Reconstruction of elevation information from radar data

Abstract

The invention relates to a method (100) for reconstructing elevation information (2a) from measurement data (2), which is recorded with at least one radar device (1) and contains a two-dimensional spatial distribution (3) of at least one physical measurement variable, wherein the measurement data (2) is delivered (110) as an input variable (5) to at least one generator module (4) configured as a neural network, and wherein at least one output variable (6) is called up (120) from the generator module (4), which is a measure of how great an elevation angle (6a) at which radar radiation is reflected by at least one object to the radar device (1). A method (200) for training a generator module (4) for use in the method (100) is also described. A method (300) with a complete chain of actions up to the maneuvering of a vehicle (50) is also described.

Description

Technical Field

[0001] This invention relates to the analysis of measurement data recorded using vehicle-mounted radar devices, particularly synthetic aperture radar. Background Technology

[0002] For vehicles to move at least partially automatically in road traffic, it is necessary to detect the vehicle's surroundings and introduce countermeasures if a collision with an object in the surrounding environment is imminent. For safe automated driving, it is also necessary to create an environment representation and localization.

[0003] Detecting objects with radar is independent of lighting conditions and is possible, for example, at large distances even at night, without dazzling oncoming traffic with high beams. However, similar to optical imaging, positional and angular resolution is naturally limited by diffraction due to the interaction between the wavelength used and the physical size of the aperture from which radar radiation emanates.

[0004] To improve resolution, GB 2 564 648 A proposes combining radar signals measured at different observation angles based on vehicle motion when observing the environment around a moving vehicle using radar. This combination of radar signals has an effect similar to observing the surrounding environment through a long aperture. Therefore, this technology is also known as Synthetic Aperture Radar (SAR). Summary of the Invention

[0005] Within the framework of this invention, a method for reconstructing elevation information based on measurement data recorded using at least one radar device has been developed.

[0006] A radar device defines a plane by the spatial arrangement of one or more antennas, such as an antenna array. Within this plane, the azimuth angle indicates the lateral direction relative to an object detected by the radar device. The elevation angle indicates the height of the object above the plane, or how high an extended object is raised within the plane. A vehicle-mounted radar device can be defined by the spatial arrangement of antennas, such as the plane of vehicle movement.

[0007] The measurement data contains a two-dimensional spatial distribution of at least one physical measurement parameter. Examples of such measurement parameters include the intensity of reflected radar radiation, the radar cross-section of an object in the scene, and the velocity component of the object along the propagation direction of the radar radiation.

[0008] For example, when a mobile synthetic aperture radar (SAR) device is used to record measurement data and this data is analyzed in a conventional manner, this two-dimensional spatial distribution is formed. Since land vehicles can only move within a plane, the angle of elevation observed by the radar device does not change due to the movement of the land vehicle. Therefore, without a specially constructed antenna system for elevation measurement, only two-dimensional information is obtained during analysis, and the elevation angles of the radar radiation reflected to the radar device are ignored. Consequently, three-dimensional objects such as houses, trees, or trucks (LKW) are projected distortedly onto the plane. This results in additional errors in the azimuth angle at which objects are identified within this plane. Correspondingly, in the three-dimensional representation of the scene in Cartesian coordinates determined based on these measurement data, the spatial distances of objects are also distorted.

[0009] This method is employed here. The measurement data is fed as input parameters to at least one generator module constructed as a neural network. At least one output parameter is retrieved from the generator module, which is a measure of the elevation angle at which radar radiation is reflected back to the radar device by at least one object.

[0010] The projection from three-dimensional space to a two-dimensional plane can be described mathematically. In this case, a common approach is to express the reconstruction as a mathematical problem inversely related to the projection and solve it. Since only one angular component is detected, the solution to this inverse problem is often unknown, and multiple mathematically equivalent solutions exist. A unique solution can only be found under certain assumptions, such as an elevation angle of zero. However, it has been recognized that when an object is measured at multiple points, there is a correlation between the sought elevation angle information and the available two-dimensional spatial distribution of the measured information. Machine learning can utilize this correlation.

[0011] In other words, the elevation angle information, supplemented by a limited amount of training measurements known as "ground truth," generally allows the generator module to determine the appropriate elevation angle information based on the existing two-dimensional spatial distribution. This method offers different advantages compared to the mathematical inversion of projection.

