Universal transformation of traffic scenario data into simulation input data

Abstract

The invention relates to a system for generating input data for a vehicle component (1), to be tested by means of a simulation, from a description of a traffic scenario, having a computing unit (3) which is designed to transform information relating to a traffic scenario into input data compatible with the vehicle component (1) by executing a processing module with a pre-trained artificial neural network, wherein the processing module can transform a multiplicity of items of information relating to a particular traffic scenario, which differ in terms of nature and / or data format, into input data, and wherein the computing unit (3) is designed to execute a physics module for ensuring the physical plausibility of the input data generated by the processing module.

Description

[0001] UNIVERSAL TRANSFORMATION OF TRAFFIC SCENARIO DATA INTO SIMULATION INPUT DATA

[0002] The invention relates to a system for generating input data for a vehicle component to be tested by means of a simulation from a description of a traffic scenario, and to a method for producing a system for generating input data for a vehicle component to be tested by means of a simulation from a description of a traffic scenario.

[0003] Modern and future vehicles, especially passenger cars and trucks, typically feature a multitude of sensors and data processing devices. These are usually part of driver assistance systems that support manual vehicle control or of automation systems for highly automated or fully autonomous vehicles (collectively referred to as "advanced driver assistance systems" or "autonomous driving systems," collectively abbreviated to ADAS / AD). To ensure the safe operation of a vehicle with such a driving control system, its hardware and the software required for its operation must be adequately tested.

[0004] For vehicles controlled by a human driver with regard to the so-called "Dynamic Driving Task" (DDT) and with regard to the so-called "Object and Event Detection and Response" (OEDR), vehicle control and thus the determination of the vehicle's movement behavior are essentially the responsibility of the human driver and are thus largely based on human perception, possibly supported by available sensor data from the vehicle and its automated driving control function (ADAS systems). However, with the advancing development of automated driving systems, the vehicle is increasingly taking over the task of object detection in the sense of OEDR and the initiation of driving maneuvers in the sense of DDT. Thus, starting with SAE J3016, at Level 3 of the automation of the driving control function, responsibility for controlling the vehicle is transferred from a human user to the manufacturer of the respective automated driving control system.

[0005] In the development and validation of automated driving control systems, simulation-based methods are used and tested, particularly with regard to the operating ranges critical for automated driving control systems. For example, scenario-based development and testing can be used to estimate how safely the future real vehicle will operate in the open world (possibly within a limited operating range) even during early development phases by using the digital twin of the automated vehicle. Such simulations can be used for system analysis with a very wide variety of parameter variations and can include, for example, behavioral simulations of other road users. They can also be used for component testing using hardware-in-the-loop tests, software-in-the-loop tests, model-in-the-loop tests, prototype-in-the-loop tests, etc.

[0006] So-called "adaptive replay-to-sim" methods are also known in the state of the art. In such methods, real-life scenarios are recorded and information about them is made available for simulation, allowing tests of driving control systems (or their sensors) to be carried out based on realistic data. The advantage of transferring the data to a simulation is the possibility of safe virtual testing, as well as the possibility of deliberately modifying the real-life data.

[0007] DE 102019206 908 B4 relates to a method for training at least one algorithm for a control unit of a motor vehicle, wherein the control unit is provided for implementing an automated or autonomous driving function by intervening in assemblies of the motor vehicle on the basis of input data using the at least one algorithm, wherein the algorithm is trained by a self-learning neural network, comprising the following steps: a) providing a computer program product module for the automated or autonomous driving function, wherein the computer program product module contains the algorithm to be trained and the self-learning neural network, b) embedding the trained computer program product module in the control unit of the motor vehicle, c) driving the motor vehicle in a real traffic environment by a human driver, wherein the driving determines a driven trajectory,d) feeding data from an environmental sensor system and a motor vehicle sensor system to the control unit and calculating a virtual trajectory using the algorithm, e) deriving a metric from a comparison of the driven trajectory and the virtual trajectory and storing the data from the environmental sensor system and the motor vehicle sensor system in a memory if certain metric criteria are met for a traffic situation, f) providing information regarding the traffic situation to a traffic simulation, g) analyzing the traffic situation using the traffic simulation, whereby a virtual image of the traffic situation is created using the map data, the environmental sensor system and the motor vehicle sensor system, whereby the data of the traffic situation are varied using the traffic simulation,and h) training the algorithm by varying the traffic situation. The traffic simulation is carried out as a simulation of the traffic situation from the perspective of the traffic environment. Behavioral and technological models are provided for each individual road user in the traffic simulation, so that a future traffic situation is simulated based on a variation of the provided behavior and technological models. Furthermore, past environmental and vehicle data are also varied in the traffic simulation.

