Synthetic Generation of Radar, LIDAR, and Ultrasonic Measurement Data

Abstract

Synthesis generation of radar, LIDAR, and ultrasonic measurement data. A method for generating synthetic measurement data indistinguishable from actual measurement data captured by a first physical measurement mode, where the first physical measurement mode is based on emitting interrogation waves towards an object and recording reflected waves from the object in a manner that allows determination of the flight time between the emission of the interrogation beam and the arrival of the reflected wave, the method comprising the steps of: ● obtaining a first compressed representation of the synthetic measurement data in a first latent space, where the first latent space is associated with a first decoder that maps each element of the first latent space to a recording of synthetic measurement data indistinguishable from the recording of actual measurement data of the first physical measurement mode; and ● applying the first decoder to the first compressed representation in order to obtain the sought-after synthetic measurement data. A method for training the first encoder and decoder and for training the domain transformation.

Description

Technical Field

[0001] The present invention relates to the synthetic generation of measurement data, in particular for data corresponding to radar, LIDAR (Light Detection and Ranging), ultrasound, and similar physical measurement modalities. Background Art

[0002] In order to guide a vehicle through road traffic in at least a partially automated manner, it is necessary to capture physical measurement data from the vehicle's surroundings and evaluate this data for other traffic participants, lane boundaries, or any other kind of object whose presence may require a change in the vehicle's trajectory.

[0003] Objects can be captured by radar regardless of the lighting conditions. Moreover, radar data immediately yields the distance to the object and the object's velocity. This is crucial information for assessing whether the vehicle may collide with the detected object.

[0004] When a machine learning module is to be trained to identify objects based on radar measurements, the training data required for the training is a scarce resource. As detailed in German Patent DE 10 2018 204 494 B3, when the training data needs to be marked by humans for supervised learning, this task is more difficult than for images because identifying objects from radar signals is far from intuitive. Moreover, since there are many factors that affect the propagation of radar waves, object recognition based on radar data tends to require more training data compared to object recognition based on optical images. Therefore, Patent DE 10 2018 204 494B3 proposes using a generative adversarial network (GAN) to generate synthetic radar data. Summary of the Invention

[0005] The inventors have developed a method for generating synthetic measurement data that is indistinguishable from actual measurement data captured by a first physical imaging modality. This physical imaging modality is based on transmitting interrogation waves towards an object and recording the reflected waves from the object in a way that allows determination of the flight time between the emission of the interrogation beam and the arrival of the reflected wave. In the form of a directed beam, the interrogation wave can be transmitted and the reflected wave can be received.

[0006] In particular, the interrogation wave can be a radar wave, a LIDAR wave, or an ultrasonic wave. Specifically, the measurement data of the first physical measurement modality can include

[0007] ● a combination of the angle and distance to the site on the object that contributes to the emission of the reflected wave; and / or

[0008] ● a point cloud of the site on the object.

[0009] What these measurement modalities have in common is that the raw data is more difficult for humans to interpret than the image data. Thus, when a machine learning model is trained to classify the objects indicated by the measurement data or obtain a regression quantity such as the speed of an object, it is more expensive and time-consuming to label the records of the training data with "ground truth" relevant to the task at hand. The possibility of obtaining synthetic measurement data allows more training data to be made available for the training of the machine learning module without excessive human effort being expended on labeling the data.

[0010] The measurement data can be in any form suitable for the intended use. For example, the measurement data can include time series data or a transformation of that time series data into the frequency space, such as a fast Fourier transform.

[0011] The method begins with obtaining a first compressed representation of the synthetic measurement data in a first latent space. The first latent space is associated with a first decoder that is trained to map each element of the first latent space to a record of synthetic measurement data that is indistinguishable from a record of actual measurement data of a first physical measurement modality. For example, a tandem of an encoder and a decoder can be trained such that when a record of actual measurement data is transformed by the encoder into the first compressed representation and then transformed back into a record of synthetic measurement data, the record of synthetic measurement data best corresponds to the original record of actual measurement data. Even with such tandem training, only the first decoder in its trained state is required to implement the method.

