Apparatus, computer-implemented method and technical system for machine learning

Abstract

Apparatus, computer-implemented methods and technical systems for machine learning are provided wherein the method comprises providing a first model, in particular a neural network comprising weights, configured to map data of a first radar spectrum, in particular data from a region of interest of the first radar spectrum, to a first feature representing the data; providing data and physical properties of the data, in particular distance, azimuth, velocity or polarization indications of the transmitted and received radar signals on which the data is based; providing a first output configured to map the first feature to a prediction of a physical property, in particular a prediction of distance, a prediction of azimuth, a prediction of velocity or a prediction of an indication of polarization of the transmitted and received radar signal; mapping the data having the first model to a first feature; mapping the first feature with the first output to a prediction of the physical property; and learning the first model, in particular a learning weight, depending on a difference between the prediction of the physical property and the physical property.

Description

Technical Field

[0001] The present invention relates to an apparatus and a computer-implemented method for machine learning, as well as a technical system comprising the apparatus. Background Art

[0002] While there is a large amount of data available for machine learning, a limitation to machine learning is the small fraction of that data that is labeled, or the cost of labeling a large enough fraction of that data for successful machine learning. Summary of the invention

[0003] A computer-implemented method for machine learning comprises providing a first model, in particular a neural network including weights, configured to map data of a first radar spectrum, in particular data from a region of interest of the first radar spectrum, to a first feature representing the data; providing the data and at least one physical property of the data, in particular a distance, an azimuth, a speed and / or a polarization indication of a transmitted and received radar signal on which the data is based; providing a first output, the first output being configured to map the first feature to a prediction of at least one attribute, in particular a prediction of a distance, a prediction of an azimuth, a prediction of a speed and / or a prediction of a polarization indication of a transmitted and received radar signal; mapping the data with the first model to the first feature; mapping the first feature with the first output to a prediction of at least one attribute; and learning the first model, in particular learning weights, depending on the difference between the prediction of at least one physical property and the at least one physical property. The method exploits physical properties that are readily available and provided by a radar sensor. The physical properties include radar-specific information acquired through prior radar knowledge. The method learns a radar spectrum representation, i.e., a first feature, by feature extraction from unlabeled data. Predicting physical properties is a natural auxiliary task that is useful for feature extraction. Since these auxiliary tasks are realistic and highly relevant to understanding the radar spectrum (e.g., classification or detection), the newly learned first features are highly informative, which would otherwise be difficult or impossible to obtain.

[0004] The first output may include a first neural network having first weights, wherein the method includes learning the first weights depending on a difference between the prediction of the physical property and the physical property. This further improves the results of the machine learning.

[0005] The learning of the first model depending on the difference between the prediction of at least one physical property and the at least one physical property may comprise determining the difference between the prediction of the distance and the distance, determining the difference between the prediction of the azimuth and the azimuth, determining the difference between the prediction of the velocity and the velocity, and / or determining the difference between the prediction of the polarization indication and the polarization indication. The radar spectrum may comprise different physical properties. The use of one or more physical properties further improves the quality of the extracted features with respect to the corresponding physical properties.

[0006] The method may comprise learning a first model or a first output dependent on at least two differences.The use of different physical properties further improves the quality of the extracted features with respect to these physical properties.

[0007] Providing a first output may include providing the first output to include a branch for mapping the first feature to a prediction of a physical property, in particular a branch for mapping the first feature to a prediction of a distance, a branch for mapping the first feature to a prediction of an azimuth, a branch for mapping the first feature to a prediction of a velocity, and / or a branch for mapping the first feature to a prediction of a polarization indication. Machine learning may use one or more auxiliary tasks. Using separate branches for tasks further improves machine learning.

[0008] The method may include mapping data of a second radar spectrum, in particular data from an area of ​​interest of the second radar spectrum, to a second feature with a learned first model; providing a second output, the second output being configured to map the second feature to a prediction of a task, in particular a prediction of classification or a prediction of object detection; providing a reference for the prediction of the task, in particular a label indicating the ground truth of the classification or object detection; mapping the second feature with the second output to the prediction of the task; and training the learned first model, in particular updating the learned weights, depending on the difference between the prediction of the task and the reference. The first model has learned to extract features based on physical properties associated with data from the radar spectrum. The second output uses these features for the task. Further label-based training of the first model for the task results in further improvement of the first model. In addition, with respect to the prediction of the task, the amount of labeled data required to achieve the same confidence as achieved by training the first model without any physical property-based learning is reduced.

