Typical dangerous scene generalization deduction method and device for automatic driving test and readable storage medium

Abstract

The invention relates to an automatic driving test-oriented typical dangerous scene generalization deduction method and device, and a readable storage medium. The typical dangerous scene generalization deduction method comprises the steps that S1, an original accident scene data set is acquired, and accident classification is carried out; designing and simplifying a probability distribution function obeyed by driver behaviors; s2, on the basis of the accident scene data set and the accident classification thereof, determining parameters of a probability distribution function, so that the probability distribution function fits driver behaviors in a typical accident scene; s3, selecting feature items in the probability distribution function as a driver behavior reward function; s4, training a dangerous driver behavior random decision model based on the driver behavior reward function; and S5, embedding a dangerous driver behavior random decision model for the environmental vehicle in the simulation platform, and generating a typical dangerous scene database. According to the invention, expansion of the automatic driving test scene set can be realized, and the test level of automatic driving is improved.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving testing technology, and in particular to a generalized deduction method, device and readable storage medium for typical dangerous scenarios for autonomous driving testing. Background Art

[0002] As the level of autonomous driving increases, the complexity of the automotive system further increases. Changeable weather, complex traffic environment, diverse driving tasks and dynamic driving status have all posed new challenges to the evaluation and testing of autonomous driving vehicles. In particular, in terms of test objects, the focus of autonomous driving tests has shifted from traditional vehicle dynamic performance and collision safety performance tests to the operational safety of autonomous driving vehicles in complex driving environments. The test objects have shifted from vehicle performance to driving ability and behavioral safety, that is, the human-vehicle binary independent system has been transformed into a human-vehicle-environment-task strongly coupled system. Data shows that in the face of complex environment testing, due to the low probability and occasional characteristics of traffic accidents, it is necessary to conduct 8.8 billion miles of public road testing to verify that autonomous driving vehicles have higher safety performance than drivers. In theory, 100 vehicles need to be tested continuously 24 hours a day for 400 years. In addition, factors such as massive test mileage and ultra-long test cycles have made traditional mileage-based vehicle testing no longer applicable, and have been replaced by scenario-based virtual testing technology.

[0003] The scenario-based virtual testing technology has flexible experimental scenario configuration, high test efficiency, strong test repeatability, and low test cost. It can realize automated accelerated testing and save a lot of manpower and material resources. Therefore, scenario-based virtual testing has become an indispensable part of autonomous driving test evaluation, and has important practical significance for accelerating the application of autonomous driving technology and improving autonomous driving test standards.

[0004] However, the current scenario-based virtual testing technology still has the following obvious defects:

[0005] 1) The virtual test has a low degree of automation. Each type of test scenario requires in-depth manual participation in the construction, and the scene automatic loading test cannot be completed;

[0006] 2) It is difficult to derive and generalize test scenarios from existing data, resulting in the inability to achieve full and multi-dimensional coverage of test scenarios;

[0007] 3) A large number of non-critical and atypical specific scenarios are used for autonomous driving function testing, resulting in redundancy in test scenarios and low efficiency of virtual testing. Summary of the invention

[0008] In response to the above-mentioned problems in the prior art, the present invention proposes a generalized deduction method, device and readable storage medium for typical dangerous scenarios for autonomous driving tests, which expand the set of autonomous driving test scenarios and improve the testing level of automated driving.

[0009] Specifically, the present invention proposes a generalization deduction method for typical dangerous scenarios for autonomous driving testing, comprising the steps of:

[0010] S1, obtaining an original accident scene data set, classifying the accident scene data set based on the accident type; designing a probability distribution function that the driver's behavior obeys, and simplifying the probability distribution function;

[0011] S2, based on the accident scene data set and its accident classification, determining the parameters of the probability distribution function so that the probability distribution function fits the driver behavior in a typical accident scene;

[0012] S3, selecting a characteristic item in the probability distribution function as a driver behavior reward function;

[0013] S4, training a dangerous driver behavior random decision model based on the driver behavior reward function;

[0014] S5, embedding the dangerous driver behavior random decision model into the environment vehicle in the simulation platform to generate a typical dangerous scenario database.