[0012] The trained first decoder is applied to the first compressed representation. This yields the sought-after synthetic measurement data.

[0013] The inventors have found that in this way, the rather complex task of obtaining synthetic measurement data can be reduced to the much simpler task of finding an appropriate compressed representation. Specifically, if the compression is lossy, the compressed representation can be one-hundredth or less of the record of the measurement data in terms of the amount of data it contains. It is much easier to search for things in a space that is 100 times less dimensional. The raw data generated by radar, LIDAR, and ultrasonic sensors is sparse, i.e., it contains far less information than what the corresponding signal could represent. This allows for lossy compression of the data without losing any critical information about the sampled scene.

[0014] This effect is particularly significant if the first compressed representation includes a vector or tensor of discrete variables and the number of these variables is less than the number of variables in the record of the measurement data of the first physical measurement mode. For example, the first compressed representation can be a vector quantization representation, i.e., a vector or tensor with quantized components. This narrows the space in which the first compressed representation is further sought: since there are a finite number of dimensions and a finite number of possible discrete values along each dimension, there are a finite number of possible compressed representations.

[0015] Any suitable technique can be used to seek the first compressed representation in the narrowed space. For example, parameter optimization techniques can be used to optimize the variables in the compressed representation in order to find a representation that maps to a suitable record of the synthetic measurement data (i.e., a record that is indistinguishable from the actual measurement data). In the case of vector quantization, even a brute-force search of the first latent space may be feasible if candidate compressed representations can be tested quickly enough to determine whether they map to a suitable record of the synthetic measurement data.

[0016] In the following, two exemplary methods for obtaining the first compressed representation are disclosed. These exemplary methods also take into account that the first latent space is typically a subspace of a complete vector space spanned by all the variables that the decoder takes as input. For example, if an encoder-decoder tandem is trained and the compressed representation produced by the encoder is a vector with 100 components, not all vectors from this 100-dimensional vector space will be decoded into meaningful results. However, additional machine learning can be used to train the ability to find vectors that do map to suitable records of the synthetic measurement data.

[0017] In a first exemplary embodiment, samples are extracted from the input space of a prior transformation that is trained to map each element of the input space to an element of the first latent space. The prior transformation is then applied to the samples, and this produces the sought-after first compressed representation. Then, this first compressed representation will map to a suitable record of the synthetic measurement data. In this way, the overall task of obtaining the synthetic measurement data is split into two tasks that can be carried out sequentially: training the encoder-decoder tandem to form the first latent space, and then training the prior transformation to convert samples (e.g., random samples) from the input space into samples that are members of the first latent space.

[0018] For example, the prior transformation can specifically include at least one trained autoregressive neural network. For example, the neural network can be a convolutional neural network. An example of a prior transformation that can convert arbitrary input data into a member of the latent space created by an encoder-decoder tandem is known in the art as "PixelCNN".

[0019] Preferably, the prior transformation can include multiple parts such that different parts of the prior transformation map elements of their respective input spaces to respective outputs, and then these outputs are stacked to form a first compressed representation. For example, different parts of the prior transformation can map their respective inputs to different parts of the first compressed representation, and all parts together form the complete first compressed representation. This is particularly advantageous in cases where the first compressed representation is organized into a hierarchical structure with multiple levels. Then different parts of the prior transformation can be trained to produce different levels of the first compressed representation.

[0020] The advantage of splitting the prior transformation in this way is twofold.

[0021] First, there is greater flexibility in attaching conditions to the synthetic measurement data being sought. For example, the task at hand may not be to find just any synthetic radar image, but to find an image that would indicate the presence of certain objects. The presence of such certain objects can form a "class label" in a classification task. In an example where the first compressed representation includes three levels (called the highest, medium, and lowest levels), a first autoregressive neural network (e.g., PixelCNN) can be trained to find each level of the compressed representation. The obtaining of the highest level may depend on the class label; the obtaining of the lower levels may depend on the class label and the results of the previous levels.

[0022] Second, since the individual parts of the prior transformation can be trained separately, the training can be parallelized. In this way, all available computing power and memory on a hardware accelerator can be utilized.