[0009] The method may include: providing a second model comprising the same architecture as the learned first model, in particular providing the second model with a learned neural network comprising learned weights; mapping data of the second radar spectrum, in particular data from a region of interest of the second radar spectrum, to a second feature with the second model; providing a second output, the second output being configured to map the second feature to a prediction of a task, in particular a prediction of classification or a prediction of object detection; providing a reference for the prediction of the task, in particular a label indicating the ground truth of the classification or the ground truth of the object detection; mapping the second feature with the second output to the prediction of the task; and training the second model, in particular updating the learned weights, depending on the difference between the prediction of the task and the reference. This means that the learned first model is transferred to the second model. The second output uses features generated by the transferred first model of the task.

[0010] The second output may include a second neural network having second weights, wherein the method includes learning the second weights depending on a difference between the prediction of the task and the reference.

[0011] The method may include capturing a radar spectrum with a radar system; determining a prediction of a task with a trained first model or a trained second model depending on the captured radar spectrum; and actuating a technical system depending on the prediction of the task. The trained first model or the trained second model may be used in the operation of the technical system.

[0012] The method may include learning a first model or a second model, in particular learning weights, with unlabeled data from a plurality of first radar spectra, in particular range-azimuth spectra, range-velocity spectra and / or range-polarization spectra; and training the learned first model with data from a plurality of second radar spectra, in particular range-azimuth spectra, range-velocity spectra and / or range-polarization spectra, wherein the data from the plurality of second radar spectra are labeled with corresponding references. Labeling radar spectra is expensive in terms of computational cost and time. Using unlabeled data and physical properties associated therewith improves feature extraction without expensive labeling. Fine-tuning the learned first model or second model with labeled data requires less expensive labeling.

[0013] A device for machine learning comprises at least one processor and at least one memory, wherein the at least one memory comprises instructions executable by the at least one processor and, when executed by the at least one processor, causes the device to perform the method. The device has advantages corresponding to those of the method.

[0014] A technical system may include the device and a radar system. The technical system has advantages corresponding to those of the device.

[0015] A computer program may include instructions executable by a computer and, when executed by a computer, causes the computer to perform the method. The computer program has advantages corresponding to those of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Further advantageous embodiments can be derived from the following description and the accompanying drawings. In the drawings:

[0017] Figure 1 The technical system is schematically depicted,

[0018] Figure 2 Schematically depicting a device used for machine learning,

[0019] Figure 3 A first model for machine learning is schematically depicted,

[0020] Figure 4 Schematically depicting the region of interest from the radar spectrum,

[0021] Figure 5 A flowchart depicting a method for machine learning. DETAILED DESCRIPTION

[0022] Figure 1 A technical system 100 is schematically depicted. The technical system 100 is a physical system. The technical system 100 may be a robot. In this example, the technical system 100 is a vehicle. The technical system 100 includes a radar system 102. The radar system 102 is configured to capture a radar spectrum. In this example, the radar system 102 is configured to capture a radar spectrum in a field of view 104.

[0023] Radar system 102 is configured to capture physical properties associated with a radar spectrum for the radar spectrum. The radar spectrum may include reflections of an object 106. In this example, radar system 102 is configured to capture a distance 108 of object 106 relative to a sensor of radar system 102. In this example, radar system 102 is configured to capture an azimuth 110 of object 106 relative to the radar sensor of radar system 102. In this example, radar system 102 is configured to capture a velocity 112 of object 106 relative to the radar sensor of radar system 102. In this example, radar system 102 is configured to capture an indication 114 of polarization of transmitted and received radar signals of object 106 relative to the radar sensor of radar system 102.

[0024] For radar signals transmitted and received with horizontal polarization, indication 114 may be HH. For radar signals transmitted with horizontal polarization and received with vertical polarization, indication 114 may be HV. For radar signals transmitted with vertical polarization and received with horizontal polarization, indication 114 may be VH. For radar signals transmitted and received with vertical polarization, indication 114 may be VV.

[0025] Figure 2 Schematically depicted is a device 200 for machine learning. The device 200 comprises at least one processor 202 and at least one memory 204.

[0026] The at least one processor 202 is configured to execute instructions, which, when executed by the at least one processor 202, enable the at least one processor 102 to perform a method for machine learning. The at least one memory 204 is configured to store instructions.

