed based on real traffic data to obtain natural driving behavior distribution models for vehicles, bicycles, and pedestrians. Finally, the natural traffic scenarios are generalized and converted into natural driving simulation test scenarios on a simulation platform.

[0116] II. Scene Complexity Model

[0117] To better quantify the complexity and hazard of test scenarios and identify key test scenarios to accelerate testing, a driving scenario complexity model is established. This model consists of two main parts: a static scenario complexity model and a dynamic scenario complexity model. In the static model, scenario complexity is quantified from three dimensions: the area of ​​the driving area, weather visibility, and the road friction coefficient. In the dynamic model, complexity is quantified based on three key variables: the encounter angle, relative distance, and relative speed between the tested autonomous vehicle and other traffic participants. Then, based on potential field theory and other methods, the scenario complexity and hazard assessment of the entire test scenario are performed.

[0118] The proposed scene complexity model consists of static scene complexity and dynamic scene complexity, where the static scene complexity is determined by C1. S The complexity of static traffic scenarios is quantified from three dimensions: drivable area, weather visibility, and road friction coefficient. The complexity of dynamic scenarios is quantified by C. D This indicates that the dynamic scene complexity model primarily focuses on the impact of various traffic participants on the tested autonomous vehicle. The scene complexity C is derived from the static scene complexity Cr. s and dynamic scene complexity C D Based on comprehensive analysis, the following formula can be used for calculation:

[0119] C = α s C s +α D C D (1)

[0120] Where, α S and α D These are the weighting coefficients for static scene complexity and dynamic scene complexity, respectively.

[0121] The complexity of static traffic scenarios is influenced by static traffic elements and environmental factors. Static traffic elements refer to the geometry of the road, while environmental factors include weather visibility and the road friction coefficient. Given the unique and diverse nature of traffic environments, we consider the drivable area A of the tested autonomous vehicle as a factor. dWeather visibility W v and road friction coefficient F r Define static traffic scenario C S The complexity can be calculated as follows:

[0122]

[0123] Among them, A max A represents the maximum area of ​​the drivable zone. min W represents the minimum area of ​​the drivable zone. max W represents the visibility distance at which lighting and clarity are optimal. min F represents the visibility distance at which illumination and transparency are at their worst. max F represents the coefficient of road friction under the driest conditions. min This represents the coefficient of road friction under the most wet road conditions. Static traffic scenarios, such as... Figure 3 As shown, the minimum area A containing the drivable area min The actual area A of the drivable zone d The maximum area A of the drivable zone max Static obstacles are represented by black boxes. For the autonomous vehicle under test, the smaller the drivable area, the smaller the required turning radius, and the higher the driving complexity. In addition to the impact of static obstacles on the drivable area of ​​the autonomous vehicle under test, the effects of fog and darkness on environmental visibility, as well as the impact of road surface friction coefficient after snowfall, are also considered. By adjusting the parameters of the static traffic scenario, the complexity of the static traffic scenario can be quantified to test the performance boundaries of autonomous vehicles in critical boundary test scenarios.

[0124] Dynamic traffic scenarios are highly variable and prone to causing accidents involving autonomous vehicles. This invention comprehensively considers the complex interactions of mixed traffic flows comprised of autonomous vehicles, human-driven vehicles, bicycles, and pedestrians. To more comprehensively assess the complexity of dynamic traffic participants around the autonomous vehicle, an effective range of potentially complex traffic participants around the autonomous vehicle is first defined, and then the complexity of dynamic interaction pairs is calculated. Each interaction pair consists of the tested autonomous vehicle and surrounding traffic participants. Finally, the complexity of all interaction pairs is comprehensively considered to calculate the complexity of the dynamic traffic scenario.

[0125] During driving, considering that not all traffic participants contribute to the dynamic scene complexity of the tested autonomous vehicle, it is first necessary to quantify the influence range of the area surrounding the autonomous vehicle to reduce its computational load. The influence area of ​​the autonomous vehicle can be described by a semicircle centered on the rear axle center point. The autonomous vehicle needs to ensure that it will not collide with other vehicles or rear-end the vehicle in front. Based on the responsibility-sensitive safety theory, the formula for the safe zone radius r of the tested autonomous vehicle is... safe Defined as:

[0126] Where v0 represents the initial speed of the autonomous vehicle, and τ represents the response lag time of the autonomous vehicle. and These represent the maximum acceleration and minimum deceleration of an autonomous vehicle, respectively.

