Method for determining target object extent data from raw radar data using at least one machine learning model

Abstract

A method for determining target object extent data (40), which characterizes at least one spatial extent of at least one target object detected in at least one radar measurement with at least one radar system, a method for determining at least one point cloud (34), a method for determining a model (52) by machine learning, a radar system, and a model (52) are described. In the method for determining target object extent data (40), at least one model (52) determined by machine learning is executed. At least a portion of the raw radar data (50) acquired in at least one radar measurement is fed to at least one model (52) determined by machine learning, and corresponding target object extent data (40) are determined by the at least one model (52) based on the fed radar raw data (50).

Description

Technical field

[0001] The invention relates to a method for determining target object extent data, which characterize at least a spatial extent of at least one target object detected in at least one radar measurement with at least one radar system, in which at least one model determined by machine learning is executed.

[0002] Furthermore, the invention relates to a method for determining at least one point cloud which contains target object data of at least one target object detected by at least one radar system, in which at least a part of the raw radar data determined by at least one radar measurement with at least one radar system is processed into data of a point cloud.

[0003] Furthermore, the invention relates to a method for determining a model by machine learning, wherein the model is configured to assign raw radar data obtained in at least one radar measurement with at least one radar system to target object extent data, wherein the target object extent data characterize at least one spatial extent of at least one target object detected in the at least one radar measurement, wherein in the method a method for determining target object extent data is carried out for at least one target object and at least a part of the target object extent data obtained thereby is compared with basic truth data for the spatial extent of the at least one target object and the model is adapted accordingly.

[0004] Furthermore, the invention relates to a radar system, in particular a radar system for a vehicle, with at least one antenna system for transmitting radar signals and receiving radar echo signals, and with at least one evaluation device for processing raw radar data obtained during radar measurements, at least to target object extent data, which characterize spatial extents of target objects detected during radar measurements.

[0005] Furthermore, the invention relates to a model generated by machine learning and designed to determine target object data that characterizes at least one target object detected during radar measurements, based on raw radar data obtained during radar measurements. State of the art

[0006] From CN 113409381 A, a CNN-based dual-polarization-channel fusion ship target estimation method from the technical field of radar equipment is known, comprising the following steps: first, the input of a training pattern into a ship target estimation frame of a convolutional neural network CNN for training in order to obtain a trained convolutional neural network frame, and the input of a test pattern into the trained convolutional neural network frame to obtain a ship target size estimation result; the method can improve the performance of target estimation under the condition of medium size and low resolution, thereby improving the precision of the target estimation; the information of the polarized channel can be merged with two sizes of the SAR ship image and can be used for SAR image interpretation and the identification and classification of the SAR image ship target.

[0007] The invention is based on the objective of designing a method for determining target object extent data, a method for determining at least one point cloud, a method for determining a model, a radar system and a model with which target object extent data for target objects can be determined better, in particular faster, with less effort and / or more accurately. Disclosure of the invention

[0008] The object of the invention is achieved in the method for determining target object extent data by feeding at least a part of the raw radar data obtained in at least one radar measurement to at least one model determined by machine learning and using the at least one model to determine corresponding target object extent data on the basis of the supplied raw radar data.

[0009] According to the invention, the corresponding target object extent data for the at least one detected target object are determined directly from the raw radar data using a model. The model allows the target object extent data to be determined from the phase and magnitude patterns of the raw radar data. Prior conversion of the raw radar data is not required. This has the advantage that all information from the raw radar data can be used to determine the target object extent data. Information can be lost during prior conversion of the raw radar data, particularly to data points of a point cloud.

[0010] The model is determined through machine learning. This allows the model to be adapted to recognize differences in the phase and magnitude patterns of the raw radar data and to derive corresponding target object extent data from them. In particular, information about the spatial extent of the detected target objects is extracted from the raw radar data.

[0011] Advantageously, the model can be a mathematical model. Mathematical models allow for the implementation of mapping functions that can assign corresponding target object extent data to the raw radar data.

