Processing of Radar Signals in the Presence of Suppressed Motion Artifacts

Abstract

A method for training a trainable module for analyzing radar signals, comprising the following steps: feeding an actual radar signal and / or an actual representation to the module and having it converted into a processed radar signal and / or a processed representation; evaluating, by means of a cost function, to what extent the processed radar signal is suitable for reconstructing the movement of an object or to what extent the processed representation contains artifacts; optimizing parameters characterizing the properties of the module according to the cost function. A method for analyzing a moving object based on radar signals. A method comprising the following steps: training the module by using the method according to any one of claims 1 to 13; measuring radar signals from the surroundings of the vehicle by using at least one radar sensor arranged on the vehicle; analyzing the radar signals by using the trained module; forming at least one control signal based on the result of the analysis; and controlling the vehicle by using the control signal.

Description

Field of the Invention

[0001] The present invention relates to the processing of radar signals, in particular to the processing of radar signals in combination with the application of synthetic radar apertures when observing the surrounding environment of a vehicle. Background Art

[0002] In order for a vehicle to move at least partially autonomously in road traffic, it is necessary to detect the surrounding environment of the vehicle and to initiate countermeasures if a collision with an object in the surrounding environment of the vehicle is imminent. For safe automated driving, it is also necessary to create a surrounding environment representation and localization.

[0003] Detecting objects by means of radar is independent of lighting conditions and is possible, for example, even at large distances at night, and oncoming traffic is not dazzled by high beams. However, similar to optical imaging, the position and angular resolution are naturally diffraction-limited due to the interaction of the wavelength used and the physical size of the aperture from which the radar radiation emanates.

[0004] To improve the resolution, GB 2 564 648 A proposes that when observing the surrounding environment of a moving vehicle by means of radar, the radar signals measured at different observation angles based on the vehicle movement are combined with each other. The combination of radar signals has the effect of observing the surrounding environment as if through a long aperture. Thus, this technique is also referred to as synthetic aperture radar (Synthetische-Apertur-Radar), i.e., SAR. Summary of the Invention

[0005] Within the scope of the present invention, a method for training a trainable module for analyzing radar signals has been developed. In this method, actual radar signals and / or the accordingly determined actual representation of the scene observed using the actual radar signals are fed to the trainable module and are converted by the trainable module into processed radar signals and / or into a processed representation of the corresponding scene.

[0006] The trainable module is in particular regarded as a module that embodies a function parameterized with adaptable parameters with great effort. These parameters can be adapted in particular when training the trainable module such that the associated learning output parameter values are reproduced as well as possible when learning input parameter values are input into the module. The trainable module can in particular comprise an artificial neural network KNN and / or the trainable module can be a KNN.

[0007] It is evaluated by means of a cost function to what extent the processed radar signals are suitable for reconstructing the movement of the object or to what extent the processed representation contains artifacts caused by moving objects in the scene. The parameters characterizing the properties of the trainable module are optimized with the goal of obtaining processed radar signals or processed representations that are better evaluated by the cost function. This is synonymous with the situation that the trainable module increasingly suppresses artifacts caused by moving objects in the processed radar signals or in the processed representation as the training progresses.

[0008] Here, there is sufficient flexibility with respect to the spatial aspect of the transformation of the trainable module and thus also sufficient flexibility with respect to the possible integration with the analysis by means of the SAR algorithm.

[0009] For example, the trainable module can be configured to convert the radar signals into processed radar signals existing in the same space as these radar signals. If these processed radar signals are better suitable for reconstructing the movement of the object, then for example the possible tendency of the downstream SAR algorithm to generate motion artifacts is suppressed.

[0010] However, the trainable module can also be configured, for example, to convert the representation of the scene determined from the radar signals into a processed representation existing in the same space as these representations. For example, these representations can already be created using the SAR algorithm based on the measured radar signals. In this way, the suppression of motion artifacts can be tailored to the specific SAR algorithm used.