[0008] The publication "A Needle in a Haystack - How to Derive Relevant Scenarios for Testing Automated Driving Systems in Urban Areas" by Nico Weber, Dr.-lng. Christoph Thiem, and Prof. Dr.-lng. Ulrich Konigorski, a preprint for the "30th Aachen Colloquium Sustainable Mobility 2021," also demonstrates the use of rigid scenarios for traffic simulation and their flexibility. Based on an "adaptive replay-to-sim" approach, it presents a simulation-based toolchain for the development and testing of vehicles equipped with automated driving control systems in urban environments. Multimodal interactions between different road users are considered and, among other things, provided for feeding back into a scenario database.

[0009] For testing and / or validating automated driving control systems, the use of scenario libraries is also known. These ideally contain descriptions of a large number of (ideally as widely as possible) varying scenarios, so that a vehicle does not encounter fundamentally untested situations during subsequent regular operation. In such scenario libraries, scenario descriptions are typically stored in a machine-readable and automatically storable format. Scenario description languages ​​that employ special documentation and programming languages ​​are usually used for this purpose. One example of this is the established scenario description language of the "ASAM eV" association called 'OpenSCENARIO®' for use in scenario libraries. There are also scenario libraries with data in machine-readable formats that also have a human-understandable context, such as "ADScene" and "SafetyPool."

[0010] In order to be able to use such a scenario library to investigate the behavior of a driving control system or its sensor unit in a reactive manner to the environment of a road user under consideration, in particular an automated vehicle (typically called a "vehicle under test"), the journey of this selected real road user under consideration can be simulated, while the behavior of the other real road users as well as the environmental parameters can be taken from the scenario library and, in this case, are based on real data.

[0011] Due to the general difficulty of translating information from data sources such as the aforementioned real traffic scenarios (compare the "Adaptive Replay to Sim" approaches introduced above) into a data structure and / or data quality suitable for simulation, there is a need to make simulation-appropriate data more easily available. This is because not all data sources, such as reality-based data from real traffic scenarios and variations from information in scenario libraries, are necessarily suitable for use unmodified for system testing in simulations, be it in hardware-in-the-loop tests or software-in-the-loop tests, or model-in-the-loop tests of individual vehicle components such as sensors or computing units of an automated driving control system, or be it for systemic tests of the behavior of automated vehicles as a whole and with all their subsystems, including automated vehicles in interaction with one another.Depending on the degree of substitution of the vehicle component to be tested (through hardware-in-the-loop, prototype-in-the-loop, software-in-the-loop, model-in-the-loop, etc.), it may be necessary to artificially synthesize input data from the vehicle component integrated into the simulation. However, if such scenario libraries are already capable of containing adequate data as system input for testing a vehicle component, the problem arises of having to generate such adequate data from other data sources for the scenario libraries.

[0012] It is therefore an object of the invention to solve these problems and to more easily provide compatible data from (diverse) data sources for a simulation test of a vehicle component.

[0013] The invention is based on the features of the independent claims. Advantageous developments and refinements are the subject of the dependent claims.

[0014] A first aspect of the invention relates to a system for generating input data for a vehicle component to be tested by means of a simulation from a description of a traffic scenario, wherein the vehicle component to be tested has an input interface for receiving the input data, wherein the system has a computing unit with a first interface for receiving a specification with information about a necessary quality and / or a necessary data format of the input data for compatibility with the vehicle component and for receiving information about a traffic scenario intended for the simulation, wherein the computing unit is designed to transform the information about the traffic scenario into the input data by executing a processing module with a pre-trained artificial neural network such that the input data are compatible with the vehicle component to be tested according to the specification,wherein the processing module can transform a plurality of information about a respective traffic scenario, which differ in terms of nature and / or data format, into input data, and wherein the computing unit is designed to execute a physics module to ensure the physical plausibility of the input data generated by the processing module in order to transform input data describing at least physically impossible conditions for the vehicle component to be tested into physically possible conditions.