[0015] According to one embodiment of the present invention, in step S1, the acquired accident scene data set is preprocessed; the preprocessing includes eliminating irrelevant variables and redundant data, and calculating the longitudinal and lateral spacings and speed differences between vehicles.

[0016] According to an embodiment of the present invention, in step S1, the accident types include rear-end collision, cut-in / cut-out, and collision within an intersection.

[0017] According to one embodiment of the present invention, in step S1, a probability distribution of driver behavior is designed based on maximum entropy, and the probability distribution function is simplified by Laplace approximation, including the following steps:

[0018] S11, assuming that the trajectory of the driver's behavior obeys a certain conditional probability distribution:

[0019] ζ~P(ζ|β);

[0020] Among them, ζ is the driver's behavior trajectory, is an undetermined parameter, P represents the probability that the driver chooses ζ under the condition that the parameter β is constant, and the trajectory ζ is defined as:

[0021] ζ=[s0,a0,…,s k ,ak ,…,s T ] T ;

[0022] Among them, s k It is represented as the driving state of the environment vehicle in the scene at time k; a k It is represented as the driving control input of the environment vehicle in the scene at time k;

[0023] S12, establish a maximum entropy model to derive the probability distribution function that the trajectory of the driver's behavior obeys. The maximum entropy model is defined as follows:

[0024]

[0025] Where H is the entropy function, Represented as driving behavior trajectory ζ i The j-th characteristic function value under the action of Expressed as the jth expected characteristic function value;

[0026] S13, solving the maximum entropy model to obtain the probability distribution of the driver's behavior trajectory:

[0027]

[0028] Among them, f(ζ i )=[f1(ζ i ),f2(ζ i ),…f j (ζ i )…f M (ζ i )] T The trajectory ζ i The characteristic function vector under the action; Z is the partition function, which needs to traverse all feasible driving trajectories, specifically expressed as:

[0029]

[0030] S14, note The probability distribution function of the driver's behavior trajectory is reasonably simplified by using the Laplace approximation, so the probability distribution function P(ζ i |β) can be approximately expressed as:

[0031]

[0032] Where d is the random variable ζ i The dimension, g i and H i They are respectively expressed as the function R(ζ,β) at ζ i The gradient vector and Hessian matrix at this point are:

[0033]

[0034] According to one embodiment of the present invention, in step S2, based on the accident scene data set and its accident classification, the parameters of the probability distribution function are calibrated using maximum likelihood estimation, including the steps of:

[0035] S21, construct the likelihood function and omit the constant term that is independent of the parameter β:

[0036]

[0037] Where L is the likelihood function, Represented as a typical accident scenario dataset;

[0038] S22, taking the first-order derivative of the likelihood function with respect to the parameter β, we can obtain the gradient function:

[0039]

[0040] in, tr(·) represents the trace of the matrix;

[0041] S23, input typical accident scene data set Linearized vehicle system dynamic equation, characteristic function f(·), convergence threshold ε, gradient descent step α, parameter initial value β0;

[0042] S24, set the number of iterations iter = 0, the gradient G = 0;

[0043] S25, traversing the typical accident scene data set The trajectory in i , calculate the function R(ζ,β) at ζ i The gradient g at this point i and the Hessian matrix H i ;

[0044] S26, update gradient

[0045] S27, based on maximum likelihood estimation, with the goal of maximizing the likelihood function L, calculate β using the gradient ascent method: β = β0 + αG;

[0046] S28, convergence judgment, if it satisfies ‖β-β0‖≤ε, the optimal probability distribution function parameter β is obtained and the process ends; if it does not satisfy ‖β-β0‖≤ε, let iter=iter+1, β0=β, and return to step s25.

[0047] According to one embodiment of the present invention, in step S3, a characteristic term R(ζ,β)=β in the probability distribution function is selected. Tf(ζ) is used as the driver behavior reward function.

