Computing device for evaluating radar data from at least one MIMO radar sensor with a non-uniform antenna array

Abstract

Computing device for evaluating radar data from at least one MIMO radar sensor (2) with a non-uniform antenna array (3), comprising at least one interface (4) for receiving the radar data, and at least one computing module (5) designed to split the radar data into first radar data from a uniform portion (6) of a virtual antenna array (7) and second radar data from a non-uniform portion (8) of the virtual antenna array (7), to process the first and second radar data separately to obtain a uniform beam pattern (9) and a non-uniform beam pattern (10), to combine the uniform and non-uniform beam patterns (9, 10) to obtain an overall beam pattern (11), and to evaluate the overall beam pattern (11) using a trained artificial neural network (12) to obtain a unique angular position of at least one radar target (13).wherein the artificial neural network (12) is designed as a two-dimensional convolutional neural network comprising a convolution arrangement which is block-wise trained for evaluation in at least two dimensions by means of a convolution kernel and in which another convolution kernel reduces a depth of an output of the convolution kernel.

Description

[0001] The invention relates to a computing device for evaluating radar data from at least one MIMO radar sensor with a non-uniform antenna array. Furthermore, the invention relates to a control unit for a vehicle, a radar system, and a corresponding vehicle. The invention also relates to a computer-implemented method for evaluating radar data from at least one MIMO radar sensor with a non-uniform antenna array and to a corresponding computer program. Finally, the invention relates to a method for training an artificial neural network.

[0002] MIMO (Multiple Input Multiple Output) radar (Radio Detection and Ranging) sensors with non-uniform antenna arrays are known from the prior art. These MIMO radar sensors typically provide ambiguous results regarding the angular positions of detected radar targets. These ambiguities cannot be reliably resolved using known methods. The present invention overcomes this problem and enables the unambiguous determination of the angular positions of the radar targets.

[0003] German patent DE 10 2023 204 002 A1 discloses an antenna array for a radar sensor. German patent DE 10 2023 201 585 A1 discloses a method for radar angle estimation. In Nguyen MQ [et al.]: High Angular Resolution Method Based on Deep Learning for FMCW MIMO Radar. In: IEEE Transactions on Microwave Theory and Techniques, Vol. 71, No. 12, pp. 5413-5427, 2023. - ISSN 1557-9670, the possibilities of deep learning for high-resolution angle estimation with an automotive FMCW radar sensor are investigated. In Schuessler C., Hoffmann M., Vossiek M.: Super-Resolution Radar Imaging With Sparse Arrays Using a Deep Neural Network Trained With Enhanced Virtual Data. In: IEEE Journal of Microwaves, Vol. 3, No. 3, pp. 980-993, 2023. - ISSN 2692-8388 describes a deep learning network that is capable of providing high-resolution angle estimates based on very thin antenna arrays.

[0004] A computing device for evaluating radar data from at least one MIMO radar sensor with a non-uniform antenna array is proposed. The computing device includes at least one interface for receiving the radar data. The computing device includes at least one computing module designed to split the radar data into first radar data from a uniform portion of a virtual antenna array and second radar data from a non-uniform portion of the virtual antenna array. The computing module is designed to process the first and second radar data separately to obtain a uniform beam pattern and a non-uniform beam pattern. The computing module is designed to combine the uniform and non-uniform beam patterns to obtain a complete beam pattern.The computing module is designed to evaluate the entire beam pattern using a trained artificial neural network to obtain a unique angular position of at least one radar target. The artificial neural network is implemented as a two-dimensional convolutional neural network. It comprises a convolution array that is block-wise trained for evaluation in at least two dimensions using a convolution kernel, and in which another convolution kernel reduces the depth of an output from the convolution kernel.

[0005] The computing device is preferably intended for use in a vehicle. In particular, the radar target can be an object located in the vicinity of the vehicle, such as another vehicle, a pedestrian, a lane marking, a traffic light, a building, a road sign, or the like. Preferably, a control unit, in particular an electronic control unit (ECU), or a MIMO radar sensor of the vehicle can comprise the computing device. "Intended" is understood to mean, in particular, specially programmed, specially equipped, and / or specially designed. The phrase "an object is intended for a function" is understood to mean, in particular, that the object performs the function in at least one operating state.

