Method, system, device and medium for generating test scenarios for autonomous driving vehicles

Abstract

The present invention discloses a method, system, device and medium for generating a test scenario for an autonomous driving vehicle, and relates to the technical field of dangerous scenario generation. The method comprises: generating an initial scenario based on a lane coordinate system; constructing a one-way single-lane following scenario; specifically comprising: creating roads and vehicles based on the initial scenario; the vehicles comprise an autonomous driving vehicle and an intelligent vehicle based on an IDM model; constructing a one-way two-lane cutting scenario; specifically comprising: creating a two-lane and vehicles based on the initial scenario; the vehicles comprise an autonomous driving vehicle and an intelligent vehicle based on an IDM model; generating dangerous scenarios for autonomous driving vehicle testing based on the constructed one-way single-lane following scenario and one-way two-lane cutting scenario; the method efficiently and accurately generates dangerous scenarios for autonomous driving systems to accelerate testing and verifying the performance and safety of autonomous driving systems in complex and high-risk environments.

Description

Technical Field

[0001] The present invention relates to the technical field of dangerous scenario generation, and in particular to a method, system, device and medium for generating a test scenario for an autonomous driving vehicle. Background Art

[0002] Autonomous vehicles are generally considered to be a promising means of transportation to reduce traffic accidents, improve traffic efficiency, and reduce fuel consumption and pollution emissions in the future. However, there are still many challenges before autonomous vehicles can be widely used. One of the important challenges is the testing and evaluation of autonomous driving systems. Scenario-based testing methods are considered to be a promising method for testing and evaluating autonomous driving systems.

[0003] The scenario-based test method verifies whether the autonomous driving system can safely perform its expected functions in different traffic scenarios by defining different traffic participants (such as motor vehicles, pedestrians, bicycles, etc.) and their movement trajectories, weather factors, road topology, etc. in the driving environment. At present, the research on virtual simulation test scenario generation is mainly divided into two directions. One is to solve the problem of scenario diversification, mainly using data-driven or expert knowledge-based methods to establish a scenario library with high coverage of real traffic scenarios; the other is the boundary scenario generation method for solving the problem of accelerated testing. To establish a scenario library based on the data-driven method, it is first necessary to obtain real driving data or traffic accident data, identify scenarios from the data, extract implicit information related to the scenarios, then parameterize the scenario information and model it, and finally use combination generation, probability density sampling and other methods to generate a new test scenario library to achieve the purpose of covering a large number of natural driving scenarios. At present, the research on boundary scenario generation is mainly focused on boundary scenario generation based on large deviation theory and boundary scenario optimization generation. Boundary scene generation methods based on large deviation theory usually use large deviation theories such as importance sampling and cross entropy to sample boundary scenes from existing scene libraries. Such methods are mostly used in simple scenes involving a limited number of vehicles and very short duration, without considering the interaction process between the driving system under test and other dynamic traffic elements and the spatiotemporal continuity of the driving environment. Boundary scene optimization generation describes the interaction of traffic participants in the scene as an optimization problem. Through optimization iteration, the deviation between the scene collision risk and the expected collision risk during the test is continuously reduced, and finally a boundary scene that can test the vehicle's extreme capabilities is obtained.

[0004] However, the current method for generating test scenarios for autonomous vehicles does not take into account the real-time interaction between the vehicle under test and the environment and the spatiotemporal continuity of the generated scenarios, resulting in the inability to accurately test the safety of autonomous vehicles in high-risk environments. Summary of the invention

[0005] In view of the shortcomings of the prior art in that the real-time interaction between the tested vehicle and the environment and the spatiotemporal continuity of the generated scenes are not considered, the present invention proposes a method, system, equipment and medium for generating an autonomous driving vehicle test scene, which solves the problems of the prior art by constructing a one-way single-lane following scene and a one-way two-lane cutting scene.