[0012] Machine learning is not bound by the principle that projections always contain simplified mathematical models. Rather, machine learning can also automatically grasp effects that are not included in such models and may be fundamentally difficult to model.

[0013] In methods using machine learning, there is no known tendency when inverting mathematical problems that noise in the input measurement data is significantly amplified and detrimental to the quality of the reconstruction.

[0014] Finally, there is also arbitrary flexibility in precisely querying elevation angle information. To enable the generator module to provide new elevation angle information, the module only needs to be retrained by the machine. The mathematical model, however, must be modified by a human.

[0015] In a particularly advantageous design, the generator module's output parameters regarding the location in a two-dimensional spatial distribution specify the elevation angle at which radar radiation contributing to the value of the physical measurement parameters at the corresponding location is reflected back to the radar device. In this way, for example, a two-dimensional radar image can be converted into an elevation image, in which each pixel value specifies the elevation angle (e.g., the elevation angle) with respect to the location represented by the corresponding pixel in the radar image. The generator module can be trained particularly well to provide this type of output parameter because, for example, the elevation angle represented numerically can be compared particularly clearly with the target elevation angle from the training measurement data.

[0016] In another particularly advantageous design, the output parameters comprise a three-dimensional spatial distribution of at least one sought physical measurement parameter. This distribution can be, for example, a three-dimensional radar image in arbitrary coordinates, such as Cartesian coordinates (x, y, z) or polar coordinates (range, azimuth, and elevation). Such a three-dimensional image can be analyzed with particular clarity in specific applications, such as the classification of objects in the environment surrounding a vehicle. For this purpose, a suitable metric is needed during training, which allows for a numerical evaluation of the difference between the image provided by the generator module and the target image generated based on "ground reality" of the training measurement data.

[0017] The cost of memory and computing power for generating three-dimensional spatial distributions can be reduced by breaking down these tasks into multiple subtasks that are processed in parallel or sequentially. For this purpose, the two-dimensional spatial distribution is divided into multiple interconnected regions. For example, an image can be divided into "patches" of arbitrary shapes, where whether these "patches" overlap at their edges and, if so, to what extent, is optional. For the input parameters belonging to each region ("patch"), a generator module is used to determine the corresponding output parameters, i.e., the values ​​that should be assigned to certain points in three-dimensional space. The generator module can be designed to be correspondingly smaller. The generator module only needs to be able to process one "patch" at a time.

[0018] All determined output parameters are aggregated together to obtain the sought three-dimensional spatial distribution of at least one of the sought physical measurement parameters.

[0019] In this way, an initially empty 3D image can be filled with other values ​​of the physical measurement parameters being sought each time another "patch" is processed, where the entries caused by connected "patches" do not need to be located in connected areas of the 3D image. Here, values ​​that have already been assigned to a location can also be modified retroactively using new entries.

[0020] In particular, processing performed in a "patch" manner can be well parallelized on graphics processing units (GPUs) with multiple processing units.

[0021] In another particularly advantageous design, the generator module includes:

[0022] • An encoder that transforms input parameters into latent variables in a space whose dimension is less than the dimension of the input parameter space and less than the dimension of the output parameter space; and

[0023] • Decoder, which transforms these latent variables into output parameters.

[0024] By compressing the input parameters into latent variables in a low-dimensional space, the number of parameters characterizing the generator module is significantly reduced. This also simplifies training.

[0025] The use of an encoder-decoder structure should not be misunderstood as the generator module being an autoencoder. The goal of the generator module is not to reproduce the input parameters identically by transforming them into latent variables and then back to the output parameters. Rather, these input parameters in a first space, where the input parameters do not contain elevation information, should be transformed into a second space, where elevation information is also present. The commonality with autoencoders is that this structure only has a "bottleneck" in the form of latent variables.

[0026] The encoder and decoder can each be organized, for example, into multiple consecutive layers. Thus, the neural network can also include, for example, direct connections from layers of the encoder to layers of the decoder, these direct connections bypassing the space of latent variables and optionally bypassing one or more other layers of the encoder and / or decoder. In this way, certain information in the layers from which the direct connections of the encoder originate can be protected from deletion during compression into latent variables. Thus, the network becomes a "U-Net".