[0011] The trainable module can also be configured, for example, to directly convert the radar signals into a processed representation. In this way, the trainable module can, for example, during the working process not only learn to convert the radar signals into the searched-for representation of the scene, for example by means of the SAR algorithm, but also learn the processing of these radar signals with respect to better suitability for reconstructing the movement of the object. However, the trainable module can also already be pre-trained for this conversion, for example using the SAR algorithm, and further trained to improve motion reconstruction.

[0012] It has been found that, using the trainable module trained in the described manner, on the one hand, the image quality of the representation of the scene obtained by means of the SAR algorithm can be improved. However, on the other hand, these representations are also quantitatively more correct in that the moving objects appear in the correct position in these representations and there is no trailing. This improvement is caused in a self-consistent manner by the learning process.

[0013] The cost function can evaluate the suitability for motion reconstruction or the presence of motion artifacts according to any metric, which can in particular be adapted to the processed radar signals or the intended use. If, for example, it is known that certain interferences, artifacts or other features have a particularly adverse effect on the intended further processing of these radar signals or representations (such as using SAR algorithms), this can be taken into account using corresponding penalty terms in the cost function. However, the cost function can also work, for example, with a trained classifier that classifies these radar signals according to how suitable they are for motion reconstruction or that classifies these representations according to whether motion artifacts are present and, if necessary, to what extent. Such classification is much simpler to implement than, for example, classifying the objects recognizable in the representation or semantic segmentation of the representation according to object type.

[0014] The optimization of the parameters can be achieved using any algorithm, such as using the gradient descent method. However, alternatively or also in combination with this, for example, the search space spanned by multiple parameters can also be searched according to a fixed grid. Such a search does not depend on whether the cost function is continuous. Different optimization methods can be executed simultaneously or can also be executed, for example, alternately.

[0015] In a particularly advantageous design, the cost function contains a GAN term, and the better the value of this GAN term is, the less distinguishable the processed radar signals or processed representations are from a pre-given set of target radar signals or from a pre-given set of target representations according to the discriminator module. Here, the discriminator module is additionally trained to distinguish the processed radar signals or processed representations from the target radar signals or from the target representations.

[0016] In this design, the trainable module acts as the generator G in a "conditional generative adversarial network", abbreviated as "conditional GAN". This generator G takes as input a radar signal or representation x, and optionally also a sample z drawn from a multi-dimensional random variable (such as a normal distribution), and attempts to generate a processed radar signal or representation y' therefrom, which is as indistinguishable as possible from the target radar signal or target representation y. In this context, "conditional" means that the generator G maps the input x (optionally with sample z) to an output G(x,z)=:y' that relates to the same scene. That is, in the pursuit of providing a result with as few motion artifacts as possible in any one of the target spaces, it does not, for example, convert a representation of an object with multiple motions and thus trailing into a representation in which these objects completely disappear. The discriminator module D is only needed during training and is no longer used during actual operation for the subsequent processing of radar signals or representations.

[0017] The GAN term can, for example, take the following form L cGAN (G, D):

[0018] . Where E x,y denotes the expected value ("sample mean") with respect to the pairs of x and y. Correspondingly, E x,z denotes the expected value with respect to the pairs of x and z. The generator G seeks to minimize L cGAN (G, D), while the discriminator D seeks to maximize L cGAN (G, D). Then, the optimal generator G* is the solution to the optimization problem . Here, denotes the maximization of L cGAN (G, D) with respect to the parameters of the discriminator D. Correspondingly, denotes the minimization with respect to the parameters of the generator G.

[0019] In another particularly advantageous design, the cost function additionally includes a similarity term, and the better the value of this similarity term is taken, the more similar the processed radar signal or processed representation is to the target radar signal or target representation according to a pre-given metric. This also more strongly counteracts the possible tendency of the generator to search for "easy ways" that are not clearly targeted for the specified application in order to eliminate motion artifacts. For example, if most of the target radar signals or target representations relate to a scene in which moving vehicles can be seen, the generator "hides" by making all the vehicles disappear to minimize L cGANAttempts at better values of (G, D) are frustrated by this similarity phase. An example of a similarity term is . In this case, in principle, any other metric can also be used as the L1 metric.