[0012] Radar raw data is data output after radar measurements and, if necessary, after analog-to-digital conversion. The radar raw data still contains all the information from the radar measurements. During further processing of the radar raw data, particularly into point cloud data, some of the information from the radar raw data can be lost. For example, using pattern recognition, such as that performed with a model according to the invention, corresponding target object extent data can be determined more quickly and accurately from the information in the radar raw data, especially from phase and magnitude patterns, than is possible from the data in a point cloud.

[0013] A target object is an object that is detected by the radar system. Target objects can be, in particular, stationary or moving objects, especially vehicles, people, animals, plants, obstacles, road surface irregularities, especially potholes or stones, lane boundaries, (road) markings, (traffic) signs, open spaces, especially parking spaces, precipitation, or the like.

[0014] Target object data is data that characterizes a target object detected by radar measurements. This target object data can include information about the distances, directions, and / or velocities of the detected target object relative to the radar system or a corresponding reference system, such as a reference point or a coordinate system. Furthermore, the target object data can also include target object extent data, which characterizes the spatial extent of the target object.

[0015] A target object's extent can be either horizontal or vertical. This method can determine both horizontal and vertical target object extents. When used in conjunction with a vehicle radar system, the method can determine the vertical height and / or horizontal width of target objects such as vehicles.

[0016] Advantageously, the target object's extent data can characterize its spatial angular extent. This spatial angular extent is described by the target object's opening angle. Therefore, knowing the target's distance from the radar system is not necessary to characterize its spatial extent. The absolute extent of the detected target can be determined, and in particular calculated, from the target object's extent data in conjunction with distance data, which can also be acquired by the radar system. This allows for the characterization of a detected target based on absolute extent data. From the absolute extent of the target, the type of vehicle can be determined.In particular, a distinction can be made between two-track vehicles, such as passenger cars, trucks or the like, and single-track vehicles, such as two wheels.

[0017] The invention can be used in conjunction with radar systems for vehicles. A key functional characteristic of a vehicle is its ability to move. Vehicles can be motor vehicles. Advantageously, the invention can be used in land vehicles, in particular cars, trucks, buses, motorcycles, drones, mobile robots, mobile machinery, in particular construction or transport machinery such as cranes, excavators, or the like, aircraft, in particular drones, and / or (under)water vehicles, in particular (under)water drones. The invention can also be used in vehicles that can be operated autonomously or semi-autonomously. However, the invention is not limited to vehicles. It can also be used in stationary operation.

[0018] Advantageously, at least one radar system can be a vehicle-mounted radar system. Vehicle-mounted radar systems can monitor areas inside or around the vehicle.

[0019] In an advantageous embodiment of the method, at least one model can be implemented using at least one artificial neural network. An artificial neural network can be trained efficiently. Furthermore, artificial neural networks can very efficiently distinguish phase patterns and magnitude patterns from raw radar data and assign them to corresponding target object extent data. Artificial neural networks are particularly fast at pattern recognition. Furthermore, artificial neural networks can be designed for optimized use of processor power. This can reduce the required processor power. For example, the method for determining the target object extent data can be executed in parallel or overlapping time with other methods, particularly those for determining target object distance and / or direction data.

[0020] In a further advantageous embodiment of the method, at least a portion of the target object extent data can be assigned to at least one point cloud that describes the at least one target object detected by the at least one radar system. In this way, the target object extent data can be combined with other target object data, such as target object distance data and / or target object direction data. The at least one point cloud can be used to describe the monitoring area detected by the at least one radar system, in particular target objects within that monitoring area. Advantageously, the at least one point cloud can be used by a driver assistance system for the autonomous or semi-autonomous operation of vehicle functions.

[0021] A point cloud is a set of points, particularly in Cartesian or spherical coordinates, representing the detections of the radar system. These points represent areas of the detected scene within the surveillance area. The points can contain information about their velocity and 3D coordinates relative to the radar system, especially directional and distance coordinates.

[0022] In a further advantageous embodiment of the method, raw radar data can be used, which are obtained from radar measurements with a multiple multiple-input multiple-output radar system, a frequency-modulated continuous-wave radar system or a MIMO-FMCW radar system.

