Computer-implemented method and system for generating a virtual vehicle environment

Abstract

This invention relates to a computer-executed method for generating a virtual vehicle environment (U2) using previously detected video image data (10a), radar data, and / or lidar point clouds of a real vehicle environment (U1) to test highly automated driving functions of a motor vehicle (1). Furthermore, the invention relates to a system (2) for generating the virtual vehicle environment (U2) to test highly automated driving functions of a motor vehicle. Additionally, the invention relates to a computer program and a computer-readable data carrier.

Description

Technical Field

[0001] This invention relates to a computer-executed method for testing highly automated driving functions of motor vehicles by generating a virtual vehicle environment using previously detected video image data, radar data, and / or lidar point clouds of a real vehicle environment.

[0002] Furthermore, the present invention relates to a system for testing highly automated driving functions of motor vehicles by generating a virtual vehicle environment using previously detected video image data, radar data, and / or lidar point clouds of a real vehicle environment.

[0003] Furthermore, the present invention relates to a computer program and a computer-readable data carrier. Background Technology

[0004] Graphical user interfaces used to test highly automated driving functions of motor vehicles typically have multiple components that enable the management of parameter sets, the establishment of virtual vehicle environments, and test management.

[0005] The creation of virtual vehicle environments, namely the definition of static and dynamic objects in the scene, is done by manually configuring and importing objects stored in the object library.

[0006] CN000205121971U discloses a method for testing autonomous vehicles. According to this method, an autonomous vehicle is generated in a simulated environment. Based on previously recorded state information of the virtual vehicle, a traffic environment is created in the simulated environment to enable the autonomous vehicle to drive within that traffic environment.

[0007] However, what these methods have in common is the high cost of generating virtual vehicle environments to test the highly automated driving functions of motor vehicles, which leads to high costs in terms of personnel and resources.

[0008] Therefore, there is a need to improve existing methods and systems for generating virtual vehicle environments to test the highly automated driving functions of motor vehicles, so that virtual vehicle environments can be created simply, more efficiently and cost-effectively. Summary of the Invention

[0009] Therefore, the object of the present invention is to provide a computer-executable method, system, computer program, and computer-readable data carrier that enables the creation of virtual vehicle environments simply, more efficiently, and more cost-effectively.

[0010] According to the present invention, the task is solved by a computer-executed method according to the present invention, a system according to the present invention, a computer program product according to the present invention, and a computer-readable data carrier according to the present invention. The computer-executed method is used to generate a virtual vehicle environment to test the highly automated driving functions of a motor vehicle using previously detected video image data, radar data, and / or lidar point clouds in a real vehicle environment. The system is used to generate a virtual vehicle environment to test the highly automated driving functions of a motor vehicle using previously detected video image data, radar data, and / or lidar point clouds in a real vehicle environment.

[0011] This invention relates to a computer-executed method for generating a virtual vehicle environment to test highly automated driving functions of a motor vehicle using previously detected video image data, radar data, and / or lidar point clouds of a real vehicle environment.

[0012] The method includes receiving previously detected video image data, radar data, and / or lidar point clouds of a real vehicle environment, the real vehicle environment comprising multiple real objects characterized previously or subsequently using at least one first machine learning algorithm.

[0013] A real-world vehicle environment corresponds to the vehicle environment of a motor vehicle in road traffic, particularly under various traffic conditions. Real-world objects can be static objects, such as traffic signs, buildings, vegetation, and / or parked motor vehicles. Furthermore, real-world objects can be dynamic objects, such as moving motor vehicles.

[0014] The method includes generating a first feature vector representing the corresponding real object by applying a second machine learning algorithm to the corresponding real object and storing the first feature vector.

[0015] Furthermore, the method includes providing a plurality of stored second feature vectors, each representing a synthesized object. The method also includes identifying a second feature vector that has the highest similarity to a first feature vector.

[0016] The similarity can be defined based on a predetermined feature definition of the first feature vector and the second feature vector. The synthesized objects are classified into multiple object categories and represent real objects characterized in video image data, radar data, and / or lidar point clouds of real vehicle environments.