[0027] Figure 3 A first model 300 for machine learning is schematically depicted. The first model 300 is configured to map data 302 of the radar spectrum, in particular a region of interest, to features 304.

[0028] In the method, a first model 300 is learned. To learn the first model 300, a first output 306 is provided. The first output 306 is configured to map a feature 304 having a first branch 308 of the first output 306 to a prediction 310 of a distance. The first output 306 is configured to map a feature 304 having a second branch 312 of the first output 306 to a prediction 314 of an azimuth. The first output 306 is configured to map a feature 304 having a third branch 316 of the first output 306 to a prediction 318 of a velocity. The first output 306 is configured to map a feature 304 having a fourth branch 320 to a prediction of a polarization indication. In this example, the first prediction 322 is an indication of HH. In this example, the second prediction 324 is an indication of HV. In this example, the third prediction 326 is an indication of VH. In this example, the fourth prediction 328 is an indication of VV. This means that the fourth branch 320 is a single channel for polarization prediction.

[0029] exist Figure 3 In the example depicted in , the radar spectrum includes different regions of interest. One region of interest includes data 302, while the other regions of interest include data 330 associated with indication HH, data 332 associated with indication VH, and data 334 associated with indication VV, respectively. In this example, data 302 is associated with indication HV.

[0030] The data 302 mapped by the first model 300 is randomly selected in this example. In this example, the corresponding region of interest is randomly selected from the region of interest of the radar spectrum.

[0031] In machine learning, data for learning the first model 300 may be sampled from one radar spectrum or from different radar spectra. In one example, different regions of interest are sampled as data for learning the first model 300. The data of the radar spectrum may be associated with no physical attribute, one physical attribute, or multiple physical attributes.

[0032] Figure 4 An example of a region of interest from a radar spectrum is schematically depicted. More specifically, Figure 4 A region of interest is depicted that includes a first object 402 (eg, a car), a second object 404 (eg, a motorcycle), a third object 406 (eg, an architectural obstacle), a fourth object 408 (eg, a pedestrian), or a fifth object 410 (eg, a stop sign).

[0033] Figure 4 The region of interest at different distances relative to the radar system 102 is depicted, particularly the first range distance 412

[0034] (e.g., up to 15 m), a second range distance 414 (e.g., 15 m to 25 m), and a third range distance 416 (e.g., 25 m to 35 m). Figure 4 The region of interest depicted in is sampled from a range-azimuth spectrum or from several different range-azimuth spectra.

[0035] The radar spectrum or region of interest includes an indication of the strength of radar reflections. In this example, the strength of the reflections decreases from a first strength 418 (eg, 10 dB) to a second strength 420 (eg, -40 dB).

[0036] Figure 5 A flow chart of the method is depicted. The method is described by way of example with data 302 associated with polarization indication HV. Other radar spectrum data associated with at least one physical property are processed similarly.

[0037] The method includes step 502 .

[0038] In step 502 , a first model 300 is provided.

[0039] The first model 300 is configured to map data 302 to features 304 representing the data 302 .

[0040] The first model 300 is, for example, a neural network of an architecture for mapping data 302 to first features 306. The neural network includes weights. The weights can be randomly initialized, for example.

[0041] The method includes step 504 .

[0042] In step 504, a first output 306 is provided.

[0043] In this example, the first output 306 includes at least one branch for mapping the feature 304 to a prediction of a physical property. In this example, the first output 306 includes one branch 308 for mapping the feature 304 to a prediction 310 of a range, one branch 312 for mapping the feature 304 to a prediction 314 of an azimuth, one branch 316 for mapping the feature 304 to a prediction 318 of a velocity, and one branch 320 for mapping the feature 304 to a prediction of an indication of polarization.

[0044] The first output 306 may include more or fewer branches.

[0045] The method includes step 506 .

[0046] In step 506 , data 302 and a polarization indication HV of the data 302 are provided.

[0047] The data 302 is associated with the polarization indication HV. For example, the data 302 is associated with the indication HV in metadata of the radar spectrum. The metadata may be added to the radar system 102 when the radar spectrum is acquired.

[0048] The method includes step 508 .

[0049] In step 508 , data 302 is mapped to features 304 using first model 300 .

[0050] The method includes step 510 .

[0051] In step 510 , the first output 306 is used to map the features 304 to a prediction of a polarization indication.