[0023] Advantageously, the radar system used to acquire the raw radar data can be designed as a multiple-input multiple-output (MDI) radar system. A MDI radar system has multiple transmitting antennas and multiple receiving antennas. It offers increased spatial resolution. In a MDI radar system, the raw radar data can consist of detections from complex channels assigned to the transmitting and receiving antennas.

[0024] Alternatively or additionally, the radar system used to acquire the raw radar data can advantageously be designed as a frequency-modulated continuous-wave radar system, a so-called FMCW radar system. With a frequency-modulated continuous-wave radar system, radar signals are continuously emitted, the frequency of which is modulated. In this way, the measurement function of the radar system can be extended.

[0025] Advantageously, the raw radar data can be acquired using a MIMO-FMCW radar system. A MIMO-FMCW radar system has multiple transmitting and receiving antennas and uses frequency-modulated continuous-wave radar signals. Thus, the MIMO-FMCW radar system combines the advantages of a frequency-modulated continuous-wave radar system with those of a multiple-input multiple-output radar system. A MIMO-FMCW radar system is therefore both a multiple-input multiple-output radar system and a frequency-modulated continuous-wave radar system.

[0026] In MIMO or MIMO-FMCW radar systems, the radar scene can be sparse in the sense that only a small portion of the point cloud cells, particularly range-direction cells, range-Doppler cells, range-Doppler-direction cells, or the like, are occupied by target objects of interest. The point cloud cells derived from raw radar data from a MIMO or MIMO-FMCW radar system typically do not contain information about the target object's extent. Using the model according to the invention, the target object extent of isolated target objects can also be determined from raw radar data acquired with a MIMO or MIMO-FMCW radar system.

[0027] In a further advantageous embodiment of the method, target object extent data for isolated target objects can be determined. In this way, each target object can be individually characterized by its target object extent data. This allows, in particular, differentiation between vehicles of different types and sizes, especially between two-track and single-track vehicles.

[0028] Advantageously, target object extent data for multiple target objects detected by radar measurements within the monitored area can be determined separately. This allows for the differentiation of target objects of varying sizes within the monitored area.

[0029] In a further advantageous embodiment of the method, digital radar raw data can be used instead of radar raw data. In this way, the radar raw data can be digitally processed with the model.

[0030] The raw digital radar data can be determined from received echo signals using analog-to-digital conversions.

[0031] Furthermore, the object of the invention is achieved with the method for determining at least one point cloud by carrying out a method according to the invention for determining target object extent data in the method for determining at least one point cloud and by adding at least a part of the target object extent data determined thereby to the data of the at least one point cloud.

[0032] According to the invention, the at least one point cloud is supplemented with the target object extent data, which are directly determined from the raw radar data using the inventive method for determining the target object extent data. In this way, the information about target objects contained in the at least one point cloud can be further expanded.

[0033] In an advantageous embodiment of the method, at least some of the raw radar data can be processed into target direction data that characterize the reception directions from which radar echo signals received by the at least one radar system originate, and at least some of the target direction data can be added to the data of the at least one point cloud. and / or At least a portion of the raw radar data can be processed into target range data that characterizes the distances to targets detected by the at least one radar system, and at least a portion of the target range data can be added to the data of the at least one point cloud. In this way, the at least one point cloud can be augmented with further data that characterizes the direction and / or distance of the detected targets.

[0034] The direction from which the radar echo signals are received can correspond to the direction in which the target of a target object is located, from which the radar signals are reflected.

[0035] In a further advantageous embodiment of the method, target direction data and / or target distance data and / or target extent data can be determined from at least a portion of the raw radar data in a temporally overlapping manner. This allows the overall generation of the at least one point cloud to be accelerated. The determination of the target direction data, the target distance data, and / or the target extent data can be performed at least partially in parallel.

[0036] In a further advantageous design of the procedure Target direction data and / or target distance data can be determined from at least a portion of the raw radar data by applying at least one model determined by machine learning. and / or Target object direction data and / or target object distance data can be generated from at least part of the raw radar data using signal processing.