[0017] Furthermore, the method includes: selecting a second feature vector based on the similarity of the identification and calling a stored synthetic object corresponding to a real object configured for the second feature vector or programmatically generating a synthetic object corresponding to a real object, and integrating the synthetic object into a predetermined virtual vehicle environment.

[0018] Virtual vehicle environments are computer-generated representations of the real vehicle environment based on sensor detection.

[0019] Programmatic generation describes a method for generating 3D objects in real time, even during the execution of a computer program.

[0020] 3D objects are not generated randomly, but rather according to a deterministic algorithm so that the same content can be generated again and again under the same initial conditions.

[0021] Furthermore, the present invention relates to a system for testing highly automated driving functions of a motor vehicle by generating a virtual vehicle environment using previously detected video image data, radar data, and / or lidar point clouds of a real vehicle environment. The system includes means for receiving previously detected video image data, radar data, and / or lidar point clouds of a real vehicle environment, the real vehicle environment comprising a plurality of real objects characterized prior to or subsequently using at least one first machine learning algorithm.

[0022] Furthermore, the system includes means for applying at least one first machine learning algorithm to video image data, radar data, and / or lidar point clouds to detect and classify the plurality of real objects, and the system includes means for generating a first feature vector representing a corresponding real object, the means being configured to apply a second machine learning algorithm to the corresponding real object, and means for storing the first feature vector.

[0023] Furthermore, the system includes means for providing a plurality of stored second feature vectors, the second feature vectors representing synthesized generated objects, and means for identifying a second feature vector that has the highest similarity to a first feature vector.

[0024] Furthermore, the system includes devices for selecting a second feature vector for identification and recalling a stored synthetic object corresponding to the real object, which is assigned to the second feature vector, and devices for programmatically generating a synthetic object corresponding to the real object, wherein the devices for selecting the second feature vector for identification and recalling the stored synthetic object assigned to the second feature vector or the devices for programmatically generating a synthetic object corresponding to the real object are used based on the similarity of the identification.

[0025] In addition, the system includes devices for integrating synthetic objects into a predetermined virtual vehicle environment.

[0026] Furthermore, the present invention relates to a computer program product with program code, which, when executed on a computer, implements the method according to the present invention.

[0027] Furthermore, the present invention relates to a computer-readable data carrier having program code of a computer program that, when executed on a computer, implements the method according to the present invention.

[0028] The concept of this invention is to automatically construct a synthetic 3D scene from real video image data of motor vehicle test driving.

[0029] Since a first feature vector representing the corresponding real object is generated by applying a machine learning algorithm to the corresponding real object, it is advantageous that direct object classification and configuration regarding synthetic objects are not required.

[0030] In the realm of direct object classification, artificial neural networks are required to classify each object. This necessitates the costly training of these artificial neural networks using a large amount of training data.

[0031] Instead, within the scope of this invention, existing data on typical objects stored in a database or a saved traffic scene can be used. Therefore, the second feature vector can be easily assigned to the corresponding synthetic object.

[0032] By comparing the first feature vector representing the corresponding real object with the existing second feature vector representing the synthetic generated object, it is possible to achieve in an advantageous manner the automatic configuration of the corresponding real object and the synthetic generated object, thereby enabling the integration of the thus identified synthetic generated object into the virtual vehicle environment.

[0033] By automating the construction of virtual vehicle environments using data obtained from real vehicle environments, the following advantages can be achieved: a significant simplification of virtual vehicle environment construction, along with considerable efficiency improvements and associated cost reductions.

[0034] Other embodiments of the present invention are described below with reference to the accompanying drawings.

[0035] According to one aspect of the invention, the method further includes: performing representation, particularly detection and classification, of the plurality of real objects by a supervised learning algorithm or by an unsupervised learning algorithm, particularly an artificial neural network.

[0036] Therefore, appropriate algorithms can be used in an advantageous manner according to the data type and / or other system settings.