[0127] The radius of influence area of ​​an autonomous vehicle can be set as follows:

[0128] r range =max{r safe ,r l} (4)

[0129] Where, r l Indicates the length of the intersection area.

[0130] To characterize the scenario complexity caused by combinations of different traffic participants, a directed graph G centered on autonomous vehicles is used. N+1 ={V N+1 E N+1 Let} represent the influence of different surrounding traffic participants on the complexity of the tested autonomous vehicle. In the directed graph G... N+1 In the diagram, N represents the number of traffic participants in the area of ​​influence surrounding the autonomous vehicle, and V is the set of points. N+1 (V N+1 ={0,1,2,…,N}) represents the set of all traffic participants, including autonomous vehicles, where the number of the tested autonomous vehicles is represented by 0, and the set of directed edges. This represents the complexity relationship between the interactive traffic participant i and the tested autonomous vehicle 0.

[0131] Three key variables—meeting angle, relative distance, and relative speed—are used to quantify the spatiotemporal interactions between dynamic traffic participants. The correlation between the meeting angle and the interaction complexity is considered fundamental. Considering that different dynamic traffic participants have different relative speeds and relative distances, two additional correction functions are then introduced using a multiplication operator. Therefore, the complexity C between dynamic traffic participants is as follows.i,0 Definition:

[0132] C i,0 =Γ(θ) i,0 ,d i,0 ,v i,0 )=f1(θ i,0 )×f2(d i,0 )×f3(v i,0 (5)

[0133] Where, θ i,0 d represents the meeting angle between dynamic traffic participants. i,0 v represents the relative distance between dynamic traffic participants. i,0 f1(θ) represents the relative speed between dynamic traffic participants. i,0 ) represents the meeting angle θ i,0 With dynamic interaction complexity C i,0 The relationship between f2(d) i,0 ) represents the relative distance d i,0 With dynamic interaction complexity C i,0 The relationship between f3(v) i,0 () represents relative velocity v i,0 With dynamic interaction complexity C i,0 The relationship between them.

[0134] The encounter angle between autonomous vehicles and surrounding traffic participants near intersections varies significantly. To better describe the impact of different encounter angles on the complexity of tested autonomous vehicles, an improved method for calculating encounter angle complexity is proposed as follows:

[0135]

[0136] Besides the meeting angle, the relative distance between the surrounding traffic participant i and the test autonomous vehicle 0 also significantly impacts the scenario complexity. Given the difficulty in accurately and effectively describing the relative distances between different traffic participants at an intersection, the concept of virtual formation is used to calculate the relative distances between the test autonomous vehicle and surrounding traffic participants, such as... Figure 3 As shown. Based on the conflict points P between the autonomous vehicle and surrounding traffic participants, the set V N+1 The positions of surrounding traffic participants are mapped to the lane where the test autonomous vehicle is located. The relative distances between the surrounding traffic participants and the conflict point P remain constant, satisfying Δs′. f =Δs f ,like Figure 4 As shown.

[0137] Based on information entropy theory, the nonlinear functional relationship between the relative distance and complexity among different traffic participants is proposed as follows:

[0138]

[0139]

[0140] Where, d′ i,0 d represents the standardized relative distance. i,0 This indicates the relative distance between different traffic participants. min This represents the relative distance in the least complex case, equal to the radius r of the influence region in formula (4). range . d max This represents the relative distance in the most complex case; zero indicates that the two vehicles are infinitely close.

[0141] Furthermore, the relative speeds between different traffic participants also significantly impact scene complexity. The influence of the relative speeds between different traffic participants on interaction complexity can be expressed as follows:

[0142]

[0143]

[0144] Where v′ i,0 v represents the standardized relative velocity. i,0 This indicates the relative speed between different traffic participants. min This represents the least complex relative velocity, equal to a divergent relative velocity of 3.5 m / s. max This represents the relative velocity under the most complex condition, which is equal to the convergent relative velocity of 3.5 m / s.