[0015] The vehicle component to be tested is either integrated into the simulation as a physical device using a hardware-in-the-loop method, or is virtually mapped and also simulated.

[0016] Therefore, such a vehicle component (as a physical device, a virtual image, or in the form of an implementation of an algorithm) is provided for testing in a simulation, which has an input interface to receive input data. If the vehicle component is a sensor, this input data is, in particular, information carriers, such as images (in which case the sensor corresponds to a camera) or radio waves (in which case the sensor corresponds to a radar sensor).If the vehicle component is an algorithm or a computing module with the algorithm implemented on it, the input data can correspond to the output data of a sensor and can be transmitted, for example, as an image data stream in a data-technical video format to the vehicle component as input data in the vehicle, so that exactly this (generated) interface is deliberately used in the simulation to send artificial video data, which was generated by the processing module, to the vehicle component to be tested in the sense of the computing module processing the video images.The specification with the information on the necessary compatibility includes, for example, the specification that the vehicle component to be tested has an input interface that has an Ethernet adapter or is an interface to a vehicle CAN bus, so that the input data must also be transmitted to the vehicle component to be tested via a CAN bus (physically or virtually represented in the simulation). It is therefore the particular task of the artificial neural network of the processing module, depending on the specification regarding the necessary nature or the necessary data format of the input data, to generate compatible and thus suitable data for the input interface and the vehicle component as a whole. This advantageously ensures that the vehicle component receives data via the input interface that is fully compatible for its intended use.This includes, in particular, as already described above, a corresponding data format, since typically all electronic vehicle components are designed for one or a limited number of data formats. Cameras are designed to capture visual information, lidar sensors for detecting reflections in the vicinity of an emitted laser beam, radar sensors for reflected radar waves in the environment, ultrasonic sensors for reflected ultrasound, etc.; so that these sensors receive their corresponding information, which they would otherwise naturally capture in reality, based on the result of the artificial neural network's calculations (in the above examples, corresponding to environmental images (e.g., projected onto a screen), ultrasonic waves, radio waves, light waves in the frequency range of the lidar laser, etc.).); any interfaces downstream of a sensor can also be used with the corresponding formats of the data generated by the artificial neural network, coordinating a time synchronization (preferably real-time) of the relevant vehicle component. For example, at any interface of a vehicle's data bus, the vehicle component to be tested can be embedded in the simulation as a physical element (or its virtual image in the simulation), and the protocol provided on this data bus can be adhered to in all its aspects in the data generated by the artificial neural network. Thus, in the simulation, the vehicle component to be tested is always supplied with data that it could in principle also receive during a real journey of the vehicle with this vehicle component.For this purpose, the embedding of the vehicle component to be tested into the simulation with the input data generated by the artificial neural network in a compatible manner opens up the possibility of arbitrarily varied data in order to create a simulation volume that would not be practically possible in real test drives using a real vehicle due to the large amount of test time required.

[0017] The processing module is advantageously designed such that it can receive a multitude of different descriptions of a respective desired traffic scenario in order to transform these descriptions into the above-described compatible form for the vehicle component. Such different descriptions are advantageously at least one of image sequences (from the first-person perspective and / or from a bird's-eye view), internal vehicle communication (such as that taken from a real journey on a vehicle's CAN bus), descriptions in human language, descriptions in flowcharts, text in natural language, data from internal vehicle communication, V2X data, specified description languages ​​(e.g., OpenScenario), silent testing data, symbolic scenario descriptions, databases, image descriptions, video sequences from entertainment films or documentaries, etc.What these data sources have in common is that they contain information about a traffic situation or parts of it, either explicitly or implicitly. By executing the pre-trained artificial neural network, the computing unit generates an abstracted internal representation of the information about the intended traffic scenario and, derived from this, appropriate input data for the input interface.

[0018] The processing module therefore has a scenario representation model, particularly internally. This scenario representation model is preferably based on a translation model similar to a transformer network (known from: Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, AN, . . . Polosukhin, I. (2017). "Attention is All you Need." Proceedings of the 31st International Conference on Neural Information Processing Systems (pp. 6000-6010). Red Hook, NY, USA: Curran Associates Inc.), which, due to its network architecture, is capable of converting input data, which is available in various formats, into an internal representation (compare autoencoders for encoding and subsequent decoding of data).