[0037] According to another aspect of the invention, the method further includes: if the similarity between the identified second feature vector and the first feature vector is greater than or equal to a predetermined threshold, then selecting the identified second feature vector and calling a stored synthetic object corresponding to the real object assigned to the second feature vector.

[0038] Therefore, it can be achieved in an advantageous way: to invoke the synthetic object that has the highest similarity to the selected real object.

[0039] According to another aspect of the invention, the method further includes: if the similarity between the identified second feature vector and the first feature vector is less than a predetermined threshold, then programmatically generating the synthetic object.

[0040] Even if there is insufficient consistency between the first feature vector corresponding to the real object and one of the second feature vectors corresponding to multiple synthesized objects, it is still possible to generate a synthesized object corresponding to the real object in an advantageous manner by procedurally generating the synthesized object.

[0041] According to another aspect of the invention, the method further includes the step of programmatically generating the synthetic object using a first feature vector generated by a second machine learning algorithm. Therefore, it can be advantageously ensured that the synthetic object generated in this manner corresponds to a high degree of real-world object on which it is based.

[0042] According to another aspect of the invention, the method further includes: a second machine learning algorithm for generating, and particularly for encoding, the first feature vector, is constructed using a first variational autoencoder. The advantage of the variational autoencoder is that it can generate a dimensionality-reduced or complexity-reduced representation based on real data.

[0043] According to another aspect of the invention, the method further includes: generating a second feature vector representing the corresponding synthetic object by applying a third machine learning algorithm to the corresponding synthetic object, and storing the second feature vector in a data storage device.

[0044] Therefore, the configuration or feature comparison of the first and second feature vectors can be performed, and thus the real object can be configured to the synthetic object.

[0045] According to another aspect of the invention, the method further includes: classifying real objects contained in video image data, radar data, and / or lidar point clouds into a main category using a first machine learning algorithm, wherein real objects assigned to a main category are classified into subcategories using a fourth machine learning algorithm.

[0046] Therefore, objects contained in the real data can be progressively classified in a favorable manner. Here, in the first step, the real objects are roughly classified into main categories. Then, in the second step, the corresponding real objects assigned to the main categories are further classified into their respective subcategories.

[0047] According to another aspect of the invention, the method further includes: when the first feature vector and the second feature vector are of the same type, especially when the first feature vector and the second feature vector are of the same data structure type, in order to identify similarity, performing a direct comparison between the first feature vector and the plurality of second feature vectors.

[0048] Therefore, it is possible to efficiently identify a suitable second feature vector or compare a first feature vector with a second feature vector in an advantageous manner without any other intermediate steps.

[0049] According to another aspect of the invention, the method further includes: when the first feature vector and the second feature vector are of different types, especially when the data structure types of the first feature vector and the second feature vector are different, using a classifier that establishes a correlation between the first feature vector and the second feature vector in order to identify the similarity between the second feature vector and the first feature vector.

[0050] Therefore, even when using encodings of different types, it is also possible to establish correlations between different feature vectors in an advantageous manner and thus assign the corresponding first feature vector to the corresponding second feature vector.

[0051] According to another aspect of the invention, the method further includes: the plurality of stored first feature vectors and / or second feature vectors representing 2D image data, 3D image data, radar data, and / or lidar point clouds.

[0052] Therefore, the source data on which the second feature vector is based can be of different types in an advantageous manner, so that a large number of different actually obtained source data can be used and associated with the synthesized generated object or object data, or can be configured to the synthesized generated object or object data.

[0053] According to another aspect of the invention, the method further includes: generating a notification if the similarity between the identified second feature vector and the first feature vector is less than a predetermined threshold, i.e., the stored plurality of second feature vectors do not have sufficient similarity with the identified first feature vector.

[0054] Therefore, the user is informed of the situation in an advantageous manner, and thus the data storage of the synthesized object on which the second feature vector is based can be expanded, for example.

[0055] The features of the method described in this paper can also be applied to other virtual environments, such as testing other vehicle types in different environments. Attached Figure Description

[0056] To better understand the present invention and its advantages, reference is now made to the following description in conjunction with the accompanying drawings.

