Conversion of measurement data between measurement modes

Abstract

To provide conversion of measured data between measurement modes.SOLUTION: Provided is a method (100) for converting measured data (1) of at least one source measurement mode (Q) into realistic measured data (2) of at least one target measurement mode (Z). The method includes: a step (110) in which the measured data (1) of the source measurement mode (Q) are mapped onto representations (6) in a latent space by using an encoder (3) of a trained encoder-decoder configuration (5); and a step (120) in which the representations (6) are mapped onto the realistic measured data (2) of the target measurement mode (Z) by using a decoder (4) of the encoder-decoder configuration (5), the amount of information of the representations (6) of measured data (1) in the latent space being smaller than the amount of information of the measured data (1).SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to the conversion of measurement data from a source measurement mode to measurement data of a target measurement mode, thereby enabling, for example, the inspection of control logic for a driving assistance system or a system for at least partially autonomous driving of a vehicle.

Background Art

[0002] Systems for driving assistance and for at least partially autonomous driving of a vehicle make decisions about driving maneuvers based on observations of the vehicle surroundings by sensors. The observation results are recorded, for example, in the form of camera images, video images, thermal images, ultrasonic images, radar data, and lidar data. These raw signals are further processed and partially adjusted at various levels of abstraction.

[0003] For example, the raw signals can first be processed in the form of a "position", i.e., the location where an object was recognized based on the corresponding sensor. The "positions" generated from the measurement data of multiple sensors can be fused, thereby ultimately creating an overall image of the vehicle's surroundings in the form of a list of all objects present there.

Summary of the Invention

Problems to be Solved by the Invention

[0004] Realistic signals at each intermediate stage are required so that the individual components of this processing chain can be tested and optimized individually. Therefore, it is desirable to be able to generate these realistic signals from a given test scenario.

Means for Solving the Problems

[0005] In the framework of the present invention, a method has been developed for converting measurement data of at least one source measurement mode into realistic measurement data of at least one target measurement mode.

[0006] Here, the concept of "measurement mode" is not limited to the physical contrast mechanism from which measurement data can be acquired. For example, various processing or abstraction stages of measurement data representing a particular contrast mechanism can also be considered various measurement modes.

[0007] Furthermore, the concept of "measurement data" is not necessarily related to the actual recording of physical data. Data obtained through simulations of physical measurement processes, as well as detailed descriptions of realistic scenes that can be observed in such measurement processes, can also be considered measurement data.

[0008] Measurement data in source measurement mode is mapped to a representation in latent space by the encoder of a trained encoder-decoder configuration. The representation is then mapped to realistic measurement data in target measurement mode by the decoder of the encoder-decoder configuration.

[0009] Here, the amount of information in the representation of the measurement data in the latent space is smaller than the amount of information in the measurement data itself. Therefore, the measurement data is compressed to contain less information. The concept of information can be understood in the sense of Shannon's information theory as parameters such as the minimum number of bits required to transmit the measurement data or its representation.

[0010] For example, the dimensions of the latent space, • The encoder makes the dimension of the space from which it acquires measurement data in source measurement mode smaller than the dimension of the space, • Make the dimension smaller than the spatial dimension in which the decoder maps the measurement data for the target measurement mode. This allows measurement data to be compressed into a smaller amount of information.

[0011] The latent space may have, for example, the same or larger dimensions as the space in which the encoder acquires measurement data in the source measurement mode, but nevertheless, the representation in the latent space may have smaller dimensions than the measurement data in the source measurement mode. For example, many of the parameters characterizing the representation in the latent space can be restricted in the form of being related to each other by quadratic conditions, boundary conditions, and / or optimization criteria.

[0012] If the amount of information in the representation in latent space is small compared to the measured data, a "bottleneck" is formed that the information must pass through on its way from the encoder to the decoder, similar to known automatic encoders where the decoder reconstructs the data originally supplied to the encoder. Here, the decoder no longer reconstructs the original data but provides the measured data for the target measured mode, resulting in an unexpectedly advantageous division of labor between the encoder and decoder during the training of the encoder-decoder configuration. That is, the encoder compresses the measured data for the source measured mode in the representation into important informational content about the scene to which these measured data relate, and the specifications regarding the source measured mode, i.e., the "filter (Brille)" through which this scene is seen in the measured data, are removed. The decoder adds the specifications regarding the target measured mode to the compressed representation. Thus, the realistic measured data for the target measured mode provided by the decoder shows how the scene, characterized by the compressed representation and de-specified for the source measured mode, looks through the "filter" of the target measured mode.

