Training method, device and electronic equipment for autonomous driving system

Through the instantiation of multiple virtual driver models and shadow layer group linkage simulation of driving, diversified training data is generated, which solves the problem of insufficient adaptability and robustness of the autonomous driving system in the prior art, and improves the face of complex environments Adaptability.

CN118350407BActive Publication Date: 2024-11-22CHONGQING SELIS PHOENIX INTELLIGENT INNOVATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410359136.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-27
Publication Date
2024-11-22
Estimated Expiration
2044-03-27

AI Technical Summary

Technical Problem

Existing autonomous driving system training methods usually adopt a single shadow mode, resulting in insufficient adaptability and robustness, and the inability to effectively deal with complex and changeable driving scenarios.

Method used

By selecting multiple virtual driver models from the pre-built virtual driver model library, instantiating and linking different shadow layer groups for simulated driving, generating diversified simulated driving data to improve the adaptability and resilience of the autonomous driving system. Great.

Benefits of technology

A comprehensive consideration of the interaction between different driving scenarios, behaviors and decision-making processes is achieved, and the adaptability and robustness of autonomous driving systems in complex environments can be improved, so as to better handle the rich and changeable drivers in the real world. Habits and needs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118350407B_ABST
    Figure CN118350407B_ABST
Patent Text Reader

Abstract

The present application relates to a training method, device and electronic device for an automatic driving system, the method comprising: selecting N virtual driver models from a pre-built virtual driver model library, and instantiating each virtual driver model in the N virtual driver models to obtain N virtual driver instances, wherein the N virtual driver models correspond to the N virtual driver instances one by one, and the N virtual driver instances are used to simulate the driving characteristics and decision logic of the N drivers, and N is an integer greater than 1; determining N shadow layer groups corresponding to the N virtual driver instances; starting the N virtual driver instances, and respectively linking the shadow layer groups corresponding to each virtual driver instance to simulate driving, to obtain N simulated driving data; using the N simulated driving data as training data to train the automatic driving system. In this way, the purpose of improving the adaptability and robustness of the automatic driving system can be achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of autonomous driving technology, and in particular to a training method, device and electronic equipment for an autonomous driving system. Background Art

[0002] Autonomous driving, also known as unmanned driving, computer driving or wheeled mobile robots, is a cutting-edge technology that relies on computers and artificial intelligence technology to complete complete, safe and effective driving without human control. However, autonomous driving also faces many challenges, such as bad weather, complex traffic conditions, etc., so its performance in some scenarios may not match that of human drivers. In order to solve this problem, more testing and iterative updates of the autonomous driving system are needed to improve its processing capabilities in complex scenarios.

[0003] However, the existing technology usually uses a single shadow mode to train a single virtual driver model when training an autonomous driving system. This method is prone to lack of comprehensive consideration of the interactions between different driving scenarios, behaviors, and decision-making processes, resulting in insufficient adaptability of the trained autonomous driving system when facing actual complex environments. It is also easy to ignore the importance of different driving styles and decision-making processes in driving behaviors, making the trained autonomous driving system incapable of coping with the rich and varied habits and needs of drivers in the real world. Therefore, how to improve the adaptability and robustness of the autonomous driving system has become a technical problem that needs to be solved urgently. Summary of the invention

[0004] The present application provides a training method, device and electronic equipment for an autonomous driving system to solve the problem in the prior art that a single shadow mode is usually used to train a single virtual driver model, resulting in poor adaptability and robustness of the trained autonomous driving system.

[0005] In a first aspect, an embodiment of the present application provides a training method for an autonomous driving system, the method comprising:

[0006] Selecting N virtual driver models from a pre-built virtual driver model library, and instantiating each virtual driver model in the N virtual driver models to obtain N virtual driver instances, wherein the N virtual driver models correspond one-to-one to the N virtual driver instances, and the N virtual driver instances are used to simulate driving characteristics and decision logic of the N drivers, where N is an integer greater than 1;

[0007] Determine N shadow layer groups corresponding to the N virtual driver instances, wherein the N virtual driver instances correspond one-to-one to the N shadow layer groups, each shadow layer group includes at least one shadow layer, and the at least one shadow layer is selected from a pre-built multi-level shadow mode architecture, the multi-level shadow mode architecture includes multiple shadow layers, and different shadow layers are used to simulate different functional modules in the autonomous driving system;

[0008] Starting the N virtual driver instances, and respectively linking the shadow layer groups corresponding to the virtual driver instances to perform simulated driving, to obtain N simulated driving data;

[0009] The N simulated driving data are used as training data to train the automatic driving system.

[0010] Optionally, instantiating each virtual driver model in the N virtual driver models to obtain N virtual driver instances includes:

[0011] Acquire a model category and a driving attribute corresponding to each virtual driver model in the N virtual driver models, wherein each model category has two mutually opposing driving attributes;

[0012] Determining N initialization parameter vectors corresponding to the N virtual driver models according to the model category and the driving attribute, wherein the N virtual driver models correspond one-to-one to the N initialization parameter vectors, and the initialization parameter vectors of two virtual driver models belonging to the same model category have the same parameter category, and the parameter values ​​corresponding to the same parameter category are different;

[0013] The N virtual driver models are instantiated using the N initialization parameter vectors to obtain the N virtual driver instances.

[0014] Optionally, before selecting N virtual driver models from a pre-built virtual driver model library, the method further includes:

[0015] Constructing M virtual driver model groups, wherein each of the M virtual driver model groups corresponds to a model category, each of the M virtual driver model groups includes two virtual driver models with the same model category and different driving attributes, M is an integer greater than or equal to 1, and M is equal to N / 2;

[0016] Each virtual driver model in the M virtual driver model groups is named according to a preset naming rule to obtain the virtual driver model library, wherein the preset naming rule is set based on the model category and driving attributes corresponding to each virtual driver model group.

[0017] Optionally, the shadow layer groups corresponding to the virtual driver instances are linked to perform simulated driving to obtain N simulated driving data, including:

[0018] Obtaining a target shadow layer group corresponding to a target virtual driver instance, wherein the target virtual driver instance is any instance of the N virtual driver instances, and the target shadow layer group is a shadow layer group corresponding to the target virtual driver instance among the N shadow layer groups;

[0019] Determine, according to a preset corresponding relationship, a parameter set of each shadow layer of the target shadow layer group for the initialization parameter vector corresponding to the target virtual driver instance, wherein the preset corresponding relationship is used to characterize the corresponding relationship between the initialization parameter vector of each virtual driver instance and the parameter set of each shadow layer;

[0020] Assigning a parameter set corresponding to each shadow layer of the target shadow layer group;

[0021] The target shadow layer group after linkage assignment is subjected to simulated driving to obtain target simulated driving data corresponding to the target virtual driver instance.

[0022] Optionally, the using the N simulated driving data as training data to train the automatic driving system includes:

[0023] Taking the first type of simulated driving data among the N simulated driving data as negative samples, and taking the second type of simulated driving data among the N simulated driving data as positive samples, wherein the first type of simulated driving data is simulated driving data corresponding to a virtual driver model with negative driving attributes, and the second type of simulated driving data is simulated driving data corresponding to a virtual driver model with positive driving attributes;

[0024] The automatic driving system is trained using the negative samples and the positive samples.

[0025] Optionally, after using the N simulated driving data as training data to train the autonomous driving system, the method further includes:

[0026] Performing a performance evaluation on the trained autonomous driving system to obtain a performance evaluation result, wherein the performance evaluation result includes at least one of accuracy, recall, and F1 score;

[0027] According to the performance evaluation result, the model parameters of the N virtual driver models are adjusted.

