Automatic driving key test scene generalization generation system and generation method

Abstract

The invention relates to the technical field of vehicle testing, in particular to an automatic driving key test scene generalization generation system and method. The scene generation system includes an encoder, a decoder, a generator, and a discriminator. Wherein the encoder is used for performing feature extraction on input real scene information and encoding the extracted feature information to obtain encoded real scene information; the generator and the discriminator are used for generating potential scene information through mutual gaming of the generator and the discriminator according to the input noise information; and the decoder is used for decoding the real scene information and the potential scene information to obtain a test evaluation scene. Therefore, the test scene close to the real scene can be constructed according to the real scene information and the potential random noise signal to meet the simulation test of the autonomous vehicle, and different and diversified scenes can be constructed by adjusting the scene construction information, so that the test scene of the autonomous vehicle is more comprehensive.

Description

Technical Field

[0001] The present application relates to the field of vehicle testing technology, and in particular to a generalized generation system and method for key test scenarios for autonomous driving. Background Art

[0002] Before being put on the market, self-driving cars must undergo sufficient and rigorous testing to verify their safety and reliability. Testing based on open roads is not only time-consuming and costly, but also faces the ever-changing road traffic environment and many uncontrollable factors, resulting in greater safety risks in testing self-driving vehicles on real roads. In contrast, although closed-field testing has a lower safety risk, it is difficult to cover a variety of different driving scenarios, resulting in some test scenarios being restricted, thus affecting the comprehensiveness of the test results.

[0003] It can be seen that the relevant technologies have certain limitations in testing autonomous driving vehicles in real road scenarios or closed scenarios, which makes the test results inaccurate and incomplete. Summary of the invention

[0004] The present application provides a generalized generation system and generation method for key test scenarios for autonomous driving to solve the technical problem in related technologies that there are certain limitations in testing autonomous driving vehicles in real road scenarios or closed scenarios.

[0005] In a first aspect, the present application provides a generalization generation system for key test scenarios for autonomous driving, including: an encoder, a decoder, a generator, and a discriminator;

[0006] The encoder is used to extract features from the input real scene information and encode the extracted feature information to obtain encoded real scene information;

[0007] The generator and the discriminator are used to generate potential scene information through the mutual game between the generator and the discriminator according to the input noise information;

[0008] The decoder is used to decode the real scene information and the potential scene information to obtain a test evaluation scene.

[0009] In one possible design, the encoder includes an autoencoder and a Transformer encoder;

[0010] The autoencoder includes a feature extraction layer, a gated recurrent unit, a maximum pooling layer, an average pooling layer, a fully connected layer, an activation function layer, and a Transformer layer;

[0011] The Transformer encoder includes a multi-head attention mechanism module, a residual connection and layer normalization module, and a feedforward neural network module;

[0012] The decoder includes a fully connected layer, a Sigmoid function module and a gated recurrent unit.

[0013] The generator is implemented using a multi-layer perceptron, and the discriminator consists of multiple fully connected layers.

[0014] In one possible design, the real scene information includes the main vehicle speed, the lane-changing vehicle speed, the relative speed, the longitudinal distance, the lateral distance, the main vehicle acceleration, the lane-changing vehicle acceleration, the main vehicle lateral speed, the lane-changing vehicle lateral speed, the lane-changing vehicle average speed, the main vehicle average speed, the lane-changing vehicle average acceleration and the main vehicle average acceleration.

[0015] In a second aspect, the present application also provides a method for generalizing and generating key test scenarios for autonomous driving, including:

[0016] Acquire real scene information, and input the real scene information into an encoder, wherein the encoder extracts features from the input real scene information and encodes the extracted feature information to obtain encoded real scene information;

[0017] Acquire noise information, and transform the noise information through the game between the generator and the discriminator to generate potential scene information;

[0018] A decoder is used to decode the real scene information and the potential scene information to obtain a test evaluation scene.

[0019] In one possible design, the real scene information includes the main vehicle speed, the lane-changing vehicle speed, the relative speed, the longitudinal distance, the lateral distance, the main vehicle acceleration, the lane-changing vehicle acceleration, the main vehicle lateral speed, the lane-changing vehicle lateral speed, the lane-changing vehicle average speed, the main vehicle average speed, the lane-changing vehicle average acceleration and the main vehicle average acceleration.