[0020] In another advantageous design, the cost function additionally includes an application term that measures the desirable properties themselves of the processed radar signal or the processed representation for the intended application. This application term is also referred to as the "Perceptual Loss" L P (G). This application term is not limited to depending only on the result of the transformation. Rather, this application term can also depend, for example, on intermediate results of the transformation. If the generator includes, for example, a multi-layer neural network, these intermediate results can also be extracted on the hidden layers between the input layer and the output layer.

[0021] This application term can measure, for example: whether the scene presented by the processed radar signal or by the processed representation is reasonable for the corresponding application. In this way, for example, representations where a car is three times as high or three times as wide as normal or moving at approximately the speed of sound in an urban area are discarded. Alternatively or in combination with this, for example, a comparison with the target radar signal or the target representation can also be inserted in the application term. In this way, for example, the presentation generated by an autoencoder or other KNN based on the processed radar signal or the processed representation can be compared with the presentation generated by the same KNN based on the target radar signal or the target representation.

[0022] Using the similarity term and the application term, the optimization problem can be written, for example, as . Where λ and γ are hyperparameters that weight the different cost function terms.

[0023] In another particularly advantageous design, the target radar signal or the target representation is determined based on the measured radar signal. The actual radar signal or the actual representation is determined based on a reduced information version of the measured radar signal. In this way, on the one hand, it is ensured that both the actual radar signal or the actual representation and the target radar signal or the target representation relate to the same scene respectively. On the other hand, the ability of the trainable module can be enhanced in a targeted manner to compensate for certain mechanisms that are expected to blur the information content in the radar signal or the representation in the corresponding application.

[0024] For example, radar sensors for vehicles that have multiple channels in the plane of vehicle movement (i.e., for example, along the direction of movement) are relatively expensive. Thus, it is desirable that for the final mass-produced version of radar-assisted environmental observation, fewer channels or even just one channel can be accommodated. And in order to record the actual radar signals, a radar sensor with multiple channels can be used, since only a small number of vehicles need to be equipped with this radar sensor.

[0025] Thus, it is advantageous to use a radar sensor for measuring radar signals, which is mounted on a vehicle and has multiple channels with a layout that is not only oriented perpendicular to the vehicle movement. Accordingly, in particular, a target representation of a scene can be determined, in which moving objects are correctly positioned and largely free of motion artifacts. In addition, the speed of the moving objects can also be determined based on this data. Then, for example, an information-reduced version of the measured radar signal can be formed by masking at least one channel.

[0026] In another particularly advantageous design, additionally, another generator module is trained to convert a target radar signal or a target representation back into a radar signal or a representation of the type of the actual radar signal or the actual representation. That is, the converted-back radar signal or representation exists in the space of the actual radar signal or the actual representation. Then, the cost function additionally includes a return GAN term. The better the converted-back radar signal or representation cannot be distinguished from the actual radar signal or the actual representation by another discriminator module, the better the value of this return GAN term.

[0027] This other discriminator module is trained to distinguish the converted-back radar signal or representation from the actual radar signal or the actual representation. The cost function also additionally includes a consistency term. This consistency term is a measure of how identically the actual radar signal or the actual representation is reproduced in the case of conversion by a trainable module and conversion back by this other generator module.

[0028] In the case of this type of training, the architecture in which the trainable module is incorporated as a generator in terms of function is extended from a "conditional GAN" to a CycleGAN. An important advantage is that the target radar signal or the target representation no longer has to relate to the same scene as the actual radar signal or the actual representation. A major strength of CycleGAN is that the CycleGAN can convert data between domains that are each characterized by an unpaired set of examples.

[0029] This can significantly simplify and cost less in training. In the previously described example, in which the trainable module is trained based on, on the one hand, an actual radar signal or actual representation with less information content and, on the other hand, a target radar signal or target representation with more information content, training is also possible in the absence of a radar sensor with multiple channels having a layout oriented not only perpendicular to the vehicle movement for determining the target radar signal or target representation. As an alternative, for example, the target radar signal or target representation can be purchased from outside and only the actual radar signal and actual representation can be collected by means of one's own measurements.