[0145] Finally, by integrating the complexity over all interactions within the affected area, the dynamic scene complexity C of the tested autonomous vehicle can be obtained. D The complexity is transformed into an estimate based on the sum of all interactions. This complexity metric is updated at a frequency consistent with the data collection frequency. The complexity of the driving environment is then smoothed out using an appropriate update frequency. The dynamic scene complexity C of an autonomous vehicle... D It can be calculated as follows:

[0146]

[0147]

[0148] Among them, C t Let N represent the dynamic scene complexity of the integration before smoothing at time step t, and let λ represent the number of vehicle pairs within the affected region. i C represents the influence of traffic participant i. i,0Represents the set of interaction pairs E N+1 The complexity is given by k, where k represents the length of the sliding window.

[0149] III. Autonomous Vehicle Evaluation and Scenario Adaptive Adjustment

[0150] Based on the generated natural driving simulation test scenarios and scenario complexity models, comprehensive intelligent driving tests are conducted on autonomous driving systems. The intelligent driving automated testing system mainly includes scenario import, intelligent driving algorithm embedding, simulation implementation, and automated evaluation of the autonomous driving system. In this system, the impact of the behavior of other dynamic traffic participants on the tested autonomous vehicle is considered, as well as the impact of the tested autonomous vehicle on traffic coordination with surrounding traffic participants. Since future truly deployable autonomous driving systems need to improve their own driving efficiency while also considering the impact of their own decision-making behavior on overall traffic coordination, the tested autonomous vehicle should be tested simultaneously on multi-dimensional performance indicators, including safety, driving comfort, driving performance, and traffic coordination. The main purpose of comprehensively evaluating the autonomous driving system is to identify its performance defects and then optimize it until an autonomous driving system that meets multi-dimensional performance requirements, including safety, is obtained.

[0151] After evaluating and testing the autonomous driving system, the test results are analyzed to determine whether the current test scenario belongs to a safety-critical test scenario, a critical-edge test scenario, or a test scenario requiring further adaptive optimization. If it belongs to a safety-critical test scenario, the generated scenario is added to the safety-critical test scenario library. If it belongs to a critical-edge test scenario, the generated scenario is added to the critical-edge test scenario library. If it does not meet the requirements of the desired scenario, the current test scenario is further adjusted and optimized online using a critical-edge test scenario adaptive generation algorithm. The critical-edge test scenario adaptive generation algorithm performs adaptive optimization of the scenario from both static and dynamic perspectives. The adaptive optimization of the static scenario is performed from three aspects: drivable area, weather visibility, and road friction coefficient. The adaptive optimization of the dynamic traffic scenario is achieved by selecting other major adversarial traffic participants and optimizing their natural and adversarial behaviors. During the adaptive optimization of static and dynamic traffic scenarios, a scenario complexity model is used to quantify the complexity of the optimized scenario. Based on the scenario complexity improvement coefficient, the complexity and danger of the test scenario are adaptively improved. Ultimately, a critical test scenario library with different complexity levels, suitable for performance testing in different dimensions and covering various scenarios, can be generated. The entire process of adaptive generation and optimization of critical edge scenes is as follows: Figure 5 As shown.

[0152] IV. Static Scene Generation and Adaptive Optimization Algorithm

[0153] Static traffic scenarios refer to roads, traffic facilities, and weather. For adaptive generation and optimization of static traffic scenarios, the focus is on the drivable area A for autonomous vehicles. d And the impact of weather conditions on visibility range W v and road friction coefficient F r In terms of optimizing the impact.

[0154] First, based on the geographical coordinates G of the actual traffic environment c and distance threshold d th Road network data is extracted from OpenStreetMap (OSM). After data extraction, a static traffic scene is generated and post-processed. Then, the parameter vector R is used... it The static traffic scene is parameterized, where the parameter vector is defined as R. it =[P s V s W v ,F r ] T P s and V s Let R and R represent the position and size of the static obstacle, respectively. First, the static traffic scene is initialized using an initial state vector R0. Then, the autonomous vehicle is tested based on the generated scene, and trajectory data T is obtained. s And the corresponding test results (e.g., PET, TTC). Finally, the area of ​​the drivable region is calculated using the reachable state set, as shown in equation (13). x(R it Let u(t) represent the motion state of the tested autonomous vehicle at time t, where u(t) represents the vehicle's action at time t. proj(x) projects the vehicle's state x onto the two-dimensional position domain. lane \O(R it O(R) indicates that it does not contain static obstacles. it (a two-dimensional space).