[0006] A method for generating an autonomous driving vehicle test scenario comprises the following steps:

[0007] Build the initial vehicle driving test scenario based on the lane coordinate system;

[0008] Build a one-way single lane following scenario; specifically including: creating roads and vehicles based on the initial vehicle driving test scenario; the vehicle includes an autonomous driving vehicle V based on the IDM model 11 and the intelligent vehicle V 12 ; Among them, V 11 Follow V 12 ; V 11 The influence of acceleration rate of change Jerk and collision time TTC on the one-way single lane following scene is used as the reward function of the one-way single lane following scene; among them, the larger the Jerk and the smaller the TTC, the higher the risk factor of the scene;

[0009] Build a one-way two-lane cutting scene; specifically include: creating a two-lane and a vehicle based on the initial vehicle driving test scene; the vehicle includes an autonomous driving vehicle V based on the IDM model 21 and the intelligent vehicle V 22 , and V 22 At V 21 rear; when V 21 A V is detected in the current lane. 22 Close to V 21 When V 21 Select to change lanes; when V 21 Stay away from V 22 When V 21 Choose to keep the current lane and slow down; by introducing the responsibility-sensitive safety model RSS, the judgment of V 21 With V 22 The actual distance between them and the minimum longitudinal safety distance of RSS are used as the reward function of the one-way two-lane cutting scene; where, for any moment, if V 21 With V 22 The actual distance between d m in Less than the minimum longitudinal safety distance of RSS d m in _ RSS , consider the scene at this moment as a dangerous scene:

[0010] Based on the constructed one-way single-lane following scenario and one-way two-lane cutting scenario, dangerous scenarios for autonomous driving vehicle testing are generated.

[0011] Furthermore, the initial vehicle driving test scenario is constructed based on the lane coordinate system, specifically including the following steps:

[0012] Using the spatial logic method, the lane-based coordinate system is used to generate the initial positions of all objects according to the rationality rule. Assuming that there are N elements in the scene, N∈N+, the position of the element is represented by p, p=(p x , p y ) T ∈R 2 , where p x Indicates the projection position of the element on the x-axis in the lane coordinate system, p y Represents the projection position of the element on the y-axis in the lane coordinate system, so the two elements item i and item j The distance between items is expressed as dist(item i , item j )=|item i .p-item j .p| 2 , i, j∈{1, 2, ..., N}; the set of all elements in the scene is represented by X=(item 1 , item 2 ,...,item N ) T , T is the matrix transpose; item i .p represents the element item i Projected position in the lane coordinate system; item j .p represents the element item j Projected position in the lane coordinate system;

[0013] The spatial relationship of elements based on projection on a two-dimensional plane is front, back, left, and right, which are represented by FrontOf, BackOf, LeftOf, and RightOf respectively. Then the two elements item i and item j The position is expressed as:

[0014] FrontOf(item i , item j ):=proj y (item i .p)>proj y (item j .p)

[0015]

[0016] LeftOf(item i , item j ):=proj x (item i .p) <proj x (item j .p)

[0017]

[0018] where proj x Indicates x-axis projection, proj y Indicates y-axis projection. Indicates negation;

[0019] The spatial relationship on the scene element set X is quantitatively expressed as The spatial relationship of the elements can be quantitatively expressed as:

[0020]

[0021]

[0022] Where M and ε are a large number and a small number defined by the user, respectively.

[0023] Furthermore, the V 11 The influence of acceleration rate of change Jerk and collision time TTC on the one-way single lane following scene is used as the reward function of the one-way single lane following scene. The reward function of the following scene is expressed as:

[0024]

[0025] Among them, XL (t) represents the position of the preceding vehicle at time t, X F(t) represents the position of the following car at time t, D L Indicates the length of the vehicle in front, V F(t) -V L(t) It represents the speed difference between the rear vehicle and the front vehicle at time t. Jerk represents the rate of change of vehicle acceleration over time. The larger the Jerk is, the higher the risk factor of the scene is. TTC represents the collision time. The smaller the TTC is, the greater the probability of a vehicle collision is. When TTC is less than or equal to 2, the two objects are in a critical state of collision.

[0026] Furthermore, when the absolute value of the Jerk is greater than or equal to the absolute value of 4, the rate of change of the vehicle over time reaches a critical state; when the TTC is less than or equal to 2, the two vehicles are in a critical state of collision.

[0027] Furthermore, by introducing the RSS responsibility-sensitive security model, the judgment of V 21 With V 22 The actual distance between the two vehicles and the minimum longitudinal safety distance of RSS are used as the reward function of the cutting scene, and its reward function is expressed as:

[0028] f(v cut_in ) = d min_RSS -d min

[0029]

[0030] d min =x r -x f

[0031] Among them, v r , v f Respectively represent the longitudinal speed of the rear vehicle and the front vehicle; ρ represents the response time; a max,accel Indicates the maximum acceleration; a min,brake， a max,brake Represent the minimum deceleration and the maximum deceleration respectively; by introducing the RSS responsibility-sensitive safety model, for any time, if V 21 With V 22 The actual distance d min Less than the minimum longitudinal safety distance d of RSS min_RSS , regard the scene at this moment as a dangerous scene.