[0023] In a second exemplary embodiment, the method is specifically configured for domain transfer of measurement data. That is, starting from the actual measurement data of a second physical measurement mode, synthetic measurement data of a first physical measurement mode representing a scene with substantially similar content is sought.

[0024] For example, the second physical measurement mode can specifically include: recording the spatially resolved distribution of the intensity and / or wavelength of light waves incident on a sensor. Such a sensor produces an image that can be easily interpreted by humans. Thus, it is very common practice to obtain labels representing the types of objects contained in the image by assigning the task to a large number of human laborers who will label the objects they identify in the image. There are also many existing sets of such labeled images. If such a labeled image domain is transferred into a synthetic record of radar data, it is known from the start which objects the radar data will represent. In other words, any "ground truth" labels attached to the image can be reused for the radar data. This is much easier than manually labeling the radar data from scratch. Manually labeling such radar data requires more expertise and more time compared to manually labeling images.

[0025] To accomplish the domain transfer, a second trained encoder is applied to the recording of the actual measurement data of a second physical measurement mode, which is different from the first physical measurement mode. For example, the recording may include an image.

[0026] Applying the second trained encoder results in a second compressed representation of the second actual measurement data (e.g., an image) in a second latent space. Similar to the first latent space, this second latent space is associated with a second decoder that is trained to map each element of the second latent space to a recording of synthetic measurement data of a second measurement mode that is indistinguishable from the recording of the actual measurement data of the second physical measurement mode. For example, the second encoder and the second decoder may be trained in an encoder-decoder tandem manner such that when an input recording of the actual measurement data of the second mode (e.g., an image) is encoded by the encoder into a compressed representation and subsequently decoded by the decoder, the input recording is best reproduced. Even when the second encoder and the second decoder are trained in this tandem manner, during the course of the method, only the second encoder needs to be in its training state.

[0027] Apply a domain transformation to the second compressed representation. The domain transformation is trained to map each element of the second latent space to an element of the first latent space. In this way, the sought-after first compressed representation is obtained. Then the first decoder can be applied to this first compressed representation to obtain the final result, i.e., the sought-after synthetic measurement data of the first physical measurement mode.

[0028] Similar to the first embodiment, the task of domain transfer is split into on the one hand the generation of the second compressed representation and on the other hand the actual transfer of this second compressed representation to the first latent space. Training for the two tasks can again be carried out sequentially and is thus easier to accomplish than a single training of the overall mapping that directly leads from the recording of the actual measurement data of the second mode to the first compressed representation in the first latent space. Here, the real-world analogy is that it is much easier to jump 1 m high from the ground onto the first step and then 1 m high from there onto the next step than to jump 2 m high in one go.

[0029] In a further particularly advantageous embodiment, regardless of whether the task is "from scratch" (e.g., based on arbitrary samples randomly extracted from the input space) or to obtain synthetic measurement data by domain transfer from another measurement modality, the first latent space from which the first compressed representation is obtained can be selected such that the first decoder maps the elements of the first latent space to synthetic measurement data that is consistent with at least one predetermined condition. As discussed previously, the predetermined condition may include a certain class label such that synthetic measurement data belonging to a specific class of classification can be obtained. For example, it may be particularly desirable to have radar data showing two vehicles and a stop sign during a collision process.

[0030] Preferably, the predetermined condition specifically includes: the interaction of the interrogation wave with one or more specific objects, and / or one or more environmental conditions that affect the propagation of the interrogation wave and / or the reflected wave. For example, since the microwave radiation used for radar imaging is partially absorbed by water, the radar data of the same scene may change when heavy rain comes. An object detection system for vehicles should work reliably in various environmental conditions, so training data with a certain variability regarding these environmental conditions is needed.

[0031] By appropriately setting the predetermined condition, the method can be used to obtain training data representing various conditions and combinations, even if not all such combinations have been part of the training of any of the encoders, decoders, prior transforms, and domain transforms used. For example, such training may be based on radar data of various types of vehicles and various types of weather conditions, but there may be no data of a Lamborghini in heavy snowfall because the sensible owners of such expensive cars will not risk an accident under adverse winter driving conditions. Using the above method, the radar data can be synthetically generated, thus expanding the data pool that can be used to train machine learning modules for object detection.