[0027] Advantageously, the output parameters of the generator module include multiple two-dimensional spatial distributions of at least one physical measurement parameter. Each of these distributions then corresponds to the radar radiation portion from which the corresponding distribution was generated, along with the elevation angle range from which it is reflected to the radar device. These two-dimensional spatial distributions can be explicitly combined into a three-dimensional spatial distribution of the physical measurement parameter. If the two-dimensional spatial distributions exist, for example, in coordinates—range and azimuth—then these two-dimensional spatial distributions can be stacked overlappingly in the coordinate—elevation direction, thereby obtaining a three-dimensional tensor.

[0028] Of particular advantage, for the purpose of performing the described analysis, the measurement data is selected from radar devices whose antenna arrays cannot directly measure elevation angles. That is, if the specific task is to detect radar data with a defined level of elevation angle information, this opens up a method that first records the measurement data without elevation angle information using a structurally simpler and cheaper antenna array, and then reconstructs the elevation angle information using a generator module. This results in cost savings, especially when many structurally identical copies of the radar device should be manufactured and sold. The generator module only needs to be trained once and is then effective for all structurally identical radar devices.

[0029] As previously mentioned, it is particularly advantageous to select measurement data recorded using a mobile synthetic aperture radar device. This is especially beneficial in the case of land vehicles that can only move within a single plane, as it eliminates the need to compensate for the lack of an elevation angle due to the movement, which would otherwise change the elevation angle. In other words, the radar sensor does not need additional channels, for example, outside the plane defining the azimuth angle.

[0030] In another particularly advantageous design, analysis is performed based on the output parameters of the generator module.

[0031] • At least one category of a pre-defined classification of traffic signs, other road users, lane markings, signaling devices, or other traffic-related objects in the vehicle's surrounding environment; and / or

[0032] • The minimum location, spatial dimensions, and / or velocity of such an object.

[0033] This is the most important information that at least partially automated vehicles need in order to join mixed traffic with vehicles controlled by human drivers.

[0034] This invention also relates to a method for training a generator module for use in the previously described method. In this method, training measurement data is acquired by recording using a radar device whose antenna array enables direct elevation angle measurement. Based on the training measurement data, a two-dimensional spatial distribution of at least one physical measurement parameter, serving as an input parameter for the generator module, is determined. For example, this two-dimensional spatial distribution can be obtained by simulating the previously described projection onto a two-dimensional plane.

[0035] Based on these training measurement data, target output parameters are also determined. These target output parameters are measures of the elevation angle at which radar radiation is reflected back to the radar device by at least one object. Depending on the type of radar device with elevation capability, the elevation angle may be read directly from these training measurement data.

[0036] The input parameters are processed into output parameters using the generator module to be trained. The generator parameters, which characterize the generator module, are optimized with the goal of ensuring that the output parameters reproduce the target output parameters well according to a pre-defined generator cost function.

[0037] As mentioned earlier, in this way, the generator module can be customized not only with respect to the specific radar device used, especially the antenna array used, but also with respect to the specific desired output parameters.

[0038] In a particularly advantageous design, the output parameters are additionally fed to a discriminator module constructed as another neural network. The discriminator parameters, which characterize the discriminator module, are trained alternately with the generator parameters. The goal of this optimization is for the discriminator module to effectively distinguish the output parameters generated by the generator module from the target output parameters according to the discriminator cost function.

[0039] The inclusion of the discriminator module results in more realistic elevation information generated by the generator module. Specifically, it suppresses the tendency for the generator module to produce output parameters that, while close to the target output parameters in terms of the metric used, are plagued by visible artifacts that do not appear in the same form in real radar images.

[0040] To train the generator and discriminator modules together, specifically by combining the generator cost function and the discriminator cost function into a total cost function, such that the value of the total cost function is...

[0041] • It gets better when the output parameters are close to the target output parameters, but

[0042] • It gets worse when the discriminator module can distinguish these output parameters well from the true target output parameters.