[0032] The present invention also includes a system for generating a test scenario for an autonomous driving vehicle, comprising:

[0033] The initial scene generation module is used to build the initial vehicle driving test scene based on the lane coordinate system;

[0034] The following scenario building module is used to build a one-way single-lane following scenario; specifically, it includes: creating a road and a vehicle based on the initial vehicle driving test scenario; the vehicle includes an autonomous driving vehicle V based on the IDM model 11 and the intelligent vehicle V 12 ; Among them, V 11 Follow V 12 ; V 11 The influence of acceleration rate of change Jerk and collision time TTC on the one-way single lane following scene is used as the reward function of the one-way single lane following scene; among them, the larger the Jerk and the smaller the TTC, the higher the risk factor of the scene;

[0035] The vehicle cutting scenario building module is used to build a one-way two-lane vehicle cutting scenario; specifically, it includes: creating a two-lane and a vehicle based on the initial vehicle driving test scenario; the vehicle includes an autonomous driving vehicle V based on the IDM model 21 and the intelligent vehicle V 22 , and V 22 At V 21 rear; when V 21 A V is detected in the current lane. 22 Close to V 21 When V 21 Select to change lanes; when V 21 Stay away from V 22 When V 21 Choose to keep the current lane and slow down; by introducing the responsibility-sensitive safety model RSS, the judgment of V 21 With V 22 The actual distance between them and the minimum longitudinal safety distance of RSS are used as the reward function of the one-way two-lane cutting scene; where, for any moment, if V 21 With V 22 The actual distance d min Less than the minimum longitudinal safety distance d of RSS min_RSS , regard the scene at this moment as a dangerous scene;

[0036] The dangerous scenario generation module is used to generate dangerous scenarios for autonomous driving vehicle testing based on the constructed one-way single-lane following scenario and one-way two-lane cutting scenario.

[0037] The present invention also includes a computer device for generating an autonomous driving vehicle test scenario, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor implements the steps of the method for generating an autonomous driving vehicle test scenario when executing the computer program.

[0038] Furthermore, a readable storage medium is provided, wherein the readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, they are used to execute the steps of the method for generating a test scenario for an autonomous driving vehicle.

[0039] The present invention provides a method, system, device and medium for generating a test scenario for an autonomous driving vehicle, which has the following beneficial effects:

[0040] The present invention constructs a one-way single-lane following scenario by setting a driving vehicle and a following intelligent vehicle, and uses the influence of the acceleration change rate and collision time of the autonomous driving vehicle on the one-way single-lane following scenario as its reward function; at the same time, a one-way two-lane cutting scenario is constructed by setting the intelligent vehicle to be behind the autonomous driving vehicle, and using the actual distance between the autonomous driving vehicle and the intelligent vehicle and the RSS minimum longitudinal safety distance as the reward function; by considering the real-time interaction problem between the tested vehicle and the environment and the spatiotemporal continuity problem of the generated scenario, a one-way single-lane following scenario and a one-way two-lane cutting scenario are set to efficiently and accurately generate dangerous scenarios of the autonomous driving system, so as to accelerate the testing and verification of the performance and safety of the autonomous driving system in complex and high-risk environments; the method effectively improves the reliability and safety of autonomous driving vehicles in complex and dangerous situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of a method for generating dangerous scenarios for an autonomous driving vehicle in an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of overtaking in an embodiment of the present invention;

[0043] Figure 3 It is a reinforcement learning training graph in an embodiment of the present invention;

[0044] Figure 4 1. A diagram of a car-following and car-cutting scenario in an embodiment of the present invention;

[0045] Figure 5 Schematic diagram of the criticality of each moment in the car-following scenario in an embodiment of the present invention;

[0046] Figure 6 Schematic diagram of the criticality of each moment in the car-cutting scenario in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0048] The present invention proposes a method for generating dangerous scenarios for an autonomous driving vehicle based on reinforcement learning, which specifically includes the following steps:

[0049] S1. Conduct an in-depth analysis of the autonomous driving system test tasks and scenario elements, and divide the scenarios into three parts: road structure, dynamic elements, and meteorological elements.