[0028] Optionally, before determining the N shadow layer groups corresponding to the N virtual driver instances, the method further includes:

[0029] Dividing the autonomous driving system according to functions to obtain multiple functional modules;

[0030] Determining input data and output data of each functional model in the plurality of functional modules, and standardizing input interfaces and output interfaces of each functional model;

[0031] Construct the network layer corresponding to each functional model in the multiple functional modules to obtain the multi-level shadow mode architecture, wherein the network layer corresponding to each functional model corresponds to a shadow layer, and the data between the network layers corresponding to different functional models interact according to a preset interaction mechanism.

[0032] Optionally, the multi-level shadow mode architecture includes an environment scene layer, a basic driving layer, a thinking decision layer, an emergency response layer, and a human-computer interaction layer;

[0033] The environment scene layer is used to simulate driving strategies under various road conditions and different environments based on the environment data in the real driving scene;

[0034] The basic driving layer is used to simulate basic driving decisions according to real-time road information and vehicle status, and the basic driving decisions include control of the vehicle power system and control of the steering system;

[0035] The thinking decision layer is used to simulate high-level driving decisions based on the output data of the basic driving layer, the preset traffic rules database and the real-time road information, wherein the high-level driving decisions include at least one of determining a safe overtaking window, determining the best time and location for implementing lane changes, and avoiding potential risk areas in advance;

[0036] The emergency response layer is used to simulate driving decisions in emergency situations based on the output data of the thinking decision layer, emergency event information and a preset traffic rules database, wherein the driving decisions in emergency situations include at least one of emergency braking, emergency obstacle avoidance and path correction in abnormal situations;

[0037] The human-computer interaction layer is used to simulate human-computer interaction decision-making based on user interaction data.

[0038] In a second aspect, an embodiment of the present application further provides a training device for an autonomous driving system, the device comprising:

[0039] an instantiation module, used to select N virtual driver models from a pre-built virtual driver model library, and instantiate each virtual driver model in the N virtual driver models to obtain N virtual driver instances, wherein the N virtual driver models correspond one-to-one to the N virtual driver instances, and the N virtual driver instances are used to simulate the driving characteristics and decision logic of the N drivers, where N is an integer greater than 1;

[0040] A first determination module is used to determine N shadow layer groups corresponding to the N virtual driver instances, wherein the N virtual driver instances correspond to the N shadow layer groups one by one, each shadow layer group includes at least one shadow layer, and the at least one shadow layer is selected from a pre-built multi-level shadow mode architecture, and the multi-level shadow mode architecture includes multiple shadow layers, and different shadow layers are used to simulate different functional modules in the autonomous driving system;

[0041] A simulated driving module, used to start the N virtual driver instances, and respectively link the shadow layer groups corresponding to the virtual driver instances to perform simulated driving, so as to obtain N simulated driving data;

[0042] A training module is used to train the automatic driving system by using the N simulated driving data as training data.

[0043] In a third aspect, an embodiment of the present application further provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0044] Memory, used to store computer programs;

[0045] The processor is used to implement the training method of the autonomous driving system described in any one of the first aspects when executing the program stored in the memory.

[0046] The above technical solution provided by the embodiment of the present application has the following advantages over the prior art: the method provided by the embodiment of the present application, by selecting N virtual driver models from a pre-built virtual driver model library, and instantiating each virtual driver model in the N virtual driver models, to obtain N virtual driver instances, wherein the N virtual driver models correspond one-to-one to the N virtual driver instances, and the N virtual driver instances are used to simulate the driving characteristics and decision logic of N drivers, and N is an integer greater than 1; determining N shadow layer groups corresponding to the N virtual driver instances, wherein the N virtual driver instances correspond one-to-one to the N shadow layer groups, each shadow layer group includes at least one shadow layer, and the at least one shadow layer is selected from a pre-built multi-level shadow mode architecture, and the multi-level shadow mode architecture includes multiple shadow layers, and different shadow layers are used to simulate different functional modules in the automatic driving system; starting the N virtual driver instances, and respectively linking the shadow layer groups corresponding to each virtual driver instance to perform simulated driving, to obtain N simulated driving data; using the N simulated driving data as training data to train the automatic driving system. Through the above method, when training multiple different virtual driver instances, different shadow layer groups can be selected for linkage training according to the training focus, so that different functional modules in the autonomous driving system can be simulated to achieve comprehensive consideration of the interactions between different driving scenarios, behaviors and decision-making processes, so that the trained autonomous driving system has improved adaptability when facing actual complex environments; and, multiple virtual driver instances can be trained at the same time to obtain diversified simulated driving data, so that the autonomous driving system trained based on diversified simulated driving data can handle the rich and varied driver habits and needs in the real world with ease, thereby achieving the purpose of improving the adaptability and robustness of the autonomous driving system. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0049] One or more embodiments are exemplarily described by pictures in the corresponding drawings, and these exemplified descriptions do not constitute limitations on the embodiments. Elements with the same reference numerals in the drawings represent similar elements, and unless otherwise stated, the figures in the drawings do not constitute proportional limitations.

[0050] Figure 1 A flowchart of a training method for an autonomous driving system provided in an embodiment of the present application;

[0051] Figure 2 A schematic diagram of a training process of an autonomous driving system provided in an embodiment of the present application;

[0052] Figure 3 A schematic diagram of the structure of a training device for an autonomous driving system provided in an embodiment of the present application;

[0053] Figure 4 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0055] The disclosure below provides many different embodiments or examples to implement different structures of the present invention. In order to simplify the disclosure of the present invention, the parts and settings of specific examples are described below. Of course, they are only examples, and the purpose is not to limit the present invention. In addition, the present invention can repeat reference numbers and / or letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed.

[0056] In order to solve the technical problem that in the prior art, a single shadow mode is usually used to train a single virtual driver model, resulting in poor adaptability and robustness of the trained autonomous driving system, the present application provides a training method, device and electronic equipment for an autonomous driving system, which can improve the adaptability and robustness of the autonomous driving system.

[0057] See also Figure 1 , Figure 1 A flowchart of a training method for an autonomous driving system provided in an embodiment of the present application. Figure 1 As shown, the training method of the automatic driving system may include the following steps:

[0058] Step 101: select N virtual driver models from a pre-built virtual driver model library, and instantiate each of the N virtual driver models to obtain N virtual driver instances, wherein the N virtual driver models correspond one-to-one to the N virtual driver instances, and the N virtual driver instances are used to simulate the driving characteristics and decision-making logic of the N drivers, and N is an integer greater than 1.

[0059] Specifically, the value of N can be set according to actual needs, and this embodiment does not specifically limit it, such as 2, 3, 4, etc. Each virtual driver model is constructed based on the behavioral statistics, physiological and psychological characteristics, and driving habits of real-world drivers. After the construction, the initialization parameters of each virtual driver model can be assigned to obtain virtual driver instances with different driving characteristics and decision logics.

[0060] Step 102: determine N shadow layer groups corresponding to the N virtual driver instances, wherein the N virtual driver instances correspond one-to-one to the N shadow layer groups, each shadow layer group includes at least one shadow layer, and the at least one shadow layer is selected from a pre-built multi-level shadow mode architecture, and the multi-level shadow mode architecture includes multiple shadow layers, and different shadow layers are used to simulate different functional modules in the autonomous driving system.