[0057] The present invention will now be described in detail with reference to exemplary embodiments. Wherein:

[0058] Figure 1 This is a flowchart of a method for generating a virtual vehicle environment to test the highly automated driving function of a motor vehicle according to a preferred embodiment of the present invention;

[0059] Figure 2 A schematic diagram illustrating the application of the method to a real vehicle environment to generate a virtual vehicle environment according to a preferred embodiment of the invention is shown; and

[0060] Figure 3 A diagram is shown of a system for generating a virtual vehicle environment to test the highly automated driving functions of a motor vehicle, according to a preferred embodiment of the present invention.

[0061] Unless otherwise stated, the same reference numerals denote the same elements in the figures. Detailed Implementation

[0062] Figure 1 A flowchart is shown of a method for generating a virtual vehicle environment to test the highly automated driving functions of a motor vehicle according to a preferred embodiment of the present invention.

[0063] The method includes: receiving previously detected video image data 10a, radar data, and / or lidar point cloud of a real vehicle environment U1, S1A and S1B, wherein the real vehicle environment includes a plurality of real objects 12a, 12b, 12c, 12d, 12e that previously characterize S1A or subsequently characterize S1B using at least one first machine learning algorithm A1.

[0064] Furthermore, the method includes: generating a first feature vector M1 representing the corresponding real objects 12a, 12b, 12c, 12d, 12e by applying a second machine learning algorithm A2 to the corresponding real objects 12a, 12b, 12c, 12d, 12e, and storing the first feature vector M1 in S3.

[0065] Furthermore, the method includes: providing multiple stored representations of the synthesized objects 14a, 14b, 14c, 14d, 14e in S4, and identifying the second feature vector M2 that has the highest similarity MA with the first feature vector M1 in S5.

[0066] Furthermore, the method includes: selecting a second feature vector M2 identified by S6A for generating synthetic objects 14a, 14b, 14c, 14d, 14e corresponding to real objects 12a, 12b, 12c, 12d, 12e.

[0067] Alternatively, procedurally generate (prozedurales Generieren) S6B corresponding to the real objects 12a, 12b, 12c, 12d as synthetic objects 14a, 14b, 14c, 14d.

[0068] Step S6A or S6B is selected based on the identified similarity MA. Then, the synthesized objects 14a, 14b, 14c, 14d, and 14e are integrated into the predetermined virtual vehicle environment U2 in step S7.

[0069] Figure 2 A schematic diagram is shown illustrating the application of the method to a real vehicle environment to generate a virtual vehicle environment according to a preferred embodiment of the present invention.

[0070] In the embodiment shown here, video image data is used accordingly not only in the virtual vehicle environment U2 but also in the real vehicle environment U1.

[0071] Alternatively or additionally, a virtual vehicle environment can be generated using radar data and / or lidar point clouds from a real vehicle environment.

[0072] The previously detected video image data 10a is first characterized using a first machine learning algorithm A1.

[0073] The representation includes: probing multiple real objects using a first machine learning algorithm A1 and classifying the multiple real objects.

[0074] In the current embodiment, the first machine learning algorithm A1 is formed by a supervised learning algorithm, particularly an artificial neural network.

[0075] Alternatively, the first machine learning algorithm A1 is formed, for example, by an unsupervised learning algorithm, particularly an artificial neural network. Using the first machine learning algorithm A1, the various real objects 12a, 12b, 12c, and 12d contained in the video image data 10a are classified into main categories C1, C2, C3, and C4.

[0076] The real objects 12a, 12b, 12c, and 12d, assigned to a main category C1, C2, C3, and C4, are further classified into subcategories 12d1 and 12d2 using a fourth machine learning algorithm A4.

[0077] In the current embodiment, object 12a is a tree; object 12b is a pedestrian; object 12c is a street lamp; multiple objects 12d are different vehicles; and object 12e is a traffic sign.