[0037] Advantageously, target direction data and / or target distance data can be determined using at least one model. This at least one model can be generated by machine learning. In this way, the target direction data and / or the target distance data can be efficiently and quickly derived from the raw radar data.

[0038] Alternatively or additionally, target direction data and / or target distance data can be generated from at least a portion of the raw radar data using signal processing. This allows the use of existing components designed for signal processing. Preparation through machine learning, as required when using a model, is therefore unnecessary.

[0039] Furthermore, the object of the invention is solved in the method for determining a model by carrying out a method according to the invention for determining target object extent data for at least one isolated target object.

[0040] According to the invention, the model used to determine target object extent data according to the inventive method is trained using raw radar data acquired from isolated target objects. For this purpose, the baseline data for at least one isolated target object are determined. The raw radar data are then fed into the model. The target object extent data determined by the model are compared with the baseline data, thus validating the model.

[0041] Ground truth data, also known as "ground truth data" in the context of machine learning, is data that allows for the verification of model quality. The required outcome of ground truth data is well understood. For the method according to the invention, the actual dimensions of the respective target objects can be used as ground truth data. In particular, the ground truth data can also be determined beforehand using other measurement methods.

[0042] In an advantageous embodiment of the method, at least a portion of the raw radar data can be normalized, and then the model can be used to assign corresponding target object extent data to the normalized raw radar data. This allows for further improvement of the model's training.

[0043] Furthermore, the problem is solved according to the invention in the radar system by the fact that the at least one evaluation device is designed to carry out the method according to the invention for determining target object extent data.

[0044] According to the invention, at least one evaluation device is designed to carry out the method according to the invention. In this way, the corresponding target object extent data can already be obtained from raw radar data using means of the radar system.

[0045] In an advantageous embodiment, at least one evaluation unit, at least one machine learning-generated model, in particular an artificial neural network, designed to determine target object extent data from raw radar data, and / or at least one evaluation unit must be designed to process raw radar data into target direction data and / or target distance data.

[0046] Advantageously, at least one evaluation unit can include at least one model. The model can be used to derive target object extent data from raw radar data, in particular from phase patterns and magnitude patterns of the raw radar data.

[0047] Alternatively or additionally, at least one evaluation unit can be configured to process raw radar data into target direction data and / or target range data. In this way, the radar system can additionally determine target direction data and / or target range data from the raw radar data. The target direction data can characterize the directions of targets detected by the radar system, and the target range data can characterize the distances of detected targets.

[0048] In a further advantageous embodiment, the radar system can include at least part of a learning algorithm for training at least one model, in particular at least one artificial neural network, which can be used in the method according to the invention. In this way, the model used can be trained with means that are already contained in the radar system. Thus, the model can also be further trained during the regular operation of the radar system.

[0049] Furthermore, the problem is solved according to the invention in the model by the fact that the model is designed for use in the inventive method for determining target object extent data.

[0050] Advantageously, the model can be an artificial neural network. An artificial neural network allows for the efficient comparison of patterns, particularly phase and magnitude patterns, in raw radar data.

[0051] Advantageously, the model can be trained using multi-layered learning (deep learning). This allows even complex patterns of raw radar data to be analyzed.

[0052] Advantageously, the model can incorporate an artificial neural network using a transformer architecture. This allows for further improvement of the model.

[0053] Advantageously, the model can be integrated into the radar system. In this way, the method according to the invention can be carried out using the radar system.

[0054] Advantageously, the model can be implemented using software and / or hardware. Advantageously, the model can be implemented using software. Software can be easily integrated into a processor of the radar system.