[0061] Specifically, the shadow layer group corresponding to the virtual driver instance can be understood as a shadow layer or a combination of multiple shadow layers associated with the virtual driver instance. The shadow layer groups corresponding to different virtual driver instances can be the same or different. For example, assuming that the multi-level shadow mode architecture includes shadow layer 1, shadow layer 2, shadow layer 3, and shadow layer 4, then the shadow layer group corresponding to each virtual driver instance can be any one of the four shadow layers or a combination of multiple shadow layers.

[0062] Step 103: Start N virtual driver instances, and respectively link the shadow layer groups corresponding to the virtual driver instances to perform simulated driving, to obtain N simulated driving data.

[0063] After obtaining the virtual driver instances and the shadow layer groups corresponding to each virtual driver instance, each virtual driver instance can be started, and the shadow layer groups corresponding to each virtual driver instance can be linked to perform simulated driving to obtain simulated driving data corresponding to each virtual driver instance.

[0064] It should be noted that since different virtual driver models are constructed based on the behavioral statistics, physiological and psychological characteristics, and driving habits of different drivers in the real world, the simulated driving data obtained using different virtual driver instances can reflect the behavioral statistics, physiological and psychological characteristics, and driving habits of different drivers, such as experienced, energetic, and psychologically confident, inexperienced, not energetic, and not psychologically confident, etc.

[0065] Step 104: Use the N simulated driving data as training data to train the automatic driving system.

[0066] After obtaining the simulated driving data corresponding to each virtual driver instance, the simulated driving data corresponding to each virtual driver instance can be used as training data to train the automatic driving system.

[0067] It should be noted that the autonomous driving system may be a model constructed using algorithms such as deep learning and reinforcement learning. In the process of training the autonomous driving system using simulated driving data corresponding to each virtual driver instance, the decision-making and control capabilities of the autonomous driving system can be continuously optimized.

[0068] In this embodiment, when training each virtual driver instance, different shadow layer groups can be selected for linkage training according to the training focus, so that different functional modules in the autonomous driving system can be simulated to achieve comprehensive consideration of the interactions between different driving scenarios, behaviors and decision-making processes, so that the trained autonomous driving system has improved adaptability when facing actual complex environments; and, it is also possible to train multiple virtual driver instances at the same time to obtain diversified simulated driving data, so that the autonomous driving system trained based on diversified simulated driving data can handle the rich and varied driver habits and needs in the real world with ease, thereby achieving the purpose of improving the adaptability and robustness of the autonomous driving system.

[0069] Furthermore, the above step 101, instantiating each virtual driver model in the N virtual driver models to obtain N virtual driver instances, includes:

[0070] Obtaining a model category and a driving attribute corresponding to each virtual driver model in the N virtual driver models, wherein each model category has two mutually opposing driving attributes;

[0071] Determine N initialization parameter vectors corresponding to the N virtual driver models according to the model category and the driving attribute, wherein the N virtual driver models correspond to the N initialization parameter vectors one-to-one, and the initialization parameter vectors of two virtual driver models belonging to the same model category have the same parameter category, and the parameter values ​​corresponding to the same parameter category are different;

[0072] N virtual driver models are instantiated using N initialization parameter vectors to obtain N virtual driver instances.

[0073] Specifically, the number of the above-mentioned model categories can be set according to actual needs, and the embodiments of the present application are not specifically limited. Specifically, the model categories can be defined from different dimensions such as driving experience, energy level, psychological state, driving style, road adaptability, driving concentration, night driving experience, etc. Each model category has two opposing driving attributes. For example, the model category for driving experience includes two driving attributes: rich driving experience and insufficient driving experience; the model category for energy level includes two driving attributes: energetic and insufficient energy.

[0074] In the process of instantiating each virtual driver model, the model category and driving attributes corresponding to each virtual driver model can be obtained, and then the initialization parameter vector corresponding to each virtual driver model can be determined according to the model category and driving attributes corresponding to each virtual driver model. Then, the initialization parameter vector corresponding to each virtual driver model is used to instantiate the corresponding virtual driver model to obtain the virtual driver instance corresponding to each virtual driver model.

[0075] For example, assume that the nth type of driver model among N virtual driver models is Pn, where n is a positive integer less than or equal to N, and different n represents different model categories. Under this model category, a virtual driver model with negative driving attributes (such as dangerous / inexperienced, etc.) is Pn-a, and a virtual driver model with positive driving attributes (such as safe / experienced, etc.) is Pn-b, where a and b represent different driving attributes, respectively.

[0076] For the virtual driver model Pn-a, set its initialization parameter vector to Θ n a , which will initialize the parameter vector Θ n a By assigning the initialization parameters in to the virtual driver model Pn-a, a virtual driver instance of the virtual driver model Pn-a can be obtained.

[0077] For the virtual driver model Pn-b, set its initialization parameter vector to Θ n b , which will initialize the parameter vector Θ n b By assigning the initialization parameters in to the virtual driver model Pn-b, a virtual driver instance of the virtual driver model Pn-b can be obtained.

[0078] It should be noted that the initialization parameter vector may include features such as driving experience En, energy level Fn, and psychological state Mn. n a and θ n b The same parameter category, but the corresponding parameter values ​​of the same parameter category are different. For example, for a virtual driver model with rich driving experience in the model category of driving experience, the driving experience En value in its initialization parameter vector is higher; for a virtual driver model with insufficient driving experience in the model category of driving experience, the driving experience En value in its initialization parameter vector is lower. For another example, for an energetic virtual driver model in the model category of energy level, the energy level Fn value in its initialization parameter vector is higher; for a virtual driver model with insufficient energy in the model category of energy level, the energy level Fn value in its initialization parameter vector is lower.

[0079] In this implementation, by setting multiple model categories, and the two virtual driver models under each model category have two mutually opposing driving attributes, each virtual driver instance can be made more diversified, such as an experienced driver instance accustomed to long-distance driving, a novice driver instance, a driver instance good at night driving, and a driver instance trapped in a fatigued state, etc. These virtual driver instances perform realistic simulated driving in their respective corresponding shadow layer groups based on their unique driving experience and driving style, forming a variety of training scenarios and realizing randomized multimodal training.

[0080] Furthermore, before the above step 101, selecting N virtual driver models from the pre-built virtual driver model library, the method further includes:

[0081] Constructing M virtual driver model groups, wherein each virtual driver model group in the M virtual driver model groups corresponds to a model category, each virtual driver model group in the M virtual driver model groups includes two virtual driver models with the same model category and different driving attributes, M is an integer greater than or equal to 1, and M is equal to N / 2;

[0082] Each virtual driver model in the M virtual driver model groups is named according to a preset naming rule to obtain a virtual driver model library, wherein the preset naming rule is set based on the model category and driving attributes corresponding to each virtual driver model group.

[0083] Specifically, the value of M can be set according to actual needs, and this embodiment does not specifically limit it, such as 1, 2, 3, etc. When the value of M is 1, a group of two virtual driver models with the same model category and different driving attributes are constructed; when the value of M is greater than 1, multiple groups of multiple virtual driver models with the same model category and different driving attributes are constructed.