[0048] According to one embodiment of the present invention, in step S4, taking the driver behavior reward function as input, based on the actor-critic algorithm, training a dangerous driver behavior random decision model specifically includes:

[0049] S401, initialize the policy network π θ The parameters θ of (a|s) and the value network V ω (s) parameter ω, policy network π θ (a|s) is defined as:

[0050] π θ (a|s)=P(a k =a|s k =s,θ);

[0051] It is expressed as the probability that the driver chooses action / operation a in state s;

[0052] S402, initialize the training environment k+1 =g(s k ,a k );

[0053] S403 , input the driver behavior reward function R(ζ,β) selected in step S3 , the dimension n of the state s, and the dimension m of the discretized action a.

[0054] S404, given the initial state s0, learning rates α, μ, discount factor γ, and the maximum number of interactions between the driver and the environment Max_Episode, set the termination condition for a single interaction;

[0055] S405, set Episode = 0, time step k = 0;

[0056] S406, input state s j , using the policy network π θ (a|s) Calculate the probability P of selecting each action; randomly select the current action a according to the probability distribution k ;

[0057] S407, action a k Act on the environment and get the state s at the next moment k+1 , and the current reward R(s k ,a k ,β);

[0058] S408, calculate TD error, input current state s k and the next state s k+1 , using the value network to calculate the value function Vω (s k ) and V ω (s k+1 ), calculate the TD error value δ:

[0059] δ=R(s k ,a k ,β)+γV ω (s k+1 )-V ω (s k );

[0060] S409, using the gradient descent method to update the value network parameter ω:

[0061]

[0062] S410, update the policy network parameters θ using the policy gradient method:

[0063]

[0064] S411, if the termination condition is not met, set k = k + 1, s k =s k+1 , return to step S406; if the termination condition is met, set k=0, Episode=Episode+1, and return to step S406;

[0065] S412, when Episode = Max_Episode, output the optimal network parameters ω and θ, and obtain the converged strategy network π θ (a|s), the policy network π θ (a|s) as a random decision model of dangerous driver behavior.

[0066] According to one embodiment of the present invention, in step S5, the dangerous driver behavior random decision model is embedded in the simulation platform for the environmental vehicle, and given the initial scene state of the simulation environment, the simulation is run and the accident scene driving trajectory is randomly generated to generate a typical dangerous scene database.

[0067] The present invention also provides a generalized deduction device for typical dangerous scenarios for autonomous driving testing, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the generalized deduction method for typical dangerous scenarios described in any one of the above items are implemented.

[0068] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method for generalizing and deducing typical dangerous scenarios described in any of the above items are implemented.

[0069] The present invention provides a generalized deduction method, device and readable storage medium for typical dangerous scenarios for autonomous driving testing. The method trains a random decision-making model for dangerous driver behavior through an original accident scenario data set and a probability distribution function that the designed driver behavior obeys. The model is embedded in a simulation platform to simulate the driving behavior of environmental vehicles. By running the simulation, a typical dangerous scenario database can be automatically generated, thereby improving the testing level of automated driving.

[0070] It is to be understood that both the foregoing general description and the following detailed description of the present invention are exemplary and explanatory and are intended to provide further explanation of the invention as claimed. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] The accompanying drawings are included to provide further explanation of the present invention, and are incorporated into and constitute a part of this application. The accompanying drawings illustrate embodiments of the present invention and together with the description serve to explain the principle of the present invention.

[0072] In the attached figure:

[0073] Figure 1 A flowchart of a method for generalizing and deducing typical dangerous scenarios for autonomous driving testing according to an embodiment of the present invention is shown.

[0074] Figure 2 A schematic diagram showing the training convergence results of the dangerous driver behavior random decision model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0075] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application may be combined with each other.

[0076] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present application and its application or use. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0077] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0078] Unless otherwise specifically stated, the relative arrangement, numerical expressions and numerical values ​​of the parts and steps set forth in these embodiments do not limit the scope of the application. Meanwhile, it should be understood that, for ease of description, the sizes of the various parts shown in the accompanying drawings are not drawn according to actual proportional relationships. The technology, methods and equipment known to those of ordinary skill in the relevant art may not be discussed in detail, but in appropriate cases, the technology, methods and equipment should be considered as a part of the specification. In all examples shown and discussed here, any specific value should be interpreted as being merely exemplary, rather than as a limitation. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters represent similar items in the following drawings, and therefore, once a certain item is defined in an accompanying drawing, it does not need to be further discussed in subsequent drawings.