[0055] Furthermore, the features and advantages identified in connection with the inventive method for determining target object extent data, the inventive method for determining at least one point cloud, the inventive method for determining a model by means of machine learning, the inventive radar system, and the inventive model and their respective advantageous embodiments apply to each other accordingly, and vice versa. The individual features and advantages can, of course, be combined with one another, potentially resulting in further advantageous effects that go beyond the sum of the individual effects. Brief description of the drawings

[0056] Further advantages, features, and details of the invention will become apparent from the following description, in which exemplary embodiments of the invention are explained in more detail with reference to the drawing. The person skilled in the art will expediently consider the features disclosed in the drawing, the description, and the claims individually and combine them into meaningful further combinations. The drawing schematically illustrates Fig. 1 a vehicle with a driver assistance system comprising a radar system and a control unit; Fig. 2 a functional diagram of the driver assistance system with the radar system and the control unit Fig. 1; Fig. 3. A flowchart of a procedure for determining a point cloud from raw radar data, which is obtained from the radar system during radar measurements. Fig. 1 and Fig. 2. can be determined using a model determined by machine learning; Fig. 4. A method for determining the model used in the procedure from the Fig. 3 is used.

[0057] In the figures, identical components are labelled with the same reference symbols. embodiment(s) of the invention

[0058] In Fig. Figure 1 shows a vehicle 10. The vehicle 10 includes a driver assistance system 12 with a radar system 14 and a control unit 16.

[0059] The driver assistance system 12 allows functions of the vehicle 10, such as driving functions, to be controlled autonomously or partially autonomously.

[0060] The radar system 14 can monitor a surveillance area 18 for target objects 20. The radar system 14 is, for example, located at the front of the vehicle 10 and directed in the direction of travel. In this way, the surveillance area 18 in front of the vehicle 10 can be monitored. Alternatively, the radar system 14 can also be located elsewhere on the vehicle 10 and oriented differently. Multiple radar systems 14 can also be provided.

[0061] Information obtained by the radar system 14 over the monitoring area 18, for example, information on target objects 20 detected by the radar system 14, can be transmitted from the radar system 14 to the control unit 16. The control unit 16 can then control the functions of the vehicle 10 based on the information obtained by the radar system 14.

[0062] Fig. Figure 2 shows a functional representation of the driver assistance system 12 with the radar system 14 and the control unit 16.

[0063] Radar system 14 is designed as an exemplary MIMO-FMCW radar system.

[0064] The radar system 14 comprises an antenna system 24 and a control and evaluation unit 26. The antenna system 24 includes several receiving antennas 30 and several transmitting antennas 28 (multiple-in-multiple-out). This is the origin of the designation as a MIMO radar system. By way of example, the antenna system 24 has two transmitting antennas 28 and two receiving antennas 30. However, the invention is not limited to two transmitting antennas and two receiving antennas. The antenna system 24 can also include more or fewer transmitting and receiving antennas.

[0065] By combining the transmitting antennas 28 and receiving antennas 30 during operation of the radar system 14, a virtual receiving antenna array is created. The transmitting antennas 28 send radar signals 22, which are generated by the control and evaluation unit 26, into the monitoring area 18. The radar system 14 uses frequency-modulated continuous waves (FMCW) as radar signals 22. This is the origin of the designation of the radar system 14 as an FMCW radar system.

[0066] If the radar signals 22 are reflected by one or more target objects 20, the reflected radar signals 22 are received as echo signals 32 by the receiving antennas 30. The electromagnetic echo signals 32 received by the receiving antennas 30 are converted into electrical signals in a manner known per se. The electrical signals are transmitted to the control and evaluation unit 26 and processed by it.

[0067] The control and evaluation unit 26 processes the electrical received signals originating from received echo signals 32 into data for point clouds 34. For this purpose, the control and evaluation unit 26 contains hardware, such as processors, and software with which algorithms are implemented.

[0068] The point clouds 34 contain information in the form of target object data about individual targets of detected target objects 20. The targets of a target object 20 are locations, for example on the surface of the target object 20, where radar signals 22 can be reflected. The targets can be assigned to individual points in the point cloud 34. The points can be implemented as so-called cells (bins), such as distance-direction-Doppler cells or the like.

[0069] The target object data in the described embodiment includes target object distance data 36, ​​target object direction data 38 and target object extent data 40.

[0070] The target object distance data 36 characterize the distance 42 of a detected target of the target object 20 relative to the radar system 14.