[0020] In a possible design, the scene generation method further includes:

[0021] Taking hazard and exposure frequency as criticality functions, a search optimization function is set, and an optimization algorithm is used to search for key scenes in the logical scene space.

[0022] In a possible design, the generalization generation method for key test scenarios for autonomous driving also includes:

[0023] Based on the key scenarios found, a simulation test scenario including road elements, meteorological elements and traffic elements is constructed in the simulation platform;

[0024] An adaptive cruise control algorithm is used to implement a simulation test of the vehicle in the simulation test scenario.

[0025] In a third aspect, the present application also provides a computer-readable storage medium storing a computer program, which, when running on a processor, executes the generalization generation method for key test scenarios for autonomous driving as described in any one of the above items.

[0026] In a fourth aspect, the present application also provides a computer program product, comprising: a computer program code, which, when executed on a computer, enables the computer to execute a generalization generation method for key test scenarios for autonomous driving as described in any one of the above items.

[0027] The generalized generation system of key test scenarios for autonomous driving provided by the first aspect above includes: an encoder, a decoder, a generator, and a discriminator. Among them, the encoder is used to extract features from the input real scene information, and encode the extracted feature information to obtain the encoded real scene information; the generator and the discriminator are used to generate potential scene information through the mutual game between the generator and the discriminator according to the input noise information; the decoder is used to decode the real scene information and the potential scene information to obtain the test evaluation scene. In this way, a test scene close to the real scene can be constructed based on the real scene information and the potential random noise signal to meet the simulation test of the autonomous driving vehicle, and different and diverse scenes can be constructed by adjusting the scene construction information, so that the test scene of the autonomous driving vehicle is more comprehensive, and the test results are comprehensive and universal.

[0028] The beneficial effects provided in the above-mentioned second aspect and various possible designs of the above-mentioned second aspect can refer to the beneficial effects brought about by the above-mentioned first aspect and various possible implementation methods of the first aspect, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 A schematic diagram of the structure of a scene generation system provided in an embodiment of the present application;

[0030] Figure 2 A flowchart of a generalized generation method for key test scenarios for autonomous driving provided in an embodiment of the present application;

[0031] Figure 3 A flowchart of a key scene search method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0032] In the description of the embodiments of the present application, unless otherwise specified, “ / ” means or. For example, A / B can mean A or B. The “and / or” in this article is merely a way to describe the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0033] In the following, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this embodiment, unless otherwise specified, "plurality" means two or more.

[0034] In related technologies, the testing methods for autonomous vehicles are usually divided into three categories: public road testing, closed field testing, and scenario-based virtual testing. Public road testing is time-consuming and inefficient, while closed field testing has a small scope and is difficult to implement testing in multiple scenarios. Scenario-based virtual simulation testing overcomes the above shortcomings and has become the most effective supplementary solution in the development and verification stages of autonomous driving functions.

[0035] The autonomous driving test scenario is a combination of the autonomous driving vehicle and its driving environment. It describes the development of a series of scenarios over a period of time, and can specify the actions and events of the autonomous driving vehicle as test targets. Scenario generation is the focus and premise of the test, and is usually divided into data-driven and knowledge-based methods. The knowledge-based method cannot derive the performance of the system under test in natural traffic, while the data-driven scenario generation framework is divided into four main steps: obtaining data sources, scenario identification, scenario generation, and scenario evaluation. After the scenario generation and evaluation are completed, the safety test of the autonomous driving car can be implemented based on the simulation test software. It is necessary to import the generated scenario in the simulation environment, add vehicle models and sensor models, etc., execute the test, and obtain the test results of the autonomous driving system.

