Automatic Generation Method and System for Extreme Test Scenarios of Intelligent Vehicles Based on Accident Information

Abstract

The present invention discloses a method and system for automatically generating extreme test scenarios of intelligent vehicles based on accident information. The method includes: obtaining the investigation information text of actual traffic accidents, first performing preprocessing, and then using natural language processing technology to extract the static elements of the actual traffic accident scenarios, including scenario road network information, weather environment information, scenario start and end point information, and collision basic information; initializing the scenario to be generated based on the extracted static elements of the scenario, and using the deep deterministic gradient policy algorithm to search for the dynamic parameter combination set of the scenario to be generated; taking the extracted static elements of the actual traffic accident scenarios as the static elements of the scenario to be generated, and combining them with the searched dynamic parameter combination set to generate an extreme test scenario of intelligent vehicle traffic accident similar to the actual traffic accident. The present invention supports the construction of the extreme test scenario library of intelligent vehicles and accelerates the implementation of its tests.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous driving simulation, and particularly to a method and system for automatically generating extreme test scenarios of intelligent vehicles based on accident information. Background Art

[0002] Simulation testing is an important means to verify the safety performance of intelligent vehicles, and building a large test scenario library is the primary basis for testing. For the construction of test scenarios, current research mainly focuses on natural driving scenarios of vehicles, standard regulation scenarios, and accident extreme scenarios. To make up for the defect that the overall test efficiency is low due to the low triggering frequency of dangerous scenarios in the natural driving scenario-based test method, the method of extracting extreme scenarios from accident data to accelerate the overall test process has attracted wide attention in the industry. Existing research on the extraction of accident extreme scenarios focuses on the clustering extraction of key types of scenarios, some scenario difference information will be ignored, and the automation degree of the final construction process of extreme scenarios is not high. To solve the above problems, there is an urgent need for a method to automatically extract extreme test scenarios from existing accident investigation information. Summary of the Invention

[0003] The present invention provides a method and system for automatically generating extreme test scenarios of intelligent vehicles based on accident information, which supports the construction of an extreme test scenario library for intelligent vehicles and accelerates the implementation of its testing.

[0004] To achieve the above technical objectives, the present invention adopts the following technical solutions:

[0005] A method for automatically generating extreme test scenarios of intelligent vehicles based on accident information, comprising:

[0006] Obtain the investigation information text of actual traffic accidents, first perform preprocessing, and then use natural language processing technology to extract the static elements of the actual traffic accident scenarios therefrom, including scenario road network information, weather environment information, scenario start and end point information, and collision basic information;

[0007] Initialize the scenario to be generated based on the extracted static elements of the scenario, and use the deep deterministic gradient policy algorithm to search for the dynamic parameter combination set of the scenario to be generated;

[0008] Use the extracted static elements of the actual traffic accident scenarios as the static elements of the scenario to be generated, and combine them with the searched dynamic parameter combination set to generate an extreme test scenario of an intelligent vehicle traffic accident similar to the actual traffic accident.

[0009] Further, the scenario start and end point information includes the relative position information and relative driving orientation information of the traffic accident participants, and the collision basic information includes the collision position and collision angle of the traffic accident. The collision position includes the geographical location of the accident point and the position of the collision point on the vehicle body.

[0010] Furthermore, the preprocessing of the investigation information text includes: eliminating secondary information, stemming and lemmatization, and correcting spelling mistakes; and extracting the static scene elements of the actual traffic accident by using natural language processing technology, specifically using a trained natural language processing model to extract the static scene element information from the preprocessed accident description information text.

[0011] Furthermore, the deep deterministic gradient policy algorithm is adopted to search for the dynamic parameter combination set of the to-be-generated scene, specifically including: taking the speed, acceleration, and position of the participants in the to-be-generated scene together as the state in the deep deterministic gradient policy algorithm, taking the acceleration / deceleration and steering operations of the participants in the to-be-generated scene together as the action in the deep deterministic gradient policy algorithm, and using the deep deterministic gradient policy algorithm and based on the initial speed, acceleration, and position of the participants in the to-be-generated scene, gradually searching for the dynamic parameter combination set of the to-be-generated scene until the generated dynamic parameter combination meets the same collision basic information as the actual traffic accident, and then ending the search.