[0013] The measurement data obtained in this way for the target measurement mode can be used in various formats. For example, if labeled measurement data for the target measurement mode is required for training a neural network such as an image classification function, this can be generated from labeled measurement data for the source measurement mode while retaining the labels. Here, it is not necessary to directly label the measurement data for the target measurement mode. In particular, for example, measurement data for multiple different target measurement modes can be obtained from the same amount of labeled measurement data for the source measurement mode.

[0014] For example, traffic images can be manually labeled to indicate the location and type of objects present. These images can then be converted to radar or LiDAR data, which can also be readily and appropriately labeled. Manual labeling does not need to be repeated multiple times. Furthermore, image labeling can be performed much faster and with less expertise than labeling radar or LiDAR data.

[0015] However, the measurement data in target measurement mode can also be used, for example, in driver assistance systems or systems for fully or partially autonomous driving of vehicles, to inspect specific parts or components of the processing chain described above. Therefore, in particularly advantageous embodiments, the measurement data in source measurement mode includes a description of a scene containing one or more objects. The measurement data in target measurement mode includes the spatial locations assigned to each object in the scene by the actual sensors during the physical observation of the scene. These locations correspond to the “positions” mentioned at the beginning. It is relatively easy to create a realistic description of a scene containing specific objects. On the other hand, predicting “positions” based on such descriptions is not so intuitive. However, the relationship between scene descriptions and “positions” can be used for machine learning.

[0016] Measurement data in target measurement mode can, in particular, include locations in space where an actual radar sensor or ridiculous sensor records radar or ridiculous reflections during a physical observation of the scene. This is the primary format in which these sensors output information about the scene.

[0017] A scene description can, in particular, include a clear description of multiple geometric shapes in a plane or space, each of which is considered to be occupied by an object. These geometric shapes may be, for example, rectangles or cuboids, so-called "bounded boxes." The attribution of locations to such "bounded boxes" can be checked computationally, particularly quickly and easily.

[0018] In a particularly advantageous configuration, the measurement data from the source measurement mode is converted into an input image or input point cloud. The input image or input point cloud is mapped to a representation in latent space by an encoder. This representation is then mapped to an output image or output point cloud by a decoder. The output image or output point cloud is finally converted into measurement data for the target measurement mode.

[0019] This means that the encoder-decoder configuration operates based on an image or point cloud. In particular, existing configurations that operate based on an image or point cloud can continue to be used or modified.

[0020] Furthermore, by converting the output image or output point cloud into measurement data for a target measurement mode in this way, it is possible to set, for example, the number of ultimately determined locations ("positions") that present this measurement data for the target measurement mode. This number is an important criterion for determining the degree to which the downstream algorithms perceive the measurement data for the target measurement mode as realistic. When the decoder provides an output image, a variable number of points can also be sampled from this output image, for example. Alternatively, the output point cloud can be thinned out later by starting with the maximum number of points and randomly subtracting a specific number of points from these points.

[0021] Scene descriptions based on geometric shapes such as "bounded boxes" can be converted into input images by, for example, discretizing the scene using a 2D or 3D grid, and checking whether each grid point belongs to the geometric shape defined in the description.

[0022] In this grid, for example, a numerical value of 0 can represent that the grid point is empty, and a numerical value of 1 can represent that the grid point is occupied. The belonging of an object to a particular class can also be encoded numerically. Furthermore, additional numerical values ​​related to additional information such as the object's material and / or movement state can be assigned to the grid points.

[0023] The description of a scene based on geometric shapes such as "bounding boxes" can be converted into an input point cloud, for example, by deriving the coordinates of points from a distribution and adding each point belonging to one of the geometric shapes to the input point cloud. The points may be, for example, on the edges and / or surfaces of the geometric shapes and / or may completely fill these geometric shapes. Similar to the grid points described above, numerical additional information regarding the class of the object, material (e.g., surface material), and / or movement state can be added and encoded as numerical values to these points. For example, this additional information can be encoded in the additional coordinates of the points, and thus these points are in a space with four or more spatial dimensions.