[0030] In particular, it also facilitates the subsequent transformation of the radar sensors to be used in batches. With the new radar sensors, it is sufficient to record the new actual radar signals and, if necessary, determine the new actual representations accordingly. The target radar signals and target representations used so far can continue to be used.

[0031] CycleGAN learns a mapping G from the space X in which there is an actual radar signal or actual representation x to the space Y in which there is a target radar signal or target representation y, like "conditional GAN". Additionally, it learns the reverse mapping F from space Y to space X. A first discriminator D is learned x , which attempts to distinguish between the generated data F(y) and the real radar signal or representation x. A second discriminator Dy is learned, which attempts to distinguish between the generated data G(x) and the real target radar signal or target representation y. This can be expressed, for example, by cost function terms and . Where z1 and z2 are samples of the random variables Z1 and Z2. The use of the random variables Z1 and Z2 is optional.

[0032] An exemplary consistency term for monitoring compliance with the consistency conditions and is .

[0033] Then, the overall cost function for CycleGAN can be written, for example, as

[0034] .

[0035] This cost function can also be extended with an application term L P , similar to "conditional GAN" (cGAN), where this application term now depends not only on G but also on F: L P = L(G,F). This term can be added, for example, with a weight γ.

[0036] Similarly, a similarity term can be added to the cost function for CycleGAN. Different from cGAN, there are now two terms for the two generators G and F: and . These terms can be added, for example, with a weight μ.

[0037] In another particularly advantageous design, the hyperparameters are optimized according to a predefined optimization criterion, and these hyperparameters specify the relative weights of the terms in the cost function with respect to each other. These hyperparameters represent additional degrees of freedom that the trainable module can use to adapt to a specific task. For example, a search space spanned by multiple hyperparameters can be searched in a predefined grid. This does not require the optimization criterion to depend stably on these hyperparameters.

[0038] As already introduced in these formulas, in another particularly advantageous design, at least one actual radar signal includes not only the measured radar signal but also a sample drawn from a random variable. This sample can be added to the measured radar signal, for example. The random variable can in particular be multi-dimensional and, for example, normally distributed. The noise added in this way has a twofold effect: on the one hand, multiple other variables can be generated according to a predefined reserve of the actual radar signal or a predefined reserve of the actual representation generated therefrom, in order to increase the variability of the training. On the other hand, other features in the latent space can also be learned.

[0039] In another particularly advantageous design, the trainable module is additionally trained to determine processed representations that contain position-resolved information about the speed of the objects contained in these processed representations. If the processed representation exists in the form of an image, for example, the image can be extended with other channels (similar to color channels), and each of these other channels contains speed information of the object to which the pixel belongs. The speed information here is the radial component of the absolute speed of the object in a coordinate system centered on the radar sensor. The speed information can be used, for example, to check the following question: whether the host vehicle (this vehicle), from which the surrounding environment is observed, has to change its behavior in order to avoid a collision with other moving objects.

[0040] In another particularly advantageous design, a trainable module with a hierarchically structured artificial neural network KNN is selected. In this case, the number of neurons and / or other processing units in each layer of the first layer sequence monotonically decreases, while the number of neurons and / or other processing units in each layer of the second layer sequence monotonically increases. Thereby, a "bottleneck" is formed between the end of the first layer sequence and the start of the second layer sequence, in which intermediate results exist that have a significantly reduced dimensionality compared to the input radar signal or the input representation of the scene. This "bottleneck" enables the KNN to learn and compress relevant features. Thereby, better performance can be achieved and the computational effort can be reduced.

[0041] In another particularly advantageous design, the KNN has at least one direct connection between the first layer of the first layer sequence and the second layer of the second layer sequence. In this way, certain information can be selectively passed through the mentioned "bottleneck", such that the information content in the processed radar signal or in the processed representation is increased overall. That is, the direct connection is particularly understood as bypassing at least one layer that would otherwise have to be passed through in the first and / or second layer sequence.

[0042] Once the trainable module is trained, the final state is reflected in a parameter set that has parameters characterizing the properties of this final state. In the case of the KNN, these parameters can include, for example, weights with which the inputs for the activation of the neuron or the processing unit fed to the neuron or other processing unit are calculated. The parameter set enables the trainable module to be arbitrarily replicated without further training and is thus a marketable product on its own.