[0155]

[0156] To improve the criticality of static scenes, the parameter vector R is optimized. it Obtain the critical static traffic scenario S static (it+1) is used for the (it+1)th intelligent driving test:

[0157]

[0158] Using the static scene parameters R calculated in the it-th intelligent driving test it And scalar β∈[0,1] to quantize and improve the complexity of static scenes: C S,crit (it+1)=(1+β)·CS (it).

[0159] V. Dynamic Traffic Scene Generation and Adaptive Optimization Algorithm

[0160] To generate critical boundary test scenarios that satisfy the distribution of natural driving data, we first collect natural driving data D. n Then, the behaviors of different traffic participants in the natural driving data are identified and analyzed, and the motion states s and actions u of the traffic participants are obtained. To generate a dynamic traffic scene that meets the requirements, the state frequencies P(s) and action frequencies P(u|s) of different traffic participants in the corresponding states are calculated. The dynamic scene parameters (s(t), u(t), P(s), P(u|s)) are initialized and compared with the static traffic scene S. static The scenarios are combined to generate an initial scene. Then, the autonomous vehicle is tested based on each generated test scene. To generate critical boundary test scenes efficiently and reasonably, it is necessary to identify the main other traffic participants (LDPs) who have a significant impact on the autonomous vehicle under test from the background traffic participants. The behavior of the main LDPs is dynamically adjusted using dynamic scene generation and adaptive optimization algorithms, and a model based on natural driving environment data is used to control the behavior of other background traffic participants, so that the generated critical boundary test scenes conform as closely as possible to the distribution of natural driving data.

[0161] To identify the major other traffic participants who pose the greatest threat to the tested autonomous vehicle, the concept of the autonomous vehicle's influence area, as defined in formula (4), is used to traverse and select traffic participants within that area. First, natural traffic flow data in the test environment is analyzed, and the state frequency P(s) and action frequency P(u|s) of different traffic participants under different states are calculated. At time steps t = 0, ..., T, the states and actions are expressed as follows:

[0162] s(t)=[s0(t),s1(t),…,s M+N+L (t)] (15)

[0163] u(t) = [u0(t), u1(t), ..., u M+N+L (t)] (16)

[0164] Where s0(t) and u0(t) represent the state and action of the autonomous vehicle being tested at time t, respectively. i (t)(i=1,2,…,M+N+L) represents the state of the i-th background traffic participant at time t, where M, N, and L represent the number of background vehicles, background pedestrians, and background bicycles within the affected area, respectively. i (t) represents the action of the i-th background traffic participant at time t.

[0165] Then, dynamic scene interaction is used to adjust the complexity C. i,0 This is used to measure the threat level posed by surrounding traffic participants to the tested autonomous vehicle. State-action pair (s) i ,u i The interaction complexity under ) is C i,0 (s i ,u i It can be calculated as:

[0166]

[0167] Where P(u0|s0) represents the probability of the vehicle taking action u0 in state s0. Based on the above obtained action frequencies P(u0|s0) of background traffic participants... i |s i ) and interaction complexity C i,0 (s i ,u i ), and comprehensively calculate the critical value q(u) of the corresponding background traffic participants. i |s i ), which can be represented as:

[0168] q(u i |s i )=P(u i |s i )×C i,0 (s i ,u i (18)

[0169] Then, the threshold value for each background traffic participant is calculated as the sum of the action threshold values ​​for all actions of the corresponding background traffic participant:

[0170]

[0171] Based on the calculated threshold value for each background traffic participant, the background vehicles, pedestrians, and bicycles with the largest threshold values ​​are selected as the primary other traffic participants. The primary other vehicles, pedestrians, and bicycles can be identified as follows:

[0172]

[0173]

[0174]

[0175] Where, q v q p and q bLet M, N, and L represent the numbers of the main other vehicles, pedestrians, and bicycles, respectively. Let Q represent the sets of background vehicles, pedestrians, and bicycles within the affected area. v Q P and Q b These represent predetermined critical thresholds used to identify primary other vehicles, pedestrians, and bicycles. Primary other traffic participants here can be adversarial vehicles, adversarial pedestrians, adversarial bicycles, or any combination thereof, to maximize the efficiency of generating critical boundary test scenarios.