[0019] The processing module comprises the artificial neural network for the internal representation of scenarios based on different formats of information about a selected traffic scenario. To train this network, two-element tuples are used as data pairs, for example. The tuples comprise, in particular on the input side, one or more images of a driving scene in the sense of a snapshot or frame of a traffic scenario and the associated description, or from a text in natural language and simultaneously recorded vehicle communication data. The required information of the input part is, in particular, generated synthetically in an upstream simulation and / or is obtained from silent testing recordings or real driving records, or from data freely available on the Internet. This approach can be used to obtain the required data volume for machine learning models.The amount of data obtained is further increased through appropriate permutations. The processing module therefore either comprises a very large artificial neural network or is formed by interconnecting several domain-specific artificial neural networks.

[0020] While the artificial neural network itself merely ensures adequate transformation into compatible input data for the vehicle component and its interface under test, this could still result in physically absurd scenarios for the vehicle component under test. The physics module takes care of preventing such cases (since they cannot occur in reality by definition and therefore do not need to be safeguarded for a vehicle component under test).

[0021] The physics module specifically checks the current internal representation by making the scenario representation model available to the physical model, which evaluates the physical plausibility of the internal representation and can thus, for example, monitor predefined limits of the vehicle's physics. If the metrics of the physical model (e.g., maximum acceleration, speed) are below a threshold, the input data can be output.

[0022] For this purpose, the artificial neural network was pre-trained, in particular, with corresponding predefined pairs of input and output data, preferably using the "supervised learning" method known in the broad field of machine learning. In this case, an extremely large number of such pairs of various combinations of information about a traffic scenario intended for simulation, with a multitude of different vehicle components and their respective input interfaces, are preferably pre-defined and used to design the parameters of the artificial neural network. These parameters are designed in such a way that an execution of the artificial neural network, based on the input data, also generates the corresponding and predefined output data, so that usable output data can also be obtained in the future with additional input data.

[0023] Advantageously, the artificial neural network is designed and pre-trained in such a way that the transformation of information about a traffic scenario intended for simulation into the input data occurs with a high degree of realism. The degree of realism is defined here such that the differences between real recordings and synthetically generated data are minimized.

[0024] Compatible with the input interface or with other modules of the vehicle component to be tested means in particular that the input data is in a form that can be processed by a computer module of the vehicle component to be tested (e.g. as text, image file, video file, ...), and within these categories is in the correct format with the corresponding conventions for the file structure or the structure of the data stream.

[0025] It is therefore an advantageous effect of the invention that it is possible to generate realistic data with compatibility and suitability as system input for the vehicle component to be tested from any type of scenario description for a simulation to test the vehicle component, be they implicit or explicit scenario descriptions, and that this data can be varied to a high degree, taking the degree of realism into account. The data, in particular sensor data, can advantageously be generated automatically from such scenario descriptions. Automated parameter variation and a sensible parameter sampling rate of the input data of the vehicle component can also be used to directly manipulate realistic recordings. The data for the vehicle component to be tested can, for example, be video sequences from the front camera of a vehicle, wave patterns for radar orUltrasonic sensors, WLAN communication packets, messages from internal vehicle communication (CAN, LIN, Ethernet, ...), or other data formats depending on the desired application and the vehicle component to be tested. The proposed invention makes it possible to generate domain-specific input data. For example, an automated vehicle on a test bench would require realistic image data for the camera. In this way, information about a traffic scenario intended for a simulation is used to generate physically plausible and temporally coordinated images, which can be recorded by the vehicle camera, for example, as the vehicle component to be tested. Another example is the generation of radar data for the situation-appropriate embedding of the vehicle's radar system in a simulation. The road users present in the described scenario thus receive a radar echo that matches the current situation.The input data also includes control signals for simulation hardware, provided the input data cannot directly represent the target format of the input interface. This includes, for example, the display of output image sequences using a projector or screen, or, for example, modulation signals for radio network simulation devices. It is therefore also advantageous that the complexity of the required simulation hardware can be gradually reduced, depending on the performance of the computing unit used. Thus, corresponding input data for the vehicle component to be tested, along with its input interface, is available for testing embedded in a simulation. Such testing of the vehicle component includes, in particular, a large number of scenarios in order to be able to test the vehicle component for safe operation and subsequently validate it.