[0012] Furthermore, the model training reward function of the deep deterministic gradient policy algorithm includes a rationality constraint evaluation model and a danger evaluation model; the rationality constraint evaluation model includes traffic rule constraints and vehicle stability constraints; the danger evaluation model is used to guide the traffic accident participants in the to-be-generated scene to gradually drive to the collision position of the actual traffic accident, and the rationality constraint evaluation model is used to prevent the irrationality of the generated dynamic parameter combination; the reward function expression is:

[0013] w = c1w1 + c2w2 + c3w3

[0014] In the formula, w represents the reward function value; w1, w2, and w3 respectively represent the traffic rule constraint reward, the stability constraint reward, and the scene danger reward, and c1, c2, and c3 are the distribution coefficients of the corresponding rewards.

[0015] Furthermore, for the traffic rule constraints and vehicle stability constraints, the corresponding constraint switch fields are used to control whether the traffic rule constraints and vehicle stability constraints participate in the rationality constraint evaluation model respectively. If the actual traffic accident is caused by the participants not observing traffic rules, then the traffic rule constraints do not participate in the rationality constraint evaluation model. If the actual traffic accident is caused by the vehicle stability of the participants, then the vehicle stability constraints do not participate in the rationality constraint evaluation model.

[0016] Furthermore, the vehicle stability constraint performs dynamic constraints on the yaw angular velocity lateral velocity v y , acceleration a, and steering angle θ of the vehicle:

[0017]

[0018]

[0019] |θ| ≤ θ lim

[0020] |a| ≤ a lim

[0021] where α r represents the rear wheel sideslip angle, v x and v y respectively represent the longitudinal speed and the lateral speed of the vehicle's center of mass. is the yaw rate, l r and l f are respectively the distances from the vehicle's center of mass to the rear and front axles; α r,lim is the rear wheel sideslip angle threshold, is the linear sideslip stiffness of the rear wheel; θ lim represents the steering angle limit of the vehicle, a lim represents the vehicle acceleration limit.

[0022] Furthermore, the risk assessment model considers the distance between the involved parties and the distances from each involved party to the actual traffic accident collision location, and is negatively correlated with the respective distance change values:

[0023]

[0024] where w3 represents the reward for the scene risk constraint, ΔD represents the distance change from the involved party to the collision location, Δd represents the distance change amount between the involved parties, n represents the number of scene involved parties, z1 and z2 respectively represent the distribution coefficients of each part, and f and g respectively represent the negative correlation functions with the distance change between the involved parties and the distance change from the involved party to the collision point as independent variables.

[0025] Furthermore, the static elements are combined with the dynamic parameter combination set to generate the intelligent vehicle traffic accident extreme test scenario, specifically by converting the static elements and the dynamic parameter combination set into the scene general format OpenX series files based on the XML format.

[0026] An intelligent vehicle extreme test scenario automatic generation system based on accident information, comprising:

[0027] A text preprocessing module for preprocessing the text of the investigation information of the actual traffic accident obtained;

[0028] A natural language processing module for extracting the static elements of the actual traffic accident scene from the preprocessed investigation information text, including scene road network information, weather environment information, scene start and end point information, and collision basic information;

[0029] A dynamic parameter search module, configured to: initialize a to-be-generated scenario based on the extracted static scene elements, and search for a set of dynamic parameter combinations of the to-be-generated scenario by using a deep deterministic gradient policy algorithm;

[0030] An extreme test scenario generation module, configured to: interface with the output results of the natural language processing module and the dynamic element search module, i.e., the static and dynamic elements of the extreme scenario, and combine them to generate an intelligent vehicle traffic accident extreme test scenario similar to an actual traffic accident, i.e., convert it into a general format file for the test scenario;

[0031] A storage module, configured to: store the model hyperparameters of the natural language processing module and the deep deterministic gradient policy algorithm, as well as various types of extreme test scenario files obtained by conversion.