[0006] The computing device, in particular the computing module, can be configured, at least partially, as a microprocessor, a microcontroller, an integrated circuit, in particular an FPGA (Field Programmable Gate Array), an accelerator for artificial neural networks, an application-specific integrated circuit (ASIC), or the like. The interface is preferably wired. Alternatively, a wireless interface is conceivable. The interface is specifically designed to connect the MIMO radar sensor and the computing module for data transmission. In particular, the interface is designed to provide the radar data received from the MIMO radar sensor to the computing module.

[0007] The radar data are preferably at least digital tube radar data. In particular, analog radar signals received by the MIMO radar sensor have undergone at least one analog-to-digital conversion within the MIMO radar sensor to obtain digital tube radar data. The MIMO radar sensor is preferably configured as an FMCW (frequency modulated continuous wave) radar sensor. Preferably, the MIMO radar sensor, and in particular the transmitter of the MIMO radar sensor, is / are designed to emit radiation with a ramp-modulated frequency, so-called frequency ramps. A frequency ramp represents, in particular, a linear change in the frequency of the emitted radiation over a period of time. Other waveforms of the radar radiation are also possible, for example, waveforms based on phase-modulated continuous wave (PMCW) waveforms or orthogonal frequency division multiplexing (OFDM) waveforms.

[0008] The MIMO radar sensor preferably comprises at least one physical antenna array. The antenna array is formed, in particular, by a plurality of antennas, which are arranged at intervals from one another, especially in a common plane. A portion of the antennas serves, in particular, as transmitters for emitting radar signals. Another portion of the antennas serves, in particular, as receivers for receiving the radar signals emitted by the transmitters and reflected in the vicinity of the MIMO radar sensor, especially by radar targets. The radar data is generated, in particular, from the received radar signals. The antennas can be configured, in particular, as planar (patch) antennas, as waveguide antennas, or the like. The antenna array is preferably configured as a non-uniform antenna array."Non-uniform" means, in particular, that the phase centers of transmitters and / or receivers adjacent along at least one direction are not equidistant from each other. Preferably, the phase centers of transmitters and / or receivers of the antenna array adjacent along at least one direction are arranged partially equidistant from each other.

[0009] A virtual antenna array of the MIMO radar sensor results in particular from the antenna array when several, preferably all, receivers receive radar signals originally transmitted by several, preferably all, transmitters, especially at the same time. The virtual antenna array comprises, in particular, a plurality of virtual antennas. A virtual antenna specifies, in particular, which combination of transmitter and receiver is responsible for a radar signal detected at the respective position in the virtual antenna array.

[0010] The non-uniform antenna array results in, in particular, a virtual antenna array that has a uniform component and a non-uniform component. The uniform component includes, in particular, virtual antennas resulting from combinations of antennas arranged equidistantly within the antenna array. The non-uniform component includes, in particular, virtual antennas resulting from combinations with antennas not arranged equidistantly within the antenna array.

[0011] The processing module is specifically designed to process the first and second radar data separately, preferably using different methods. For this processing, generally known methods such as digital beamforming (DBF), fast Fourier transform (FFT), or the like can be used. The processing of the radar data results in a beam pattern. This beam pattern preferably indicates several hypotheses for the angular position of a radar target, of which a single hypothesis corresponds to the correct angular position. The uniform beam pattern results particularly from the processing of the first radar data. The non-uniform beam pattern results particularly from the processing of the second radar data. An angular position preferably comprises at least one angular position in an azimuth dimension.An angular position can, in particular, consist of an angular position in an azimuth dimension and an angular position in an elevation dimension. In a simplified antenna array without elevation measurement capability, the elevation angle is specifically assumed to be zero.