[0071] The target direction data 38 characterize a direction 44, for example, a directional angle, in which the detected target of the target object 20 is located relative to the radar system 14. The target direction data 26 can also characterize the reception direction of the received echo signals 32. The reception direction is the direction from which the echo signals 32 originate. The reception direction is denoted in English as the Direction of Arrival (DoA).

[0072] The target object extent data 40 characterize an extent 46 of the target object 20. In the described embodiment, the target object extent data 40 characterize, by way of example, the spatial horizontal extent 46. Alternatively or additionally, target object extent data 40 can also be determined that characterize a spatial vertical extent of the detected target object 20. The extent 46 is, by way of example, an angular extent. The angular extent can also be referred to as angular size or apparent size. It corresponds to the viewing angle that the detected target object 20 covers in the horizontal direction. The angular extent depends on the actual extent of the target object 20 and its distance 42 from the radar system 14. From the angular extent, the actual horizontal extent of the target object 20 can be determined.The actual extent can be calculated, for example, from the angular extent and the distance 42.

[0073] The point clouds 34 are transmitted to the control unit 16. As already mentioned, the information can be used with the control unit 16 to control functions of the vehicle 10.

[0074] The control and evaluation unit 26 includes means for converting electrical received signals, which originate from the electromagnetic echo signals 32, into raw radar data 50. These means may, for example, include analog-to-digital converters with which the analog received signals can be converted into digital raw radar data 50.

[0075] Radar raw data 50 are data output after radar measurements before being processed into point cloud data 34. The radar raw data 50 are, for example, the digitized data after an analog-to-digital conversion of the electrical received signals of the received echo signals 32.

[0076] Furthermore, the control and evaluation unit 26 includes a direction and distance determination unit 48 for determining the target object distance data 36 and the target object direction data 38 from the raw radar data 50. The direction and distance determination unit 48 may include signal processing means known per se. Alternatively or additionally, the direction and distance determination unit 48 may include a model determined by machine learning, for example, an artificial neural network.

[0077] Furthermore, the control and evaluation unit 26 includes a model 52 for determining the target object extent data 40 from the raw radar data 50. The model 52 is a mathematical model generated by machine learning. For example, the model 52 is an artificial neural network. The model could, for instance, be a so-called Vision Transformer model. The target object extent data 40 can be determined from the raw radar data 50, for example, the phase patterns and magnitude patterns of the raw radar data 50, using the model 52.

[0078] Furthermore, the control and evaluation unit 26 includes a point cloud algorithm 54 for generating point clouds 34 from the target object distance data 36, ​​the target object direction data 38, and the target object extent data 40. The point cloud algorithm 54 allows the target object distance data 36, ​​the target object direction data 38, and the target object extent data 40 to be assigned to corresponding points in the point cloud 34.

[0079] Furthermore, the control and evaluation unit 26 includes a learning algorithm 56 for training the model 52. The learning algorithm 56 is described below in connection with the Fig. 4 explained in more detail.

[0080] In the Fig. Figure 3 shows a flowchart of an exemplary procedure for determining a point cloud 34. The points of the point cloud 34 characterize the target object 20 detected by the radar system 14.

[0081] First, radar measurements 58 are carried out using the radar system. During the radar measurements 58, radar signals 22 in the form of frequency-modulated continuous wave signals are transmitted by the transmitting antennas 28 into the monitoring area 18. The echo signals 32 reflected from a target object 20 are received by the receiving antennas 30 and converted into corresponding electrical receive signals. The electrical receive signals of the received echo signals 32 are converted into raw radar data 50 by the means of the control and evaluation unit 26.

[0082] The raw radar data 50 are fed to the direction and distance determination device 48 and the model 52.

[0083] The direction and distance determination device 48 determines the target object distance data 36 and the target object direction data 38 from the radar raw data 50.

[0084] Temporally overlapping, essentially in parallel, with the determination of the target range data 36 and the target direction data 38, a procedure for determining the target extent data 40 from the raw radar data 50 is carried out. In the procedure for determining the target extent data 40, the target extent data 40 are determined from the raw radar data 50 using the model 52.