[0032] Beneficial effects

[0033] On the one hand, the present invention promotes the deep extension of the accident investigation field to the intelligent vehicle test field, and can expand the existing intelligent vehicle test scenario library; on the second hand, the present invention focuses on building an extreme test scenario highly similar to an actual traffic accident, and the obtained test scenario has a relatively significant danger, which can, to a certain extent, accelerate the test implementation of intelligent vehicles and related safety products; on the third hand, the extreme test scenario finally generated by the present invention is an OpenX series file in the general format of the scenario based on the XML format, which can directly convert the extracted extreme test scenario information into the standard format common to simulation software, improving the generality of the results of the present invention. Therefore, when the present invention is applied to the test application link of intelligent vehicles and related safety products, it can test their safety performance in extreme dangerous scenarios, timely discover the deficiencies of the vehicle under test, and accelerate the upgrade and improvement of safety functions. Description of the drawings

[0034] Figure 1 An automated generation method for an intelligent vehicle extreme test scenario based on accident information.

[0035] Figure 2 A method for extracting static scene elements based on natural language processing technology.

[0036] Figure 3 An adversarial generation method for extreme scenario dynamic elements based on the DDPG algorithm.

[0037] Figure 4 A schematic diagram of a constraint model in the DDPG algorithm.

[0038] Figure 5 A framework of the DDPG algorithm.

[0039] Figure 6 A schematic diagram of scene dynamic parameter search.

[0040] Figure 7Automatically generate system architecture diagrams for extreme test scenarios of smart cars based on accident information. DETAILED DESCRIPTION

[0041] In order to make the above-mentioned purposes, features and advantages of the application more clearly understood, the specific embodiments of the present application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than described herein, and those skilled in the art can make similar improvements without violating the connotations of the present application. Therefore, the present application is not limited to the specific embodiments disclosed below.

[0042] Example 1

[0043] This embodiment provides a method for automatically generating extreme test scenarios for smart cars based on accident information. Figure 1 As shown, it includes the following main steps:

[0044] Step S10, obtain the investigation information text of the actual traffic accident, pre-process it, and then use natural language processing technology to extract the static elements of the actual traffic accident scene, including scene road network information, weather environment information, scene start and end point information and basic collision information.

[0045] In this embodiment, before the original complex accident investigation information text is imported into the natural language processing model, the text information must be preprocessed. The specific steps of scene static element extraction refer to Figure 2 , mainly including:

[0046] (1) Obtain and organize accident investigation information text, which generally involves thousands of fields of text information with a wide range of content, while only some key information is needed for extreme scenario construction;

[0047] (2) Text secondary information removal, mainly for the next stage of scene element analysis to delete unimportant element information, such as some spaces, line breaks and other punctuation marks that cannot convey any effective information, and some descriptive information such as the year, month, day and city of the accident that has no practical meaning for static scene extraction, so that the subsequent natural language processing model can better identify key information;

[0048] (3) Text stemming and lemmatization: To address the problem that some text abbreviations in accident investigation records may cause natural language processing models to distort the meaning of the text, stemming and lemmatization processing is performed on the existing accident information text to obtain more accurate and standardized accident description text information;

[0049] (4) Text spelling error correction. Usually, due to the incorrect input by relevant investigators, it becomes more difficult to understand the relevant scenario elements in the accident description information. Therefore, detecting and correcting these spelling errors will help improve the accuracy of static scenario element extraction.