[0012] In particular, the ambiguity arising from the uniform beam pattern or the non-uniform beam pattern alone cannot be resolved. Preferably, the computing module is designed to combine the uniform and the non-uniform beam patterns, in particular to superimpose them, in order to obtain the entire beam pattern. In particular, the computing module is designed to use the entire beam pattern as input data for the artificial neural network. Preferably, the computing module is designed to decompose the entire beam pattern, especially a complex one, into a real part and an imaginary part and to use the real part and the imaginary part together as input data for the artificial neural network.

[0013] The artificial neural network can, in particular, be configured as a convolutional neural network (CNN). Preferably, the artificial neural network is configured as a two-dimensional artificial neural network. The computing module is specifically designed to execute the artificial neural network. The artificial neural network is specifically trained, particularly by supervised training, to select the correct hypothesis for the angular position of a radar target from among the hypotheses contained in the entire beam pattern and to output this angular position as the unique angular position of the radar target. Preferably, the artificial neural network is designed to output a unique angular position of each radar target, depending on the number of radar targets contained in the radar data. The computing device can, in particular, comprise a plurality of computing modules.In particular, several of the data processing steps described above can be performed by different computing modules.

[0014] In particular, the input data for the artificial neural network are configured as two-dimensional matrices. The output data of the artificial neural network are also configured as a two-dimensional matrix. Preferably, one dimension of the matrices is proportional to the azimuth dimension and another dimension is proportional to the elevation dimension, for example, in a UV representation. The beam patterns are, in particular, two-dimensional power distributions, where the hypotheses for an angular position correspond to power peaks in the power distribution. Preferably, the artificial neural network is configured to output a beam pattern that comprises only a single power peak for each radar target, corresponding to the unique angular position.

[0015] The convolution arrangement is, in particular, an arrangement of convolution kernels. Preferably, the convolution arrangement is trained block-wise for evaluation in the azimuth and elevation dimensions. The depth reduction of the output of the convolution kernel can, in particular, be described as 1×1 convolution. In particular, points that represent the same position of the uniform portion of the virtual antenna array, whose ambiguity has not yet been resolved, are processed together in the convolution arrangement.

[0016] The design of the computing device according to the invention advantageously enables the determination of unambiguous angular positions of radar targets using MIMO radar sensors with non-uniform antenna arrays. Advantageously, a computationally efficient computing device for angular position determination can be provided. Advantageously, the angular positions can be determined with particular precision. Advantageously, particularly accurate target separation can be enabled. Advantageously, false detections can be suppressed.

[0017] Furthermore, it is proposed that the artificial neural network has a plurality of hidden layers that differ from one another at least with respect to their parameters. In particular, the artificial neural network has at least three, preferably at least five, and most preferably at least seven hidden layers. Most preferably, the artificial neural network has between three and seven hidden layers. Preferably, each hidden layer has several, for example, five, filter kernels. Preferably, each hidden layer has an activation function, for example, a Rectified Linear Unit (ReLU). The parameters by which the hidden layers differ from one another are, in particular, the weights of neurons in the hidden layers. In particular, the parameters are set by training the artificial neural network.Advantageously, unambiguous angular positions can be determined with particular reliability.

[0018] It is further proposed that the artificial neural network is trained using a multitude of radar data containing known angular positions of at least one radar target to recognize a correct angular position of the at least one radar target in a plurality of ambiguous angular positions. Preferably, the artificial neural network is trained using supervised training. In particular, the radar data contains a multitude of different angular positions of at least one radar target, especially different combinations of angular positions in the azimuth and elevation dimensions. In particular, the radar data can contain multiple radar targets, especially with different angular positions.In particular, the radar data can contain a multitude of different angular positions in combination with different and / or identical velocities and / or distances of the at least one radar target to the MIMO radar sensor. Preferably, the MIMO radar sensor is configured as a 4D MIMO radar sensor that can detect, in particular, an angular position in the azimuth dimension, an angular position in the elevation dimension, a velocity, and a distance of radar targets. The radar data is preferably simulated radar data. In particular, the radar data is simulated as a function of a configuration of the MIMO radar sensor, especially the antenna array. Alternatively or additionally, it is conceivable that the artificial neural network is trained using real radar data.Preferably, the training process adjusts the weights of the artificial neural network such that the trained network can correctly identify the angular position of at least one radar target in a plurality of ambiguous angular positions. Advantageously, this can result in a particularly reliable artificial neural network.