[0078] The aforementioned objects therefore correspond to the respective main categories. Subcategories are defined by the object's subtypes—such as different tree types and / or different types of pedestrians selected according to gender and age—and other classifiable features.

[0079] The traffic sign 12e is particularly noteworthy here because, in the current embodiment, the traffic sign is neither generated by assigning a second feature vector to a first feature vector nor is it generated programmatically. Instead, it is generated from other data sources or by other methods, in which case the accuracy, detail, and reliability of the traffic sign information depend on the presentation of the information.

[0080] Alternatively, traffic signs can be generated, for example, by a synthetic object stored in a data storage device, when the first feature vector and the corresponding second feature vector have sufficiently high consistency.

[0081] Then, real objects 12a, 12b, 12c, and 12d contained in the real vehicle environment U1 are selected, and a first feature vector M1 representing the corresponding real objects 12a, 12b, 12c, and 12d is generated by applying the second machine learning algorithm A2 to the corresponding real objects 12a, 12b, 12c, and 12d.

[0082] The method is executed for all objects 12a, 12b, 12c, and 12d contained in the real vehicle environment U1.

[0083] Furthermore, for each of the first feature vectors M1 thus generated, a comparison is made with or a configuration is performed with the second feature vector M2, and the first feature vector is stored in the corresponding data memory.

[0084] To this end, multiple stored second feature vectors M2 are provided, each representing a synthesized object 14a, 14b, 14c, 14d. A second feature vector M2 with the highest similarity MA to the first feature vector M1 is then identified.

[0085] If the similarity MA between the identified second feature vector M2 and the first feature vector M1 is greater than or equal to a predetermined threshold T, then the identified second feature vector M2 is selected.

[0086] Subsequently, or as a result thereof, the synthesized objects 14a, 14b, 14c, 14d, and 14e corresponding to the real objects 12a, 12b, 12c, 12d, and 12e, which are stored in the data storage and assigned to the second feature vector M2, are retrieved from the data storage.

[0087] Prior to this, a second feature vector M2 representing the corresponding synthetic objects 14a, 14b, 14c, 14d, 14e is generated by applying the third machine learning algorithm A3 to the corresponding synthetic objects 14a, 14b, 14c, 14d, 14e, and the second feature vector is stored in the data memory by S3.

[0088] If the similarity MA between the second feature vector M2 and the first feature vector M1 is less than a predetermined threshold T, then synthetic objects 14a, 14b, 14c, and 14d are generated programmatically.

[0089] Programmatic generation can be performed, for example, by a software application that can be used for this purpose, such as Unreal Engine. The programmatic generation of composite objects 14a, 14b, 14c, and 14d is performed using the first feature vector M1 generated by the second machine learning algorithm A2.

[0090] The second machine learning algorithm A2 used to generate, and in particular to encode, the first feature vector M1 is constructed in the current embodiment by a first variational autoencoder.

[0091] In the current embodiment, the object 14e formed by traffic signs is generated from traffic signs 12e contained in the real vehicle environment U1 by other suitable methods.

[0092] In order to identify the similarity MA, when the first feature vector M1 and the second feature vector M2 are of the same type, especially when the first feature vector and the second feature vector have the same data structure type, a direct comparison between the first feature vector M1 and the plurality of second feature vectors M2 is performed as shown in the current embodiment.

[0093] When the first feature vector M1 and the second feature vector M2 are of different types, especially when the data structure types of the first feature vector and the second feature vector are different, alternatively, a classifier that establishes a correlation between the first feature vector M1 and the second feature vector M2 can be used to identify the similarity MA between the second feature vector M2 and the first feature vector M1. In this embodiment, the plurality of stored first feature vectors M1 and second feature vectors M2 are constructed from 3D image data.

[0094] Alternatively, the first feature vector M1 and the second feature vector M2 can be represented, for example, by 2D image data, radar data, and / or lidar point clouds.

[0095] If the similarity MA between the identified second feature vector M2 and the first feature vector M1 is less than a predetermined threshold T, then according to the current embodiment, a notification is also generated that the stored multiple second feature vectors M2 do not have sufficient similarity MA with the identified first feature vector M1.