[0043] That is, during training, there is a continuous "competition" between the generator module and the discriminator module. Therefore, the device consisting of the generator module and the discriminator module is a Generative Adversarial Network (GAN). Here, the generator module generates elevation angle information regardless of its form under the boundary condition that the elevation angle information relates to the same scene represented by input parameters (such as a deformed 2D radar image). Specifically, it does not generate, for example, a 3D radar image, which is very realistic but does not perfectly correspond to the scene discussed here. Therefore, based on this boundary condition, the device is a "conditional GAN," or cGAN. Thus, during training, the discriminator module should be provided with the same input parameters upon which the generator module operates.

[0044] Advantageously, for recording training measurement data, a radar device with multiple channels outside the plane defining its azimuth angle is selected. The training measurement data can then be clearly divided into input parameters of a type obtainable even in measurements without elevation capability, and additional information about elevation angle available only during training. For example, input parameters can be formed based on measurement data from only one channel of the radar device, while measurement data from all channels are incorporated into the formation of the target output parameters.

[0045] In another particularly advantageous design, for each channel, a range-relative velocity spectrum of the corresponding training measurement data is determined. This range-relative velocity spectrum is used not only for the input parameters of the generator module but also for the target output parameters. For this purpose, for example, an angle estimate of the electrical angle can be determined for each pair of range and relative velocity contained in this spectrum, which in turn depends on the azimuth and elevation angles. That is, the azimuth and elevation angles can then be calculated from the relative velocity combined with the electrical angle.

[0046] The present invention also relates to another method for closing the behavioral chain leading up to the control of the vehicle.

[0047] In this method, the generator module is first trained using the training method described earlier. Measurement data from the vehicle's surrounding environment is recorded using at least one radar device mounted on or within the vehicle. Based on this measurement data, elevation angle information is reconstructed using the method described earlier. Using this elevation angle information, a control signal is generated. The vehicle is then controlled using the control signal.

[0048] As mentioned earlier, the reconstruction of elevation angle information results in more accurate interpretation of measurement data regarding the traffic conditions in which the vehicle is located. In particular, the size of objects and the distances between them are determined more precisely. That is, the "shadowing" of objects in radar images caused by non-zero elevation angles is not mistaken for physically existing objects. Therefore, the overall probability of the vehicle responding appropriately to the traffic conditions when controlled using control signals is increased.

[0049] As previously mentioned, the training results of the generator module are reflected in the generator parameters. Even without prior training, possessing these generator parameters allows the generator module to supplement elevation information into measurement data from radar devices lacking elevation capability. Therefore, these generator parameters are independently marketable products with customer benefits. Consequently, this invention also relates to a set of parameters with generator parameters obtained by training the generator module using the previously described training method.

[0050] These methods can be implemented entirely or partially by a computer. Therefore, the present invention also relates to a computer program having machine-readable instructions that, when implemented on one or more computers, cause the computers to perform one of the described methods. In this regard, control devices for vehicles and embedded systems for technical devices, which are also capable of implementing machine-readable instructions, should also be considered as computers.

[0051] Similarly, the present invention also relates to a machine-readable data carrier and / or a downloadable product having the parameter set and / or the computer program. The downloadable product is a digital product that can be transmitted via a data network, that is, downloaded by a user of the data network, and such digital product may, for example, be sold in an online store for immediate download.

[0052] In addition, the computer may be equipped with the parameter set, computer program, machine-readable data carrier, or downloadable product. Attached Figure Description

[0053] In the following description of preferred embodiments of the invention, other improvements to the invention are presented in common with reference to the accompanying drawings.

[0054] in:

[0055] Figure 1 An embodiment of a method 100 for reconstructing elevation angle information 2a based on measurement data 2 is shown;

[0056] Figure 2 An illustration shows multiple two-dimensional distributions 6g-6j of the parameters to be reconstructed aggregated into a three-dimensional distribution 6b;

[0057] Figure 3 An illustration showing the region-based processing of the two-dimensional distribution 3 of the measured parameters is shown;

[0058] Figure 4 An embodiment of a method 200 for training generator 4 is shown;

[0059] Figure 5 An embodiment of method 300 with a complete chain of behaviors is shown. Detailed Implementation

[0060] Figure 1 This is a schematic flowchart of a method 100 for reconstructing elevation information 2a based on measurement data 2, which is recorded using at least one radar device 1. In step 110, measurement data 2 of a spatial distribution 3 containing at least one physical measurement parameter is fed as input parameter 5 to at least one generator module 4. In step 120, at least one output parameter 6 is retrieved from the generator module 4, which is a measure of the elevation angle 6a at which radar radiation is reflected to radar device 1 by at least one object. In other words, at least one output parameter 6 contains elevation information 2a. According to block 121, the output parameter 6 of the generator module 4, particularly regarding position 3a in the spatial distribution 3, can indicate the specific elevation angle 6a at which radar radiation contributing to the value of the physical measurement parameter at the corresponding position 3a is reflected to radar device 1.