[0043] The invention also provides a method for analyzing moving objects based on radar signals. In this method, a processed representation of the scene observed using the radar signals is determined using the trainable module trained as previously described. A second representation of the scene observed using these radar signals is determined based on the input fed to the trained module in this case. Position-resolved information about the speed of the objects contained in these representations is determined based on the difference between the processed representation and the second representation. This task can in particular be taken on by another trained generator.

[0044] This is based on the following consideration: Suppressing motion artifacts requires knowledge of the underlying motion and the motion can thus be reconstructed based on the results of this suppression.

[0045] As described previously, an important application of a module trained using the method described at the beginning is to appropriately react to objects identified in the vehicle's surroundings for the purpose of a driving assistance system and for the purpose of at least partially automated driving. Thus, the present invention relates to another method.

[0046] In this method, first, a trainable module is trained using the method described at the beginning. Using at least one radar sensor arranged on the vehicle, radar signals from the vehicle's surroundings are measured. At least one control signal is formed based on the result of the analysis. The vehicle is controlled using the control signal.

[0047] In this case, the formation of the control signal can in particular include checking whether, based on the result of the analysis in combination with the current or planned trajectory of the own vehicle, there is a concern that the trajectory of an object in the surroundings of the own vehicle intersects with this current or planned trajectory of the own vehicle. If this is the case, the control signal can in particular be aimed at modifying the trajectory of the own vehicle such that this trajectory no longer intersects with the trajectory of the identified object.

[0048] These methods can in particular be implemented fully or partially by a computer. Thus, the present invention also relates to a computer program having machine-readable instructions which, when implemented on one or more computers, cause the one or more computers to implement one of the methods described. In this regard, a control device for a vehicle and an embedded system for a technical device which are also capable of implementing the machine-readable instructions should also be regarded as computers.

[0049] Likewise, the present invention also relates to a machine-readable data carrier and / or a download product having the parameter set and / or having the computer program. A download product is a digital product that can be transmitted via a data network, that is to say can be downloaded by a user of the data network, and this digital product can for example be sold in an online store for immediate download.

[0050] Furthermore, a computer can be equipped with the parameter set, the computer program, the machine-readable data carrier or the download product. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In the following, other measures for improving the present invention are further presented in conjunction with the description of the preferred embodiments of the present invention based on the drawings.

[0052] wherein:

[0053] Figure 1 shows an embodiment of a training method 100;

[0054] Figure 2 shows an exemplary scenario 2 to be analyzed by a trainable module 1;

[0055] Figure 3 shows an exemplary neural network 14 for use in a trainable module 1;

[0056] Figure 4 shows an embodiment of a method 200 for analyzing a radar signal 10;

[0057] Figure 5 shows an embodiment of a method 300 having a complete behavior chain up to controlling a vehicle 50. Detailed Description

[0058] Figure 1 shows a flowchart of an embodiment of a training method 100. In step 105, a trainable module 1 is selected for training, which trainable module includes a KNN 14 having an encoder-decoder architecture ("U-Net"). Further shown in Figure 3 is this KNN 14.

[0059] In step 120, a scene 2 is observed by means of radar. The radar signal 10 is measured using a radar sensor 51, which has a plurality of channels 51a - 51c. According to block 121, a reduced information version 10a of the measured radar signal 10 is determined by masking at least one of the channels 51a - 51c.

[0060] Based on this reduced information version 10a, an actual radar signal 11 and / or an actual representation 12 of the observed scene 2 is determined. In step 110a, the actual radar signal 11 or the actual representation 12 is fed to the trainable module 1. In step 110b, the trainable module 1 maps the actual radar signal 11 or the actual representation 12 onto a processed radar signal 11* or onto a processed representation 12*. The goal of this processing is: to improve the suitability of the actual radar signal for the motion reconstruction of an object; or to reduce motion artifacts in the actual representation.