[0176] After identifying the major other traffic participants (NOTs), the next challenge is how to generate critical boundary test scenarios reasonably and effectively. The main idea is to optimize the behavior of critical boundary test scenarios by gradually increasing the complexity of the dynamic scenarios. The behavior of the major NOTs has a crucial impact on the complexity of the dynamic scenarios. Variations in scenario complexity can accelerate the evaluation of the performance boundaries of autonomous vehicles and identify critical boundary test scenarios.

[0177] During the adaptive adjustment of scene parameters, scene complexity is optimized. If the autonomous vehicle passes the current test, the dynamic traffic scene complexity threshold is further increased, calculated as follows:

[0178] C crit =(1)C D (it)(23)

[0179] Where β∈(0,1] is the scene complexity enhancement coefficient.

[0180] After determining the complexity threshold of the scenario, the action parameters u of the main other traffic participants are... q Optimization is performed to generate critical boundary test scenarios that meet complexity requirements and minimize deviation from the distribution of natural driving data. The behavioral trajectories of the main other traffic participants obtained from the natural driving traffic flow are represented as T. nat (t,u q,nat ), and represent the optimized behavioral trajectories of the main other traffic participants as T crit (t,u q The objective function and constraints of the proposed algorithm are shown in Equation (24).

[0181]

[0182] Where t0 represents the start time of behavior optimization by other major traffic participants, t f TTC(t) represents the duration during which major other traffic participants influence the behavior of the tested autonomous vehicle. TTC(t) represents the minimum collision time between the major other traffic participants and the nearest obstacle or other traffic participant at time t.min This represents the minimum safe collision time. The constraints in Equation (24) ensure that no other traffic participants actively collide with the tested autonomous vehicle.

[0183] The most crucial part of this technical solution is generating critical boundary test scenarios for intelligent driving testing of autonomous vehicles. Essentially, the main objective is to increase the complexity of the environment surrounding the autonomous vehicle while ensuring the generated test scenarios conform to a natural distribution. To demonstrate the effectiveness of this technical solution, this embodiment conducted tests in both natural traffic scenarios and the critical boundary test scenarios generated in this invention, and calculated the relative distance between the tested autonomous vehicle and other surrounding traffic participants, as well as the time difference (PET) for passing through conflict points. To further investigate the impact of the generated critical boundary test scenarios on autonomous vehicles, two types of autonomous vehicle models were also developed: a black-box VTD autonomous driving model and a rule-based DPA autonomous driving model. Figure 6a , 6b The data shows that for the VTD (Vehicle-Driven Defender) autonomous vehicle model, the distribution of critical boundary test scenarios is very similar to that of natural driving scenarios, but the PET (Peak Point of Interest) and relative distance are smaller, meaning that VTD autonomous vehicles are more dangerous in these scenarios. This is the same as the DPA (Driving-Action-Defender) autonomous vehicle model. Figure 6c , 6d As shown. It is also demonstrated that the VTD autonomous vehicle model is more dangerous than the DPA autonomous vehicle model because the VTD model has a smaller PET and relative distance in both natural traffic scenarios and critical boundary test scenarios. This is not surprising, as the rule-based DPA autonomous vehicle model is relatively conservative and designed to prevent collisions. Furthermore, the different accident probabilities encountered by different autonomous driving models in natural traffic scenarios and critical boundary test scenarios were compared based on performance metrics. In addition to accident events, driving comfort, driving performance, and traffic coordination were also defined, such as... Figure 6e , 6f As shown.