[0061] In step 130, the elevation angle information 2a is further analyzed in order to...

[0062] • Determine at least one category 7a of a pre-given classification for traffic signs, other traffic members, lane markings, signaling devices, or other traffic-related objects in the environment surrounding the vehicle; and / or

[0063] • Determine at least the position 7b, spatial location, size 7c, and / or velocity 7d of such an object.

[0064] Here, step 105 can be followed to select measurement data 2, which is recorded by radar device 1 whose antenna array cannot achieve direct elevation angle measurement. Step 106 can be followed to select measurement data 2, which is recorded by a moving synthetic aperture radar device.

[0065] According to block 111, generator module 4 may include: encoder 41, which transforms input parameter 5 into low-dimensional latent variables 5*; and decoder 42, which transforms these latent variables 5* into output parameters 6 with increased dimension.

[0066] According to box 112, the two-dimensional spatial distribution 3 of the physical measurement parameters used as input parameter 5 can be divided into multiple interconnected regions 3c-3f. Next, according to box 123, the input parameter 5 can be processed into corresponding output parameters 6c-6f according to the region divisions 5c-5f, and these output parameters can then be aggregated according to box 124. In this way, a three-dimensional spatial distribution 6b of at least one sought physical measurement parameter is obtained.

[0067] Commonly, according to block 122, the output parameters of generator module 4 may include a three-dimensional spatial distribution of at least one of the sought physical measurement parameters. Here, according to block 125, generator module 4 may output multiple two-dimensional spatial distributions 6g-6j. Then, each of these distributions 6g-6j corresponds to the radar radiation portion from which the corresponding distribution 6g-6j was generated, and the elevation angle range from which it is reflected to radar device 1.

[0068] exist Figure 2 The diagram illustrates how this distribution 6g-6j can be aggregated into a three-dimensional distribution 6b for the sought measurement parameters. The input parameter 5 is first fed to a generator 4, which consists of an encoder 41 that reduces the dimension and a decoder 42 that increases the dimension again. The data exists as a latent variable 5* at the interface between the encoder 41 and the decoder 42.

[0069] Decoder 42 for different elevation angles Using coordinates—distance r and azimuth angle respectively. This outputs the two-dimensional distribution 6g-6j of the physical measurement parameters being sought. These two-dimensional distributions 6g-6j can be combined into a tensor in a certain organization, and this tensor describes the three-dimensional distribution 6b of the physical measurement parameters being sought.

[0070] Figure 3 The processing of input parameter 5 by region is explained. The two-dimensional spatial distribution 3 of the physical measurement parameters, which should be used as input parameter 5 and are embodied in the measurement data 2, is divided into regions. Figure 3 The diagram illustrates four regions 3c-3f. Following box 123, the input parameters 5c-5f belonging to each region 3c-3f are transformed into output parameters 6c-6f. The generator module 4 used for this purpose can be designed to be smaller than a generator module that transforms all existing input parameters 5 into output parameters 6 in one step.

[0071] According to box 124, the output parameters 6c-6f can be aggregated into a three-dimensional distribution 6b of the sought physical measurement parameters. For example, in Figure 3As illustrated in [the document], the output parameters 6c-6f provide information about completely different locations within the three-dimensional distribution 6b. Not all locations need to lie within a contiguous region at once; conclusions can be drawn about individual portions 6c-6f of the output parameters 6 at all these locations.

[0072] Figure 4 A schematic flowchart illustrating an embodiment of the method 200 for training generator module 4 is shown. In step 210, training measurement data 2* is acquired, which is recorded using radar device 1*. The antenna array of the radar device 1* enables direct elevation angle measurement. In particular, in step 205, a radar device 1* having multiple channels outside the plane defining its azimuth angle can be selected.