[0079] In the description of the present application, it should be understood that the directions or positional relationships indicated by directional words such as "front, back, up, down, left, right", "lateral, vertical, perpendicular, horizontal" and "top, bottom" are usually based on the directions or positional relationships shown in the drawings, and are only for the convenience of describing the present application and simplifying the description. Unless otherwise specified, these directional words do not indicate or imply that the device or element referred to must have a specific direction or be constructed and operated in a specific direction, and therefore cannot be understood as limiting the scope of protection of the present application; the directional words "inside and outside" refer to the inside and outside relative to the outline of each component itself.

[0080] In addition, it should be noted that the use of words such as "first" and "second" to define components is only for the convenience of distinguishing the corresponding components. If not otherwise stated, the above words have no special meaning and cannot be understood as limiting the scope of protection of this application. In addition, although the terms used in this application are selected from well-known and commonly used terms, some terms mentioned in the specification of this application may be selected by the applicant at his or her discretion, and their detailed meanings are explained in the relevant parts of the description of this article. In addition, it is required to understand this application not only by the actual terms used, but also by the meaning implied by each term.

[0081] Figure 1 A flowchart of a typical dangerous scenario generalization deduction method for autonomous driving test according to an embodiment of the present invention is shown. As shown in the figure, a typical dangerous scenario generalization deduction method for autonomous driving test comprises the following steps:

[0082] S1, obtain the original accident scene data set, which is formed by collecting the original accident scene data. For example, the real scene when the accident occurs is captured from all directions and angles based on the roadside camera. Then, the accident scene data set is classified based on the accident type, and the accident classification can be performed based on the captured video information and vehicle trajectory data features. At the same time, the probability distribution function that the driver's behavior obeys is designed and simplified.

[0083] S2, based on the accident scenario data set and its accident classification, determine the parameters of the probability distribution function so that the probability distribution function fits the driver's behavior in a typical accident scenario. The typical accident scenario refers to a common accident scenario, such as rear-end collision, lane cutting, intersection turn and straight conflict, etc. This step helps to deeply understand the driver's reaction mode in a specific situation and prepare for subsequent operations.

[0084] S3, select the characteristic item in the probability distribution function as the driver behavior reward function. Conventionally, the driver behavior reward function can be used to evaluate the performance of the autonomous driving system. During the test, the behavior of the autonomous driving system can be compared with the behavior of the driver, and the calculation result of the reward function can be used to judge whether the performance of the autonomous driving system meets expectations. If the behavior of the autonomous driving system is closer to the behavior of the human driver, the value of the reward function is higher, indicating that the performance of the autonomous driving system is better.

[0085] S4, based on the driver behavior reward function, train the random decision model of dangerous driver behavior. The driver behavior reward function provides a clear direction and criteria for the training of the random decision model, enabling the model to more accurately simulate the random decision process under dangerous driving behavior.

[0086] S5, embeds the random decision model of dangerous driver behavior for the environmental vehicles in the simulation platform and generates a typical dangerous scenario database.

[0087] In some examples, in step S1, the acquired accident scene data set is preprocessed. The preprocessing includes eliminating irrelevant variables and redundant data, and calculating the longitudinal and lateral spacing and speed difference between vehicles. The purpose of the preprocessing is to form structured data for subsequent use.

[0088] In some examples, in step S1, the accident type includes rear-end collision, cut-in / cut-out, and collision in an intersection. As an example and not limitation, the accident type may also include collision, scrape, rollover, etc.