[0184] The following criteria are used to query these events: (a) Driving comfort: The absolute value of the acceleration of the autonomous vehicle is greater than 3 m / s². 2 (b) Driving performance: The autonomous vehicle failed to successfully navigate the intersection area within 60 seconds without a collision; (c) Traffic coordination: The autonomous vehicle turned left into straight traffic, causing the straight-going vehicles to brake suddenly (the absolute value of the background vehicle's acceleration was greater than 3 m / s²). 2 ).like Figure 6e and Figure 6fAs shown, compared to natural traffic scenarios, critical boundary test scenarios generated more driving comfort and performance issues in both types of autonomous vehicles. The impact of the two scenarios on traffic coordination was not significantly different, mainly because these algorithms focus on their own driving performance and rarely consider the impact on traffic coordination. In fact, in the simulation verification of the two autonomous vehicles, the frequency of safety, driving comfort, and driving performance issues in natural traffic scenarios was lower. All these results indicate that critical boundary test scenarios can more effectively evaluate the overall performance of autonomous vehicles compared to natural traffic scenarios.

[0185] In summary, the goal of this technical solution is to research an online method for generating critical boundary test scenarios that satisfy the distribution, diversity, and multiple interactions of natural traffic flow behaviors, in order to test the performance boundaries of autonomous vehicles. Therefore, a flexible, complex, and diverse online adaptive critical test scenario generation architecture is designed; an autonomous driving test scenario complexity model is proposed to quantify the complexity of different test scenarios; the complex interactive behaviors of autonomous vehicles, human-driven vehicles, bicycles, and pedestrians are comprehensively considered, along with their impact on the tested autonomous vehicles. Test scenarios are systematically described, generated, selected, optimized, and used, combining virtual scenarios with real traffic flow data, thereby effectively improving the flexibility, complexity, and diversity of critical boundary test scenarios.

Claims

1. A method for online generation of critical edge test scenarios for autonomous driving acceleration testing, characterized in that, Includes the following steps: S1. Generalize and generate natural driving test scenarios. The specific process is as follows: S11. Perform environmental recognition on the real-world driving environment and convert static traffic environment and weather environment into static traffic scene; S12. Perform behavior recognition on real-world driving environments to obtain a distribution model of the natural driving behavior of dynamic traffic participants; S13. Generalize the natural traffic scenario on the simulation platform and convert it into a natural driving test scenario; S2. Construct a scene complexity model, including static scene complexity and dynamic scene complexity. The scene complexity model is specifically as follows: in, and These are the static scene complexity and dynamic scene complexity Weighting coefficients; The quantification dimensions of the static scene complexity include the area of ​​the drivable area, weather visibility, and road friction coefficient; The key variables for quantifying the complexity of the dynamic scenario include the encounter angle, relative distance, and relative speed between the tested autonomous vehicle and other traffic participants. S3. Based on the generated natural driving test scenarios and the constructed scenario complexity model, evaluate and test autonomous vehicles. S4. Based on the test results obtained in step S3, perform scene adaptive adjustment, including static scene adaptive adjustment and dynamic scene adaptive adjustment, and output the key boundary test scene. Specifically, the following steps are included: S41. Based on the test results obtained in step S3, determine whether the current test scenario belongs to a safety-critical test scenario, a critical edge test scenario, or a test scenario that requires further adaptive optimization. If it belongs to a safety-critical test scenario, the generated scenario will be added to the safety-critical test scenario library; If it belongs to a critical edge test scenario, the generated scenario will be placed in the critical edge test scenario library; If the requirements of the desired scenario are not met, proceed to step S42; S42. The current test scenario is further adjusted and optimized online through the critical edge test scenario adaptive generation algorithm. The critical edge test scenario adaptive generation algorithm performs adaptive optimization of the scenario from both static and dynamic perspectives simultaneously. The adaptive optimization of static scenes is carried out from three aspects: drivable area, weather visibility, and road friction coefficient; Adaptive optimization of dynamic traffic scenarios is achieved by selecting other major adversarial traffic participants and optimizing their natural and adversarial behaviors. In the process of adaptive optimization of static and dynamic traffic scenarios, the scenario complexity model is used to quantify the complexity of the optimized scenario. Based on the scenario complexity improvement coefficient, the complexity and danger of the test scenario are adaptively improved, and finally a key test scenario library with different complexity levels, used for performance testing in different dimensions, and covering various scenarios is generated.