[0036] In the related art, the scene generation methods mainly include: sampling method, combination method, deep learning, optimization search and reinforcement learning. The sampling-based method generates typical scenes or key scenes by random sampling or importance sampling. Among them, one sampling method is to use Markov chain Monte Carlo (MCMC) to generate lane change cut-in scenes with high coverage and similar feature distribution, but the smoothness of the generated trajectory is poor. Another sampling method is to use Gibbs sampling to generate 20,000 cut-in scenes and perform comfort evaluation. Due to the presence of a large number of low-risk scenes in the logical scene space, the efficiency of Monte Carlo sampling is affected. Zhou Wenshuai et al. used importance sampling to generate cut-in scene test cases, which significantly increased the number and proportion of dangerous scenes and high-risk scenes. The sampling-based method is simple and direct. If the logical scene space is given, all specific scenes can be generated in theory. The idea of ​​combination testing comes from software testing. Scenarios are generated by discretizing scene parameters in the logical scene parameter space and designing combination rules. Chongqing University combines the relative positions and possible movement directions of the main vehicle and surrounding traffic participants, defines scenario rules under motion constraints, and uses the Pairwise Independent Combinatorial Testing (PICT) tool to generate a large number of combined scenario sets. Among them, a test matrix is ​​added to the combined test method, which improves the overall complexity of the generated scenario and realizes the scenario generation of the lane departure warning system considering the scenario complexity and test cost. When using the combination method to generate test scenarios, it is usually necessary to combine the ontology to analyze the combination parameters to reduce the complexity of the combination. In the related art, a scenario generation method based on three entities of highway, weather and vehicle is proposed to generate and verify the scenario in the ramp environment. The combined test method has high controllability and can quickly improve the scenario coverage, but it is not suitable for high-dimensional data generation. The deep learning-based method generates test scenarios through a large amount of training data iteration, automatically learns the intrinsic patterns and characteristics of the data, and constructs a high-fidelity and diverse test scenario library. It is mainly implemented by generative networks and neural networks, such as using recurrent neural networks (RNN) to generate accident scenarios and using GAN to generate cut-in scenarios. The scene can also be generated as an image. Deep learning-based methods can automatically learn and optimize models, but they require a large amount of data sources as support.The optimization search-based method uses the logical scenario parameter range as the solution space, usually considering the worst case scenario danger, converting the key scenario generation into an optimization problem, and establishing the objective function, solving the optimization algorithm such as Genetic Algorithm (GA), Bayesian Optimization (BO), Particle Swarm Optimization (PSO) and other methods to iteratively find the high-demand test scenarios in the test process, that is, the optimal solution of the objective function. The purpose of the optimization search-based method is to discover scenarios with high test requirements, which can realize the closed loop of generation and testing.

[0037] Through the above analysis, it can be seen that the generalization generation method of key test scenarios for autonomous driving in related technologies still has the following technical defects. First, the coverage of the test scenario library is low, and data collection and processing are difficult. Existing research relies on limited data sets, and the test scenarios are collected in specific geographical locations and traffic conditions, which makes it difficult to fully cover complex and diverse driving environments. Due to the high cost and difficulty of data collection, the test scenarios identified from natural driving data cannot effectively simulate rare traffic conditions, affecting the safety verification and performance of the autonomous driving system in unknown or rare environments. Therefore, it is essential to use the existing natural driving data set to generate more typical scenarios and expand the basic scenario library. Second, the "sparseness disaster" of key test scenarios for autonomous driving leads to low test efficiency. Dangerous scenarios and edge scenarios in the scenario library are the research focus of the test. The rarity of these key scenarios is an important factor leading to low sampling and testing efficiency. Although the frequency of occurrence is low, the impact on safety is extremely significant. Therefore, while ensuring coverage, searching for key scenarios and constructing reasonable importance functions for sampling are key steps to improve test efficiency. Third, scenario-based virtual simulation testing lacks a quantitative evaluation method for test results. Existing research determines whether key scenarios are challenging by building a simulation platform, executing tests, and recording whether there is a collision during the test process after the scenario is generated. However, the simulation test environment itself also has a certain impact on key scenarios. The possibility and severity of collisions of the same test case under different vehicle control system parameters or sensor accuracy are different, and its importance cannot be determined solely based on whether a collision occurs during the test process. Therefore, it is necessary to measure the severity of the collision and the danger of an imminent collision separately, and prioritize key scenarios, which will be conducive to exploring the vehicle danger boundary under different test methods and control algorithms in the future.