[0085] The target object distance data 36, ​​the target object direction data 38, and the target object extent data 40 are fed to the point cloud algorithm 54. The point cloud 34 is generated using the point cloud algorithm 54. In doing so, the target object distance data 36, ​​the target object direction data, and the target object extent data 40 are assigned to the respective points in the point cloud 34, which characterize targets of the target object 20.

[0086] The point cloud 34 can then be fed to the control unit 16. The control unit 16 can then control functions of the vehicle 10 autonomously or semi-autonomously, taking into account the data from the point cloud 34.

[0087] Fig. Figure 4 shows a method for determining model 52 using machine learning.

[0088] First, in step 60, a training dataset is determined. To determine the training dataset with raw radar data 50 from different situations with individual extended target objects 20 with different dimensions 46, this dataset is obtained using radar measurements.

[0089] For each of the situations, the basic truth data 66 for the corresponding target object 20 is further determined in a basic truth determination step 62 with respect to the target object extent data 40. The basic truth data 66 are the expected output data of the model 52, which corresponds to the respective target object 20 in the corresponding situation. The basic truth data 66 thus correspond to the extent 46 of the detected target object 20. The basic truth data 66 can be determined, for example, by separate measurements, even independently of the radar measurements.

[0090] Before the learning algorithm 56 is executed, the raw radar data 50 are normalized in a normalization step 64. In normalization step 64, the raw radar data 50 coming from the individual antenna channels of the receiving antennas 30 are normalized relative to the first antenna channel. For better differentiation, the reference symbol of the raw radar data 50 is given the index N after nomination.

[0091] The learning algorithm 56 is then executed. For this purpose, the procedure for determining the target object extent data 40, which is already used in conjunction with the Fig. As explained in section 3, the process was carried out. The standardized raw radar data (50) are used. N The data, which originate from radar measurements in a situation with a target object 20, are fed to the model 52 to be trained. The model 52 then uses the normalized raw radar data 50 to generate a model 50. N The target object extent data 40 for this situation with this target object 20 was determined.

[0092] Subsequently, in a validation step 68, it is checked whether the model 52 is sufficient. For this purpose, the target object extent data 40 determined with the model 52 and the basic truth data 66 for the same situation with the same target object 20 are compared. In the comparison, it is checked whether the target object extent data 40 agree with the corresponding basic truth data 66, possibly taking into account a predefined tolerance.

[0093] If the test in validation step 68 reveals that model 52 is insufficient, i.e., the basic truth data 66 and the target object extent data 40 do not sufficiently match, model 52 is adjusted in an adjustment step 70. After the adjustment, model 52 is run again with the normalized radar raw data 50 and checked in validation step 40.

[0094] The learning algorithm 56, in which the model 52 is used with the normalized radar raw data 50 N The process, which is carried out and checked in validation step 40, is repeated until model 52 is deemed sufficient in validation step 40.

[0095] If validation step 68 shows that model 50 is sufficiently trained, model 50 is made available for use in deployment step 72. The trained model 50 can now be used in the Fig. The 3 described methods for determining the point cloud 34 are used. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] CN 113409381 A

[0006]