[0024] In a further particularly advantageous form, the measurement data of several source measurement modes are mapped by respective encoders into representations in the latent space. These representations are mapped by the same decoder into the measurement data of the target measurement mode. Thus, the decoder only needs to be trained once in order to be able to process the measurement data of various source measurement modes.

[0025] Similarly, in a further particularly advantageous form, at least one representation in the latent space is mapped by at least two different decoders into the measurement data of at least two target measurement modes. That is, for example, when realistic measurement data provided by a new sensor regarding the same scene is generated, the encoder can be left as it is and only a new decoder specific to the new sensor can be trained.

[0026] As described above, an important application of the method described above is to generate realistic measurement data of the target measurement mode in order to inspect a driving dynamic system. Thus, the present invention also relates to a method for inspecting control logic for a driving assistance system or a system for at least partially automated driving of a vehicle.

[0027] In the framework of this method, a description of a scene including a predetermined arrangement of objects is provided as measurement data in a source measurement mode. These measurement data in the source measurement mode are converted into measurement data in a target measurement mode using the method described above.

[0028] The measurement data in the target measurement mode obtained in this way are supplied as an input to the control logic of the inspection target. The reaction presented by the control logic based on this input is compared with a predetermined target reaction regarding the scene. From the result of this comparison, the extent to which the control logic can handle the scene is evaluated.

[0029] For example, in at least partial autonomous driving of a vehicle, there are many abstract situations that the control logic needs to learn, and it is not important how those situations look in detail or in reality. For example, when a vehicle controlled by the control logic is on a collision course with another vehicle, the only important thing is that an emergency braking and / or collision avoidance operation is initiated. On the other hand, the exact passenger car model of the other vehicle, the color or surface material of this vehicle, or the current weather is not at all important. Therefore, the control logic does not function based on raw signals such as camera images, but based on the location ("position") belonging to the other vehicle based on sensors.

[0030] The advantage of the inspection procedure described here is that the control logic can be inspected separately. In a test during a driving operation where the vehicle's sensors naturally capture the scene, the finally presented reaction depends on both the capture and processing of the scene by the sensors and the further processing by the control logic. When the presented reaction does not correspond to the target reaction, in addition to suboptimal processing in the control logic, the low quality of the raw signals captured by the sensors and suboptimal processing in the sensors are also considered as error causes.

[0031] When a scene description is translated into input for control logic by a physical or phenomenological model of the sensor, uncertainty regarding physical data acquisition is eliminated. Furthermore, since the scene does not need to occur spontaneously, a much larger class of scenes can be examined. In this regard, it is difficult to guarantee that the input generated by the physical or phenomenological model of the sensor is actually realistic, as this depends on the accuracy of the modeling. On the other hand, when the measurement data of the target measurement mode is generated as input for control logic by the method described above, the realism of this input is guaranteed by prior training of the encoder-decoder configuration. The only unknown is the behavior of the control logic.

[0032] The present invention also relates to a method for training an encoder-decoder configuration for use in one of the methods described above.

[0033] In this method, training source measurement data is prepared for at least one source measurement mode, and these training source measurement data represent a given physical scene. Furthermore, training target measurement data is prepared for at least one target measurement mode for the same physical scene.

[0034] The training source measurement data is first mapped to a representation in latent space by an encoder-decoder configuration, and from there it is mapped to the measurement data of the target measurement mode. Based on a given cost function, the degree to which the measurement data of the target measurement mode matches the training target measurement data is evaluated.

[0035] The parameters characterizing the encoder behavior and / or decoder behavior are optimized with the expectation that further processing of the training source measurement data by the encoder-decoder configuration will improve the evaluation of the measurement data of the target measurement mode that occurs later.