2. The method for online generation of critical edge test scenarios for autonomous driving acceleration testing according to claim 1, characterized in that, Specifically, step S13 involves using natural driving data from a natural driving environment to generalize parameters and generate simulation test scenarios in OpenDrive and OpenScenario formats.

3. The method for online generation of critical edge test scenarios for autonomous driving acceleration testing according to claim 1, characterized in that, The specific complexity of the static scene is as follows: in, This represents the maximum area of ​​the drivable zone. The minimum area of ​​the drivable zone. The visibility distance at which lighting and clarity are optimal. The visibility distance at which lighting and transparency are at their worst. The coefficient of friction for the road under the driest conditions. The coefficient of friction is the road friction coefficient under the most wet road conditions. , , These correspond to the drivable area. Weather visibility and road friction coefficient The weighting coefficients.

4. The method for online generation of critical edge test scenarios for autonomous driving acceleration testing according to claim 1, characterized in that, The calculation process for the complexity of the dynamic scene is as follows: First, define the effective range of potentially complex traffic participants around the autonomous vehicle. Then the complexity of the dynamic interaction pair, which consists of the autonomous vehicle under test and surrounding traffic participants, is calculated. Finally, the complexity of all interaction pairs is taken into account to calculate the complexity of the dynamic traffic scenario. The effective range of potential traffic participants around the autonomous vehicle is described by a semicircle centered on the rear axle center of the vehicle, with a radius of: in, The radius of the area affected by the autonomous vehicle. The radius of the safe zone for the tested autonomous vehicle. The length of the intersection area. The initial speed of the autonomous vehicle. The response lag time for autonomous vehicles. and These represent the maximum acceleration and minimum deceleration of an autonomous vehicle, respectively. The complexity of the dynamic interaction pair is: in, For the complexity among dynamic traffic participants, The meeting angle between dynamic traffic participants. The relative distance between dynamic traffic participants. The relative speeds between dynamic traffic participants. For meeting angle Complexity of dynamic interaction The relationship between them Relative distance Complexity of dynamic interaction The relationship between them relative velocity Complexity of dynamic interaction The relationship between them The standardized relative distance. The relative distance between different traffic participants The relative distance in the least complex case is equal to the radius of the autonomous vehicle's influence area. , The relative distance in the most complex case is zero, indicating that the two cars are infinitely close. The standardized relative speed, The relative speed between different traffic participants, The relative velocity under the least complex condition is equal to the divergent relative velocity of 3.5 m / s. The relative velocity under the most complex condition is equal to the convergent relative velocity of 3.5 m / s; The complexity of the dynamic traffic scenario is: in, For dynamic scene complexity, In order to time step t The complexity of the integrated dynamic scene before smoothing N To influence the number of vehicle pairs within the affected area, For traffic participants The impact, For interactive collections The complexity, This indicates the length of the sliding window.

5. The method for online generation of critical edge test scenarios for autonomous driving acceleration testing according to claim 1, characterized in that, The specific process of static scene adaptive optimization in step S42 is as follows: First, based on the geographical coordinates of the actual traffic environment. and distance threshold Extract road network data from OpenStreetMapOSM; After data extraction, a static traffic scene is generated and post-processed. Then, it is analyzed using parameter vectors. The static traffic scene is parameterized, where the parameter vector is defined as follows: = , and Representing the position and size of a static obstacle, respectively, using the initial state vector. Initialize a static traffic scenario, then test autonomous vehicles based on the generated scenario and obtain trajectory data. and the corresponding test results; Finally, the area of ​​the drivable region is calculated using the reachable state set: Indicates the time of the test of the autonomous vehicle The state of motion, Indicates the vehicle at time The action, Vehicle status Projected onto a two-dimensional position domain, This indicates that it does not contain static obstacles. Two-dimensional space; To improve the criticality of static scenes, the parameter vector is optimized. To obtain the critical static traffic scenario For the This intelligent driving test: st Use the Static scene parameters calculated in this intelligent driving test and scalar To quantify and improve the complexity of static scenes: .