[0084] After constructing M virtual driver model groups, each virtual driver model in the M virtual driver model groups can be named according to the preset naming rules to obtain a virtual driver model library. For example, suppose there are M different model categories, each model category is identified by an integer n, where n∈{1,2,...,M}. For each category n, there are two attribute settings: a represents a driver with potential risk factors (negative samples), and b represents a model driver who drives safely (positive samples). In this way, a binary function M(n,a) and M(n,b) can be defined to represent the model that selects a specific category and attribute:

[0085] M(n,a)=P n -a for negative samples

[0086] M(n,b)=P n -b for positive samples

[0087] That is to say, when building a virtual driver model library, different types of driver behavior characteristics can be distinguished according to a set of standardized naming rules. Through this rule, a series of virtual driver models covering different driver characteristic dimensions can be systematically constructed to provide comprehensive and rich simulated driving data for the training, optimization and adaptability research of the autonomous driving system. The rule is defined as follows:

[0088] The model naming follows the format of "Pn-a" and "Pn-b", where "n" represents a continuous integer number that identifies the different categories to which the model belongs. n = 1, 2, 3... For each category "n", "a" and "b" represent the opposing driving attributes within the model category. The "Pn-a" model represents the behavior of drivers in this category who are typically dangerous, inexperienced, lack energy, and are not psychologically confident, and are considered negative samples. The "Pn-b" model represents the behavior of drivers in this category who show safe driving habits, are experienced, energetic, and psychologically confident, and are considered positive samples.

[0089] For example, assuming that n=1 represents the experience class and n=2 represents the energy class, then P1-a represents an inexperienced virtual driver model, while P1-b represents an experienced virtual driver model; P2-a represents a virtual driver model that is not energetic, and P2-b represents an energetic virtual driver model.

[0090] As an optional implementation, the virtual driver models included in the virtual driver model library may be shown in the following table:

[0091]

[0092]

[0093]

[0094] Through the above method, each virtual driver model can be named according to the preset naming rules, and a virtual driver model library can be constructed, so that the corresponding model can be quickly located according to the name of the virtual driver model. In addition, by constructing a comprehensive and detailed virtual driver model library, a series of virtual driver models with distinct characteristics and different driving decision logics can be included, which can finely portray the unique driving characteristics and decision logic of various drivers in different situations, thereby deeply simulating the driving diversity of the real world and improving the generalization ability and safety of the autonomous driving system.

[0095] Furthermore, in the above step 103, each shadow layer group corresponding to each virtual driver instance is linked to perform simulated driving to obtain N simulated driving data, including:

[0096] Obtain a target shadow layer group corresponding to a target virtual driver instance, wherein the target virtual driver instance is any instance of N virtual driver instances, and the target shadow layer group is a shadow layer group corresponding to the target virtual driver instance among the N shadow layer groups;

[0097] According to a preset corresponding relationship, determining a parameter set of each shadow layer of the target shadow layer group for the initialization parameter vector corresponding to the target virtual driver instance, wherein the preset corresponding relationship is used to characterize the corresponding relationship between the initialization parameter vector of each virtual driver instance and the parameter set of each shadow layer;

[0098] Assigning values ​​to the parameter sets corresponding to the shadow layers of the target shadow layer group;

[0099] The target shadow layer group after linkage assignment is subjected to simulated driving to obtain target simulated driving data corresponding to the target virtual driver instance.

[0100] In one embodiment, for any virtual driver instance, the shadow layer group corresponding to the virtual driver instance can be obtained first, and then the parameter set of each shadow layer of the shadow layer group of the initialization parameter vector corresponding to the virtual driver instance is determined according to the preset corresponding relationship. A , L B , L C , L D and L E Respectively represent the parameter sets of shadow layer A, shadow layer B, shadow layer C, shadow layer D and shadow layer E in the multi-level shadow mode architecture. The corresponding parameters can be pre-assigned to the selected virtual driver instance Pn-a or Pn-b in each shadow layer, and the corresponding relationship between the initialization parameter vector of each virtual driver instance and the parameter set of each shadow layer can be expressed by the following formula:

[0101] L A n =f A (Θ n a ) or L A n =f A (Θ n b );

[0102] L B n =f B (Θ n a ) or L B n =f B (Θ n b );

[0103] L C n =f C (Θ n a ) or L C n =f C (Θ n b );

[0104] L D n =f D (Θ n a ) or L D n =f D (Θ n b );

[0105] LE n =f E (Θ n a ) or L E n =f E (Θ n b );

[0106] Among them, f A represents the mapping function corresponding to the shadow layer A, f B represents the mapping function corresponding to the shadow layer B, f C Represents the mapping function corresponding to the shadow layer C, f D represents the mapping function corresponding to the shadow layer D, f E Represents the mapping function corresponding to the shadow layer E. The above preset function is used to map the common parameters of the virtual driver instance to the specific parameter set required by each shadow layer. For example, for the virtual driver instance Pn-a, the common parameter is Θ n a , through the mapping function f A , get the shadow layer A specific parameter set L A n .

[0107] After determining the parameter set of each shadow layer of the shadow layer group for the initialization parameter vector corresponding to the virtual driver instance, the parameter set corresponding to each shadow layer of the shadow layer group can be assigned, and then the shadow layer group after the assignment is linked to perform simulated driving to obtain the simulated driving data corresponding to the virtual driver instance. Specifically, assuming that the virtual driver instance Pn-a needs to link five shadow layers, namely shadow layer A, shadow layer B, shadow layer C, shadow layer D and shadow layer E, the linkage relationship can be expressed by the following formula:

[0108]

[0109] in, Indicates linkage operation, which is used to make each shadow layer interact with each other during simulation. In specific implementation, each shadow layer can be selected and combined.

[0110] For another example, assuming that the virtual driver instance Pn-a needs to link shadow layers B, C, and E, the following formula can be used to express the linkage relationship:

[0111]

[0112] in, Represents a linkage operation, which is used to make each shadow layer interact with each other during the simulation process.

[0113] In this embodiment, the shadow layers can be flexibly combined according to the training focus of each virtual driver instance. This means that in the process of training the virtual driver instance, the weight and configuration of each shadow layer can be adjusted and optimized according to actual needs, thereby ensuring that the training results are more in line with the needs of specific scenarios or tasks. This flexibility makes the training of the autonomous driving system more accurate and efficient.

[0114] Furthermore, the above step 104, using the N simulated driving data as training data to train the automatic driving system, includes:

[0115] The first type of simulated driving data among the N simulated driving data is used as a negative sample, and the second type of simulated driving data among the N simulated driving data is used as a positive sample, wherein the first type of simulated driving data is the simulated driving data corresponding to the virtual driver model with negative driving attributes, and the second type of simulated driving data is the simulated driving data corresponding to the virtual driver model with positive driving attributes;

[0116] Use negative and positive samples to train the autonomous driving system.

[0117] Specifically, the first type of simulated driving data is the simulated driving data corresponding to the virtual driver model with negative driving attributes (such as the Pn-a model in the above example), and the second type of simulated driving data is the simulated driving data corresponding to the virtual driver model with positive driving attributes (such as the Pn-b model in the above example). The simulated driving data here includes but is not limited to vehicle status, driving behavior, environmental perception information, etc.

[0118] In one embodiment, the first type of simulated driving data among N simulated driving data can be used as negative samples, and the second type of simulated driving data among the N simulated driving data can be used as positive samples to train the autonomous driving system, so that the autonomous driving system can gradually approach the positive samples and more realistically simulate the behavioral habits of the driver corresponding to the positive samples.

[0119] Furthermore, after the above step 104, using the N simulated driving data as training data to train the automatic driving system, the method further includes:

[0120] Performing a performance evaluation on the trained autonomous driving system to obtain a performance evaluation result, wherein the performance evaluation result includes at least one of accuracy, recall, and F1 score;

[0121] According to the performance evaluation results, the model parameters of the N virtual driver models are adjusted.

[0122] Specifically, the validation data set can be used to evaluate the performance of the trained autonomous driving system. The performance evaluation results may include dimensions such as accuracy, recall, and F1 score (i.e., a weighted average of accuracy and recall, with a maximum value of 1 and a minimum value of 0) to comprehensively evaluate the performance of the system. Based on the performance evaluation results, the model parameters of each virtual driver model can be adjusted. For example, if it is assessed that the autonomous driving system performs poorly in certain scenarios, the initial parameter values ​​of the relevant virtual driver instances or the parameter values ​​in the shadow layer can be adjusted to improve the training effect.