[0050] Then, use natural language processing technology to extract the static elements of the actual traffic accident scene from the preprocessed investigation information text. In this implementation, a natural language processing model is constructed to screen the information related to and representative of the static scene elements. The main steps of this process are as follows: First, perform keyword assignment. A relatively rich keyword library is predefined, generally including information about accident participants (such as vehicle type, driving intention, etc.), the relative positions of accident participants (such as oncoming vehicle, same-direction vehicle, etc.), collision directions (such as the main vehicle being hit frontally by the secondary vehicle or being hit on the right side by the secondary vehicle, etc.), accident road network information (such as key information like intersections, three-lane roads, etc.), weather environment (such as sunny, rainy, overcast, etc.) and other related feature words. Subsequently, calculate the similarity to determine the matching degree between each type of keyword in the accident description text and the keyword library. When the matching degree reaches the preset threshold, it is considered that this word segment is the key static information of the scene.

[0051] Among them, the extracted road network information includes the current road type, number of lanes, lane width, etc., the weather environment information includes illumination, rainfall, visibility, etc., the scene start and end information includes the initial relative positions of the participants, the coordinates of the collision occurrence location, the driving intentions of the participants, etc., the basic collision information includes the collision position and collision angle of the traffic accident, and the collision position includes the geographical location of the accident point and the position of the collision point on the vehicle body.

[0052] Step S20, initialize the scene to be generated based on the extracted static elements of the scene, and use the deep deterministic gradient policy algorithm to search for the set of dynamic parameter combinations of the scene to be generated.

[0053] Use the scene start and end information extracted in step S10 to set the initial state of the traffic accident participants at the start time of the scene to be generated, including the initial position and initial driving direction. Then, use the deep deterministic gradient policy algorithm (abbreviated as DDPG algorithm) to explore the possibilities of various specific scenarios similar to the actual traffic accident from the initial state to a specific end state (i.e., the state at the actual traffic accident collision position).

[0054] In this implementation, before using the DDPG algorithm to search for dynamic parameter combinations, the model of the DDPG algorithm is first trained. In the algorithm model, the Critic network and the Actor network adopt a recognized simple and efficient 4-layer fully connected network structure, and the activation function between each layer is set to ReLU. Among them, the output layer of the Actor network uses the tanh function to ensure the boundedness of the output speed and acceleration; the basic settings of the model training parameters are as follows: the sampling step size is 0.1s, the training cycle is 25s each time, and the maximum number of training times is 10,000 times.

[0055] In this embodiment, the observed state s of the scenario environment mainly includes: vehicle speed (magnitude and direction), acceleration (magnitude and direction), and vehicle coordinate position; the action parameters output by the action policy network mainly refer to: operations such as acceleration, deceleration, and steering; the value Q network is mainly used to obtain the current state s i and take the action a i corresponding reward. Refer to Figure 5 As shown, the specific training process of the DDPG algorithm model used in this embodiment is as follows:

[0056] a) Initialize the action policy network, the value Q network, and an experience replay pool with a certain capacity;

[0057] b) Initialize the scenario according to the static element information of the scenario;

[0058] c) Observe and output the current scenario state s t (vehicle speed, acceleration, position)

[0059] d) According to the current state, the online action policy network outputs the corresponding action a t and observes the next state s of the scenario t+1 ; at the same time, the online Q network outputs the reward r of the current action according to a t , s t , s t+1 and guides the update of the parameters of the online policy network; t+1

[0060] e) Store the conversion of this round (s t , a [[ID=4l]] t , r t , s t+1 ) in the experience replay pool;

[0061] f) Sample N groups of data N(s i , a i , r i , s i+1 ) from the experience replay pool as the training data of the online policy network and the online Q network;

[0062] g) According to the target policy network μ′s|θμ’ and the target Q-network Q′(s,a|θ Q′ ), the following formula is obtained:

[0063] y i = r i + γQ′(s i+1 , μ′s i+1 θ μ′ |θ Q′ )

[0064] h) Update the online Q-network parameters according to the backpropagation of the network gradient. The mean squared error loss function is as follows:

[0065]

[0066] i) Update the online policy network parameters according to the policy gradient of the selected samples:

[0067]

[0068] j) Update the parameters of the target policy network and the target Q-network regularly in a soft update manner:

[0069]

[0070] Among them, τ is the soft update parameter, which is used to adjust the update speed of the network.