[0038] In order to overcome the defects in the above-mentioned related technologies, the present application provides a generalization generation system for key test scenarios for autonomous driving. Figure 1 For a schematic diagram of the scene generation system structure provided in the embodiment of the present application, please refer to Figure 1As shown in the figure, the scene generation system includes: encoder, decoder, generator and discriminator. The encoder is used to extract features from the input real scene information, remove redundant information, add Transformer encoder to achieve effective learning of time series data, and obtain low-dimensional scene information, which is called real potential representation; the generator and discriminator are used to generate scene information based on the input noise information through the mutual game between the generator and the discriminator, which is called generated potential representation; the decoder is used to decode the compressed real scene information and potential scene information to obtain the test evaluation scene.

[0039] Please continue to see Figure 1 As shown, the encoder in this embodiment includes an autoencoder and a Transformer encoder. This embodiment combines the autoencoder and the Transformer to encode the cut-in scene parameters into low-dimensional potential representations, simulates the long time series information in multiple cut-in scene parameters in real driving data, and reconstructs the data through the decoder. The improved Wasserstein GAN algorithm is used to learn the real potential representation and generate the most realistic potential representation, and finally decoded into the cut-in scene parameters through the decoder to expand the highway cut-in scene library.

[0040] The autoencoder includes a feature extraction layer, a gated recurrent unit, a maximum pooling layer, an average pooling layer, a fully connected layer, an activation function layer, and a Transformer layer. The Transformer encoder includes a multi-head attention mechanism module, a residual connection and layer normalization module, and a feedforward neural network module; the decoder includes a fully connected layer, a sigmoid function module, a feature extraction layer, and a gated recurrent unit. The generator is implemented using a multi-layer perceptron, and the discriminator includes multiple fully connected layers.

[0041] Among them, the real scene information includes 13 types of information, including the main vehicle speed, lane-changing vehicle speed, relative speed, longitudinal distance, lateral distance, main vehicle acceleration, lane-changing vehicle acceleration, main vehicle lateral speed, lane-changing vehicle lateral speed, lane-changing vehicle average speed, main vehicle average speed, lane-changing vehicle average acceleration and main vehicle average acceleration.

[0042] In some embodiments, since the ODD (Operational Design Domain) has been determined, static scene elements can be coordinated in simulation experiments. In order to simulate the parameter distribution of each time step and describe the average state of the vehicle at the same time, the speed, acceleration, relative distance and other parameters of the two vehicles are selected as features in this embodiment, and the average speed of the main vehicle and the lane-changing vehicle is used as an attribute to design a model that can simultaneously process static attributes and dynamic features to capture the distribution, correlation and time trend of scene parameters. The entire algorithm process is divided into three parts: encoding, decoding and generation. In the encoding and decoding process, the original data is compressed into a latent representation (LatentRepresentation) of a fixed dimension through an encoder, and the data is reconstructed from it through a decoder. In the generation part, GAN (GenerativeAdversarial Networks) is used to generate a new latent representation, synthesize a latent representation that is as realistic as possible, and then use the decoder to obtain the generated scene information.

[0043] Please continue to see Figure 1 As shown in the figure, this embodiment is improved on the basis of AEGAN, and the overall framework includes an autoencoder, a Transformer encoder, and an improved WGAN network. A Transformer encoder is also added to the encoder part of the autoencoder, so that the algorithm can better understand the feature distribution of long time series, and use the encoding result as a latent space vector, which the decoder can reconstruct into the original feature. The improved WGAN network adds a gradient penalty term during the training process to ensure the stability of the generation.

[0044] Figure 2 For a flowchart of a generalized generation method for key test scenarios for autonomous driving provided in an embodiment of the present application, please refer to Figure 2 As shown, the method includes:

[0045] S201, obtaining real scene information, inputting the real scene information into the encoder, the encoder extracts features from the input real scene information, and compresses the extracted feature information to obtain a low-dimensional potential representation to obtain important information in the high-dimensional scene. The decoder is used to reconstruct the real scene information, and a trained encoding-decoding model is obtained by minimizing the difference between the real data and the reconstructed data.