[0123] In this embodiment, the above-mentioned method can be used to continuously iterate and improve the autonomous driving system, continuously optimize the instantiation and configuration of the virtual driver, the selection and combination of the shadow layer, and the training algorithm of the autonomous driving system, so as to improve the performance and safety of the autonomous driving system.

[0124] Furthermore, before the above step 102 of determining the N shadow layer groups corresponding to the N virtual driver instances, the method further includes:

[0125] Divide the autonomous driving system according to its functions to obtain multiple functional modules;

[0126] Determine the input data and output data of each functional model in multiple functional modules, and standardize the input interface and output interface of each functional model;

[0127] Construct the network layer corresponding to each functional model in multiple functional modules to obtain a multi-level shadow mode architecture, wherein the network layer corresponding to each functional model corresponds to a shadow layer, and the data between the network layers corresponding to different functional models interact according to the preset interaction mechanism.

[0128] In one embodiment, before determining the N shadow layer groups corresponding to the N virtual driver instances, a multi-level shadow mode architecture needs to be constructed. Specifically, the different functional modules of the autonomous driving system can be decoupled and independent network layers can be created for each functional module. These layers should be able to work together while maintaining sufficient independence so that they can be trained and optimized separately. Specifically, the following steps may be included:

[0129] Step A: Functional analysis;

[0130] Conduct an in-depth analysis of the overall functions of the autonomous driving system and clarify the role, input and output of each functional module. For example, it can include core modules such as environmental perception, decision planning, control execution, and human-computer interaction. In actual application, the functional modules can be expanded according to actual conditions.

[0131] Step B: module division;

[0132] Based on the results of functional analysis, the autonomous driving system is divided into several independent functional modules. Each functional module is responsible for processing specific tasks and data, and has clear input and output interfaces. For example, the environmental perception module is responsible for collecting and processing sensor data and outputting environmental information; the decision-making planning module is responsible for formulating driving strategies based on environmental information and outputting control instructions; the control execution module is responsible for receiving control instructions and controlling the vehicle to perform corresponding actions; the human-computer interaction module is responsible for processing the interaction information between the driver and the vehicle, etc.

[0133] Step C: interface design;

[0134] Design standardized interfaces for each functional module, including data format, communication protocol, and interaction mode, etc. Ensure seamless connection between functional modules to achieve accurate information transmission and collaborative work.

[0135] Step D: Network layer construction;

[0136] For each functional module, a corresponding network layer is constructed. These network layers can use different deep learning models or algorithms and be customized and optimized according to the functional requirements of the module. In actual application, the network structure and algorithm of each network layer can be selected according to the actual situation.

[0137] Step E: Inter-layer interaction and integration;

[0138] Design the inter-layer interaction mechanism to ensure that information between network layers can be transmitted smoothly. This can be achieved by defining a unified data exchange format, implementing inter-layer communication protocols, and building middleware. At the same time, it is necessary to consider the coordination and optimization issues between the layers to ensure the stability and performance of the entire system.

[0139] Through the above method, a multi-level shadow mode architecture can be constructed to fully consider the interactions between different driving scenarios, behaviors and decision-making processes, so that the trained system has better adaptability when facing actual complex environments.

[0140] Furthermore, the multi-level shadow mode architecture includes the environmental scenario layer, basic driving layer, thinking and decision-making layer, emergency response layer, and human-computer interaction layer;

[0141] The environment scene layer is used to simulate driving strategies under various road conditions and environments based on environmental data in real driving scenarios.

[0142] The basic driving layer is used to simulate basic driving decisions based on real-time road information and vehicle status. Basic driving decisions include control of the vehicle power system and control of the steering system.

[0143] The thinking decision layer is used to simulate high-level driving decisions based on the output data of the basic driving layer, the preset traffic rules database and real-time road information, wherein the high-level driving decisions include at least one of determining a safe overtaking window, determining the best time and location for lane change and avoiding potential risk areas in advance;

[0144] The emergency response layer is used to simulate driving decisions in emergency situations based on the output data of the thinking decision layer, emergency event information and a preset traffic rules database, wherein the driving decisions in emergency situations include at least one of emergency braking, emergency obstacle avoidance and path correction in abnormal situations;

[0145] The human-computer interaction layer is used to simulate human-computer interaction decisions based on user interaction data.

[0146] In one embodiment, the multi-level shadow mode architecture covers five interrelated hierarchical structures with different focuses: environmental scene layer (mainly dedicated to simulating various real-world traffic environments and dynamic situations), basic driving layer (mainly focusing on accurate simulation of basic vehicle control behaviors, including but not limited to speed control, direction adjustment, etc.), thinking decision layer (focusing on simulating the driver's high-level decision-making process when facing road conditions, traffic rules and emergencies), emergency response layer (focusing on simulating the rapid response mechanism of different drivers when encountering emergencies), human-computer interaction layer (focusing on the interaction characteristics between human drivers and vehicle systems during driving, including interface operations, voice command responses, etc.).

[0147] Specifically, the above-mentioned environmental scene layer is mainly responsible for simulating and adaptively training the diverse driving scenes faced by the autonomous driving system. Its function is to ensure that the autonomous driving algorithm has a high degree of scene adaptation ability by collecting and deeply analyzing data covering a wide range of driving environments, whether it is cruising on highways, driving on busy city streets, or facing severe weather conditions such as rain, snow, fog and haze, it can accurately execute the corresponding driving strategy. The input data of the above-mentioned environmental scene layer is comprehensive sensor data, including GPS information, weather conditions, road types and real-time traffic conditions. The output data of the above-mentioned environmental scene layer is scene-adaptive driving strategies and parameter adjustments to cope with different road conditions and environmental changes. The network structure of the above-mentioned environmental scene layer adopts an advanced multi-task learning network architecture, which integrates a series of parallel processing units, each of which specializes in identifying and processing specific types of driving scene features. By sharing the underlying feature representation and weight optimization, the network can learn and understand multiple scene modes at the same time. At the top layer of the network, a decision fusion mechanism is adopted, which can comprehensively analyze the optimal response under different scenarios, weigh risks and efficiency, generate the best driving strategy adapted to the current scenario, and then guide the real-time behavioral decision of the autonomous driving system.

[0148] The function of the above basic driving layer is to focus on simulating the basic operating skills of human drivers in the actual driving process, covering such as vehicle acceleration, deceleration, smooth driving, precise steering and other necessary basic control actions. The core goal of this layer is to ensure that the autonomous driving system can accurately and efficiently execute the corresponding physical control instructions in various basic driving scenarios. The input data of the above basic driving layer is the real-time road information and vehicle status from the sensor. The output data of the above basic driving layer is the control instructions for the vehicle power system and steering system, such as the accelerator pedal opening, braking force and steering wheel angle. The network structure of the above basic driving layer can use a convolutional neural network (CNN) to extract road features, and combine the long short-term memory network (LSTM) to process time series data to generate basic driving operations.