[0032] Therefore, in a further particularly advantageous embodiment, the method may further include: using the generated synthetic measurement data of the first physical measurement modality to train at least one machine learning module. The machine learning module is to map the actual measurement data captured from the vehicle to at least one classification and / or regression value. The classification and / or regression value is related to operating the vehicle in road traffic in at least a partially automated manner. In particular, the generated synthetic measurement data can be used to augment an existing set of actual measurement data of the first physical measurement modality such that the final data set used for training has a desired variability with respect to different situations and conditions.

[0033] As discussed previously, the categories of classification can relate to the type of object. In particular, measurement data obtained from a vehicle can be "semantically segmented" into contributions from different objects. For example, the regression values can include: the speed and direction of the object, the coefficient of friction between the tires and the road, or the maximum range in the direction of travel in front of the vehicle that can be surveyed by the vehicle's sensors under the current conditions.

[0034] After training it in this way, the machine learning module can be put into use in the vehicle. Thus, in a further particularly advantageous embodiment, the method further comprises:

[0035] ● Obtaining actual measurement data from the vehicle using a first physical measurement mode;

[0036] ● Processing the obtained actual measurement data by the trained machine learning module to obtain at least one classification and / or regression value;

[0037] ● Calculating at least one actuation signal for at least one system of the vehicle from the classification and / or regression value; and

[0038] ● Actuating the system using the actuation signal.

[0039] For example, when it is determined that the trajectory of the currently envisaged vehicle intersects the trajectory of another vehicle on the way, the steering system and / or the braking system can be actuated to slow the vehicle to a stop before reaching the other vehicle, or to follow a path around the other vehicle. In this safety-critical application and other safety-critical applications, the possibility of generating synthetic measurement data allows the expansion of the training data used to train the machine learning module, thereby improving the variability of the training data. This improves the result of the training and thus also improves the likelihood that the machine learning module will cause the vehicle to perform the correct action in a given specific traffic situation.

[0040] As discussed previously, a first decoder in its training state can be obtained using training in the style of a variational autoencoder, based on unlabeled actual measurement data. Thus, the present invention also relates to a method for training a first encoder and a decoder. The method comprises the following steps:

[0041] ● Obtaining a record of a set of actual measurement data by a first physical measurement mode;

[0042] ● Mapping each record of the actual measurement data to a first compressed representation by means of a first trainable encoder;

[0043] ● Mapping the first compressed representation to a record of synthetic measurement data in a first physical measurement mode by means of a first trainable decoder;

[0044] ● Optimize the parameters that characterize the behavior of the first encoder and decoder, with the goal of minimizing the difference between the recorded synthetic measurement data and the corresponding record of the actual measurement data.

[0045] In this document, as discussed previously, the first physical measurement mode is based on emitting an interrogation wave towards an object and recording the reflected wave from the object in a way that allows determination of the time of flight between the emission of the interrogation beam and the arrival of the reflected wave. The interrogation wave is a radar wave, a LIDAR wave, or an ultrasonic wave.

[0046] As discussed previously, the training of the first decoder is independent of the training of any other means used to obtain the first compressed representation, such as a prior transformation or a domain transformation. This means that if a prior transformation is to be made into a transformation that maps between a new desired input space and the first latent space, the training of the first decoder remains valid and does not have to be repeated. Similarly, if it is desired to perform a domain transfer from a new physical measurement mode to the first physical measurement mode, a new domain transformation will have to be trained, but the first decoder will not need to be changed.

[0047] The present invention also provides a method for training a domain transformation that can be used to map a second compressed representation of the actual measurement data of a second physical measurement mode to the first latent space.

[0048] During the process of this method, a set of records of actual measurement data is obtained through the second physical measurement mode. Each record of the actual measurement data is mapped to a second compressed representation by means of a second trained encoder, which, as discussed previously, can be serially trained with a second decoder in the style of a variational autoencoder.