[0073] In step 220, a two-dimensional spatial distribution 3 of at least one physical measurement parameter 5, serving as input parameter 5 for generator module 4, is determined based on training measurement data 2*. In step 230, target output parameters 6* are also determined based on training measurement data 2*, which are measures of the elevation angle 6a at which radar radiation is reflected to radar device 1* by at least one object.

[0074] The input parameter 5 is processed into the output parameter 6 using generator module 4. In step 250, the generator parameters 4a, which characterize the features of generator module 4, are optimized with the goal that the output parameter 5 reproduces the target output parameter 6* well according to the pre-given generator cost function 4b. The result of this training is the training completion state 4a* of the generator parameters 4a.

[0075] exist Figure 4 In the example shown, the neural network of generator module 4 is extended into a conditional GAN, i.e., cGAN, by an additional discriminator module 8. Output parameter 6, target output parameter 6*, and input parameter 5 are additionally fed to discriminator module 8. In step 270, discriminator parameters 8a, which characterize the features of discriminator module 8, are optimized alternately with generator parameters 4a, with the goal that discriminator module 8 effectively distinguishes the output parameter 6 generated by generator module 4 from the target output parameter 6* according to the discriminator cost function 8b. As a result, this training provides a training completion state 8a* for discriminator parameters 8a and a training completion state 4a* for generator parameters 4a.

[0076] However, after the training is complete, in order to analyze the real measurement data 2, usually only the generator 4 is needed, whose characteristics are characterized by the parameters 4a* of its completed training.

[0077] According to box 211, the distance-relative velocity spectrum 211a of the corresponding training measurement data 2* can be determined for each channel. Next, according to box 221, the input parameter 5 for the generator module 4 can be determined based on the distance-relative velocity spectrum, and according to box 231, the target output parameter 6* can also be determined based on the distance-relative velocity spectrum.

[0078] Figure 5 This is a schematic flowchart of an embodiment of method 300 with a complete chain of actions. In step 310, utilizing... Figure 4 The method 200 shown is used to train the generator module 4, resulting in the parameters 4a of the generator module exhibiting their training completion state 4a*. In step 320, at least one radar device 1 installed on or in the vehicle 50 is used to record measurement data 2 from the surrounding environment of the vehicle 50.

[0079] In step 330, using in Figure 1 The method 100 shown reconstructs elevation angle information 2a based on measurement data 2. This elevation angle information is in the form of output parameters 6 of the generator module 4 used, for example, in the form of elevation angle 6a and / or a three-dimensional distribution 6b of the sought physical measurement parameters. Thus, a control signal 9 is formed in step 340. In step 350, the vehicle 50 is controlled using this control signal 9.

Claims

1. A method (100) for reconstructing elevation information (2a) based on measurement data (2), the measurement data being a two-dimensional spatial distribution (3) recorded using at least one radar device (1) and containing at least one physical measurement parameter, wherein the measurement data (2) is fed (110) as an input parameter (5) to at least one generator module (4) configured as a neural network, and wherein at least one output parameter (6) is retrieved (120) from the generator module (4), the output parameter being a measure of the elevation angle (6a) at which radar radiation is reflected to the radar device (1) by at least one object, wherein the output parameter (6) of the generator module (4) contains (122) a three-dimensional spatial distribution (6b) of at least one sought physical measurement parameter, wherein • Divide the two-dimensional spatial distribution (3) into (112) multiple interconnected regions (3c-3f); • For each input parameter (5c-5f) belonging to region (3c-3f), determine the output parameter (6c-6f) to which (123) belongs; and • Gather all determined output parameters (6c-6f) together (124) to obtain the sought three-dimensional spatial distribution (6b) of the at least one sought physical measurement parameter.

2. The method (100) according to claim 1, wherein the output parameter (6) of the generator module (4) with respect to the position (3a) in the two-dimensional spatial distribution (3) describes (121) the radar radiation that contributes to the value of the physical measurement parameter at the corresponding position (3a) is reflected to the radar device (1) at what elevation angle (6a).