[0046] S202. Use the GAN network to transform noise into various data distributions to simulate the characteristics of high-dimensional data. A WGAN model is proposed, and the training method is improved. Through the confrontation between the generator and the discriminator, a more realistic latent space representation vector is generated.

[0047] S203: Use a trained decoder to decode the generated latent space representation vector to obtain a test evaluation scenario.

[0048] Among them, in this embodiment, the real scene information includes all of the main vehicle speed, lane-changing vehicle speed, relative speed, longitudinal distance, lateral distance, main vehicle acceleration, lane-changing vehicle acceleration, main vehicle lateral speed, lane-changing vehicle lateral speed, lane-changing vehicle average speed, main vehicle average speed, lane-changing vehicle average acceleration, and main vehicle average acceleration.

[0049] Figure 3 For a flow chart of the key scene search method provided in the embodiment of the present application, please refer to Figure 3 As shown, in one embodiment, dangerousness and exposure frequency are used as criticality functions, a search optimization function is set, and an optimization algorithm is used to search for key scenes in the logical scene space. For example, in order to evaluate the criticality of a scene, a criticality assessment method is usually required to determine the importance of the scene in the autonomous driving test. Currently, most studies only consider dangerousness, and use a series of proxy indicators such as TTC to reflect the criticality of the scene. The probability of occurrence of the scene is not considered, which makes the generation of key scenes more idealized. In this embodiment, the scene criticality V(x) is measured by exposure frequency and dangerousness.

[0050] The exposure frequency of the scene P(x) is obtained by analyzing the extracted and generated data, and the danger of the scene P(s|x) is divided into dangerous scene areas and safe scene areas by judging whether there is a collision in the agent model, and distinguished by the variables of the indicator function. The criticality function is defined as the product of the two:

[0051] V(x)=P(s|x)P(x) (1)

[0052] The algorithm in this embodiment uses the longitudinal relative speed and distance x = [d0, d0] of the two vehicles at the time of cut-in. T is a decision variable. Danger is an evaluation criterion that is considered more in the autonomous driving test scenario. The danger of subsequent time steps is unknown and cannot be directly iterated and calculated through some intuitive indicators. Therefore, an agent model is needed to simulate the traffic behavior after the vehicle cuts in and calculate the subsequent danger situation. The present invention uses the IDM model to simulate the acceleration behavior of the following vehicle during the following process:

[0053]

[0054] In the above formula (2), a is the acceleration of the vehicle, a max is the maximum acceleration of the vehicle, v is the speed of the vehicle at the current moment, and v dis the expected speed of the vehicle, L is the length of the vehicle, usually L is 4m, δ is a constant used to adjust the nonlinearity of acceleration. s*(v,d0) is the dynamic safety distance, which is used to calculate the minimum safety distance to the vehicle in front.

[0055]

[0056] In the above formula, s0 represents the distance between the vehicle and the vehicle in front when the vehicle is stationary, T is the safe headway, and b is the comfortable deceleration.

[0057] In some embodiments, in order to avoid unreasonable vehicle speed and acceleration and make the behavior of the host vehicle more consistent with human driving characteristics, this embodiment also sets the following restrictions during the test process:

[0058] v min ≤v≤v max ,a min ≤a≤a max (4)

[0059] Among them, v min The value is 1m / s, v max The value is 40m / s, a min and a max The values ​​are -4m / s 2 and 2m / s 2 In the IDM model, if the distance between two vehicles is less than the threshold ddcc, it is considered that the vehicles have collided, and the danger indicator function P(s|x) is 1. Specifically, the danger indicator function can be expressed as:

[0060]

[0061] In this embodiment, the exposure frequency of the scene refers to the probability of occurrence of the scene analyzed by reference from the data set. The speed difference and relative distance in the scene library are analyzed, and discrete steps of 1km / h and 1m are taken to form a logical scene space. The probability of occurrence of each discretized parameter combination in the logical scene space can be counted, thereby analyzing the scene exposure probability. After the criticality function is established, it is necessary to find the key scene through an optimization algorithm. However, since most scenes in the scene space are not critical, if the key scene is directly searched according to the criticality function, the optimization problem will degenerate into a random search, which is less efficient. Therefore, in order to efficiently search for key scenes, it is necessary to provide a search direction for the optimization algorithm, that is, to design an objective function, and to find all key scenes according to the set search direction.