[0061] In step 130, based on a cost function 13, it is evaluated to what extent the processed radar signal 11* is suitable for the reconstruction of the motion 21a of an object 21; or to what extent motion artifacts are actually reduced in the processed representation 12*. For this purpose, the cost function compares the processed radar signal 11* or the processed representation 12* with a target radar signal 11' or with a target representation 12'. In the Figure 1 example shown, in step 125a, a target radar signal 11' or a target representation 12' is determined based on the measured radar signal 10, and the actual radar signal 11 or the actual representation 12 has been determined based on this reduced information version 10a of the measured radar signals.

[0062] In step 140, the evaluation 130a performed by means of the cost function is used to optimize the parameter 1a that characterizes the properties of the trainable module 1. This optimization is used to pursue the following: in the case where the actual radar signal 11 or the actual representation 12 is further processed by the trainable module 1, the evaluation 130a of the obtained processed radar signal 11* or the obtained processed representation 12* by means of the cost function 13 becomes better. If this optimization ends according to a predefined interruption criterion, then the value of the parameter 1a obtained then characterizes the properties of the trainable module 1 in the training completed state 1*.

[0063] Additionally, in step 150, the hyperparameters 13a can also be optimized, which specify the relative weights of the terms in the cost function 13 with respect to one another. Here, it can be freely chosen whether this optimization is carried out before, during, or after the optimization 140 of the parameter 1a.

[0064] Some exemplary paths of this optimization are drawn within the box 140.

[0065] In this way, according to block 141, the discriminator module can be trained to distinguish the processed radar signal 11* or the processed representation 12* from the target radar signal 11' or from the target representation 12'. Then, the discriminator module and the trainable module 1 together form a "conditional generative adversarial network" cGAN.

[0066] According to block 142, another generator module can be trained, which works exactly in the opposite direction to the actually trainable module 1 to be trained. Then, another discriminator module can be trained according to block 143 to distinguish the radar signal or the representation provided by this other generator module from the actual radar signal 11 or from the actual representation 12. The cGAN is extended to a CycleGAN in this way. The CycleGAN no longer depends on the fact that the actual radar signal 11 or the actual representation 12 relates to the same scene 2 as the target radar signal 11' or the target representation 12'. That is, then the target radar signal 11' or the target representation 12' from any source can be used.

[0067] Figure 2As an example, a scene 2 in the surroundings 50a of a vehicle 50 is shown. A trainable module 1 is trained using method 100 to analyze this type of scene 2. Vehicle 50 is equipped with a radar sensor 51 having multiple channels 51a-51c along the direction of travel of vehicle 50. Each channel 51a-51c is equipped with a separate receiving unit. This means that equipping radar sensor 51 with multiple channels 51a-51c increases the hardware costs accordingly. Therefore, the goal is to accommodate as few channels 51a-51c as possible, or even just one channel 51a-51c, during actual operation.

[0068] The scene 2 is observed by emitting a signal 52a and analyzing the signal 52b reflected by the object 21. In this case, in addition to the position, the speed 21a of the object 21 is also particularly important, because it also determines whether the vehicle 50 must change its own behavior to avoid a collision.

[0069] Radar measurements always measure a velocity component that is radially oriented toward or away from the radar sensor 51. Figure 2 In the case shown in , this component is small. Since this small component can be determined more accurately using the trained module 1, the motion prediction of the object 21 is significantly improved.

[0070] Figure 3 An exemplary KNN 14 is schematically shown, which can be used in a trainable module 1. In this example, the KNN 14 consists of seven layers 15a-15g, each of which includes neurons or other processing units 16. In this case, the layers 15a-15c form a first layer sequence 17a, in which the number of neurons 16 in each layer 15a-15c decreases monotonically. The layers 15e-15g form a second layer sequence 17b, in which the number of neurons 16 in each layer 15e-15g increases monotonically. Between them lies the layer 15d, in which there is a maximally compressed representation of the actual radar signal 11 or the actual representation 12. In addition, the KNN 14 contains three direct connections 18a-18c between the layers 15a-15c in the first layer sequence 17a and the layers 15e-15g in the second layer sequence 17b, which are connected in series. Figure 3 In the examples shown in FIG, there are in each case the same number of neurons 16 .