[0050] Triggered by the test task, the test scenarios are defined at three levels: functional scenario, logical scenario and specific scenario according to different levels of abstraction. The specific solutions are as follows: Figure 1 As shown. Functional scenarios are the most abstract in scenario descriptions. They use natural language to abstractly describe the entities involved in the test task and their behaviors and relationships, helping people to have a consistent understanding of the meaning of the scenarios. Logical scenarios constrain the boundaries of scenario parameters based on functional scenarios. Parameter ranges can be specified using probability distributions (Gaussian distribution, uniform distribution, etc.), related functions, etc., or using numerical values ​​or numerical conditions (for example, the speed of the overtaking vehicle must be greater than the speed of the overtaken vehicle). Specific scenarios are instantiations of logical scenarios, and the specific values ​​of the parameters are specified within the parameter range defined by the logical scenario. For logical scenarios with continuous value ranges, any number of specific scenarios can be derived.

[0051] S2. Describe the positional relationship of scene elements, adopt spatial logic methods, use the lane-based coordinate system, and generate the initial positions of all objects according to rationality rules to ensure the rationality of the initial positions.

[0052] Assume there are N elements in the scene, N∈N x , the position of the element is represented by p, p = (p x , p y ) T ∈R 2 , where p x Indicates the projection position of the element on the x-axis in the lane coordinate system, p y Represents the projection position of the element on the y-axis in the lane coordinate system, so the two elements item i and item j The distance between them can be expressed as dist(item i , item j )=|item i .p-item j .p| 2 , i, j∈{1, 2, ..., N}. The set of all elements in the scene is represented by X=(item 1 , item 2 ,...,item N ) T Therefore, according to the given coordinate system, the present invention can define the positional relationship of the elements in the scene according to spatial logic.

[0053] The spatial relationship of elements based on projection on a two-dimensional plane is front, back, left, and right, which are represented by FrontOf, BackOf, LeftOf, and RightOf respectively. Then the two elements item i and item j The position is defined as follows:

[0054] FrontOf(item i , item j ):=proj y (item i .p)>proj y (item j .p)

[0055]

[0056] LeftOf(item i , item j ):=proj x (item i .p) <proj x (item j .p)

[0057]

[0058] where proj x Indicates x-axis projection, proj y Indicates y-axis projection. Indicates negation.

[0059] by Figure 2 As an example, the vehicle V 0 In the environment vehicle V 1 The right side of V 0 , V 1 ), the tested vehicle V 0 In the environment vehicle V 1 In front of, its positional relationship can be described as FrontOf(V 0 , V 1 ), the tested vehicle V 0 In the environment vehicle V 2 The back of the relationship is described as BackOf(V 0 ,V 2 ). You can also use distance-based spatial relationships to define the far and near relationships between two elements, represented by FarFrom and CloseTo, respectively, as defined below:

[0060] FarFrom(item i , item j ): = dist(item i , item j )≥M

[0061] CloseTo(item i , item j): = dist(item i , item j )≤ε

[0062] The completion status of the test task and the scene hazard assessment also need to be expressed by quantitative positional relationships. The present invention adopts a spatial logic method, uses a lane-based coordinate system, and generates the initial positions of all objects according to rationality rules to ensure the rationality of the initial positions. Then, during the scene evolution process, spatial logic is used to constrain the scene search space to generate a structured effective scene.

[0063] Where M and ε They are a large number and a decimal number defined by the user respectively.

[0064] Combining spatial relationships with Boolean operators can form more complex positional relationship descriptions. For example, complex positional relationships: Where T represents Boolean true, rel 1 and rel 2 Represent univariate and bivariate spatial relationships respectively. The present invention quantitatively represents the spatial relationship on the scene element set x as Then the spatial relationship of the elements can be quantitatively expressed as:

[0065]

[0066] S3. In the virtual environment of the highway, design the following scenario and the cutting scenario and the corresponding reward functions respectively.