[0049] The second compressed representation is mapped to a first compressed representation by a trainable domain transformation. The first compressed representation is mapped to a record of synthetic measurement data by means of a first trained decoder. Optimize the parameters that characterize the behavior of the domain transformation, with the goal of making the record of the synthetic measurement data indistinguishable from the record resulting from processing the actual measurement data of the first physical measurement mode into a compressed representation using the first trained encoder and passing this compressed representation to the first trained decoder.

[0050] In other words, the optimal criterion for optimizing the parameters that characterize the behavior of the domain transformation measures the degree of "mixing" of the ultimately obtained record of synthetic measurement data among the records that have been generated from known members of the first latent space, that is, the first compressed representation has been generated from the actual measurement data of the first physical measurement mode by means of a first encoder corresponding to the first decoder.

[0051] In any of the training methods, the parameters can include, for example: weights by which the inputs to neurons or other processing units in a neural network are aggregated to form an activation of the neuron or other processing unit. Optimization of the parameters can be performed according to any suitable method. For example, the gradient descent method can be used.

[0052] All of the above methods can be at least partially computer-implemented. Accordingly, the present invention also relates to a computer program having machine-readable instructions which, when executed by one or more computers, cause the one or more computers to implement at least one of the above methods. In this regard, the meaning of the term "computer" should include an electronic control unit for a vehicle or vehicle subsystem, as well as other embedded systems that control technical devices based on programmable instructions.

[0053] The computer program can be embodied in a non-transitory machine-readable storage medium and / or in a download product. A download product is a digitally deliverable product that can be traded and purchased online such that it can be immediately delivered to a computer without having to ship a non-transitory storage medium.

[0054] Alternatively or in combination, the storage medium and / or the download product can contain synthetic measurement data generated by the methods described above. As discussed above, anyone having this synthetic measurement data can immediately start enhancing the training of the machine learning module.

[0055] Alternatively or in combination, the storage medium and / or the download product can contain parameters characterizing the behavior of a first encoder and decoder, as well as parameters generated by a training method for such an encoder and decoder. Anyone having these parameters can immediately start using the first encoder and the first decoder without having to train them.

[0056] Alternatively or in combination, the storage medium and / or the download product can contain parameters characterizing the behavior of a domain transformation, as well as parameters generated by a training method for such a domain transformation. Anyone having these parameters can immediately start using the domain transformation without having to train it.

[0057] The present invention also relates to a computer equipped with a computer program and / or equipped with a machine-readable storage medium and / or a download product.

[0058] Additional advantageous embodiments will now be described in detail with reference to the drawings, without intending to limit the scope of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The drawings illustrate:

[0060] Figure 1: Exemplary embodiment of method 100 for generating synthetic measurement data 3;

[0061] Figure 2 : Schematic overview of the spaces and transformations involved used during the course of method 100;

[0062] Figure 3 : Exemplary embodiment of method 200 for training first encoder 1a and first decoder 1c;

[0063] Figure 4 : Exemplary embodiment of method 300 for training domain transformation 6. Detailed Description

[0064] Figure 1 is a flowchart of an exemplary embodiment of method 100. In step 110, a first compressed representation 3* of the synthetic measurement data in the first latent space 1b is obtained. As illustrated in Figure 2 and 3 , this first latent space 1b is associated with a first decoder 1c that can be trained in series with the first encoder 1a. In step 120, the first decoder 1c is applied to the first compressed representation 3* such that the sought-after synthetic measurement data is obtained.

[0065] Within box 110, two exemplary ways of obtaining the first compressed representation 3* are illustrated.

[0066] According to box 111, samples 5* can be extracted from the input space 5a of the prior transformation 5. The prior transformation 5 is trained to map each element of the input space 5a to an element of the first latent space 1b. Thus, when this mapping is effected according to box 112, the sought-after first compressed representation 3* is obtained.

[0067] In Figure 1 the example shown, the prior transformation 5 includes a plurality of parts 51 - 53. Each of these plurality of parts 51 - 53 has its own input space 51a - 53a and output contributions 51b - 53b. The contributions 51b - 53b are aggregated to form the compressed representation 3*.