[0026] According to an advantageous embodiment, dynamic equations are implemented in the physics module, which have basic kinematic equations for the relationship between inertia, accelerations, and forces or moments.

[0027] According to a further advantageous embodiment, the physics module comprises a physics model pre-trained by machine learning.

[0028] Regardless of whether dynamic equations are implemented, thus performing an analytical plausibility calculation, or whether an empirical model in the form of a physics model pre-trained by machine learning (e.g., an artificial neural network) is used, kinematic limits can be specified. In particular, maximum accelerations for starting and decelerating road users are defined, as well as maximum centrifugal forces and thus accelerations for cornering, maximum pedestrian speeds, and the like. Inertial effects are also preferably considered, according to which jump signals, for example, cannot occur naturally at speeds.

[0029] According to a further advantageous embodiment, the computing unit is designed to execute a synchronization module in order to coordinate a plurality of temporally related information about the traffic scenario for the simulation and to transform the input data into a time scale suitable for the vehicle component to be tested in a synchronized manner, in particular in real time.

[0030] According to a further advantageous embodiment, the information about the traffic scenario intended for the simulation comprises information from a scenario library.

[0031] According to a further advantageous embodiment, the information about the traffic scenario intended for the simulation comprises at least information about a past real traffic scenario recorded with a camera unit. According to a further advantageous embodiment, the computing unit has a second interface for detecting a manually specified variation of the information about the traffic scenario, so that the processing module can generate compatible input data for the vehicle component to be tested based on the manually specified variations of the information about the traffic scenario.

[0032] According to a further advantageous embodiment, the computing unit is designed to continuously generate new and current input data for the vehicle component to be tested during the runtime of the simulation.

[0033] According to a further advantageous embodiment, the computing unit has a third interface for obtaining current simulation data from the simulation for testing the vehicle component, wherein the computing unit is designed to generate the current input data depending on such current simulation data that describe the behavior of a simulated road user with the vehicle component to be tested among a plurality of other real data-based or agent model-guided road users.

[0034] A further aspect of the invention relates to a method for producing a system for generating input data for a vehicle component to be tested by means of a simulation from a description of a traffic scenario, comprising the steps:

[0035] - Creating a processing module of the system by training an artificial neural network of the processing module by specifying data pairs, wherein a predetermined input part of the data pairs comprises information about a traffic scenario and a predetermined output part of the data pairs comprises input data for (and compatible with) an input interface of a vehicle component to be tested, wherein sequentially different types of information are specified for different traffic scenarios and wherein sequentially different input data associated with the traffic scenarios are specified for different vehicle components, so that after completion of the training the processing module is able to process different types of information about traffic scenarios with regard to their nature and / or data format;

[0036] - generating a physics module of the system by training an artificial neural network of the physics module; and

[0037] - Integrating the generated processing module and the generated physics module into a common system by establishing a common interface between the processing module and the physics module.

[0038] The physics module serves to ensure the physical plausibility of the input data generated by the processing module and is generated in the aforementioned process by machine learning methods, in particular by training an artificial neural network.

[0039] Advantages and preferred developments of the proposed method result from an analogous and analogous transfer of the statements made above in connection with the proposed system.

[0040] Further advantages, features, and details will become apparent from the following description, which – where appropriate with reference to the drawings – describes at least one embodiment in detail. Identical, similar, and / or functionally equivalent parts are provided with the same reference numerals.

[0041] They show:

[0042] Fig. 1 : A method for producing a system for generating input data for a vehicle component to be tested by means of a simulation from a description of a traffic scenario according to an embodiment of the invention.

[0043] Fig. 2: A system for generating input data for a vehicle component, which is generated by the method according to Fig. 1.