[0149] The function of the above-mentioned thinking and decision-making layer is a key level to simulate and surpass the advanced strategic thinking of human drivers. Its core function is to simulate complex driving decision-making processes, including but not limited to formulating reasonable overtaking strategies, choosing the best time to change lanes, responding to sudden traffic incidents and other high-level driving decision-making tasks. This level not only needs to understand the current road traffic conditions, but also must be combined with the prediction of possible changes in the future, and make the best response on this basis. The input data of the above-mentioned thinking and decision-making layer includes the output of the basic driving layer, the traffic rules database, the road conditions, etc. That is, the input data of the thinking and decision-making layer comes from the integration of data in multiple dimensions. The primary input comes from the output information of the basic driving layer, including the current state of the vehicle and its expected future motion trajectory; at the same time, it also includes a detailed traffic rules database for real-time verification and guidance of the legality of decisions; in addition, the real-time road condition data (such as road conditions, traffic flow, road network topology, etc.), other vehicle dynamic information and environmental factors that may affect driving decisions are also important input components. The output data of the above-mentioned thinking and decision-making layer is a series of clear driving decision instructions, such as determining a safe overtaking window, the best time and location to implement lane changes, and avoiding potential risk areas in advance. The network structure of the above-mentioned thinking decision-making layer can adopt the Deep Reinforcement Learning (DRL) framework and combine it with the attention mechanism to achieve decision optimization in complex environments.

[0150] The function of the above emergency response layer is to simulate and execute in real time the rapid judgment and precise operation of human drivers in the face of emergency situations, covering such as emergency braking, emergency obstacle avoidance, and path correction in abnormal situations. The input data of the above emergency response layer includes but is not limited to the output of the decision-making layer, emergency information (such as obstacles ahead, pedestrians suddenly breaking in, etc.), and a detailed traffic rules database. That is, the input data of the above emergency response layer comes from various sources and is highly timely. First, it connects to the output results of the decision-making layer to ensure that it can quickly take over control when conventional decisions fail or are insufficient to deal with emergencies. Secondly, it captures and analyzes various types of emergency information, such as static or dynamic obstacles ahead, pedestrians or animals suddenly breaking into lanes, and other immediate threats. In addition, this layer is also connected to a detailed traffic rules database so that decisions can be made in critical moments that meet safety regulations and minimize losses. The output data of the above emergency response layer is precise vehicle control instructions for specific emergency situations, such as instantly increasing the braking force for emergency braking, or adjusting the driving direction to avoid obstacles. The network structure of the emergency response layer can use Generative Adversarial Networks (GAN) to generate various emergency scenarios, and combine with a rapid response network (such as an algorithm based on Q-learning, etc.) to achieve rapid decision-making in emergency situations. It can quickly and accurately calculate the optimal action strategy at the moment of receiving the emergency signal, and realize real-time, efficient and safe emergency response measures.

[0151] The function of the above-mentioned human-computer interaction layer is to record and analyze the interaction activities between the driver and the intelligent vehicle system. For example, various interaction methods such as touch screen control behavior, voice command interaction process, contactless gesture control, etc. are aimed at refining the habitual patterns of various users, optimizing the operation process, and thus improving the overall human-computer interaction experience and user interface design quality. The input data of the above-mentioned human-computer interaction layer is user interaction data, including but not limited to the touch points and trajectories of the touch screen, the audio data of the voice command, the visual data of gesture recognition, etc. The output data of the above-mentioned human-computer interaction layer is the improved user interface design and interaction logic, making the vehicle system more intuitive and humanized. The network structure of the above-mentioned human-computer interaction layer can integrate multimodal perception networks, such as image recognition neural networks for gestures and visual feedback, speech recognition modules for parsing voice commands, and deep learning recommendation systems for predicting and simplifying user operation processes.

[0152] In this way, the multi-level shadow mode can carry out training of multiple virtual driver instances in the same real scene at the same time, and can select and combine corresponding shadow layers according to the training focus. Each instance can accumulate rich driving data from a different driving perspective, thereby significantly improving the autonomous driving system's adaptability and processing intelligence for various complex, marginal and even extreme driving conditions, further ensuring the safety and reliability of autonomous driving.

[0153] See also Figure 2 , Figure 2 A schematic diagram of a training process of an autonomous driving system provided in an embodiment of the present application is shown in FIG. Figure 2 As shown in Figure 1, the training process includes the following aspects:

[0154] 1. Real scene selection and import;

[0155] Import real-time real-world scenarios or select representative scenarios from a variety of real driving scenarios. These scenarios should cover a variety of different road conditions, traffic flows, and weather conditions to ensure the generalizability and robustness of the training results.

[0156] 2. Virtual driver instantiation and configuration;

[0157] 1) Instantiate virtual drivers: Based on the diversified virtual driver models, multiple virtual driver models are instantiated. Each virtual driver instance has specific attribute labels, such as experience level, psychological state, etc.

[0158] 2) Configure driving tasks: Configure specific driving tasks for each virtual driver instance, such as city driving, long-distance travel, night driving, etc. These tasks should cover a variety of different driving scenarios and challenges.

[0159] 3. Selection and combination of shadow layers;

[0160] 1) Determine the focus of training: clarify the goals and focus of training, such as safety, decision-making efficiency, comfort, etc. These focuses will determine which shadow layers are selected for combination.

[0161] 2) Select and combine shadow layers: Select and combine the corresponding shadow layers from the five shadow layers according to the training focus. For example, in the specific implementation, if the focus is safety, you can combine shadow layer A (environmental scene layer), shadow layer C (thinking decision layer) and shadow layer D (emergency response layer).

[0162] 4. Implementation of virtual driver training;

[0163] 1) Start the virtual driver instance: At the same time and in the same real scene, all virtual driver instances are started for simulated driving. Each instance performs independent driving operations from its own driving perspective.

[0164] 2) Data collection and processing: Collect simulated driving data of each virtual driver instance, including vehicle status, driving behavior, environmental perception information, etc. Preprocess and extract features from these data to generate a dataset suitable for training.

[0165] 3) Training the autonomous driving system: Use the collected data sets to train the autonomous driving system. Algorithms such as deep learning and reinforcement learning can be used to optimize the decision-making and control capabilities of the autonomous driving system.

[0166] 5.Performance evaluation and optimization;

[0167] 1) Performance evaluation: Use the validation data set to evaluate the performance of the trained autonomous driving system. Evaluation indicators can include accuracy, recall, F1 score, etc. to comprehensively evaluate the performance of the system.

[0168] 2) Feedback and adjustment: According to the performance evaluation results, the driving behavior and attribute labels of the virtual driver are adjusted. For example, if the system performs poorly in certain scenarios, the driving behavior or attribute labels of the relevant virtual driver can be adjusted to improve the training effect.

[0169] 6.Continuous iteration and improvement;

[0170] The training of multi-layer shadow models is a process of continuous iteration and improvement. With the continuous advancement of technology and the accumulation of data, the instantiation and configuration of virtual drivers, the selection and combination of shadow layers, and the training algorithm of the autonomous driving system can be continuously optimized to improve the performance and safety of the autonomous driving system.

[0171] In the above manner, the multi-level shadow mode can carry out training of multiple virtual driver instances in the same real scene at the same time, and select and combine corresponding shadow layers according to the training focus. Each virtual driver instance can accumulate rich driving data from a different driving perspective, thereby significantly improving the adaptability and processing wisdom of the autonomous driving system to various complex, marginal and even extreme driving conditions. This will further ensure the safety and reliability of autonomous driving and lay a solid foundation for the future development of autonomous driving technology. Specifically, the training method of the autonomous driving system provided in the embodiment of the present application has the following beneficial effects:

[0172] 1. The multi-level shadow mode architecture consists of five interconnected layers, namely the environment scene layer, basic driving layer, thinking decision layer, emergency response layer and human-computer interaction layer. Each layer focuses on different functions and together builds a comprehensive simulated driving environment. And the various layers can be flexibly combined to form a variety of training scenarios.