[0071] Figure 4It is a flowchart of a special method 200 for analyzing a radar signal 10. In step 210 of this method, using a module 1* trained according to method 100, a processed representation 2a of the scene 2 observed using the radar signal 10 is determined based on the radar signal 10 and / or based on a representation 10' of the resulting scene 2. In this processed representation 2a, motion artifacts are reduced by the trained module 1*. In step 220, a second representation 2b of the scene 2 is determined based on the input fed to the trained module 1*, in which the motion artifacts are not reduced.

[0072] For example, in step 210, the trained module 1* can first improve the applicability of the radar signal 10 for motion reconstruction, such that based on this, a processed representation 2a with reduced motion artifacts can be obtained using the SAR algorithm. Then, in step 220, the same SAR algorithm can be applied to the original radar signal 10 in order to obtain a second representation 2b without corrected motion artifacts.

[0073] However, for example, the SAR algorithm can also be directly applied to the radar signal 10 in order to obtain a representation 10' of the scene 2. Then, in step 210, based on this representation 10', a processed representation 2a with reduced motion artifacts can be obtained by applying the trained module 1*. In step 220, the representation 10' can, for example, be delivered without change in order to obtain a representation 2b without corrected motion artifacts.

[0074] In step 230, the speed 21a of an object 21 in the scene 2 is determined based on the comparison of the representations 2a and 2b. For this purpose, for example, another generator in the form of a cGAN can be used.

[0075] Figure 5 It is a flowchart of an embodiment of a method 300, which method includes a complete behavior chain from measuring the radar signal 10 up to controlling a vehicle 50.

[0076] In step 310 of method 300, a trainable module 1 is trained using method 100. In step 320, the radar signal 10 is measured using a radar sensor 51. In step 330, these radar signals 10 are analyzed using the trained module 1*.

[0077] In step 340, a control signal 340a is determined based on the result 330a of the analysis 330. In step 350, the vehicle 50 is controlled using this control signal 340a. The control signal 340a can in particular have the property that by controlling the vehicle 50 using this control signal, the trajectory of the vehicle 50 is changed such that a collision with an object identified in the surroundings 50a of the vehicle 50 is avoided.

Claims

1. A method (100) for training a trainable module (1) for analyzing radar signals (10), the method having the following steps: · Feeding (110a) an actual radar signal (11) and / or a thus determined actual representation (12) of a scene (2) observed using the actual radar signal (11) to the trainable module (1) and converting (110b) by the trainable module (1) to a processed radar signal (11*) and / or converting (110b) to a processed representation (12*) of the corresponding scene (2); · Evaluating (130) by means of a cost function (13): to what extent the processed radar signal (11*) and / or the processed representation (12*) contain artifacts caused by moving objects (21) in the scene (2); · Parameters (1a) characterizing the properties of the trainable module (1) are optimized (140), with the aim that the trainable module (1) increasingly suppresses, as training progresses, artifacts caused by the moving objects (21) in the processed radar signal (11*) and / or in the processed representation (12*).

2. The method (100) according to claim 1, wherein · The cost function (13) contains a GAN term, and the better the value of the GAN term is, the less distinguishable the processed radar signal (11*) or the processed representation (12*) is from a set of predefined target radar signals (11') or from a set of predefined target representations (12') according to a discriminator module; and · The discriminator module is additionally trained (141) to distinguish the processed radar signal (11*) or the processed representation (12*) from the target radar signal (11') or from the target representation (12').

3. The method (100) according to claim 2, wherein the cost function (13) additionally contains a similarity term, and the better the value of the similarity term is, the more similar the processed radar signal (11*) or the processed representation (12*) is to the target radar signal (11') or to the target representation (12') according to a predefined metric.

4. The method (100) according to claim 2 or 3, wherein the cost function (13) additionally contains an application term, and the application term measures the desirable properties themselves of the processed radar signal (11*) or the processed representation (12*) for the intended application.

5. The method (100) according to claim 2 or 3, wherein the target radar signal (11') or the target representation (12') is determined (125a) based on the measured radar signal (10), and wherein the actual radar signal (11) or the actual representation (12) is determined (125b) based on a reduced information version (10a) of the measured radar signal (10).