[0089] In some examples, in step S1, a probability distribution of driver behavior is designed based on maximum entropy, and the probability distribution function is simplified using the Laplace approximation method, including the steps of:

[0090] S11, assuming that the trajectory of the driver's behavior obeys a certain conditional probability distribution:

[0091] ζ~P(ζ|β);

[0092] Among them, ζ is the driver's behavior trajectory, is an undetermined parameter, P represents the probability that the driver chooses ζ under the condition that the parameter β is constant, and the trajectory ζ is defined as:

[0093] ζ=[s0,a0,…,s k ,a k ,…,s T ] T ;

[0094] Among them, s k It is represented by the driving state of the vehicle in the scene at time k (including the longitudinal and lateral position, speed, heading angle, etc. of the vehicle); a k It is represented by the driving input (such as acceleration, steering wheel angle, etc.) of the environment vehicle in the scene at time k;

[0095] S12, establish a maximum entropy model to derive the probability distribution function that the trajectory of the driver's behavior obeys. The maximum entropy model is defined as follows:

[0096]

[0097] Where H is the entropy function, Represented as driving behavior trajectory ζ i The j-th characteristic function value under the action of Expressed as the jth expected characteristic function value;

[0098] S33, solve the maximum entropy model and obtain the probability distribution of the driver's behavior trajectory:

[0099]

[0100] Among them, f(ζ i )=[f1(ζ i ),f2(ζ i ),…f j (ζ i )…f M (ζ i )] T The trajectory ζ i The characteristic function vector under the action; Z is the partition function, which needs to traverse all feasible driving trajectories. The partition function Z is specifically expressed as:

[0101]

[0102] S34, remember The probability distribution function of the driver's behavior trajectory is reasonably simplified by using the Laplace approximation, so the probability distribution function P(ζ i |β) can be approximately expressed as:

[0103]

[0104] Where d is the random variable ζ i The dimension, g i and H i They are respectively expressed as the function R(ζ,β) at ζ i The gradient vector and Hessian matrix at this point are:

[0105]

[0106] In some examples, in step S2, based on the accident scene data set and its accident classification, the parameters of the probability distribution function are calibrated using maximum likelihood estimation, including the steps of:

[0107] S21, construct the likelihood function and omit the constant term that is independent of the parameter β:

[0108]

[0109] Where L is the likelihood function, Represented as a typical accident scenario dataset;

[0110] S22, taking the first-order derivative of the likelihood function with respect to the parameter β, we can obtain the gradient function:

[0111]

[0112] in, tr(·) represents the trace of the matrix;

[0113] S23, input typical accident scene data set Linearized vehicle system dynamic equation, characteristic function f(·), convergence threshold ε, gradient descent step α, parameter initial value β0;

[0114] S24, set the number of iterations iter = 0, the gradient G = 0;

[0115] S25, traverse the accident scene dataset The trajectory in i , calculate the function R(ζ,β) at ζ i The gradient g at this point i and the Hessian matrix H i ;

[0116] S26, update gradient

[0117] S27, based on maximum likelihood estimation, with the goal of maximizing the likelihood function L, calculate β using the gradient ascent method: β = β0 + αG;

[0118] S28, convergence judgment, if it satisfies ‖β-β0‖≤ε, the optimal probability distribution function parameter β is obtained and the process ends; if it does not satisfy ‖β-β0‖≤ε, let iter=iter+1, β0=β, and return to step s25.

[0119] In some examples, in step S3, the characteristic term R(ζ,β)=β in the probability distribution function is selected T f(ζ) is used as the driver behavior reward function.

[0120] In some examples, in step S4, taking the driver behavior reward function as input, based on the actor-critic algorithm, training the dangerous driver behavior random decision model specifically includes:

[0121] S401, initialize the policy network π θ The parameters θ of (a|s) and the value network V ω (s) parameter ω, policy network π θ (a|s) is defined as:

[0122] π θ (a|s)=P(a k =a|s k =s,θ);

[0123] It is expressed as the probability that the driver chooses action / operation a in state s;

[0124] S402, initialize the training environment k+1 =g(s k ,a k );

[0125] S403 , input the driver behavior reward function R(ζ,β) selected in step S3 , the dimension n of the state s, and the dimension m of the discretized action a.