[0067] Design a one-way single-lane following scenario based on the open source environment highway-env: First, create roads and vehicles through the create_road and create_vehicles functions respectively. The vehicles include an autonomous driving vehicle V based on the IDM (Intelligent Driving Model) model. 11 , an intelligent vehicle V 12 , V 11 Follow V 12 Design a one-way two-lane following scenario based on the open source environment highway-env: First, create roads and vehicles through the create_road and create_vehicles functions respectively. The vehicles include an autonomous driving vehicle V based on the IDM (Intelligent Driving Model) model. 21 , and an intelligent vehicle V 22 .

[0068] S3.1. The reward function for the following scenario is defined as:

[0069]

[0070] Where X L(t) represents the position of the preceding vehicle at time t, X F(t) represents the position of the following car at time t, D L Indicates the length of the vehicle in front, V F(t) -V L(t) Indicates the speed difference between the rear vehicle and the front vehicle at time t. In the following scenario, the rear vehicle will passively adjust its own state according to the state of the front vehicle, and the acceleration and braking of the front vehicle will have an impact on the rear vehicle. Jerk is the rate of change of an object's acceleration over time, which is used to describe the smoothness and comfort of motion. The larger the Jerk, the more drastic the change in motion and the higher the risk factor of the scene. When the absolute value of Jerk is greater than or equal to 4, the driver and the occupants of the car will clearly feel the bumps; TTC: Collision time, a time measurement method used to evaluate the possible collision between two moving objects. The smaller the TTC, the greater the probability of a collision between the objects; when TTC is less than or equal to 2, the two objects are in a critical state of collision. Therefore, how to control V 11 Jerk and TTC are in a dangerous range, which is one of the important rewards in the reinforcement learning process.

[0071] S3.2, the reward function for the car-cutting scenario is defined as:

[0072] f(v cut_in ) = d min_RSS -d min

[0073]

[0074] d min =x r -x f

[0075] v r , v f Respectively represent the longitudinal speed of the rear vehicle and the front vehicle; ρ represents the response time; a max,accel Indicates the maximum acceleration; a min,brake， a max,brake Represents the minimum deceleration and maximum deceleration respectively.

[0076] In the car cutting scene, the biggest difficulty is to make V 22 In order to ensure the rationality of the lane change scenario, the present invention applies a very mature mobile lane change model to V 22 ; In order to ensure the danger of the scene, the present invention introduces the RSS (Responsibility-Sensitive Safety Model) responsibility-sensitive safety model. For any time, if V21 With V 22 The actual distance d min Less than the minimum longitudinal safety distance d of RSS min_RSS , the scene at this moment is regarded as a dangerous scene, so for the car cutting scene, how to ensure d min <d min_RSS It is one of the important rewards in the reinforcement learning process.

[0077] S4. Design other elements of reinforcement learning, including state space and action space, and call reinforcement learning based on the DQN algorithm for training. The intelligent body will continuously increase the danger level of the scene according to the designed goals.

[0078] Reinforcement learning algorithms rely on direct interaction between the agent and the environment to learn optimization strategies through trial and error, such as Figure 3 Specifically, the agent obtains the current state information s from the environment and generates corresponding actions according to the preset strategy. After executing the action, the environment will feedback the next state S t+1 And the reward obtained by the current action.

[0079] The definition of state space is: a huge state space S covering global information is jointly constructed in the complex traffic network. At the same time, each intersection is an independent intelligent entity in the traffic network, and its local observation space s is a microcosm of the global state space S. This local-global connection makes the intelligent vehicle V 12 、V 22 Not only can it accurately perceive its own situation, but it can also accurately control the status of the autonomous vehicle V 11 、V 21 The present invention responds to the state of the intelligent vehicle V 12 、V 22 The state at time t is defined as S t , which comprehensively reflects the intelligent vehicle V 11 、V 21 Real-time status information. For example, the vehicle status information includes the current speed and location, and whether a collision has occurred.

[0080] The action space is defined as follows: Taking the car-cutting scenario as an example: For an intelligent vehicle using a reinforcement learning algorithm, there are two actions that can be selected in each decision: "switch lanes" and "keep the current lane". When the intelligent vehicle detects that there is an obstacle in the current lane or the vehicle behind is moving at a high speed, the intelligent vehicle should choose "switch lanes" to increase the danger level of the scenario. When there is no obstacle in the current lane and the speed of the vehicle behind is appropriate, the intelligent vehicle may need to choose "switch lanes" and slow down to increase the danger level of the scenario, where the intelligent vehicle and the autonomous vehicle are in different lanes. When the distance between the intelligent vehicle and the vehicle behind is far, the intelligent vehicle should choose "keep the current lane" and slow down to reduce the distance between the intelligent vehicle and the vehicle behind. Therefore, the action space of the intelligent vehicle is defined as the following two actions: "switch lanes" and "keep the current lane". Through the combination of these two action options, the intelligent vehicle can make decisions based on real-time traffic conditions to increase the danger level of the scenario.