[0173] 2. Each virtual driver model has specific attribute labels, such as experienced, novice, good at night driving and driving under fatigue conditions, etc. These diverse virtual driver models can perform realistic simulated driving in a multi-level shadow mode.

[0174] 3. Allowing multiple virtual drivers to be trained simultaneously in the same scenario, which means that multiple driving modes and data sets can be processed in parallel, significantly improving training efficiency and speed. By training multiple virtual drivers simultaneously, it is also possible to explore the mutual influence and integration of different driving styles, providing a more comprehensive driving experience for the autonomous driving system.

[0175] 4. The ability to flexibly combine shadow layers according to the training focus means that during the training process, the weights and configurations of each layer can be adjusted and optimized according to actual needs, thereby ensuring that the training results are more in line with the needs of specific scenarios or tasks. This flexibility makes the training of autonomous driving systems more accurate and efficient.

[0176] See also Figure 3 , Figure 3 A schematic diagram of the structure of a training device for an automatic driving system provided in an embodiment of the present application. Figure 3 As shown, the device 300 includes:

[0177] An instantiation module 301 is used to select N virtual driver models from a pre-built virtual driver model library, and instantiate each virtual driver model in the N virtual driver models to obtain N virtual driver instances, wherein the N virtual driver models correspond to the N virtual driver instances one by one, and the N virtual driver instances are used to simulate the driving characteristics and decision logic of the N drivers, where N is an integer greater than 1;

[0178] A first determination module 302 is used to determine N shadow layer groups corresponding to N virtual driver instances, wherein the N virtual driver instances correspond to the N shadow layer groups one by one, each shadow layer group includes at least one shadow layer, and the at least one shadow layer is selected from a pre-built multi-level shadow mode architecture, and the multi-level shadow mode architecture includes multiple shadow layers, and different shadow layers are used to simulate different functional modules in the autonomous driving system;

[0179] The simulated driving module 303 is used to start N virtual driver instances, and respectively link the shadow layer groups corresponding to each virtual driver instance to perform simulated driving, so as to obtain N simulated driving data;

[0180] The training module 304 is used to train the automatic driving system by using the N simulated driving data as training data.

[0181] Furthermore, the instantiation module 301 includes:

[0182] A first acquisition submodule is used to acquire a model category and a driving attribute corresponding to each virtual driver model in the N virtual driver models, wherein each model category has two mutually opposing driving attributes;

[0183] A first determination submodule is used to determine N initialization parameter vectors corresponding to N virtual driver models according to the model category and the driving attribute, wherein the N virtual driver models correspond to the N initialization parameter vectors one-to-one, and the initialization parameter vectors of two virtual driver models belonging to the same model category have the same parameter category, and the parameter values ​​corresponding to the same parameter category are different;

[0184] The instantiation submodule is used to instantiate N virtual driver models using N initialization parameter vectors to obtain N virtual driver instances.

[0185] Furthermore, the device 300 also includes:

[0186] A first construction module is used to construct M virtual driver model groups, wherein each of the M virtual driver model groups corresponds to a model category, each of the M virtual driver model groups includes two virtual driver models with the same model category and different driving attributes, M is an integer greater than or equal to 1, and M is equal to N / 2;

[0187] The naming module is used to name each virtual driver model in the M virtual driver model groups according to a preset naming rule to obtain a virtual driver model library, wherein the preset naming rule is set based on the model category and driving attributes corresponding to each virtual driver model group.

[0188] Furthermore, the driving simulation module 303 includes:

[0189] A second acquisition submodule is used to acquire a target shadow layer group corresponding to a target virtual driver instance, wherein the target virtual driver instance is any instance of N virtual driver instances, and the target shadow layer group is a shadow layer group corresponding to the target virtual driver instance among the N shadow layer groups;

[0190] A second determination submodule is used to determine, according to a preset corresponding relationship, a parameter set of each shadow layer of the target shadow layer group for the initialization parameter vector corresponding to the target virtual driver instance, wherein the preset corresponding relationship is used to characterize the corresponding relationship between the initialization parameter vector of each virtual driver instance and the parameter set of each shadow layer;

[0191] The assignment submodule is used to assign a parameter set corresponding to each shadow layer of the target shadow layer group;

[0192] The simulated driving submodule is used to perform simulated driving on the target shadow layer group after linkage assignment to obtain the target simulated driving data corresponding to the target virtual driver instance.

[0193] Furthermore, the training module 304 includes:

[0194] a processing submodule, configured to use the first type of simulated driving data among the N simulated driving data as negative samples, and use the second type of simulated driving data among the N simulated driving data as positive samples, wherein the first type of simulated driving data is simulated driving data corresponding to the virtual driver model with negative driving attributes, and the second type of simulated driving data is simulated driving data corresponding to the virtual driver model with positive driving attributes;

[0195] The training submodule is used to train the autonomous driving system using negative samples and positive samples.

[0196] Furthermore, the device 300 also includes:

[0197] A performance evaluation module, used to perform performance evaluation on the trained autonomous driving system to obtain a performance evaluation result, wherein the performance evaluation result includes at least one of accuracy, recall and F1 score;

[0198] The adjustment module is used to adjust the model parameters of the N virtual driver models according to the performance evaluation results.

[0199] Furthermore, the device 300 also includes:

[0200] A division module is used to divide the autonomous driving system according to functions to obtain multiple functional modules;

[0201] A second determination module is used to determine the input data and output data of each functional model in the plurality of functional modules, and to standardize the input interface and output interface of each functional model;

[0202] The second construction module is used to construct the network layer corresponding to each functional model in multiple functional modules to obtain a multi-level shadow mode architecture, wherein the network layer corresponding to each functional model corresponds to a shadow layer, and the data between the network layers corresponding to different functional models interact according to a preset interaction mechanism.

[0203] Furthermore, the multi-level shadow mode architecture includes the environmental scenario layer, basic driving layer, thinking and decision-making layer, emergency response layer, and human-computer interaction layer;

[0204] The environment scene layer is used to simulate driving strategies under various road conditions and environments based on environmental data in real driving scenarios.

[0205] The basic driving layer is used to simulate basic driving decisions based on real-time road information and vehicle status. Basic driving decisions include control of the vehicle power system and control of the steering system.

[0206] The thinking decision layer is used to simulate high-level driving decisions based on the output data of the basic driving layer, the preset traffic rules database and real-time road information, wherein the high-level driving decisions include at least one of determining a safe overtaking window, determining the best time and location for lane change and avoiding potential risk areas in advance;

[0207] The emergency response layer is used to simulate driving decisions in emergency situations based on the output data of the thinking decision layer, emergency event information and a preset traffic rules database, wherein the driving decisions in emergency situations include at least one of emergency braking, emergency obstacle avoidance and path correction in abnormal situations;

[0208] The human-computer interaction layer is used to simulate human-computer interaction decisions based on user interaction data.

[0209] It should be noted that the device 300 can implement the steps of the training method for the autonomous driving system provided in any of the aforementioned method embodiments and can achieve the same technical effects, which will not be described in detail here.

[0210] like Figure 4 As shown, the embodiment of the present application further provides an electronic device, including a processor 411, a communication interface 412, a memory 413 and a communication bus 414, wherein the processor 411, the communication interface 412, and the memory 413 communicate with each other through the communication bus 414.

[0211] Memory 413, used for storing computer programs;

[0212] In one embodiment of the present application, the processor 411 is used to implement the training method of the autonomous driving system provided by any one of the aforementioned method embodiments when executing the program stored in the memory 413.

[0213] An embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, a training method for an autonomous driving system provided in any of the aforementioned method embodiments is implemented.