[0068] According to box 113, the second trained encoder 2a can be applied to the recording of the actual measurement data 2 of the second physical measurement mode. This results in a second compressed representation 2* of this actual physical measurement data. According to box 114, the domain transformation 6 can be applied to this compressed representation 2* to obtain the first compressed representation 3* in the first latent space 1b.

[0069] Additionally, in Figure 1The figure illustrates the uses to which the obtained synthetic measurement data 3 can be put. In step 130, these synthetic measurement data 3 can be used to enhance the training of the machine learning module 54 in order to obtain its trained state 54*. The trained machine learning module 54* can in turn be used in step 150 to process the actual physical measurement data 1 that has been captured from the vehicle 50 in step 140.

[0070] The processing in step 150 yields at least one classification and / or regression value 8 related to the operation of the vehicle 50 in traffic. Based on this classification and / or regression value 8, in step 160, at least one actuation signal 9 for the system 55 of the vehicle 50 is calculated. In step 170, the system 55 is actuated using this actuation signal 9.

[0071] Figure 2 The figure illustrates the involved transformations and spaces. Based on the actual measurement data 1 according to the first physical measurement mode, the first encoder 1a generates a compressed representation 3* residing in the latent space 1b. The first decoder 1b is trained to map each compressed representation 3* to a record of the synthetic measurement data 3 of the first physical measurement mode. That is, the compressed representation 3* is also a representation of the record of the synthetic measurement data 3. The first encoder 1a and the first decoder 1b can be trained in series, and its optimization goal is that the synthetic measurement data 3 finally obtained from a given record of the actual measurement data 1 should best match this original measurement data 1. This optimization goal indicated by the dashed line makes the synthetic measurement data 3 indistinguishable from the actual measurement data 1.

[0072] Similarly, for a second physical measurement mode different from the first physical measurement mode, there is a second encoder 2a that maps the actual measurement data 2 of this second mode to a compressed representation 2* residing in the second latent space 2b. The second decoder 2c maps the compressed representation 2* to a record of the synthetic measurement data 2' of the second measurement mode. The second encoder 2a and the second decoder 2b can be trained in series, and its optimization goal is that the synthetic measurement data 2' finally obtained from a given record of the actual measurement data 2 should best match this original actual measurement data 2. This optimization goal indicated by the dashed line makes the synthetic measurement data 2' indistinguishable from the actual measurement data 2.

[0073] One way to obtain the synthetic measurement data 3 of the first physical measurement mode is to extract a sample 5* from the input space 5a of the prior transformation 5, and then apply this prior transformation 5 to obtain the compressed representation 3*, which is then converted by the first decoder 1c into the sought-after synthetic measurement data 3.

[0074] Another way to obtain the synthetic measurement data 3 is to perform domain transfer from a second physical measurement mode. From the actual measurement data 2 of this second mode, a second encoder 2a produces a compressed representation 2*. A trained domain transformation 6 transforms this compressed representation 2* from a second latent space 2b into a compressed representation 3* in a first latent space 1b, which can again be converted by a first decoder 1c into the sought-after synthetic measurement data 3.

[0075] The first latent space 1b can be specifically selected such that the synthetic measurement data 3 obtained by the first decoder 1c from its members satisfies the desired condition 7.

[0076] Figure 3 is a flowchart of an exemplary embodiment of a method 200 for jointly training a first encoder 1a and a first decoder 1c. In step 210, actual measurement data 1 of a first physical measurement mode is obtained. In step 220, this measurement data 1 is converted into a compressed representation 3*. In step 230, synthetic measurement data 3 is obtained based on this compressed representation 3*. In step 240, the synthetic measurement data 3 is compared with the original actual measurement data 1, and the parameters 1a*, 1c* that characterize the behavior of the first encoder 1a and the first decoder 1c are optimized for the best match between the synthetic data 3 and the original data 1.