[0081] S5. After training the following and cutting scenarios in the highway environment, save the reinforcement learning model so that it can be applied to the autonomous driving test steps and used with the autonomous driving vehicle.

[0082] The present invention proposes an embodiment:

[0083] (1) Environment setting: The environment used in this invention is based on the open source simulation environment highway-env. Figure 4 As shown in a, the following lane is 1000 meters long, a one-way lane, and vehicles travel from left to right; the cutting scene is Figure 4 b is 1000 meters long, with two lanes in one direction, and vehicles travel from left to right.

[0084] (2) Evaluation index: The algorithm evaluation index used in the present invention includes: a. Average vehicle speed: the average speed of all vehicles on the lane; b. Number of collisions: V 11 and V 21 From the beginning to the end of the simulation, the number of collisions in 10 simulations; c. Vehicle distance: In the simulation test of 10 simulations, V 11 and V 12 、V 21 and V 22 d, scenario danger level: for the following scenario: calculate the weighted sum of Jerk and 1 / TTC in 10 simulation tests to reflect the danger level of the scenario; for the cutting scenario: calculate the minimum longitudinal safety distance d of RSS in 10 simulation tests min_RSS With V 21 and V 22 Longitudinal distance d lon The difference between them reflects the danger level of the scene.

[0085] (3) Experimental process: The present invention constructs a method for generating dangerous test scenarios for autonomous vehicles, which provides dangerous test scenarios for autonomous vehicles. The basic scenarios are built based on the open source simulation environment highway-env, and the Python programming language is used to implement the construction of the following scenario and the cutting scenario. Pygame is called to implement the simulation operation of the scenario. The feasibility and danger of the scenario are realized by using the IDM model, the Mobile model and the RSS responsibility-sensitive safety model, and the simulation experiment is carried out in this environment by using reinforcement learning based on the DQN algorithm. During the experiment, the intelligent vehicle V 12 and V 22 With autonomous vehicles V 11 and V 21 The current state data is obtained through interaction. Reinforcement learning obtains the state data of the next moment based on the reward and the current state data, and finally completes the training. The trained model is imported into highway-env to complete the simulation.

[0086] The present invention uses a reinforcement learning-based method to consider the rationality and degree of danger of the scenario and generates a dangerous scenario that meets the definition of the academic community. To a certain extent, it solves the problems of insufficient scenario diversity, long testing time, high cost and inability to fully cover extreme and rare dangerous scenarios in existing testing methods.

[0087] The present invention constructs a method for generating dangerous test scenarios for autonomous driving, which provides dangerous test scenarios for autonomous driving vehicles to achieve the effect of shortening the test cycle. First, the basic scenarios are built, and typical test scenarios are designed: car-following scenarios and car-cutting scenarios. The RSS responsibility-sensitive safety model, TTC, and Jerk are used to constrain the rationality and danger of the scenarios. In the 40-second car-following scenario simulation of the present invention, the dangerous scenario moments account for about 36.3% of the total time. The results are as follows: Figure 5 As shown in the figure; in the 30s car cutting scenario simulation, the dangerous scenario moment accounts for about 51.6% of the total time. Figure 6 shown.

[0088] Based on the same inventive concept, the present invention also proposes a system for generating a test scenario for an autonomous driving vehicle, comprising:

[0089] The initial scene generation module is used to generate the initial scene based on the lane coordinate system.

[0090] The following scenario building module is used to build the following scenario; specifically, it includes: creating roads and vehicles based on the initial scenario; the vehicle includes an autonomous driving vehicle V based on the IDM model 11 and the intelligent vehicle V 12 ; Among them, V 11 Follow V12 ; V 11 The influence of Jerk and TTC on the following scenario is used as its reward function; among them, the larger the Jerk and the smaller the TTC, the higher the risk factor of the scenario.