Claims

[1] Method for determining target object extent data (40) which characterize at least one spatial extent (46) of at least one target object (20) detected by at least one radar measurement with at least one radar system (14), in which at least one machine learning model (52) is executed, characterized by , that at least a part of the radar raw data (50) obtained from at least one radar measurement is fed to at least one model (52) determined by machine learning and that corresponding target object extent data (40) are determined with the at least one model (52) on the basis of the supplied radar raw data (50). [2] Method according to claim 1, characterized by , that at least one model (52) is implemented using at least one artificial neural network. [3] Method according to claim 1 or 2, characterized by, that at least part of the target object extent data (40) is assigned to at least one point cloud (34) which describes the at least one target object (20) detected by the at least one radar system (14). [4] Method according to any of the preceding claims, characterized by , that radar raw data (50) are used which are obtained from radar measurements with a multiple multiple-input multiple-output radar system (14), a frequency-modulated continuous-wave radar system (14) or a MIMO-FMCW radar system (14). [5] Method according to any of the preceding claims, characterized by , that the method is used to determine target object extent data (40) for isolated target objects (20). [6] Method according to any of the preceding claims, characterized by , that digital radar raw data (50) are used as radar raw data. [7] Method for determining at least one point cloud (34) containing target object data (36, 38, 40) of at least one target object (20) detected by at least one radar system (14), wherein at least a part of the raw radar data (50) obtained by at least one radar measurement with at least one radar system (14) are processed into point cloud data (34), characterized by , that in the method for determining at least one point cloud (34) a method for determining target object extent data (40) according to one of claims 1 to 6 is carried out and at least a part of the target object extent data (40) determined thereby is supplied to the data of the at least one point cloud (34). [8] Method according to claim 7, characterized by, that at least a part of the raw radar data (50) is processed into target direction data (38) that characterize the reception directions (44) from which radar echo signals (32) received by the at least one radar system (14) originate, and at least a part of the target direction data (38) is added to the data of the at least one point cloud (34), and / or at least part of the raw radar data (50) is processed into target range data (36) that characterize the distances (42) to target objects (20) detected by the at least one radar system (14), and at least part of the target range data (36) is added to the data of the at least one point cloud (34). [9] Method according to claim 8, characterized by, that at least from some of the radar raw data (50) target direction data (38) and / or target distance data (36) and / or target extent data (40) are determined in a temporally overlapping manner. [10] Method according to claim 8 or 9, characterized by , that target direction data (38) and / or target distance data (36) are determined from at least a part of the radar raw data (50) by performing at least one machine learning model (52). and / or Target direction data (38) and / or target distance data (36) are generated from at least a part of the raw radar data (50) by means of signal processing. [11] Method for determining a model (52) by machine learning, wherein the model (52) is designed to assign raw radar data (50) obtained in at least one radar measurement with at least one radar system (14) to target object extent data (40), wherein the target object extent data (40) characterize at least one spatial extent (46) of at least one target object (20) detected in the at least one radar measurement, wherein the method for determining target object extent data (40) is carried out for at least one target object (20) and at least some of the target object extent data (40) obtained are compared with basic truth data (66) for the spatial extent (46) of the at least one target object (20) and the model (52) is adjusted accordingly. characterized by, that for at least one isolated target object (20) a method for determining target object extent data (40) according to one of claims 1 to 6 is carried out. [12] Method according to claim 11, characterized by , that at least a part of the radar raw data (50) is normalized and then the normalized radar raw data (50) is assigned corresponding target object extent data (40) using the model (52). [13] Radar system (14), in particular a radar system (14) for a vehicle (10), comprising at least one antenna system (24) for transmitting radar signals (22) and for receiving radar echo signals (32), and comprising at least one evaluation unit (26) for processing raw radar data (50) obtained during radar measurements into at least target object extent data (40) which characterize spatial extents (46) of target objects (20) detected during radar measurements, characterized by, that the at least one evaluation device (26) for carrying out the method for determining target object extent data (40) is configured according to one of claims 1 to 6. [14] Radar system according to claim 13, characterized by , that at least one evaluation device (26) includes at least one machine learning-generated model (52), in particular an artificial neural network, which is designed to determine target object extent data (40) from radar raw data (50), and / or at least one evaluation unit (26) is designed to process radar raw data (50) into target direction data (38) and / or target distance data (36). [15] Radar system according to claim 13 or 14, characterized by, that the radar system (14) comprises at least a part of a learning algorithm (56) for training at least one model (52), in particular at least one artificial neural network, which can be used in the method according to one of claims 1 to 6. [16] Model (52) generated by machine learning and designed to determine target object data (40) that characterize at least one target object (20) detected by radar measurements, based on radar raw data (50) obtained by radar measurements, characterized by , that the model (52) is designed for use in the method for determining target object extent data (40) according to any one of claims 1 to 10.

Images