[0214] The device embodiments described above are merely illustrative, wherein 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, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0215] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a general hardware platform, and of course, by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the relevant technology can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0216] It should be understood that the terms used herein are only for the purpose of describing specific example embodiments and are not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms "one", "an" and "said" as used herein may also be meant to include plural forms. The terms "include", "comprise", "contain", and "have" are inclusive, and therefore specify the existence of stated features, steps, operations, elements and / or parts, but do not exclude the existence or addition of one or more other features, steps, operations, elements, parts, and / or combinations thereof. The method steps, processes, and operations described herein are not interpreted as necessarily requiring them to be performed in the specific order described or illustrated, unless the execution order is clearly indicated. It should also be understood that additional or alternative steps may be used.

[0217] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A training method for an automatic driving system, characterized in that: The method comprises: Selecting N virtual driver models from a pre-built virtual driver model library, and instantiating each virtual driver model in the N virtual driver models to obtain N virtual driver instances, wherein the N virtual driver models correspond one-to-one to the N virtual driver instances, and the N virtual driver instances are used to simulate driving characteristics and decision logic of the N drivers, where N is an integer greater than 1; Determine N shadow layer groups corresponding to the N virtual driver instances, wherein the N virtual driver instances correspond one-to-one to the N shadow layer groups, each shadow layer group includes at least one shadow layer, and the at least one shadow layer is selected from a pre-built multi-level shadow mode architecture, the multi-level shadow mode architecture includes multiple shadow layers, and different shadow layers are used to simulate different functional modules in the autonomous driving system; Starting the N virtual driver instances, and respectively linking the shadow layer groups corresponding to the virtual driver instances to perform simulated driving, to obtain N simulated driving data; The N simulated driving data are used as training data to train the automatic driving system.

2. The method according to claim 1, characterized in that: The instantiating each virtual driver model in the N virtual driver models to obtain N virtual driver instances includes: Acquire a model category and a driving attribute corresponding to each virtual driver model in the N virtual driver models, wherein each model category has two mutually opposing driving attributes; Determining N initialization parameter vectors corresponding to the N virtual driver models according to the model category and the driving attribute, wherein the N virtual driver models correspond one-to-one to the N initialization parameter vectors, and the initialization parameter vectors of two virtual driver models belonging to the same model category have the same parameter category, and the parameter values ​​corresponding to the same parameter category are different; The N virtual driver models are instantiated using the N initialization parameter vectors to obtain the N virtual driver instances.

3. The method according to claim 2, characterized in that Before selecting N virtual driver models from the pre-built virtual driver model library, the method further includes: Constructing M virtual driver model groups, wherein each of the M virtual driver model groups corresponds to a model category, each of the M virtual driver model groups includes two virtual driver models with the same model category and different driving attributes, M is an integer greater than or equal to 1, and M is equal to N / 2; Each virtual driver model in the M virtual driver model groups is named according to a preset naming rule to obtain the virtual driver model library, wherein the preset naming rule is set based on the model category and driving attributes corresponding to each virtual driver model group.

4. The method according to claim 1, characterized in that The shadow layer groups corresponding to the virtual driver instances are linked to perform simulated driving to obtain N simulated driving data, including: Obtaining a target shadow layer group corresponding to a target virtual driver instance, wherein the target virtual driver instance is any instance of the N virtual driver instances, and the target shadow layer group is a shadow layer group corresponding to the target virtual driver instance among the N shadow layer groups; Determine, according to a preset corresponding relationship, a parameter set of each shadow layer of the target shadow layer group for the initialization parameter vector corresponding to the target virtual driver instance, wherein the preset corresponding relationship is used to characterize the corresponding relationship between the initialization parameter vector of each virtual driver instance and the parameter set of each shadow layer; Assigning a parameter set corresponding to each shadow layer of the target shadow layer group; The target shadow layer group after linkage assignment is subjected to simulated driving to obtain target simulated driving data corresponding to the target virtual driver instance.

5. The method according to claim 1, characterized in that The using the N simulated driving data as training data to train the automatic driving system includes: Taking the first type of simulated driving data among the N simulated driving data as negative samples, and taking the second type of simulated driving data among the N simulated driving data as positive samples, wherein the first type of simulated driving data is simulated driving data corresponding to a virtual driver model with negative driving attributes, and the second type of simulated driving data is simulated driving data corresponding to a virtual driver model with positive driving attributes; The automatic driving system is trained using the negative samples and the positive samples.

6. The method according to claim 1, characterized in that After using the N simulated driving data as training data to train the automatic driving system, the method further includes: Performing a performance evaluation on the trained autonomous driving system to obtain a performance evaluation result, wherein the performance evaluation result includes at least one of accuracy, recall, and F1 score; According to the performance evaluation result, the model parameters of the N virtual driver models are adjusted.

7. The method according to claim 1, characterized in that Before determining the N shadow layer groups corresponding to the N virtual driver instances, the method further includes: Dividing the autonomous driving system according to functions to obtain multiple functional modules; Determining input data and output data of each functional model in the plurality of functional modules, and standardizing input interfaces and output interfaces of each functional model; Construct the network layer corresponding to each functional model in the multiple functional modules to obtain the multi-level shadow mode architecture, wherein the network layer corresponding to each functional model corresponds to a shadow layer, and the data between the network layers corresponding to different functional models interact according to a preset interaction mechanism.

8. The method according to claim 7, characterized in that The multi-level shadow mode architecture includes an environmental scene layer, a basic driving layer, a thinking and decision-making layer, an emergency response layer, and a human-computer interaction layer; The environment scene layer is used to simulate driving strategies under various road conditions and different environments based on the environment data in the real driving scene; The basic driving layer is used to simulate basic driving decisions according to real-time road information and vehicle status, and the basic driving decisions include control of the vehicle power system and control of the steering system; The thinking decision layer is used to simulate high-level driving decisions based on the output data of the basic driving layer, the preset traffic rules database and the real-time road information, wherein the high-level driving decisions include at least one of determining a safe overtaking window, determining the best time and location for implementing a lane change, and avoiding potential risk areas in advance; The emergency response layer is used to simulate driving decisions in emergency situations based on the output data of the thinking decision layer, emergency event information and a preset traffic rules database, wherein the driving decisions in emergency situations include at least one of emergency braking, emergency obstacle avoidance and path correction in abnormal situations; The human-computer interaction layer is used to simulate human-computer interaction decision-making based on user interaction data.

9. A training device for an automatic driving system, characterized in that: The device comprises: an instantiation module, used to select N virtual driver models from a pre-built virtual driver model library, and instantiate each virtual driver model in the N virtual driver models to obtain N virtual driver instances, wherein the N virtual driver models correspond one-to-one to the N virtual driver instances, and the N virtual driver instances are used to simulate the driving characteristics and decision logic of the N drivers, where N is an integer greater than 1; A first determination module is used to determine N shadow layer groups corresponding to the N virtual driver instances, wherein the N virtual driver instances correspond to the N shadow layer groups one by one, each shadow layer group includes at least one shadow layer, and the at least one shadow layer is selected from a pre-built multi-level shadow mode architecture, and the multi-level shadow mode architecture includes multiple shadow layers, and different shadow layers are used to simulate different functional modules in the autonomous driving system; A simulated driving module, used to start the N virtual driver instances, and respectively link the shadow layer groups corresponding to the virtual driver instances to perform simulated driving, so as to obtain N simulated driving data; A training module is used to train the automatic driving system by using the N simulated driving data as training data.

10. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor is used to implement the training method of the automatic driving system described in any one of claims 1 to 8 when executing the program stored in the memory.