[0091] The vehicle cutting scenario building module is used to build the vehicle cutting scenario; specifically, it includes: creating a dual lane and vehicles based on the initial scenario; the vehicle includes an autonomous driving vehicle V based on the IDM model 21 and the intelligent vehicle V 22 , and V 22 At V 21 rear; when V 21 A V is detected in the current lane. 22 Close to V 21 When V 21 Select to change lanes; when V 21 Stay away from V 22 When V 21 Choose to keep the current lane and slow down; by introducing the RSS responsibility-sensitive safety model, the judgment of V 21 With V 22 The actual distance and the minimum longitudinal safety distance of RSS are used as the reward function of the cutting scene; where, for any moment, if V 21 With V 22 The actual distance d min Less than the RSSS minimum longitudinal safety distance d min_RSS , regard the scene at this moment as a dangerous scene.

[0092] The dangerous scenario generation module is used to generate dangerous scenarios for autonomous driving vehicle testing based on car-following scenarios and car-cutting scenarios.

[0093] Based on the same inventive concept, the present invention also proposes a computer device for generating an autonomous driving vehicle test scenario, comprising: a memory, a processor, and a computer program stored in the memory, and when the processor executes the computer program, the steps of the method for generating an autonomous driving vehicle test scenario are implemented.

[0094] Based on the same inventive concept, the present invention also proposes a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of a method for generating an autonomous driving vehicle test scenario.

[0095] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the field of the present invention who makes equivalent substitutions or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for generating a test scenario for an autonomous driving vehicle, characterized in that: The following steps are involved: Build the initial vehicle driving test scenario based on the lane coordinate system; Build a one-way bicycle lane following scenario; Specifically including: creating roads and vehicles based on the initial vehicle driving test scenario; the vehicle includes an autonomous driving vehicle V based on the IDM model; 11 and the intelligent vehicle V 12 ; Among them, V 11 Follow V 12 ; V 11 The influence of acceleration rate of change Jerk and collision time TTC on the one-way single lane following scene is used as the reward function of the one-way single lane following scene; among them, the larger the Jerk and the smaller the TTC, the higher the risk factor of the scene; Build a one-way two-lane cutting scene; specifically include: creating a two-lane and a vehicle based on the initial vehicle driving test scene; the vehicle includes an autonomous driving vehicle V based on the IDM model 21 and the intelligent vehicle V 22 , and V 22 At V 21 rear; when V 21 A V is detected in the current lane. 22 Close to V 21 When V 21 Select to change lanes; when V 21 Stay away from V 22 When V 21 Choose to keep the current lane and slow down; by introducing the responsibility-sensitive safety model RSS, the judgment of V 21 With V 22 The actual distance between them and the minimum longitudinal safety distance of RSS are used as the reward function of the one-way two-lane cutting scene; where, for any moment, if V 21 With V 22 The actual distance d min Less than the minimum longitudinal safety distance d of RSS min_RSS , regard the scene at this moment as a dangerous scene; Based on the constructed one-way single-lane following scenario and one-way two-lane cutting scenario, dangerous scenarios for autonomous driving vehicle testing are generated.

2. The method for generating a test scenario for an autonomous driving vehicle according to claim 1, characterized in that: The initial vehicle driving test scenario is established based on the lane coordinate system, specifically including the following steps: Using the spatial logic method, the lane-based coordinate system is used to generate the initial positions of all objects according to the rationality rule; assuming that there are N elements in the scene, N∈N + , the position of the element is represented by p, p = (p x , p y ) T ∈R 2 , where p x Indicates the projection position of the element on the x-axis in the lane coordinate system, p y Represents the projection position of the element on the y-axis in the lane coordinate system, then the two elements item i and item j The distance between items is expressed as dist(item i , item j )=|item i .p-item j .p|2, i, j∈{1, 2, ..., N}; the set of all elements in the scene is represented by X=(item1, item2, ..., item N ) T , T is the matrix transpose; item i .p represents the element item i Projected position in the lane coordinate system; item j .p represents the element item j Projected position in the lane coordinate system; The spatial relationship of elements based on projection on a two-dimensional plane is front, back, left, and right, which are represented by FrontOf, BackOf, LeftOf, and RightOf respectively. Then the two elements item i and item j The position is expressed as: FrontOf(item i ,item j ):=proj y (item i .p)>proj y (item j .p) LeftOf(item i ,item j ):=proj x (item i .p)<proj x (item j .p) where proj x Indicates x-axis projection, proj y Indicates y-axis projection. Indicates negation; The spatial relationship on the scene element set X is quantitatively expressed as Then the spatial relationship of the elements can be quantitatively expressed as: Where M and ε are a large number and a small number defined by the user, respectively.