[0071] In implementation, to guide the model learning direction of the DDPG algorithm, the design of the reward function mainly considers two aspects: the rationality constraint evaluation model and the danger evaluation model; the danger index is used to guide the participating parties to gradually drive to the collision point, and the rationality index is used to prevent the generation of invalid test scenarios caused by the irrationality of the parameter combination.

[0072] First, the dynamic parameter rationality constraint includes traffic rule constraint and vehicle stability constraint. Due to the particularity of some accidents, that is, the accidents are directly caused by the irrationality of the above constraints, a constraint switch field is set correspondingly before the rationality constraint evaluation, which is specifically used to control the on / off of the rationality constraint evaluation module; the constraint switch field is obtained by analyzing the keywords related to the accident cause in the accident basic information during the extraction process of the scene static elements. If the accident is caused by a certain party of the participating parties not complying with the traffic rules, the traffic rule switch field is regarded as in the off state, that is, during the dynamic parameter search, the algorithm model will explore in the direction of violating the traffic rules without being restricted; similarly, the working principle of the stability constraint switch field is basically the same.

[0073] The vehicle stability constraint mainly considers that the unreasonable matching of speed and acceleration in the dynamic parameters causes the vehicle to roll over. Therefore, it is necessary to measure the yaw angular velocity of the vehicle lateral velocity vy Perform dynamic constraints on the acceleration a and the rotation angle θ:

[0074]

[0075]

[0076] |θ| ≤ θ lim

[0077] |a| ≤ a lim

[0078] Where α r represents the rear wheel slip angle, v x and v y represent the longitudinal speed and the lateral speed of the vehicle's center of mass respectively, is the yaw rate, l r and l f are the distances from the vehicle's center of mass to the rear and front axles respectively; α r,lim is the threshold value of the rear wheel slip angle, is the linear cornering stiffness of the rear wheel; θ lim represents the vehicle's steering angle limit, a lim represents the vehicle acceleration limit;

[0079] In summary, the reward expression related to traffic rule constraints and vehicle stability constraints is:

[0080]

[0081]

[0082] Secondly, the risk assessment model mainly considers the distance between the participating parties and the distances from each participating party to the recorded accident collision point, and is negatively correlated with the change values of each distance respectively:

[0083]

[0084] Where ΔD represents the change in the distance from the participating party to the collision point, Δd represents the change in the distance between the participating parties, n represents the number of participating parties in the scenario design, z1 and z2 respectively represent the distribution coefficients of each part, and f and g respectively represent the negative correlation functions with the change in the distance between the participating parties and the change in the distance from the participating party to the collision point as independent variables.

[0085] The expression of the reward function of the final algorithm model is:

[0086] w = c1w1 + c2w2 + c3w3

[0087] where, w1, w2, and w3 respectively represent the traffic rule constraint reward, the stability constraint reward, and the scene danger reward, and c1, c2, and c3 are the distribution coefficients of the corresponding rewards. The specific settings need to be determined according to the specific model training effect.

[0088] The DDPG algorithm is used to search for the dynamic parameter combination set of the to-be-generated scene, which specifically includes: taking the speed, acceleration, and position of the participants in the to-be-generated scene together as the state in the deep deterministic gradient policy algorithm, taking the acceleration / deceleration and steering operations of the participants in the to-be-generated scene together as the action in the deep deterministic gradient policy algorithm, using the deep deterministic gradient policy algorithm and based on the initial speed, acceleration, and position of the participants in the to-be-generated scene, and cyclically outputting the optimal action of the current vehicle according to the current scene state to promote the conversion of the current scene to the next state, so as to gradually search for the dynamic parameter combination set of the to-be-generated scene until the generated dynamic parameter combination meets the basic collision information of the actual traffic accident, and then the search ends. As Figure 6 shown in the search schematic diagram.

[0089] In step S30, the static elements of the actual traffic accident scene extracted are used as the static elements of the to-be-generated scene, and combined with the searched dynamic parameter combination set to generate an intelligent vehicle traffic accident extreme test scene similar to the actual traffic accident.