[0126] S404, given the initial state s0, learning rates α, μ, discount factor γ, and the maximum number of interactions between the driver and the environment Max_Episode, set a single interaction termination condition (the termination condition can be understood as the occurrence of a collision);

[0127] S405, set Episode = 0, time step k = 0;

[0128] S406, input state s j , using the policy network π θ(a|s) Calculate the probability P of selecting each action; randomly select the current action a according to the probability distribution k ;

[0129] S407, action a k Act on the environment and get the state s at the next moment k+1 , and the current reward R(s k ,a k ,β);

[0130] S408, calculate TD error, input current state s k and the next state s k+1 , using the value network to calculate the value function V ω (s k ) and V ω (s k+1 ), calculate the TD error value δ:

[0131] δ=R(s k ,a k ,β)+γV ω (s k+1 )-V ω (s k );

[0132] S409, using the gradient descent method to update the value network parameter ω:

[0133]

[0134] S410, update the policy network parameters θ using the policy gradient method:

[0135]

[0136] S411, if the termination condition is not met, set k = k + 1, s k =s k+1 , return to step S406; if the termination condition is met, set k=0, Episode=Episode+1, and return to step S406;

[0137] S412, when Episode = Max_Episode, output the optimal network parameters ω and θ, and obtain the converged strategy network π θ (a|s), policy network π θ (a|s) as a random decision model of dangerous driver behavior.

[0138] In some examples, in step S5, a random decision model of dangerous driver behavior is embedded in the simulation platform for the environmental vehicle, and given the initial scenario state of the simulation environment, the simulation is run and accident scenario driving trajectories are randomly generated to generate a typical dangerous scenario database.

[0139] Figure 2 A schematic diagram showing the training convergence results of the dangerous driver behavior random decision model according to an embodiment of the present invention is shown. As shown in the figure, the horizontal axis represents the number of iterations, and the vertical axis represents the reward function value. Figure 2 It means that as the number of iterations increases, the reward function value tends to be maximized and the training tends to be stable.

[0140] The present invention also provides a generalized deduction device for typical dangerous scenarios for autonomous driving tests, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any step of the aforementioned generalized deduction method for typical dangerous scenarios for autonomous driving tests.

[0141] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the computer program implements the steps of any of the aforementioned methods for generalizing and deducing typical dangerous scenarios for autonomous driving tests.

[0142] Among them, the specific implementation methods and technical effects of the generalized deduction device and computer-readable storage medium for typical dangerous scenarios for autonomous driving tests can be found in the embodiments of the generalized deduction method for typical dangerous scenarios provided by the present invention above, and will not be repeated here.

[0143] The present invention provides a method for generalizing and deducing typical dangerous scenarios for autonomous driving tests, a device for generalizing and deducing typical dangerous scenarios for autonomous driving tests, and a readable storage medium, which have the following beneficial effects:

[0144] (1) Compared with the traditional autonomous driving test scenario design method, the original accident scenario data set is used to train the random decision-making model of dangerous driver behavior, thereby generalizing and deriving limited accident scenarios, expanding the autonomous driving test scenario set, and improving the comprehensive and multi-dimensional coverage of the test scenarios.

[0145] (2) The random decision-making model of dangerous driver behavior is embedded in the simulation platform to simulate the driving behavior of environmental vehicles. Then, by running the simulation, the accident scenario can be automatically generated, thereby reducing the degree of manual participation in scenario construction and improving the level of automation of the test.

[0146] (3) A probability distribution of driver behavior in typical dangerous scenarios was constructed, which greatly increased the frequency of occurrence of typical dangerous scenarios and eliminated a large number of mundane, non-critical, and non-typical scenarios that were invalid for testing, thereby reducing the redundancy of test scenarios and improving the efficiency of simulation testing.

[0147] Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or a combination of the two. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps are generally described above in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. The technician may implement the described functionality in different ways for each specific application, but such implementation decisions should not be interpreted as resulting in a departure from the scope of the present invention.

[0148] The various illustrative logic modules and circuits described in conjunction with the embodiments disclosed herein may be implemented or executed with a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in cooperation with a DSP core, or any other such configuration.