[0077] Figure 4 is a flowchart of an exemplary embodiment of a method 300 for training a domain transformation 6. In step 310, actual measurement data 2 of a second physical measurement mode is obtained. In step 320, this data 2 is mapped by a second encoder 2a that is already in its trained state into a second compressed representation. In step 330, the compressed representation 2* residing in the second latent space 2b is transformed by the domain transformation 6 to be trained into a compressed representation 3* residing in the first latent space 1b. In step 340, the compressed representation 3* is mapped by a first decoder 1c that is already in its trained state to a record of the synthetic measurement data 3. In step 350, the synthetic measurement data 3 is compared with the result obtained when the actual physical measurement data 1 of the first physical measurement mode is first transformed by a first decoder 1a that is already in its trained state into a compressed representation 3* and then transformed back into the synthetic measurement data 3. The parameters 6* that characterize the behavior of the domain transformation 6 are optimized such that the best match is obtained between the synthetic measurement data 3 generated via the domain transformation 6 from the actual measurement data 2 of the second physical measurement mode and the synthetic measurement data 3 generated from the actual measurement data 1 of the first physical measurement mode.

Claims

1. A method (100) for generating synthetic measurement data (3) that is indistinguishable from actual measurement data (1) captured by a first physical measurement mode, wherein, The first physical measurement mode is based on emitting an interrogation wave towards an object and recording the reflected wave from the object in a manner that allows determination of the time-of-flight between the emission of the interrogation beam and the arrival of the reflected wave. The method includes the following steps: Obtaining (110) a first compressed representation (3*) of synthetic measurement data (3) in a first latent space (1b), wherein the first latent space (1b) is associated with a first decoder (1c), and the first decoder (1c) is trained to map each element of the first latent space (1b) to a record of synthetic measurement data that is indistinguishable from a record of actual measurement data (1) of the first physical measurement mode; and Applying (120) the first decoder (1c) to the first compressed representation (3*) in order to obtain synthetic measurement data (3), wherein obtaining (110) the first compressed representation (3*) includes: Applying (113) a second trained encoder (2a) to a record of actual measurement data (2) of a second physical measurement mode, which is different from the first physical measurement mode, in order to obtain a second compressed representation (2*) of the actual measurement data (2) in a second latent space (2b), wherein the second latent space (2b) is associated with a second decoder (2c), and the second decoder (2c) is trained to map each element of the second latent space (2c) to a record of synthetic measurement data (2') that is indistinguishable from a record of actual measurement data (2) of the second physical measurement mode; and Applying (114) a domain transformation (6) to the second compressed representation (2*), the domain transformation being trained to map each element of the second latent space (2b) to an element of the first latent space (1b), in order to obtain the first compressed representation (3*).

2. The method (100) according to claim 1, wherein, Obtaining (110) the first compressed representation (3*) includes: Extracting (111) samples (5*) from an input space (5a) of a prior transformation (5), the prior transformation (5) being trained to map each element of the input space (5a) to an element of the first latent space (1b); and Applying (112) the prior transformation (5) to the samples (5*) in order to obtain the first compressed representation (3*).

3. The method (100) according to claim 2, wherein, The prior transformation (5) includes at least one trained autoregressive neural network.

4. The method (100) according to claim 2 or 3, wherein, The prior transformation (5) includes a plurality of parts (51-53) such that different parts (51-53) of the prior transformation (5) map elements of their respective input spaces (51a-53a) to respective outputs (51b-53b), and then the respective outputs are stacked to form the first compressed representation (3*).

5. The method (100) according to claim 1, wherein, The second physical measurement mode includes: recording a spatially resolved distribution of the intensity and / or wavelength of light waves incident on a sensor.

6. The method (100) according to claim 1, wherein The first compressed representation (1*) includes a vector or tensor of discrete variables, wherein the number of these variables is less than the number of variables in a record of measurement data (1, 3) of the first physical measurement mode.

7. The method (100) according to claim 1, wherein Obtaining (110) the first compressed representation (3*) includes: selecting a first latent space (1b) from which the first compressed representation (3*) is obtained such that the first decoder (1c) maps elements of the first latent space (1b) to synthetic measurement data (3) that is consistent with at least one predetermined condition (7).