[0090] In this implementation, the static and dynamic elements of the scene obtained in steps S10 and S20 have a regular format, that is, "attribute - attribute value"; the format conversion script is mainly responsible for making corresponding modifications to the attribute values of the extracted scene elements. For the element information not involved in other accident descriptions or not obtained by this method, the corresponding default values are continued to be selected without change.

[0091] After obtaining the extreme test scene standard format file according to the above method, the scene-related configuration files and the corresponding model parameters are stored in sequence with the scene number as the title; during the later scene inspection, only the standard OpenX series format file of the corresponding scene needs to be called in the simulation visualization software, and the scene quality is evaluated through the three-dimensional visualization expression of the scene, and the model is adjusted accordingly according to the evaluation results. Moreover, the generated scenes can be automatically classified and stored according to different indicators, and the indicators can be "the degree of personal injury of the scene personnel recorded in the accident information", "the type of scene participants", "the cause of the accident scene", "the specific form of the scene", etc., so as to realize the automatic classification function of the scene.

[0092] In the method for automatically generating extreme test scenarios of intelligent vehicles based on accident information provided in this embodiment, the extraction of extreme scenarios is divided into two main parts. In the first part, natural language processing technology is used to extract the static key elements of the scenario from the accident investigation text information, eliminating the manual extraction process of existing scenario features and achieving a high degree of automation. Before information extraction, necessary operations such as removing secondary information, stemming and lemmatizing words, and correcting spelling mistakes are performed on the accident investigation text, which greatly ensures the accuracy and rapidity of information extraction. In the second part, the scenario parameters are initialized according to the scenario start and end point information and collision key information extracted in the previous part, and the dynamic parameters of the scenario are explored through the DDPG algorithm. Finally, multiple extreme test scenarios similar to the accident are built. The reward function of the algorithm model mainly considers two indicators: the rationality of constraints and the danger of the scenario, which can guide the exploration of dynamic parameters towards the direction of highly restoring the accident and probing the danger, and prevent the generation of invalid test scenarios caused by unreasonable parameter combinations, ensuring the efficiency and correctness of the generation of extreme test scenarios. Finally, through a data conversion script, the static and dynamic key information of the scenario obtained in the previous two parts is converted into a common format for simulation software, improving the general applicability of the proposed method and facilitating the inspection and maintenance of the scenario in the later stage.

[0093] In summary, the method of this embodiment can automatically extract relevant extreme test scenarios from the basic information of actual traffic accident investigations, making up for the deficiencies of the existing method of manually extracting accident characteristics and then building scenarios. The rationality constraints and danger assessment items considered in the reward function can ensure the rapidity and correctness of the scenario building process. Finally, by writing a data conversion script, the applicability and maintainability of this method are enhanced.

[0094] Embodiment 2

[0095] This embodiment provides an automatic generation system for extreme test scenarios of intelligent vehicles based on accident information. Referring to Figure 7 as shown, it includes:

[0096] A text preprocessing module, which is used to: preprocess the text of the investigation information of the actual traffic accident obtained;

[0097] A natural language processing module, which is used to: extract the static elements of the scenario of the actual traffic accident from the preprocessed investigation information text, including scenario road network information, weather environment information, scenario start and end point information, and collision basic information;

[0098] A dynamic parameter search module, which is used to: initialize the scenario to be generated based on the extracted static elements of the scenario, and use the deep deterministic gradient strategy algorithm to search for the set of dynamic parameter combinations of the scenario to be generated;

[0099] An extreme test scenario generation module, configured to: interface with the output results of the natural language processing module and the dynamic element search module, i.e., the static and dynamic elements of the extreme scenario, and combine them to generate an intelligent vehicle traffic accident extreme test scenario similar to an actual traffic accident, i.e., convert it into a general format file of the test scenario;

[0100] A storage module, configured to: store the model hyperparameters of the natural language processing module and the deep deterministic gradient policy algorithm, as well as various types of extreme test scenario files obtained through conversion. Since the storage module can classify and store the generated scenarios according to different metrics, the metrics can be "the degree of personal injury of the scenario personnel recorded in the accident information", "the type of scenario participants", "the cause of the accident scenario", "the specific form of the scenario", etc., thereby realizing the automatic classification function of the scenarios.