6. The method (100) according to claim 5, wherein the radar signal (10) is measured (120) by means of a radar sensor (51) which is mounted on a vehicle (50) and has a plurality of channels (51a - 51c) with a layout oriented not only perpendicular to the vehicle movement.

7. The method (100) according to claim 6, wherein an information - reduced version (10a) of the measured radar signal (10) is formed (121) by masking at least one channel (51a - 51c).

8. The method according to claim 2 or 3, wherein additionally · Another generator module is trained (142) to convert a target radar signal (11') or a target representation (12') back into a radar signal or a representation of the type of the actual radar signal (11) or the actual representation (12), wherein the cost function (13) additionally contains a return GAN term, and the better the value of the return GAN term is, the less distinguishable the converted - back radar signal or representation is from the actual radar signal (11) or from the actual representation (12) by another discriminator module; · The another discriminator module is trained (143) to distinguish the converted - back radar signal or representation from the actual radar signal (11) or from the actual representation (12); and · The cost function (13) contains a consistency term which is a measure of the extent to which the actual radar signal (11) or the actual representation (12) is reproduced identically when converted by the trainable module (1) and converted back by the another generator module.

9. The method (100) according to claim 2 or 3, wherein hyperparameters (13a) are optimized (150) according to a predefined optimization criterion, the hyperparameters specifying the relative weights of the terms in the cost function (13) with respect to each other.

10. The method (100) according to any one of claims 1 to 3, wherein at least one actual radar signal (11) includes not only the measured radar signal (10) but also samples drawn from a random variable.

11. The method (100) according to any one of claims 1 to 3, wherein the trainable module is additionally trained (145) to determine a processed representation (12*) which contains position - resolved information about the speed (21a) of an object (21) comprised in the processed representation (12*).

12. The method (100) according to any one of claims 1 to 3, wherein a trainable module (1) with a hierarchically - structured artificial neural network KNN (14) is selected (105), wherein the number of neurons and / or other processing units (16) in each layer (15a - 15g) of a first layer sequence (17a) decreases monotonically and the number of neurons and / or other processing units in each layer of a second layer sequence (17b) increases monotonically.

13. The method (100) according to claim 12, wherein the KNN (14) has at least one direct connection (18a - 18c) between a first layer (15a - 15c) in the first layer sequence (17a) and a second layer (15e - 15g) in the second layer sequence (17b).

14. A method (200) for analyzing a moving object (21) based on a radar signal (10), the method comprising the steps of: · Using a module (1*) trained by the method (100) according to any one of claims 1 to 13, determining (210) a processed representation (2a) of a scene (2) observed by means of the radar signal (10) based on the radar signal (10); · Determining (220) a second representation (2b) of the scene (2) observed by means of the radar signal (10) based on the input (10, 10') fed to the trained module (1*) in this case; · Determining (230) position - resolved information about the velocity (21a) of an object (21) comprised in these representations based on the difference between the processed representation (2a) and the second representation (2b).

15. A method (300) comprising the steps of: · Training (310) a trainable module (1) using the method (100) according to any one of claims 1 to 13; · Measuring (320) a radar signal (10) from the surroundings (50a, 2) of a vehicle (50) using at least one radar sensor (51) arranged on the vehicle (50); · Analyzing (330) the radar signal (10) using the trained module (1*); · Forming (340) at least one control signal (340a) based on the result (330a) of the analysis (330); · Controlling (350) the vehicle (50) using the control signal (340a).

16. A computer program product comprising a computer program which contains machine - readable instructions that, when implemented on 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 13 or 14 to 15.

17. A machine - readable storage medium having a parameter set obtained using the method according to any one of claims 1 to 13 and / or having the computer program product according to claim 16, wherein the parameter set has parameters (1a) characterizing the properties of the trainable module (1).

18. A computer equipped with a parameter set obtained using the method according to any one of claims 1 to 13, equipped with the computer program product according to claim 16 and / or equipped with the machine - readable storage medium according to claim 17, wherein the parameter set has parameters (1a) characterizing the properties of the trainable module (1).

Images