[0096] Figure 3 A diagram of a system for generating a virtual vehicle environment to test the highly automated driving functions of a motor vehicle, according to a preferred embodiment of the present invention, is shown. System 2 includes receiving previously detected video image data 10a, radar data, and / or lidar point clouds of a real vehicle environment U1, the real vehicle environment comprising a plurality of real objects 12a, 12b, 12c, 12d, 12e characterized by at least one machine learning algorithm A1.

[0097] Alternatively, the real objects 12a, 12b, 12c, 12d, and 12e may be, for example, previously characterized.

[0098] The system 2 further includes a device 22 for applying at least one machine learning algorithm A1 to the video image data 10a, radar data and / or lidar point cloud to detect the plurality of real objects 12a, 12b, 12c, 12d, 12e and classify the plurality of real objects.

[0099] The system 2 further includes a device 24 for generating a first feature vector M1 representing the corresponding real objects 12a, 12b, 12c, 12d, and 12e, the device 24 being configured to apply a second machine learning algorithm A2 to the corresponding real objects 12a, 12b, 12c, 12d, and 12e. The system 2 also includes a device 26 for storing the first feature vector M1.

[0100] The system 2 also includes a device 28 for providing a plurality of stored second feature vectors M2, the second feature vectors representing synthesized objects 14a, 14b, 14c, 14d, and 14e.

[0101] The system 2 also includes a device for identifying a second feature vector M2 that has the highest similarity MA to the first feature vector M1.

[0102] The system 2 further includes a device 32 for selecting the second feature vector M2 identified by S6A and calling stored synthetic objects 14a, 14b, 14c, 14d, and 14e assigned to the second feature vector M2, the synthetic objects corresponding to the real objects 12a, 12b, 12c, 12d, and 12e. The system also includes a device 34 for programmatically generating synthetic objects 14a, 14b, 14c, 14d, and 14e corresponding to the real objects 12a, 12b, 12c, 12d, and 12e.

[0103] Based on the identified similarity MA, the device 32 is used to select the second feature vector M2 for S6A identification and to call the stored synthetic objects 14a, 14b, 14c, 14d, 14e configured for the second feature vector M2, or the device 34 is used to programmatically generate synthetic objects 14a, 14b, 14c, 14d, 14e corresponding to the real objects 12a, 12b, 12c, 12d, 12e.

[0104] The system 2 also includes a device 36 for integrating the synthetic objects 14a, 14b, 14c, 14d, and 14e into a predetermined virtual vehicle environment U2.

[0105] The device 30 for identifying a second feature vector M2 that has the highest similarity MA to the first feature vector M1 has connections not only to the device 26 for storing the first feature vector M1 but also to the device 28 for providing multiple storage of the second feature vector M2, the second feature vector representing the synthesized objects 14a, 14b, 14c, 14d, 14e.

[0106] Although specific embodiments have been illustrated and described herein, those skilled in the art will understand that various alternative and / or equivalent implementations exist. It should be noted that one or more exemplary embodiments are merely examples and are not intended to limit the scope, applicability, or configuration in any way.

[0107] More precisely, the above overview and detailed description provide convenient guidance for those skilled in the art to implement at least one exemplary embodiment. It should be understood that various changes can be made to the functional scope and arrangement of the elements without departing from the scope of the appended claims and their legal equivalents.

[0108] Generally, this application is intended to cover changes, modifications, or variations of the embodiments shown herein.