[0036] As explained above, in this type of training, the encoder learns to transfer the essential information needed to understand the scene into a compressed representation, thereby de-specifying the source measurement mode from the measurement data. The decoder, on the other hand, learns to add the target measurement mode specification to the general representation in the latent space, so that the measurement data provided by the decoder appears as if it had been captured or processed in this target measurement mode. How strongly this desired behavior is characterized may depend, for example, on the relationship between the dimensions of the latent space, the dimensions of the space in which the encoder acquires measurement data for the first measurement mode, and the dimensions of the space to which the decoder maps. Therefore, for example, another parameter that defines the dimensions of the latent space and / or the amount of information in the representation within the latent space can be used as a hyperparameter that is also optimized during training.

[0037] In particularly advantageous embodiments, the training target measurement data is converted into a training target image or training target point cloud. The degree of agreement between the output image or output point cloud generated by the decoder and the training target image or training target point cloud is evaluated as the degree of agreement between the measurement data of the target measurement mode and the training target measurement data. Here, the cost function used to train the encoder-decoder configuration can directly evaluate the image or point cloud generated by the decoder without considering the effect of the conversion of the measurement data of the target measurement mode into space.

[0038] In particular, in this configuration, the training source measurement data can be further transformed into training source images or training source point clouds, and the encoder-decoder configuration can transform these training source images or training source point clouds into images or point clouds. Here, for example, an existing encoder-decoder configuration for images or point clouds can continue to be used without modification.

[0039] The learning source measurement data and learning target measurement data can be introduced, in particular, within the framework of a test run using a vehicle equipped with at least one sensor. For this purpose, in addition to the measurement data provided by the sensor, the vehicle's GPS position and movement status can also be recorded to determine the size and position of the "bounded box".

[0040] The above methods can be implemented in whole or in part by a computer. Therefore, the present invention also relates to a computer program comprising machine-readable instructions that cause a computer to perform one of the above methods when executed on one or more computers. In this sense, a control device for a vehicle and an embedded system for a technical device that can also execute machine-readable instructions can also be considered a computer.

[0041] The present invention also relates to a downloadable product comprising a machine-readable data recording medium and / or a computer program. The downloadable product is a digital product that can be transmitted over a data network, i.e., can be downloaded by users of the data network, and can be made available for immediate download, for example, from an online shop.

[0042] Furthermore, computers can be equipped with computer programs, machine-readable data storage media, or downloaded products.

[0043] Along with a description of preferred exemplary embodiments of the present invention based on the drawings, further means of improving the present invention are described in more detail below. [Brief explanation of the drawing]

[0044] [Figure 1] This figure shows an exemplary embodiment of method 100 for converting measurement data 1 from source measurement mode Q to realistic measurement data for target measurement mode Z. [Figure 2] This figure shows an exemplary procedure of method 100 relating to input image 1a and output image 2a. [Figure 3] This figure shows an exemplary procedure of method 100 relating to the input point cloud 1b and the output point cloud 2b. [Figure 4] This figure shows an exemplary embodiment of method 200 for inspecting control logic 20. [Figure 5] This figure shows an exemplary embodiment of method 300 for training an encoder-decoder configuration 5. [Modes for carrying out the invention]

[0045] Figure 1 is a schematic flowchart of an exemplary embodiment of a method 100 for converting measurement data 1 from source measurement mode Q to realistic measurement data for target measurement mode Z.

[0046] The measurement data 1 conforms to the specifications of source measurement mode Q. In step 110, these measurement data 1 are mapped to representations 6 in latent space by encoder 3 of the trained encoder-decoder configuration 5. These representations 6 are compressed with respect to the original measurement data 1 and simultaneously de-specified according to source measurement mode Q.

[0047] In step 120, the representation 6 is mapped to realistic measurement data 2 for the target measurement mode Z by the decoder 4 of the encoder-decoder configuration 5. Thus, the measurement data 2 is, so to speak, "reconstructed" from the representation 6, with the specifications of the target measurement mode Z added.

[0048] According to block 111, the measurement data 1 of source measurement mode Q can be converted into an input image 1a or input point cloud 1b. According to block 112, the input image 1a or input point cloud 1b can be mapped to a representation 6 in latent space by the encoder 3.

[0049] According to block 121, this representation 6 can be mapped to the output image 2a or output point cloud 2b by the decoder 4. Here, for example, it is also possible that the representation 6 is generated from the input image 1a by the encoder 3 and mapped to the output point cloud 2b by the decoder 4. The encoder 3 must provide the representation 6 only in a format suitable for the decoder 4. Then, according to block 122, the output image 2a or output point cloud 2b can be converted into measurement data 2 for target measurement mode Z.