[0052] The method includes step 512 .

[0053] In step 512 , the first model 300 is learned based on the features 304 extracted by the first model 300 from the data 302 .

[0054] In this example, the first model 300 is learned depending on the difference between the prediction of the polarization indication by the first output 306 and the polarization indication HV associated with the data 302 .

[0055] The method is not limited to data 302 and physical attributes that are polarization indicators HV. The physical attribute may be an indication of range 108, azimuth 110, velocity 112, or one of the other possible polarizations 114 of the transmitted and received radar signals on which the data from the radar spectrum used for machine learning is based.

[0056] The first model 300 may be learned depending on a difference between a prediction of a physical property and a physical property associated with data from radar spectrum or depending on a difference between a prediction of a physical property and a corresponding physical property associated with data from radar spectrum.

[0057] This means that the first model 300 may be learned depending on the difference between a prediction of a distance and the distance, the difference between a prediction of an azimuth and the azimuth, the difference between a prediction of a speed and the speed, and / or the difference between a prediction of a polarization indication and the polarization indication.

[0058] In this example, the weights are learned depending on the difference between the prediction of the physical property and the physical property.

[0059] The method may include determining weights of the neural network of the first model 300 that reduce or minimize the loss including the discrepancy.

[0060] The method may include depending on at least two of the differences learning the first model 300 or the first output 306. For example, the differences are attributed to losses.

[0061] According to an example, the first output 306 includes a first neural network having first weights. The method may include learning the first weights depending on a difference between a prediction of the physical property and the physical property.

[0062] The method may include determining a first weight that reduces or minimizes a loss including the difference or differences.

[0063] Steps 502 to 512 may be repeated for different data from different radar spectra associated with one physical property or several physical properties. The first model 512 may be learned depending on the difference between the prediction of the physical property and the corresponding physical property.

[0064] The first example of a model 300 is constructed using a set of U unlabeled regions of interest. Data x i Trained feature extractor model M.

[0065] In this example, x i Indicates a region of interest from a radar spectrum. The learning of the feature extractor model M is based on the radar spectrum data x in the region of interest. i and physical information {r i , Θ i , v i , p i}, where r i is the distance, Θ i is the azimuth, v i is the speed, and p i is the same as the data x i The associated polarization indication.

[0066] The feature extractor model M is, for example, configured to output features 304:

[0067]

[0068] Where h indicates the characteristic f i In some examples, h is greater than 5,000 or greater than 10,000 or greater than 100,000.

[0069] The first output 306 includes, for example, a branch 308 where r′ i is the prediction of distance 310.

[0070] The first output 306 includes, for example, a branch 312 where θ′ i is the prediction 314 of the azimuth.

[0071] The first output 306 includes, for example, a branch 316 where v′ i It is the prediction of speed 318.

[0072] The first output 306 includes, for example, a branch 320 where p′ i is the predicted one-hot encoded vector of polarization indications.

[0073] A r (f i ), A az (f i ), A v (f i ), A p (f i ) is, for example, part of a first neural network.

[0074] The first model 300 is learned, for example, using the following loss:

[0075]

[0076] where α r , α az , α v , α p is the weight of each term. The weights can be equal or chosen to control the amount of influence the physical property has in learning. The choice of weights can be determined using a validation set through a hyperparameter search on the performance of the four physical properties or downstream predictions on the task. According to some examples, the weight α r and α az Greater than weight α v and α p , because the position of an object in the field of view relative to the radar sensor has the greatest impact on the radar spectrum.

[0077] In this example, the weights are updated via back-propagation.

[0078] The loss can contain these four terms or a subset of these four terms. Different loss functions can be used.

[0079] The weights of the model M can be learned using an optimizer (e.g., Adam or SGD optimizer) using various training techniques (batch normalization, dropout, L2 regularization). Training can continue until the desired number of epochs is reached, or the improvements on the rejected validation set begin to saturate (i.e., early stopping).

[0080] This means that the method comprises learning a first model 300, in particular weights, using unlabeled data from a plurality of radar spectra. The radar spectra may be range-azimuth spectra, range-velocity spectra and / or range-polarization spectra.

[0081] The method includes step 514 .

[0082] In step 514 , a second output is provided that is configured to map the features 304 determined by the first model 300 to a prediction of the task.

[0083] The task is, for example, classification or object detection.