6. The method for online generation of critical edge test scenarios for autonomous driving acceleration testing according to claim 1, characterized in that, The specific process of dynamic scene adaptive optimization in step S42 is as follows: First, collect natural driving data. ; Then, the behavior of different traffic participants in the natural driving data is identified and analyzed, and the motion state of the traffic participants is obtained. and actions To generate a dynamic traffic scenario that meets the requirements, the state frequencies of different traffic participants under the corresponding states are calculated. and frequency of action For dynamic scene parameters Perform initialization and integrate with the static traffic scenario. Combined to generate the initial scene; Subsequently, the autonomous vehicle is tested according to the generated test scenarios. In order to generate critical boundary test scenarios efficiently and reasonably, it is necessary to identify the main other traffic participants who have a significant impact on the autonomous vehicle under test from the background traffic participants. The behavior of the main other traffic participants is dynamically adjusted using dynamic scenario generation and adaptive optimization algorithms, and the behavior of other background traffic participants is controlled using a model based on natural driving environment data, so that the generated critical boundary test scenarios conform to the distribution of natural driving data as much as possible. To identify the major other traffic participants (OOPs) that pose the greatest challenge to the tested autonomous vehicle, the concept of the autonomous vehicle's influence zone is used to traverse and select OOPs within that zone. First, natural traffic flow data in the test environment is analyzed, and the state frequencies of different OOPs are calculated. and the frequency of actions in different states In time step The state and action are represented as follows: in, and They represent time. The state and actions of the autonomous vehicles being tested. Indicates the first Background traffic participants in time state, , and These represent the number of background vehicles, pedestrians, and bicycles within the affected area, respectively. Indicates the first Background traffic participants in time The action; Then use dynamic scene interaction to reduce complexity To measure the threat level posed by surrounding traffic participants to the tested autonomous vehicle, state-action pairings The interaction complexity for: in Indicates the vehicle is in a certain state. The following action The probability of; Based on the obtained action frequency of background traffic participants and interaction complexity The critical values ​​of the corresponding background traffic participants are calculated comprehensively. : Then, the threshold value for each background traffic participant is calculated as the sum of the action threshold values ​​for all actions of the corresponding background traffic participant: Based on the calculated threshold value for each background traffic participant, the background vehicles, pedestrians, and bicycles with the largest threshold values ​​are selected as the main other traffic participants. The identification of the main other vehicles, pedestrians, and bicycles is as follows: in, , and These respectively represent the numbers of other main vehicles, pedestrians, and bicycles. , and These represent the sets of background vehicles, pedestrians, and bicycles within the affected area. , and These represent predetermined critical thresholds for identifying major other vehicles, pedestrians, and bicycles, specifically adversarial vehicles, adversarial pedestrians, adversarial bicycles, or any combination thereof, to maximize the efficiency of generating critical boundary test scenarios. After identifying the major other traffic participants, the complexity of the dynamic scenario is gradually increased by optimizing the behavior of the critical boundary test scenario. The behavior of the major other traffic participants has a crucial impact on the complexity of the dynamic scenario. Changes in scenario complexity can accelerate the evaluation of the performance boundaries of autonomous vehicles and identify critical boundary test scenarios. Scenario complexity is optimized during the adaptive adjustment of scenario parameters. If the autonomous vehicle passes the current test, the complexity threshold of the dynamic traffic scenario is further increased, and its calculation is as follows: in, It is the scene complexity enhancement factor; After determining the complexity threshold of the scenario, the action parameters of the other major traffic participants were analyzed. Optimization is performed to generate critical boundary test scenarios that meet complexity requirements and minimize deviation from the distribution of natural driving data. The behavioral trajectories of key other traffic participants obtained from natural driving traffic flow are represented as follows: And represent the optimized behavioral trajectories of other major traffic participants as The corresponding objective function and constraints are shown below: st in Indicates the start time of behavior optimization for major other traffic participants. This indicates the duration of the influence of other major traffic participants on the behavior of the tested autonomous vehicle. Indicates the time between the main other traffic participants and the nearest obstacle or traffic participant. The minimum collision time, This represents the minimum safe collision time, ensuring, through the aforementioned constraints, that no other road users in the vicinity actively collide with the tested autonomous vehicle.

Images