[0050] In particular, for example, the description 10a of scene 10 can be converted into an input image 1a by discretizing scene 10 into a two-dimensional or three-dimensional grid according to block 111a, and then, according to block 111b, checking whether each grid point belongs to a geometric shape defined in description 10a that is considered to be occupied by objects 11-13.

[0051] For example, the description of scene 10a can be transformed into the input point cloud 1b by deriving the coordinates of points from the distribution according to block 111c, and by adding each point belonging to one of the geometric shapes to the input point cloud 1b according to block 111d.

[0052] According to block 111e, additional information related to the material, class, and / or movement state of objects 11-13 from the description 10a of scene 10 can be assigned to the grid points of input image 1a or the points of input point cluster 1b.

[0053] According to block 113, measurement data 1 from multiple source measurement modes Q can be mapped to representations 6 in latent space by their respective encoders 3. According to block 123, these representations 6 can be mapped to measurement data 2 from target measurement mode Z by the same decoder 4.

[0054] According to block 124, at least one representation 6 in the latent space can be mapped to measurement data 2 for at least two target measurement modes Z by at least two different decoders 4.

[0055] Figure 2 shows a first exemplary procedure of method 100. The measurement data 1 of source measurement mode Q includes a description 10a of scene 10, which includes three objects 11-13. In this description, objects 11-13 are represented by bounded boxes, which can be shown, for example, in the form of four corner points of the objects.

[0056] According to block 111, these measurement data 1 are converted into an input image 1a, for example, by filling in a bounded box. Processing by the encoder-decoder configuration 5 according to blocks 121 and 122 results in an output image 2a in which the measurement data 2 for the target measurement mode Z is already encoded in image format. According to block 122, these measurement data 2 for the target measurement mode Z are extracted from the output image 2a. In the example shown in Figure 2, these measurement data 2 include locations where a given sensor recognizes objects 11-13. These may be, for example, locations of radar reflections or ridiculous reflections.

[0057] As indicated by the dashed arrows, during training of the encoder-decoder configuration 5, which is described in more detail in Figure 5, the training measurement data 2* is converted to the training target image 2** according to block 341 of method 300 shown in Figure 5, and the output image 2a can be directly compared with the training target image 2**.

[0058] Figure 3 shows a second exemplary procedure of method 100. Similar to Figure 2, the measurement data 1 of source measurement mode Q includes a description 10a of scene 10 containing three objects 11-13 in the form of a bounded box. However, in contrast to Figure 2, according to block 111, these measurement data 1 are converted into an input point cloud 11b. Correspondingly, processing by the encoder-decoder configuration 5 according to blocks 112 and 121 provides an output point cloud 2b. According to block 122, this output point cloud 2b is converted into desired measurement data 2 for target measurement mode Z. In the example shown in Figure 3, this conversion involves selecting enough points from the points of the output point cloud 2b so that the measurement data 2 still looks realistic in relation to target measurement mode Z.

[0059] As indicated by the dashed arrows, during training of the encoder-decoder configuration 5, which is described in more detail in Figure 5, the training measurement data 2* is converted to the training target point cloud 2** according to block 341 of method 300 shown in Figure 5, and the output image 2a can be directly compared with the training target point cloud 2**.

[0060] Figure 4 is a schematic flowchart of an exemplary embodiment of Method 200 for inspecting control logic for a driver assistance system or a system for at least partial autonomous driving of a vehicle.

[0061] In step 210, a description 10a of scene 10, including a predetermined arrangement of objects 11-13, is prepared as measurement data 1 for source measurement mode Q, as shown as an example in Figures 2 and 3. In step 220, these measurement data 1 for source measurement mode Q are converted into measurement data 2 for target measurement mode Z using the method 100 described above. In step 230, these measurement data 2 for target measurement mode Z are supplied as input to the control logic 20 under test. In step 240, the response 20a presented by the control logic 20 based on this input is compared with a predetermined target response 20a* for scene 10. In step 250, the result 240a of this comparison 240a is evaluated to the extent that the control logic 20 can handle scene 10, such as actual traffic conditions.