8. The method (100) according to claim 7, wherein, The predetermined condition (7) includes: the interaction of an interrogation wave with one or more specific objects, and / or one or more environmental conditions that affect the propagation of the interrogation wave and / or the reflected wave.

9. The method (100) according to claim 1, further comprising: Training (130) at least one machine learning module (54) using the generated synthetic measurement data (3), the machine learning module (54) mapping actual measurement data (1) captured from a vehicle (50) to at least one classification and / or regression value, where the classification and / or regression value is related to operating the vehicle (50) in road traffic in at least a partially automated manner.

10. The method according to claim 9, further comprising: Obtaining (140) actual measurement data (1) from the vehicle (50) using a first physical measurement mode; Processing (150) the obtained actual measurement data (1) by the trained machine learning module (54*) to obtain at least one classification and / or regression value (8); Calculating (160) at least one actuation signal (9) for at least one system (55) of the vehicle (50) from the classification and / or regression value (8); And Actuating (170) the system (55) using the actuation signal (9).

11. The method (100) according to claim 1, wherein, The interrogation wave of the first physical measurement mode is a radar wave, a LIDAR wave, or an ultrasonic wave.

12. The method (100) according to claim 11, wherein, The measurement data (1) of the first physical measurement mode includes: a combination of the angle and distance to a site on an object that contributes to the emission of a reflected wave; and / or A point cloud of the site on the object.

13. A method (200) of training a first encoder (1a) and a decoder (1c) for use in the method (100) according to any one of claims 1 to 12, comprising the following steps: Obtaining (210) a record of a set of actual measurement data (1) by a first physical measurement mode; Mapping (220) each record of the actual measurement data (1) to a first compressed representation (3*) by means of a first trainable encoder (1a); Mapping (230) the first compressed representation (3*) to a record of synthetic measurement data (3) of the first physical measurement mode by means of a first trainable decoder (1c); and Optimizing (240) the parameters (1a*, 1c*) that characterize the behavior of the first encoder (1a) and the decoder (1c), with the aim of minimizing the difference between the record of the synthetic measurement data (3) and the corresponding record of the actual measurement data (1), Where The first physical measurement mode is based on emitting an interrogation wave towards an object and recording the reflected wave from the object in a manner that allows determination of the flight time between the emission of the interrogation beam and the arrival of the reflected wave, and The interrogation wave is a radar wave, a LIDAR wave, or an ultrasonic wave.

14. A method (300) of training a domain transformation (6) for use in the method (100) according to any one of claims 1 to 12, comprising the steps of: obtaining (310) a record of a set of actual measurement data (2) by a second physical measurement mode; mapping (320) each record of the actual measurement data (2) to a second compressed representation (2*) by means of a second trained encoder (2a); mapping (330) the second compressed representation (2*) to a first compressed representation (3*) by a trainable domain transformation (6); mapping (340) the first compressed representation (3*) to a record of synthetic measurement data (3) by means of a first trained decoder (1c); and optimizing (350) parameters (6*) characterizing the behavior of the domain transformation (6) such that the record of the synthetic measurement data (3) is indistinguishable from the record resulting from processing actual measurement data (1) of the first physical measurement mode into the compressed representation (3*) by using the first trained encoder (1a) and passing the compressed representation (3*) to the first trained decoder (1c).

15. A computer program product comprising machine-readable instructions which, when executed by one or more computers, cause the one or more computers to implement the method (100, 200, 300) according to any one of claims 1 to 14.

16. A non-transitory machine-readable storage medium and / or a download product having one or more of the following: the computer program product according to claim 15; synthetic measurement data (3) generated by the method (100) according to any one of claims 1 to 12; parameters (1a*, 1c*) characterizing the behavior of the first encoder (1a) and decoder (1c) and generated by the method (200) according to claim 13; and parameters (6*) characterizing the behavior of the domain transformation (6) and generated by the method (300) according to claim 14.

17. A computer equipped with the computer program product according to claim 15 and / or having the machine-readable storage medium and / or download product according to claim 16.

Images