[0101] For each of the above modules included in the intelligent vehicle extreme test scenario automatic generation system according to this embodiment, its specific implementation manner is the same as that described in the method of Embodiment 1, and will not be repeated here.

[0102] The above embodiments are the preferred embodiments of the present application. Those of ordinary skill in the art can also make various transformations or improvements on this basis. Without departing from the overall concept of the present application, these transformations or improvements should all fall within the scope of protection required by the present application.

Claims

1. An automatic generation method for extreme test scenarios of intelligent vehicles based on accident information, characterized in that, Including: Obtain the investigation information text of the actual traffic accident, first perform preprocessing, and then use natural language processing technology to extract the static elements of the actual traffic accident scene, including scene road network information, weather environment information, scene start and end point information, and collision basic information; Initialize the scene to be generated based on the extracted static elements of the scene, and use the deep deterministic gradient policy algorithm to search for the dynamic parameter combination set of the scene to be generated; The step of using the deep deterministic gradient policy algorithm to search for the dynamic parameter combination set of the scene to be generated specifically includes: taking the speed, acceleration, and position of the parties involved in the scene to be generated together as the state in the deep deterministic gradient policy algorithm, taking the acceleration / deceleration and steering operations of the parties involved in the scene to be generated together as the action in the deep deterministic gradient policy algorithm, using the deep deterministic gradient policy algorithm and based on the initial speed, acceleration, and position of the parties involved in the scene to be generated, gradually search for the dynamic parameter combination set of the scene to be generated until the generated dynamic parameter combination satisfies the same collision basic information as the actual traffic accident, and then end the search; The model training reward function of the deep deterministic gradient policy algorithm includes a rationality constraint evaluation model and a danger evaluation model; the rationality constraint evaluation model includes traffic rule constraints and vehicle stability constraints; the danger evaluation model is used to guide the traffic accident parties involved in the scene to be generated to gradually drive to the collision position of the actual traffic accident, and the rationality constraint evaluation model is used to prevent the irrationality of the generated dynamic parameter combination; the expression of the reward function is: ； In the formula, represents the reward function value; , , respectively represent the traffic rule constraint reward, the stability constraint reward, and the scene danger reward, , , are the distribution coefficients of the corresponding rewards; Use the extracted static elements of the actual traffic accident scene as the static elements of the scene to be generated, and combine them with the searched dynamic parameter combination set to generate an intelligent vehicle traffic accident extreme test scene similar to the actual traffic accident.

2. The automatic generation method for extreme test scenarios of an intelligent vehicle according to claim 1, wherein The scene start and end point information includes the relative position information and relative driving azimuth information of the traffic accident parties involved, and the collision basic information includes the collision position and collision angle of the traffic accident. The collision position includes the geographical location of the accident point and the position of the collision point on the vehicle body.

3. The automatic generation method for extreme test scenarios of an intelligent vehicle according to claim 1, wherein The preprocessing of the investigation information text includes: eliminating secondary information, stemming and lemmatization, and correcting spelling mistakes; the step of using natural language processing technology to extract the static elements of the actual traffic accident scene specifically uses a trained natural language processing model to extract the static element information of the scene from the preprocessed accident description information text.

4. The automatic generation method for intelligent vehicle extreme test scenarios according to claim 1, wherein The traffic rule constraints and vehicle stability constraints control whether the traffic rule constraints and vehicle stability constraints participate in the rationality constraint evaluation model through the corresponding constraint switch fields. If the actual traffic accident is caused by the parties involved not complying with traffic rules, the traffic rule constraints do not participate in the rationality constraint evaluation model. If the actual traffic accident is caused by the vehicle stability of the parties involved, the vehicle stability constraints do not participate in the rationality constraint evaluation model.