[0062] Figure 5 is a schematic flowchart of an exemplary embodiment of Method 300 for training an encoder-decoder configuration 5 for use in one of the methods 100, 200 described above.

[0063] In step 310, training source measurement data 1* for at least one source measurement mode Q is prepared. These training source measurement data 1* represent a given physical scene 10. In parallel, in step 320, training target measurement data 2* for at least one target measurement mode Z for the same physical scene 10 is prepared.

[0064] In step 330, the learning source measurement data 1* is first mapped to a representation 6 in latent space by the encoder-decoder configuration 5, and then further mapped to the measurement data 2 of the target measurement mode Z.

[0065] In step 340, an evaluation 340a is determined based on a predetermined cost function 7 regarding how well the measurement data 2 of the target measurement mode Z matches the training target measurement data 2*.

[0066] For this purpose, the training target measurement data 2* can be converted into a training target image 2** or training target point cloud 2***, for example, according to block 341. Then, according to block 342, the degree of agreement between the output image 2a or point cloud 2b generated by the decoder 4 and the training target image 2** or training target point cloud 2*** can be evaluated as the degree of agreement between the measurement data of the target measurement mode and the training target measurement data.

[0067] In step 350, the parameter 3a characterizing the behavior of encoder 3 and / or the parameter 4a characterizing the behavior of decoder 4 are optimized. The purpose of this optimization is to ensure that, when the learning source measurement data 1* is further processed by the encoder-decoder configuration 5, the evaluation 340a of the measurement data 2 of the target measurement mode Z that occurs later will be improved.

[0068] Training can be terminated when any termination criteria are met. The states 3a* and 4a* of parameters 3a and 4a achieved at this point are the fully trained states. [Explanation of Symbols]

[0069] 3 encoders 4 Decoders 5. Encoder-Decoder Configuration 7. Cost Function 10 scenes 11~13 Object Q Source Measurement Mode Z target measurement mode

Claims

1. A method (100) for converting measurement data (1) of at least one source measurement mode (Q) provided by a sensor into realistic measurement data (2) of at least one target measurement mode (Z), The steps include: the encoder (3) of the trained encoder-decoder configuration (5) converts the measurement data (1) of the source measurement mode (Q) into an input point cloud (1b) (111); The trained encoder-decoder configuration (5) includes the step (110) of mapping the input point cloud (1b) to a representation (6) in latent space, The decoder (4) of the encoder-decoder configuration (5) performs the step (121) of mapping the representation (6) to the output point cloud (2b), The decoder (4) of the encoder-decoder configuration (5) performs the step (122) of converting the output point cloud (2b) into the realistic measurement data (2) of the target measurement mode (Z). Includes, A method (100) wherein the amount of information in the representation (6) of the measurement data (1) in the latent space is less than the amount of information in the measurement data (1).

2. The method according to claim 1 (100), wherein the dimension of the latent space is smaller than the dimension of the space in which the encoder (3) acquires the measurement data (1) of the source measurement mode (Q), and smaller than the dimension of the space in which the decoder (4) maps the measurement data (2) of the target measurement mode (Z).

3. The measurement data (1) of the source measurement mode (Q) includes a description (10a) of a scene (10) that includes one or more objects (11 to 13), The method according to claim 1 or 2 (100), wherein the measurement data (2) of the target measurement mode (Z) includes the location in space where actual sensors are assigned to the objects (11-13) in the scene at the time of physical observation of the scene.

4. The method according to claim 3 (100), wherein the measurement data (2) of the target measurement mode (Z) includes locations in space where an actual radar sensor or actual rida sensor records radar reflection or rida reflection during physical observation of the scene.

5. The method (100) of claim 3 or 4, wherein the description (10a) of the scene (10) includes a clear description of a plurality of geometric shapes (11a to 13a) in a plane or space which are each considered to be occupied by an object (11 to 13).

6. Step (111) of converting the measurement data (1) of the source measurement mode (Q) into an input image (1a), The encoder (3) performs the step (112) of mapping the input image (1a) to a representation (6) in latent space, The decoder (4) maps the representation (6) obtained by mapping the input image (1a) to the output image (2a) in step (121), Step (122) to convert the output image (2a) into realistic measurement data (2) of the target measurement mode (Z) and The method according to any one of claims 3 to 5, including (100).