[0062] In some embodiments, ETTC (advanced time to collision) and the distance of the high-frequency scene set can be used as optimization targets, which can be specifically expressed as:

[0063]

[0064] Among them, U is the normalization factor.

[0065] In some embodiments, during driving, it is necessary to pay attention to the npETTC value at the most dangerous moment, and take the minimum npETTC value as the optimization target, which can be specifically expressed as:

[0066]

[0067]

[0068] In the above formula, x and y represent a search scenario and a common scenario respectively, dim represents the decision variable dimension, and U i Represents the normalization factor, which can be calculated by calculating the farthest distance between the scene in the dataset and the common scene set Ω.

[0069] Among them, the constructed objective function contains two index weights, and the comprehensive objective function can be expressed as:

[0070]

[0071] In order to find the optimal key scene search algorithm, different optimization search methods are used to find candidate scenes. Since the candidate scenes are searched based on ETTC, they may not collide in IDM simulation. Therefore, the criticality threshold is used to further determine whether the searched scene is a key scene. The 4-neighborhood flood filling algorithm can be selected to start from the local key scene and search for other key scenes.

[0072] In this embodiment, in order to further verify the adaptability and versatility of key scenarios in autonomous driving tests, the basic knowledge of accelerated testing based on Monte Carlo and importance sampling is used to verify the accelerated testing effect of key scenarios in IDM simulation. An entry scenario is built based on the PreScan / MATLAB / Simulink joint simulation platform, and the ACC system is configured for the main vehicle to evaluate the generalization ability of the simulation test of key scenarios on the ACC system.

[0073] Among them, non-critical scenarios in the scenario space are explored with a probability of 5%, the Monte Carlo method is used to sample from the extended scenario library, the generated key scenarios are sampled based on the importance sampling method, and the IDM model is tested in a simulation environment to estimate the accident rate and the minimum number of tests, verifying the effectiveness of the key scenario generation method that considers the probability of occurrence and dangerousness for accelerated testing.

[0074] In order to test the scalability of the key scenarios obtained for the autonomous driving ACC system, the key scenarios searched in the scene space can be tested one by one. Considering the test requirements, PreScan / MATLAB / Simulink can be selected to build the adaptive cruise control system of the main vehicle.

[0075] In order to analyze the test results of key scenarios more accurately, the test results of key scenarios are divided into collision and non-collision. When no collision occurs, the minimum TTC (Time To Collision) is used to measure the danger. When a collision occurs, the severity of the vehicle in the collision is considered. It is assumed that the masses of the two vehicles in the scenario are equal, plastic deformation occurs after the collision, and they move forward together. A completely inelastic collision is used to evaluate the final speed v of the main vehicle in the collision. m , calculate the change between the final speed and the speed before the collision to define the impact speed difference Δv. According to the research on vehicle forward collision warning, the severity is related to the speed change during the collision event, and the specific calculation is as follows:

[0076]

[0077] Δv=vv m (11)

[0078]

[0079] In the above formula, v represents the current speed of the vehicle, S represents the severity of the collision, and the larger the S value is, the more severe the collision is. β0 is generally -6.068, β1 is generally 0.1, and β2 is generally -0.623.

[0080] The test results of this embodiment can be designed as a piecewise function. The test result is a positive value when a collision occurs, otherwise it is a negative value. The larger the Y value, the higher the potential danger of the key test scenario. The more it is necessary to improve the safety of autonomous driving in these key test scenarios to avoid possible harm and thus reduce accidents. The specific test results can be expressed as:

[0081]

[0082] In one embodiment, this embodiment also provides a computer-readable storage medium storing a computer program, which, when running on a processor, executes the generalization generation method for key test scenarios for autonomous driving as described above.

[0083] In one embodiment, this embodiment also provides a computer program product, which includes: computer program code, when the computer program code is run on a computer, the computer executes the generalization generation method of key test scenarios for autonomous driving as described above.