[0084] The second output may include a second neural network having second weights. The second neural network has an architecture for determining a prediction of a task based on the first feature 304.

[0085] The method includes step 516 .

[0086] Step 516 includes providing a reference for prediction of the task.

[0087] The reference is, for example, a label indicating the ground truth for classification or object detection.

[0088] The reference for classification can be the indicator data x i The one-hot encoded vector y of the ground truth class i ∈[0, 1] C .

[0089] The method includes step 518 .

[0090] In step 518 , the radar spectrum data is mapped to features 304 representing the radar spectrum data using the learned first model 300 .

[0091] In this example, a learned first model 300 is utilized to map data from a region of interest of the radar spectrum.

[0092] The method includes step 520 .

[0093] In step 520, the second output is used to map the second feature to a prediction of a task, such as classification or object detection.

[0094] The method includes step 522 .

[0095] In step 522 , the learned first model 300 is trained depending on the difference between the prediction of the task and the reference.

[0096] For example, the learned weights of the first model 300 are updated depending on the difference between the prediction of the task and the reference.

[0097] According to an example, the second output is configured to output a classification probability vector y′ for classifying C objects i =[q1, q2, ..., q c ]∈[0,1] C As the prediction of the task, where q1, q2, ..., qc Indicates the probability of the corresponding class.

[0098] The method is not limited to using the first model 300 for learning and training.

[0099] The method may include providing a second model comprising the same architecture as the learned first model 300. For example, the second model is provided with a learned neural network comprising the learned weights of the first model 300.

[0100] The method may include mapping data of the second radar spectrum, in particular data from a region of interest of the second radar spectrum, to features representative of the data of the second radar spectrum using a second model 304 .

[0101] The method may include providing a second output configured to map the second feature to a prediction of the task.

[0102] The method may include mapping a second feature having a second output to a prediction of the task. The method may include training a second model, in particular updating learned weights in the second model, depending on a difference between the prediction of the task and a reference.

[0103] The second output may include the same branches as the first output 306 .

[0104] Steps 516 to 522 may be repeated to train the learned first model, in particular the extractor model M, or the second model using: labeled data from predictions for determining the task or for determining the prediction of the task and for predicting the corresponding physical properties Data x i The second radar spectrum may be a range-azimuth spectrum, a range-velocity spectrum and / or a range-polarization spectrum. Data from the plurality of second radar spectra are labeled with corresponding references.

[0105] The method may include learning a second weight dependent on a difference between the prediction of the task and a reference.

[0106] The training of the first model 300 or the second model, in particular the training of the weights of the corresponding neural network, can use an optimizer (e.g., Adam or SGD optimizer) that uses various training techniques (batch normalization, dropout, L2 regularization). Training can continue until the desired number of epochs is reached, or the improvement on the rejected validation set begins to saturate (i.e., early stopping).

[0107] The second output may include a single channel for predicting polarization. During inference, the second output may be used to predict four polarizations in respective channels.

[0108] The method may include determining an average softmax output probability vector as a final probability vector for the object type based on the polarization prediction of the second output prediction.

[0109] According to some embodiments, the method includes step 524 .

[0110] Step 524 includes capturing a radar spectrum with radar system 102 .

[0111] According to some embodiments, the method includes step 526 .

[0112] Step 526 includes determining a prediction for the task using either the trained first model or the trained second model depending on the captured radar spectrum.

[0113] According to some embodiments, the method includes step 528 .

[0114] Step 528 includes actuating the technical system 100 depending on the prediction of the task.

[0115] For a vehicle or robot, the method may include determining whether automatic braking or automatic steering actuation is required to achieve a vehicle response to a classified or detected object depending on the prediction of the task. The response may be a stop of the vehicle in front of the object or a trajectory to avoid the location of the object.

[0116] Steps 524 to 528 may be repeated, for example, during operation of the technical system 100 . Learning or training of the first model 300 or the second model may be performed during operation of the technical system 100 , in particular based on radar spectra captured during operation of the technical system 100 .