3. The method for generating a test scenario for an autonomous driving vehicle according to claim 1, characterized in that: The V 11 The influence of acceleration rate of change Jerk and collision time TTC on the one-way single lane following scene is used as the reward function of the one-way single lane following scene. The reward function of the following scene is expressed as: Among them, X L(t) represents the position of the preceding vehicle at time t, X F(t) represents the position of the following car at time t, D L Indicates the length of the vehicle in front, V F(t) -V L(t) It represents the speed difference between the rear vehicle and the front vehicle at time t. Jerk represents the rate of change of vehicle acceleration over time. The larger the Jerk is, the higher the risk factor of the scene is. TTC represents the collision time. The smaller the TTC is, the greater the probability of a vehicle collision is. When TTC is less than or equal to 2, the two objects are in a critical state of collision.

4. The method for generating a test scenario for an autonomous driving vehicle according to claim 3, characterized in that: It also includes that when the absolute value of the Jerk is greater than or equal to the absolute value of 4, the rate of change of the vehicle over time reaches a critical state; when the TTC is less than or equal to 2, the two vehicles are in a critical state of collision.

5. The method for generating a test scenario for an autonomous driving vehicle according to claim 1, characterized in that: By introducing the RSS responsibility-sensitive security model, the V 21 With V 22 The actual distance between the two vehicles and the minimum longitudinal safety distance of RSS are used as the reward function of the cutting scene, and its reward function is expressed as: f(v cut_in )=d min_RSS -d min d min =x r -x f Among them, v r , v f Respectively represent the longitudinal speed of the rear vehicle and the front vehicle; ρ represents the response time; a max，accel Indicates the maximum acceleration; a min，brake , a max，brake Represent the minimum deceleration and maximum deceleration respectively; by introducing the RSS responsibility-sensitive safety model, for any time, if V 21 With V 22 The actual distance d min Less than the minimum longitudinal safety distance d of RSS min_RSS , regard the scene at this moment as a dangerous scene.

6. A system for generating test scenarios for autonomous driving vehicles, characterized in that: include: The initial scene generation module is used to build the initial vehicle driving test scene based on the lane coordinate system; The car-following scenario building module is used to build a one-way single lane car-following scenario; Specifically including: creating roads and vehicles based on the initial vehicle driving test scenario; the vehicle includes an autonomous driving vehicle V based on the IDM model; 11 and the intelligent vehicle V 12 ; Among them, V 11 Follow V 12 ; V 11 The influence of acceleration rate of change Jerk and collision time TTC on the one-way single lane following scene is used as the reward function of the one-way single lane following scene; among them, the larger the Jerk and the smaller the TTC, the higher the risk factor of the scene; The vehicle cutting scenario building module is used to build a one-way two-lane vehicle cutting scenario; specifically, it includes: creating a two-lane and a vehicle based on the initial vehicle driving test scenario; the vehicle includes an autonomous driving vehicle V based on the IDM model 21 and the intelligent vehicle V 22 , and V 22 At V 21 rear; when V 21 A V is detected in the current lane. 22 Close to V 21 When V 21 Select to change lanes; when V 21 Stay away from V 22 When V 21 Choose to keep the current lane and slow down; by introducing the responsibility-sensitive safety model RSS, the judgment of V 21 With V 22 The actual distance between them and the minimum longitudinal safety distance of RSS are used as the reward function of the one-way two-lane cutting scene; where, for any moment, if V 21 With V 22 The actual distance d min Less than the minimum longitudinal safety distance d of RSS min_RSS , regard the scene at this moment as a dangerous scene; The dangerous scenario generation module is used to generate dangerous scenarios for autonomous driving vehicle testing based on the constructed one-way single-lane following scenario and one-way two-lane cutting scenario.

7. A computer device for generating a test scenario for an autonomous driving vehicle, characterized in that: include: A memory, a processor, and a computer program stored in the memory, wherein the processor implements the steps of the method for generating an autonomous driving vehicle test scenario as described in any one of claims 1 to 5 when executing the computer program.

8. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the method for generating an autonomous driving vehicle test scenario as described in any one of claims 1 to 5.

Images