[0084] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid state drive (SSD).

[0085] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the embodiments of the present application.

[0086] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0087] In the several embodiments provided in the embodiments of the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0088] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0089] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0090] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present application is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a memory (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (RandomAccess Memory, RAM), disk or optical disk and other media that can store program code.

[0091] The above is only a specific implementation of the embodiment of the present application, but the protection scope of the embodiment of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or replacements within the technical scope disclosed in the embodiment of the present application, which should be included in the protection scope of the embodiment of the present application. Therefore, the protection scope of the embodiment of the present application should be based on the protection scope of the claims.

Claims

1. A generalized generation system for key test scenarios for autonomous driving, characterized in that: include: Encoder, decoder, generator and discriminator; The encoder is used to extract features from the input real scene information and encode the extracted feature information to obtain encoded real scene information; The generator and the discriminator are used to generate potential scene information through the mutual game between the generator and the discriminator according to the input noise information; The decoder is used to decode the real scene information and the potential scene information to obtain a test evaluation scene.

2. The generalization generation system for key test scenarios for autonomous driving according to claim 1, characterized in that: The encoder includes an autoencoder and a Transformer encoder; The autoencoder includes a feature extraction layer, a gated recurrent unit, a maximum pooling layer, an average pooling layer, a fully connected layer, an activation function layer, and a Transformer layer; The Transformer encoder includes a multi-head attention mechanism module, a residual connection and layer normalization module, and a feedforward neural network module; The decoder includes a fully connected layer, a Sigmoid function module and a gated recurrent unit. The generator is implemented using a multi-layer perceptron, and the discriminator consists of multiple fully connected layers.

3. The generalization generation system for key test scenarios for autonomous driving according to claim 1 or 2, characterized in that: The real scene information includes the main vehicle speed, the lane-changing vehicle speed, the relative speed, the longitudinal distance, the lateral distance, the main vehicle acceleration, the lane-changing vehicle acceleration, the main vehicle lateral speed, the lane-changing vehicle lateral speed, the lane-changing vehicle average speed, the main vehicle average speed, the lane-changing vehicle average acceleration and the main vehicle average acceleration.

4. A generalized generation method for key test scenarios for autonomous driving, characterized in that: include: Acquire real scene information, and input the real scene information into an encoder, wherein the encoder extracts features from the input real scene information and encodes the extracted feature information to obtain encoded real scene information; Acquire noise information, and transform the noise information through the game between the generator and the discriminator to generate potential scene information; A decoder is used to decode the real scene information and the potential scene information to obtain a test evaluation scene.

5. The generalization generation method of key test scenarios for autonomous driving according to claim 4, characterized in that: The real scene information includes the main vehicle speed, the lane-changing vehicle speed, the relative speed, the longitudinal distance, the lateral distance, the main vehicle acceleration, the lane-changing vehicle acceleration, the main vehicle lateral speed, the lane-changing vehicle lateral speed, the lane-changing vehicle average speed, the main vehicle average speed, the lane-changing vehicle average acceleration and the main vehicle average acceleration.

6. The method for generalizing and generating key test scenarios for autonomous driving according to claim 4 or 5, characterized in that: Also includes: Taking hazard and exposure frequency as criticality functions, a search optimization function is set, and an optimization algorithm is used to search for key scenes in the logical scene space.

7. The generalization generation method of key test scenarios for autonomous driving according to claim 6, characterized in that: Also includes: Based on the key scenarios found, a simulation test scenario including road elements, meteorological elements and traffic elements is constructed in the simulation platform; An adaptive cruise control algorithm is used to implement a simulation test of the vehicle in the simulation test scenario.

8. A computer-readable storage medium, characterized in that: It stores a computer program, which, when running on a processor, executes the generalization generation method for key test scenarios for autonomous driving as described in any one of claims 4 to 7.

9. A computer program product, characterized in that The computer program product comprises: a computer program code, which, when executed on a computer, enables the computer to execute the method for generalizing and generating key test scenarios for autonomous driving according to any one of claims 4 to 7.

Images