3. The method (100) according to claim 1 or 2, wherein the generator module (4) comprises (111): • Encoder (41), which transforms the input parameter (5) into latent variables (5*) in a space whose dimension is less than the dimension of the space of the input parameter (5) and less than the dimension of the space of the output parameter (6); and • Decoder (42), which transforms the latent variable (5*) into the output parameter (6).

4. The method (100) according to claim 3, wherein the output parameter (6) of the generator module (4) includes (125) a plurality of two-dimensional spatial distributions (6g-6j) of at least one physical measurement parameter, wherein each of the distributions (6g-6j) corresponds to the radar radiation portion that has generated the corresponding distribution (6g-6j) so that it is reflected to the radar device (1) at an elevation angle range.

5. The method (100) according to claim 1 or 2, wherein (105) the following measurement data (2) is selected, said measurement data being recorded using a radar device (1) whose antenna array cannot achieve direct elevation angle measurement.

6. The method (100) according to claim 1 or 2, wherein (106) the following measurement data (2) is selected, the measurement data being recorded using a mobile synthetic aperture radar device.

7. The method (100) according to claim 1 or 2, wherein the analysis (130) is based on the output parameter (6) of the generator module: • At least one category (7a) of a pre-defined classification of traffic signs, other road users, lane markings, signaling devices, or other traffic-related objects in the vehicle's surrounding environment; and / or • At least the position (7b), spatial dimensions (7c), and / or velocity (7d) of such an object.

8. A method (200) for training a generator module (4) for use in the method (100) according to any one of claims 1 to 7, the method (200) comprising the following steps: • Acquire (210) the following training measurement data (2*), which is recorded using a radar device (1*) whose antenna array can achieve direct elevation angle measurement; • Determine (220) the two-dimensional spatial distribution (3) of at least one physical measurement parameter (5) for the generator module (4) based on the training measurement data (2*). • The target output parameter (6*) is also determined (230) based on the training measurement data (2*), which is a measure of the elevation angle (6a) at which radar radiation is reflected to the radar device (1*) by at least one object; • The input parameter (5) is processed (240) into an output parameter (6) using the generator module (4); • Optimize (250) the generator parameters (4a) that characterize the generator module (4) with the goal that the output parameter (6) reproduces the target output parameter (6*) well according to the pre-given generator cost function (4b).

9. The method (200) according to claim 8, wherein • The output parameter (6) and the target output parameter (6*) are additionally fed (260) to the discriminator module (8) constructed as another neural network; and • The discriminator parameter (8a), which characterizes the properties of the discriminator module (8), is optimized (270) alternately with the generator parameter (4a), with the goal that the discriminator module (8) can well distinguish the output parameter (6) generated by the generator module (4) from the target output parameter (6*) according to the discriminator cost function (8b).

10. The method (200) according to claim 8 or 9, wherein (205) is selected from the radar device (1*) having a plurality of channels outside the plane defining its azimuth angle.

11. The method (200) according to claim 10, wherein for each channel, a distance-relative velocity spectrum (211a) of the corresponding training measurement data (2*) is determined (211), and wherein not only the input parameter (5) of the generator module (4) but also the target output parameter (6*) are determined based on the distance-relative velocity spectrum (221, 231).

12. A method (300) comprising the following steps: • The generator module (4) is trained (310) using the method (200) according to any one of claims 8 to 11; • Using at least one radar device (1) installed on or in the vehicle (50), measurement data (2) from the surrounding environment of the vehicle (50) is recorded (320). • Based on the measurement data (2), the elevation angle information (2a) is reconstructed (330) using the method (100) according to any one of claims 1 to 7; • When using the elevation information (2a), a control signal (9) is generated (340); • The vehicle (50) is controlled (350) using the control signal (9).

13. A computer program product comprising a computer program, the computer program containing machine-readable instructions that, when implemented on one or more computers, cause the one or more computers to perform the method (100, 200, 300) according to any one of claims 1 to 12.

14. A machine-readable data carrier and / or downloadable product having a computer program containing machine-readable instructions that, when executed on one or more computers, cause the one or more computers to perform the method (100, 200, 300) according to any one of claims 1 to 12.

15. A computer equipped with a computer program product according to claim 13 and / or equipped with a machine-readable data carrier and / or download product according to claim 14.

Images