[0149] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. The software module may reside in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor so that the processor can read and write information from / to the storage medium. In an alternative, a storage medium may be integrated into a processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and the storage medium may reside in a user terminal as discrete components.

[0150] In one or more exemplary embodiments, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented as a computer program product in software, each function may be stored on or transmitted by a computer-readable medium as one or more instructions or codes. Computer-readable media include both computer storage media and communication media, including any medium that facilitates the transfer of a computer program from one place to another. Storage media may be any available medium that can be accessed by a computer. As an example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, disk storage or other magnetic storage device, or any other medium that can be used to carry or store the desired program code in the form of an instruction or data structure and can be accessed by a computer. Any connection is also properly referred to as a computer-readable medium. For example, if the software is transmitted from a website, a server, or other remote source using a coaxial cable, a fiber optic cable, a twisted pair, a digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwaves, the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwaves are included in the definition of the medium. Disk and disc as used herein include compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk and Blu-ray disc, wherein disk often reproduces data magnetically, while disc reproduces data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media.

[0151] It will be apparent to those skilled in the art that various modifications and variations may be made to the above exemplary embodiments of the present invention without departing from the spirit and scope of the present invention. Therefore, it is intended that the present invention covers modifications and variations of the present invention that fall within the scope of the appended claims and their equivalent technical solutions.

Claims

1. A generalization deduction method for typical dangerous scenarios for autonomous driving testing, comprising the following steps: S1, obtaining an original accident scene data set, classifying the accident scene data set based on the accident type; designing a probability distribution function that the driver's behavior obeys, and simplifying the probability distribution function; S2, based on the accident scene data set and its accident classification, determining the parameters of the probability distribution function so that the probability distribution function fits the driver behavior in a typical accident scene; S3, selecting a characteristic item in the probability distribution function as a driver behavior reward function; S4, training a dangerous driver behavior random decision model based on the driver behavior reward function; S5, embedding the dangerous driver behavior random decision model into the environment vehicle in the simulation platform to generate a typical dangerous scenario database.

2. The typical danger scenario generalization deduction method according to claim 1 is characterized in that: In step S1, the acquired accident scene data set is preprocessed; the preprocessing includes eliminating irrelevant variables and redundant data, and calculating the longitudinal and lateral spacings and speed differences between vehicles.

3. The typical danger scenario generalization deduction method according to claim 1 is characterized in that: In step S1 , the accident types include rear-end collision, cut-in / cut-out, and intersection collision.

4. The typical danger scenario generalization deduction method according to claim 1 is characterized in that: In step S1, the probability distribution of the driver's behavior is designed based on maximum entropy, and the probability distribution function is simplified by using the Laplace approximation method, including the steps of: S11, assuming that the trajectory of the driver's behavior obeys a certain conditional probability distribution: ζ~P(ζ|β) Among them, ζ is the driver's behavior trajectory, is an undetermined parameter, P represents the probability that the driver chooses ζ under the condition that the parameter β is constant, and the trajectory ζ is defined as: ζ=[s0,a0...,s k ,a k ,…,s T ] T Among them, s k It is represented as the driving state of the environment vehicle in the scene at time k; a k It is represented as the driving control input of the environment vehicle in the scene at time k; S12, establish a maximum entropy model to derive the probability distribution function that the trajectory of the driver's behavior obeys. The maximum entropy model is defined as follows: Where H is the entropy function, Represented as driving behavior trajectory ζ i The j-th characteristic function value under the action of Expressed as the jth expected characteristic function value; S13, solving the maximum entropy model to obtain the probability distribution of the driver's behavior trajectory: Among them, f(ζ i )=[f1(ζ i ),f2(ζ i ),…f j (ζ i )…f M (ζ i )] T The trajectory ζ i The characteristic function vector under the action; Z is the partition function, which needs to traverse all feasible driving trajectories, specifically expressed as: S14, note The probability distribution function of the driver's behavior trajectory is reasonably simplified by using the Laplace approximation, so the probability distribution function P(ζ i |β) can be approximately expressed as: Where d is the random variable ζ i The dimension, g i and H i They are respectively expressed as the function R(ζ,β) at ζ i The gradient vector and Hessian matrix at this point are:

5. The typical danger scenario generalization deduction method according to claim 4 is characterized in that: In step S2, based on the accident scene data set and its accident classification, the parameters of the probability distribution function are calibrated using maximum likelihood estimation, including the steps of: S21, construct the likelihood function and omit the constant term that is independent of the parameter β: Where L is the likelihood function, Represented as a typical accident scene dataset; S22, taking the first-order derivative of the likelihood function with respect to the parameter β, we can obtain the gradient function: in, tr(·) represents the trace of the matrix; S23, input typical accident scene data set Linearized vehicle system dynamic equation, characteristic function f(·), convergence threshold ε, gradient descent step size α, parameter initial value β0; S24, set the number of iterations iter = 0, the gradient G = 0; S25, traversing the typical accident scene data set The trajectory in i , calculate the function R(ζ,β) at ζ i The gradient g at this point i and the Hessian matrix H i ; S26, update gradient S27, based on maximum likelihood estimation, with the goal of maximizing the likelihood function L, calculate β using the gradient ascent method: β = β0 + αG; S28, convergence judgment, if it satisfies ‖β-β0‖≤ε, the optimal probability distribution function parameter β is obtained and the process ends; if it does not satisfy ‖β-β0‖≤ε, let iter=iter+1, β0=β, and return to step s25.

6. The typical danger scenario generalization deduction method according to claim 5 is characterized in that: In step S3, select the characteristic term R(ζ,β)=β in the probability distribution function T f(ζ) is used as the driver behavior reward function.

7. The typical danger scenario generalization deduction method according to claim 6 is characterized in that: In step S4, the driver behavior reward function is used as input, and a dangerous driver behavior random decision model is trained based on the actor-critic algorithm, specifically including: S401, initialize the policy network π θ The parameters θ of (a|s) and the value network V ω (s) parameter v, policy network π θ (a|s) is defined as: p θ (a|s)=P(a k =a|s k =s,θ); It is expressed as the probability that the driver selects action / operation a in state s; S402, initialize the training environment k+1 =g(s k ,a k ); S403 , input the driver behavior reward function R(ζ,β) selected in step S3 , the dimension n of the state s, and the dimension m of the discretized action a. S404, given the initial state s0, learning rates α, μ, discount factor γ, and the maximum number of interactions between the driver and the environment Max_Episode, set the termination condition for a single interaction; S405, set Episode = 0, time step k = 0; S406, input state s k , using the policy network π θ (a|s) Calculate the probability P of selecting each action; randomly select the current action a according to the probability distribution k ; S407, action a k Act on the environment and get the state s at the next moment k+1 , and the current reward R(s k ,a k ,β); S408, calculate TD error, input current state s k and the next state s k+1 , using the value network to calculate the value function V ω (s k ) and V ω (s k+1 ), calculate the TD error value δ: δ=R(s k ,a k ,β)+γV ω (s k+1 )-V ω (s k ); S409, using the gradient descent method to update the value network parameter ω: S410, update the policy network parameters θ using the policy gradient method: S411, if the termination condition is not met, set k = k + 1, s k =s k+1 , return to step S406; if the termination condition is met, set k=0, Episode=Episode+1, and return to step S406; S412, when Episode = Max_Episode, output the optimal network parameters ω and θ, and obtain the converged strategy network π θ (a|s), the policy network π ′ (a|s) as a random decision model of dangerous driver behavior.

8. The typical danger scenario generalization deduction method according to claim 1 is characterized in that: In step S5, the dangerous driver behavior random decision model is embedded in the simulation platform for the environment vehicle, and given the initial scene state of the simulation environment, the simulation is run and the accident scene driving trajectory is randomly generated to generate a typical dangerous scene database.

9. A generalized deduction device for typical dangerous scenarios for autonomous driving testing, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method for generalizing and deducing typical dangerous scenarios as described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for generalizing and deducing typical hazardous scenarios as described in any one of claims 1 to 8 are implemented.

Images