Claims

1. A computer-executed method for generating a virtual vehicle environment (U2) using previously detected video image data (10a), radar data, and / or lidar point clouds of a real vehicle environment (U1) to test highly automated driving functions of a motor vehicle (1), the method comprising the following steps: Receive (S1A, S1B) previously detected video image data (10a), radar data and / or lidar point cloud of a real vehicle environment (U1), the real vehicle environment including a plurality of real objects (12a, 12b, 12c, 12d, 12e) previously characterized (S1A) or subsequently characterized (S1B) by using at least one first machine learning algorithm (A1). By applying the second machine learning algorithm (A2) to the corresponding real objects (12a, 12b, 12c, 12d, 12e), a first feature vector (M1) representing the corresponding real objects (12a, 12b, 12c, 12d, 12e) is generated (S2) and stored (S3) the first feature vector (M1). Provide (S4) multiple stored second feature vectors (M2) representing the synthesized objects (14a, 14b, 14c, 14d, 14e); Identify (S5) the second feature vector (M2) that has the highest similarity (MA) to the first feature vector (M1); Based on the identified similarity (MA), select (S6A) the identified second feature vector (M2) and call the stored synthetic objects (14a, 14b, 14c, 14d, 14e) corresponding to the real objects (12a, 12b, 12c, 12d, 12e) configured for the second feature vector (M2), or programmatically generate (S6B) synthetic objects (14a, 14b, 14c, 14d, 14e) corresponding to the real objects (12a, 12b, 12c, 12d, 12e); and The composite objects (14a, 14b, 14c, 14d, 14e) are integrated (S7) into the predetermined virtual vehicle environment (U2).

2. The computer-executed method according to claim 1, characterized in that, The representations (S1A, S1B) of the plurality of real objects (12a, 12b, 12c, 12d, 12e) are performed by supervised learning algorithms or by unsupervised learning algorithms.

3. The computer-executed method according to claim 2, characterized in that, Representation (S1A, S1B) of the plurality of real objects (12a, 12b, 12c, 12d, 12e) is performed using an artificial neural network; and / or The characterization includes detection and classification.

4. The computer-executed method according to any one of claims 1 to 3, characterized in that, If the similarity (MA) between the identified second feature vector (M2) and the first feature vector (M1) is greater than or equal to a predetermined threshold (T), then the identified second feature vector (M2) is selected, and the synthesized objects (14a, 14b, 14c, 14d, 14e) corresponding to the real objects (12a, 12b, 12c, 12d, 12e) stored in (S6A) are called.

5. The computer-executed method according to any one of claims 1 to 3, characterized in that, If the similarity (MA) between the identified second feature vector (M2) and the first feature vector (M1) is less than a predetermined threshold (T), then the synthesized objects (14a, 14b, 14c, 14d, 14e) are generated programmatically (S6B).

6. The computer-executed method according to claim 5, characterized in that, The synthetic objects (14a, 14b, 14c, 14d, 14e) are generated procedurally (S6B) using the first feature vector (M1) generated by the second machine learning algorithm (A2).

7. The computer-executed method according to any one of claims 1 to 3, characterized in that, The second machine learning algorithm (A2) for generating the first feature vector (M1) is constructed by the first variational autoencoder.

8. The computer-executed method according to any one of claims 1 to 3, characterized in that, The second machine learning algorithm (A2) for encoding the first feature vector (M1) is constructed by the first variational autoencoder.

9. The computer-executed method according to any one of claims 1 to 3, characterized in that, A second feature vector (M2) representing the corresponding synthetic object (14a, 14b, 14c, 14d, 14e) is generated by applying a third machine learning algorithm (A3) to the corresponding synthetic object (14a, 14b, 14c, 14d, 14e), and the second feature vector (M2) is stored (S3) in the data memory.

10. The computer-executed method according to any one of claims 1 to 3, characterized in that, The first machine learning algorithm (A1) classifies real objects (12a, 12b, 12c, 12d) contained in video image data (10a), radar data and / or lidar point clouds into main categories (C1, C2, C3, C4), wherein real objects (12a, 12b, 12c, 12d) assigned to a main category (C1, C2, C3, C4) are classified into subcategories (12d1, 12d2) using a fourth machine learning algorithm (A4).

11. The computer-executed method according to any one of claims 1 to 3, characterized in that, When the first feature vector (M1) and the second feature vector (M2) are of the same type, in order to identify similarity (MA), a direct comparison is performed between the first feature vector (M1) and the plurality of second feature vectors (M2).