5. The automatic generation method for extreme test scenarios of an intelligent vehicle according to claim 1, wherein The vehicle stability constraint performs dynamic constraints based on the yaw rate , lateral velocity , acceleration and steering angle of the vehicle: ； ； ； ； In the formula, represents the rear wheel slip angle, and respectively represent the longitudinal speed and the lateral speed of the vehicle center of mass, is the yaw rate, and are the distances from the vehicle center of mass to the rear and front axles respectively; is the rear wheel slip angle threshold, is the linear cornering stiffness of the rear wheel; represents the steering angle limit of the vehicle, represents the vehicle acceleration limit.

6. The automatic generation method for extreme test scenarios of an intelligent vehicle according to claim 1, characterized in that The danger evaluation model considers the distance between the parties involved and the distance of each party to the collision position of the actual traffic accident, and is negatively correlated with the change values of each distance respectively: ； Wherein, represents the reward for scenario danger constraint, represents the distance change of the participant to the collision position, represents the distance change amount between participants, and n represents the number of scenario participants, 、 respectively represent the distribution coefficients of each part, and respectively represent negative correlation functions with the distance change between participants and the distance change of the participant to the collision point as independent variables.

7. The automatic generation method for extreme test scenarios of an intelligent vehicle according to claim 1, wherein Combine static elements with a set of dynamic parameters to generate an extreme test scenario for intelligent vehicle traffic accidents. Specifically, convert the combination of static elements and dynamic parameters into a general scenario format OpenX series file based on XML format.

8. An intelligent vehicle extreme test scenario automatic generation system based on accident information, characterized in that, Including: A text preprocessing module for preprocessing the text of the investigation information of actual traffic accidents obtained. A natural language processing module for extracting the static elements of the actual traffic accident scenario from the preprocessed investigation information text, including scenario road network information, weather environment information, scenario start and end point information, and basic collision information. A dynamic parameter search module for initializing the scenario to be generated based on the extracted static elements of the scenario and searching for a set of dynamic parameter combinations for the scenario to be generated using the deep deterministic gradient policy algorithm. The specific process of searching for a set of dynamic parameter combinations for the scenario to be generated using the deep deterministic gradient policy algorithm includes: taking the speed, acceleration, and position of the participants in the scenario to be generated together as the state in the deep deterministic gradient policy algorithm, taking the acceleration / deceleration and steering operations of the participants in the scenario to be generated together as the action in the deep deterministic gradient policy algorithm, using the deep deterministic gradient policy algorithm and based on the initial speed, acceleration, and position of the participants in the scenario to be generated, gradually search for a set of dynamic parameter combinations for the scenario to be generated until the generated dynamic parameter combination satisfies the same basic collision information as the actual traffic accident, and then end the search. The model training reward function of the deep deterministic gradient policy algorithm includes a rationality constraint evaluation model and a hazard evaluation model; the rationality constraint evaluation model includes traffic rule constraints and vehicle stability constraints; the hazard evaluation model is used to guide the traffic accident participants in the scenario to be generated to gradually drive to the collision position of the actual traffic accident, and the rationality constraint evaluation model is used to prevent the irrationality of the generated dynamic parameter combination; the expression of the reward function is: ； In the formula, represents the reward function value; , , respectively represent the traffic rule constraint reward, the stability constraint reward, and the scene danger reward, , , are the distribution coefficients of the corresponding rewards; An extreme test scenario generation module for docking the output results of the natural language processing module and the dynamic element search module, that is, the static and dynamic elements of the extreme scenario, and combining them to generate an extreme test scenario for intelligent vehicle traffic accidents similar to the actual traffic accident, that is, converting it into a general format file for the test scenario. A storage module for storing the model hyperparameters of the natural language processing module and the deep deterministic gradient policy algorithm, as well as various types of extreme test scenario files obtained by conversion.

Images