Claims

1. A computer-implemented method for machine learning, characterized in that The method comprises providing (502) a first model (300), in particular a neural network comprising weights, configured to map data (302) of a first radar spectrum, in particular data from a region of interest of the first radar spectrum, to a first feature (304) representing the data (302); providing (506) the data (302) and a physical property of the data, in particular a distance (108), an azimuth (110), a speed (112) or an indication of polarization (114) of transmitted and received radar signals on which the data is based; providing (504) a first output (306) configured to map the first feature (304) to a first feature (304) representing the data (302); 04) to a prediction of a physical property, in particular a prediction of a distance (310), a prediction of an azimuth (314), a prediction of a speed (316) or a prediction (322, ..., 328) of a polarization indication of a transmitted and received radar signal; mapping (508) data (302) having a first model (300) to a first feature (304); mapping (510) the first feature (304) having a first output (306) to a prediction of a physical property; and learning (512) the first model (300), in particular learning weights, depending on a difference between the prediction of the physical property and the physical property.

2. The method according to claim 1, characterized in that The first output (306) includes a first neural network having first weights, wherein the method includes learning (512) the first weights depending on a difference between a prediction of a physical property and the physical property.

3. The method according to one of the preceding claims, characterized in that Learning (512) the first model (300) depending on a difference between a prediction of the at least one physical property and the at least one physical property includes: determining a difference between a prediction of a distance and the distance, determining a difference between a prediction of an azimuth and the azimuth, determining a difference between a prediction of a velocity and the velocity, and / or determining a difference between a prediction of a polarization indication and the polarization indication.

4. The method according to claim 3, characterized in that The method includes learning (512) a first model or a first output depending on at least two differences.

5. The method according to one of the preceding claims, characterized in that Providing a first output (502) includes providing the first output to include a branch for mapping the first feature to a prediction of a physical property, in particular a branch for mapping the first feature to a prediction of a distance, a branch for mapping the first feature to a prediction of an azimuth, a branch for mapping the first feature to a prediction of a velocity, and / or a branch for mapping the first feature to a prediction of a polarization indication.

6. The method according to one of the preceding claims, characterized in that The method includes mapping (518) data of a second radar spectrum, in particular data from a region of interest of the second radar spectrum, to a second feature using a learned first model; providing (514) a second output, the second output being configured to map the second feature to a prediction of a task, in particular a prediction of classification or a prediction of object detection; providing (516) a reference for the prediction of the task, in particular a label indicating a ground truth value of the classification or object detection; mapping (520) the second feature having the second output to the prediction of the task; And training (522) the learned first model, in particular updating the learned weights, depending on the difference between the prediction of the task and the reference.

7. The method according to one of claims 1 to 5, characterized in that The method comprises: providing a second model comprising the same architecture as the learned first model, in particular providing the second model with a learned neural network comprising learned weights; mapping data of a second radar spectrum, in particular data from a region of interest of the second radar spectrum, to a second feature with the second model; providing a second output, the second output being configured to map the second feature to a prediction of a task, in particular a prediction of classification or a prediction of object detection; providing a reference for the prediction of the task, in particular a label indicating a ground truth value of the classification or a ground truth value of the object detection; mapping the second feature with the second output to the prediction of the task; and training the second model, in particular updating the learned weights, depending on the difference between the prediction of the task and the reference.

8. The method according to claim 6 or 7, characterized in that: The second output comprises a second neural network having second weights, wherein the method comprises learning (522) the second weights depending on a difference between the prediction of the task and the reference.

9. The method according to claim 6, characterized in that The method includes capturing (524) a radar spectrum with a radar system (102), determining (526) a prediction of a task based on the captured radar spectrum with a trained first model or a trained second model, and actuating (528) a technical system (100) based on the prediction of the task.

10. The method according to one of the preceding claims, characterized in that The method comprises learning (512) a first model, in particular learning weights, using unlabeled data from a plurality of first radar spectra, in particular range-azimuth spectra, range-velocity spectra and / or range-polarization spectra, and training (522) the learned first model or the second model using data from a plurality of second radar spectra, in particular range-azimuth spectra, range-velocity spectra and / or range-polarization spectra, wherein the data from the plurality of second radar spectra are labeled with corresponding references.

11. A device (200) for machine learning, characterized in that: The device comprises at least one processor (202) and at least one memory (204), wherein the at least one memory (204) comprises instructions executable by the at least one processor (202), and when executed by the at least one processor (202), the instructions cause the device to perform a method according to one of claims 1 to 10.

12. A technical system (100), characterized in that The technical system (100) comprises a device (200) according to claim 11 and a radar system (102).

13. A computer program, characterized in that The computer program comprises instructions executable by a computer and, when executed by a computer, the instructions cause the computer to perform a method according to one of claims 1 to 10 .

Images