12. The computer-executed method according to any one of claims 1 to 3, characterized in that, When the data structure types of the first feature vector (M1) and the second feature vector (M2) are the same, in order to identify the similarity (MA), a direct comparison is performed between the first feature vector (M1) and the plurality of second feature vectors (M2).

13. The computer-executed method according to any one of claims 1 to 3, characterized in that, When the first feature vector (M1) and the second feature vector (M2) are of different types, a classifier that establishes a correlation between the first feature vector (M1) and the second feature vector (M2) is used to identify the similarity (MA) between the second feature vector (M2) and the first feature vector (M1).

14. The computer-executed method according to any one of claims 1 to 3, characterized in that, When the data structure types of the first feature vector (M1) and the second feature vector (M2) are different, a classifier that establishes a correlation between the first feature vector (M1) and the second feature vector (M2) is used to identify the similarity (MA) between the second feature vector (M2) and the first feature vector (M1).

15. The computer-executed method according to any one of claims 1 to 3, characterized in that, Multiple stored first feature vectors (M1) and / or second feature vectors (M2) represent 2D image data, 3D image data, radar data, and / or lidar point clouds.

16. The computer-executed method according to claim 4, characterized in that, If the similarity (MA) between the identified second feature vector (M2) and the first feature vector (M1) is less than a predetermined threshold (T), then a notification (ME) is generated, that is, the stored multiple second feature vectors (M2) do not have sufficient similarity (MA) with the identified first feature vector (M1).

17. A system (2) for testing highly automated driving functions of a motor vehicle by generating a virtual vehicle environment (U2) using previously detected video image data (10a), radar data, and / or lidar point clouds in a real vehicle environment (U1), said system comprising: Device (20) for receiving previously detected video image data (10a), radar data and / or lidar point cloud of a real vehicle environment (U1), the real vehicle environment including a plurality of real objects (12a, 12b, 12c, 12d, 12e) characterized either prior to or subsequently by using at least one first machine learning algorithm (A1). Device (22) for applying at least one first machine learning algorithm (A1) to video image data (10a), radar data and / or lidar point cloud for detecting the plurality of real objects (12a, 12b, 12c, 12d, 12e) and classifying the plurality of real objects. A device (24) for generating a first feature vector (M1) representing the corresponding real objects (12a, 12b, 12c, 12d, 12e), the device (24) being configured to apply a second machine learning algorithm (A2) to the corresponding real objects (12a, 12b, 12c, 12d, 12e); and Device (26) for storing the first feature vector (M1); Device (28) for providing a second feature vector (M2) representing multiple stored synthetic generated objects (14a, 14b, 14c, 14d, 14e). Device (30) for identifying a second feature vector (M2) that has the highest similarity (MA) to the first feature vector (M1); Device (32) for selecting (S6A) the second feature vector (M2) for identification and recalling stored synthetic objects (14a, 14b, 14c, 14d, 14e) assigned to the second feature vector (M2), the synthetic objects corresponding to real objects (12a, 12b, 12c, 12d, 12e); and Device (34) for programmatically generating composite objects (14a, 14b, 14c, 14d, 14e) corresponding to real objects (12a, 12b, 12c, 12d, 12e). Among them, the device (32) that uses the second feature vector (M2) for selection (S6A) of the identification and calls the stored synthetic objects (14a, 14b, 14c, 14d, 14e) configured for the second feature vector (M2) according to the similarity (MA) of the identification, or the device (34) for programmatically generating synthetic objects (14a, 14b, 14c, 14d, 14e) corresponding to the real objects (12a, 12b, 12c, 12d, 12e); and Device (36) for integrating synthetic objects (14a, 14b, 14c, 14d, 14e) into a predetermined virtual vehicle environment (U2).

18. A computer program product containing program code, which, when executed on a computer, performs the method according to any one of claims 1 to 16.

19. A computer-readable data carrier having program code of a computer program that, when executed on a computer, performs the method according to any one of claims 1 to 16.

Images