[0044] Fig. 1 shows a method for producing a system for generating input data for a vehicle component 1 to be tested by means of a simulation (cf. Fig. 2) from a description of a traffic scenario.In a first step, the generation S1 of a processing module of the system takes place by training an artificial neural network of the processing module by specifying data pairs, wherein a predetermined input part of the data pairs comprises information about a traffic scenario and a predetermined output part of the data pairs comprises input data for an input interface of a vehicle component 1 to be tested, wherein sequentially different types of information are specified for different traffic scenarios and wherein sequentially different input data associated with the traffic scenarios are specified for different vehicle components 1, so that after completion of the training the processing module is able to process different types of information about traffic scenarios with regard to their nature and / or data format.The internal representation generated by this processing module to generate the input data does not have to follow the laws of physics. To achieve this, the physics module is used. In addition to learning from the training data, the physics module serves to shape internal representations based on physically plausible processes. The physics module receives an image of the internal representation and creates a reward function using a metric (e.g., the degree to which a maximum acceleration is exceeded). The reward is fed into the model for internal representation, so that physical conditions can be memorized using reinforcement learning (e.g., deep Q-learning).Accordingly, S2 of a physics module of the system is created by training an artificial neural network of the physics module; the input data generated by the processing module must also take into account the vehicle geometry as well as the physical and electronic architecture of the vehicle component to be tested. To account for the vehicle geometry, two approaches are used: A model trained without concrete vehicle geometry serves as an abstract basis for specialized variants of the proposed solution. This pre-trained network is then trained analogously using a physical unit in a further reinforcement learning process, which now takes the expanded conditions of the vehicle geometry into account. Another approach is the use of recorded data from the target vehicle, which is then used to fine-tune the pre-trained model.Finally, the integration S3 of the generated processing module and the generated physics module into a common system is carried out by establishing a common interface between the processing module and the physics module.

[0045] Fig. 2 shows a system for generating input data for a vehicle component 1 to be tested by means of a simulation from a description of a traffic scenario, wherein the vehicle component 1 to be tested has an input interface for receiving the input data. In this example of Fig. 2, the vehicle component 1 to be tested is an automated driving control system for controlling the longitudinal and lateral movement of a vehicle. For the test, the vehicle component 1 is reduced to a front camera and the processing unit. Other components of the vehicle component 1 to be tested are in a simulation. The task is now to test the ability of the vehicle component 1 to be tested to drive through a given scenario without risk.The system for generating the input data comprises a computing unit 3 with a first interface 5 for receiving a specification with information about a required quality and / or a required data format of the input data for compatibility with the vehicle component 1 and for receiving information about a traffic scenario intended for the simulation. The input data consists of pre-recorded data from at least one past real traffic scenario, obtained from data from stationary traffic space observation and from drones. These describe vehicle trajectories.The computing unit 3 is further configured to transform the information about the traffic scenario into the input data by executing a processing module with a pre-trained artificial neural network such that the input data is compatible with the vehicle component 1 under test according to the specification. The processing module can transform a multitude of information about a respective traffic scenario, differing in nature and / or data format, into input data. By executing the artificial neural network, the computing unit generates a basic scene based on the information about the specified traffic scenario, which can be displayed as a sequence of realistic images on a monitor for the camera of the system under test.In addition, the computing unit 3 executes a physics module to ensure the physical plausibility of the input data generated by the processing module, in order to transform input data describing at least physically impossible conditions for the vehicle component 1 under test into physically possible conditions. The vehicle component 1 under test is part of a virtual vehicle that replaces one of the road users from the real traffic scenario and independently drives to a predetermined target position in this traffic scenario and acts independently to reach this target position. It therefore replaces one of these vehicles and acts independently with a predetermined target position in this scenario. The behavior of the vehicle component 1 under test is also made available to the processing module. This behavior is transmitted in JavaScript object notation format (JSON).Thanks to the internal representation, the elements of the generated images can be adapted. Another software unit checks the plausibility of the created scenario and describes changes in the behavior of the road users. These changes can be transferred to the processing module in XML, for example. By feeding back the behavior of the vehicle component 1 under test and the data from the adapted behavior models, the scenario representation is continuously updated according to the data in order to map the current state of all road users at all times. A sporadically present test engineer would like to vary the parameters of the scenario in order to quickly verify an incident. He expresses this in natural language: "Please increase the speed of the vehicle in front by 20 km / h for the next 10 minutes of the simulation.Following this instruction from the test engineer, computing unit 3 adapts the generation of input data via this command received via the second interface, so that the input data is adjusted according to the command and remains compatible with the input interface of vehicle component 1. In the example specified by the test engineer, the vehicle in front would now accelerate upon input. This behavior is then represented in the realistic images. The change in behavior would also affect the generated images. The output data is physically correct in all cases and adapted to the physical conditions of the scene and the sensor of vehicle component 1. In this example, the installation position of the camera and its specifications are also taken into account.