7. The method according to claim 6 (100), wherein the scene (10) is discretized into a two-dimensional or three-dimensional grid (111a), and for each grid point, it is checked whether the grid point belongs to one of the geometric shapes defined in the description (10a) (111b), thereby converting the description (10a) of the scene (10) into an input image (1a) (111).

8. The method according to claim 5 (100), wherein the input point group (1b) is obtained by deriving the coordinates of points from the distribution of the objects (11 to 13) (111c), and adding each point belonging to one of the geometric shapes to the input point group (1b) (111d).

9. The method according to claim 7 (100), wherein information regarding the material, class, and / or movement state of the objects (11-13) from the description (10a) of the scene (10) is assigned to the grid points of the input image (1a) (111e).

10. The method according to claim 8 (100), wherein information relating to the material, class, and / or movement state of the objects (11-13) from the description (10a) of the scene (10) is assigned to the points of the input point cloud (1b) (111e).

11. The method according to any one of claims 1 to 10 (100), wherein measurement data (1) from multiple source measurement modes (Q) are mapped (113) to representations (6) in the latent space by their respective encoders (3), and the representations (6) are mapped (123) to measurement data (2) from the target measurement mode (Z) by the same decoder (4).

12. The method according to any one of claims 1 to 11 (100), wherein at least one representation (6) in the latent space is mapped (2) to measurement data (2) of at least two target measurement modes (Z) by at least two different decoders (4) (124).

13. A method (200) for testing control logic (20) for a driver assistance system or a system for at least partial autonomous driving of a vehicle, Step (210) includes preparing a description (10a) of a scene (10) including a predetermined arrangement of objects (11-13) as measurement data (1) for the source measurement mode (Q), A step (220) of converting the measurement data (1) of the source measurement mode (Q) to measurement data (2) of the target measurement mode (Z) using the method (100) described in any one of claims 3 to 12, The step (230) of supplying the measurement data (2) of the target measurement mode (Z) as input to the control logic (20) of the object to be inspected, The steps include: comparing the response (20a) presented by the control logic (20) based on the input with a predetermined target response (20a*) relating to the scene (10); From the results (240a) of the comparison (240), a step (250) is made to evaluate the extent to which the control logic (20) can handle the scene (10) (250a). A method including (200).

14. A method (300) for training an encoder-decoder configuration (5) for use in the method (100, 200) according to any one of claims 1 to 13, Step (310) is to prepare training source measurement data (1*) for at least one source measurement mode (Q), wherein the training source measurement data (1*) represents a predetermined physical scene (10), Step (320) of preparing training target measurement data (2*) for at least one target measurement mode (Z) for the same physical scene (10), The encoder-decoder configuration (5) first maps the learning source measurement data (1*) to the representation (6) in the latent space, and further maps it to the measurement data (2) of the target measurement mode (Z) (330), A step (340) to evaluate how well the measurement data (2) of the target measurement mode (Z) matches the learning target measurement data (2*) based on a predetermined cost function (7), Step (350) is to optimize the parameters (3a) that characterize the behavior of the encoder (3) and / or the parameters (4a) that characterize the behavior of the decoder (4), with the expectation that the evaluation (340a) of the measurement data (2) of the target measurement mode (Z) that occurs later will be improved when the learning source measurement data (1*) is further processed by the encoder-decoder configuration (5), and A method including (300).

15. The method according to claim 14 (300), comprising the steps of: (341) converting the learning target measurement data (2*) into a learning target image (2**) or a learning target point cloud (2***); and (342) evaluating the degree of agreement between the output image (2a) or point cloud (2b) generated by the decoder (4) and the learning target image (2**) or the learning target point cloud (2***) as the agreement between the measurement data of the target measurement mode and the learning target measurement data.

16. A computer program that, when executed on one or more computers, includes machine-readable instructions causing the computers to perform the method(100, 200) according to any one of claims 1 to 15.

17. A machine-readable data recording medium and / or download product comprising the computer program described in claim 16.

18. A computer comprising the computer program according to claim 16, and / or the machine-readable data recording medium and / or download product according to claim 17.

Images