[0046] Although the invention has been illustrated and explained in detail by preferred embodiments, the invention is not limited by the disclosed examples, and other variations may be derived therefrom by those skilled in the art without departing from the scope of the invention. It is therefore clear that a multitude of variations exist. It is also clear that exemplary embodiments are truly only examples and should not be construed as limiting the scope, possible applications, or configuration of the invention in any way.Rather, the preceding description and the description of the figures enable the person skilled in the art to implement the exemplary embodiments in concrete terms, whereby the person skilled in the art, with knowledge of the disclosed inventive concept, can make various changes, for example with regard to the function or the arrangement of individual elements mentioned in an exemplary embodiment, without departing from the scope of protection defined by the claims and their legal equivalents, such as further explanations in the description.

[0047] List of reference symbols

[0048] 1 vehicle component

[0049] 3 Computing unit

[0050] 5 first interface

[0051] 7 second interface

[0052] 51 Create

[0053] 52 Generate

[0054] 53 Integrate

Claims

Patent claims 1. A system for generating input data for a vehicle component (1) to be tested by means of a simulation from a description of a traffic scenario, wherein the vehicle component (1) to be tested has an input interface for receiving the input data, wherein the system has a computing unit (3) with a first interface (5) for receiving a specification with information about a necessary condition and / or a necessary data format of the input data for compatibility with the vehicle component (1) and for receiving information about a traffic scenario intended for the simulation, wherein the computing unit (3) is designed to transform the information about the traffic scenario into the input data by executing a processing module with a pre-trained artificial neural network such that the input data are compatible with the vehicle component (1) to be tested according to the specification,wherein the processing module can transform a plurality of information items about a respective traffic scenario, which differ in terms of their nature and / or data format, into input data, and wherein the computing unit (3) is designed to execute a physics module to ensure the physical plausibility of the input data generated by the processing module in order to transform input data describing at least physically impossible conditions for the vehicle component (1) to be tested into physically possible conditions.

2. System according to claim 1, wherein dynamic equations are implemented in the physics module, which have basic kinematic equations for the relationship between inertia, accelerations, and forces or moments.

3. System according to one of the preceding claims, wherein the physics module comprises a physics model pre-trained by machine learning.

4. System according to one of the preceding claims, wherein the computing unit (3) is designed to execute a synchronization module in order to coordinate a plurality of temporally related information about the traffic scenario for the simulation and to process the input data to be transformed into a time scale synchronized, in particular in real time, suitable for the vehicle component (1) to be tested.

5. System according to one of the preceding claims, wherein the information about the traffic scenario intended for the simulation comprises information from a scenario library.

6. System according to one of the preceding claims, wherein the information about the traffic scenario intended for the simulation comprises at least information about a past real traffic scenario recorded with a camera unit.

7. System according to one of the preceding claims, wherein the computing unit (3) has a second interface (7) for detecting a manually specified variation of the information about the traffic scenario, so that compatible input data can be generated by the processing module on the basis of the manually specified variations of the information about the traffic scenario for the vehicle component (1) to be tested.

8. System according to one of the preceding claims, wherein the computing unit (3) is designed to continuously generate new and current input data for the vehicle component (1) to be tested during the runtime of the simulation.

9. System according to claim 8, wherein the computing unit (3) has a third interface for obtaining current simulation data from the simulation for testing the vehicle component (1), wherein the computing unit (3) is designed to generate the current input data depending on such current simulation data which describe the behavior of a simulated road user with the vehicle component (1) to be tested among a plurality of other real data-based or agent model-guided road users.

10. A method for producing a system for generating input data for a vehicle component (1) to be tested by means of a simulation from a description of a traffic scenario, comprising the steps: - Creating (S1) a processing module of the system by training a artificial neural network of the processing module by specifying data pairs, wherein a predetermined input part of the data pairs comprises information about a traffic scenario and a predetermined output part of the data pairs comprises input data for an input interface of a vehicle component (1) to be tested, wherein the input data is compatible with the input interface, wherein sequentially different types of information are specified for different traffic scenarios and wherein sequentially different input data associated with the traffic scenarios are specified for different vehicle components (1), so that after completion of the training the processing module is able to process different types of information about traffic scenarios with regard to their nature and / or data format; - generating (S2) a physics module of the system by training an artificial neural network of the physics module; and - integrating (S3) the generated processing module and the generated physics module into a common system by establishing a common interface between the processing module and the physics module.

Images