[0019] Furthermore, it is proposed that the computing module be designed to process the first radar data using a two-dimensional fast Fourier transform and / or the second radar data using digital beamforming. In particular, the fast Fourier transform requires less computing time than digital beamforming. Specifically, the fast Fourier transform can only be meaningfully applied to radar data from a uniform virtual antenna array or at least a uniform portion of the virtual antenna array. In particular, processing the first radar data using the two-dimensional fast Fourier transform yields only a sub-region of the uniform beam pattern, specifically corresponding to a sub-region of the MIMO radar sensor's field of view. Preferably, the uniform beam pattern is periodically structured due to the uniformity of the virtual antenna array's portion.Preferably, the computing module is configured to multiply the portion of the uniform beam pattern obtained by the two-dimensional fast Fourier transform and to combine it into a complete uniform beam pattern with the same size as the non-uniform beam pattern. Preferably, the computing module is configured to combine the uniform beam pattern thus obtained with the non-uniform beam pattern to obtain the complete beam pattern. Preferably, the computing module is configured to further process the second radar data using conventional digital beamforming (CBF). Advantageously, this enables time-saving evaluation of the radar data.

[0020] Furthermore, it is proposed that the artificial neural network includes at least an incomplete kernel designed to jointly process points representing the same position within the uniform portion of the virtual antenna array. This processing can be specifically termed atrous convolution.

[0021] Furthermore, it is proposed that the artificial neural network be designed to process points representing the same position of the uniform portion of the virtual antenna array during the depth reduction of the convolution kernel output, particularly jointly, especially by means of the incomplete kernel. Specifically, the computational module can be designed to rearrange the uniform and non-uniform portions of the virtual antenna array such that the virtual antenna array is extended by one dimension in which points of the uniform portion of the virtual antenna array are sequential, particularly to enable processing during the depth reduction of the convolution kernel output.

[0022] Furthermore, a control unit for a vehicle is proposed. The control unit comprises at least one computing device according to the invention. The control unit can, in particular, be designed as a high-performance vehicle computer. Specifically, the control unit can be configured to process, and especially evaluate, signals from other sensors of the vehicle, such as cameras or lidar (light detection and ranging) sensors. In particular, the control unit can be configured to generate control commands for at least semi-autonomous control of the vehicle, especially based on the results of radar data evaluation. Advantageously, a control unit with high-performance radar data processing can be provided.

[0023] Furthermore, a radar system is proposed. The radar system comprises at least one MIMO radar sensor with a non-uniform antenna array. The radar system comprises at least one computing device or at least one control unit according to the invention. In particular, the computing device can be integrated into the MIMO radar sensor. In particular, the computing device can be arranged in a housing of the MIMO radar sensor. In particular, the computing device can be designed as a signal processor of the MIMO radar sensor. Alternatively, the control unit can comprise the computing device and be arranged separately from the MIMO radar sensor. In particular, the radar system can comprise a plurality of MIMO radar sensors and / or computing devices or control units. In particular, a computing device for evaluating radar data from several MIMO radar sensors can be provided.A radar system can be advantageously provided with powerful radar data processing.

[0024] Furthermore, an automated vehicle is proposed. The automated vehicle comprises at least one radar system according to the invention. An "automated vehicle" is understood to mean, in particular, a vehicle with one of the automation levels 1 to 5 of standard SAE J3016. Specifically, the automated vehicle has technical equipment required for these automation levels. This technical equipment includes, in particular, environmental sensing sensors, such as at least one MIMO radar sensor, lidar sensors, cameras and / or acoustic sensors, control units, or the like. Preferably, the vehicle is designed as a land vehicle, in particular as a road vehicle.The automated vehicle can be configured in particular as a passenger car, preferably a passenger transport vehicle, a truck, a construction vehicle, an agricultural vehicle, or any other vehicle that would be considered suitable by an expert. Alternatively, the automated vehicle can also be configured as an aircraft, for example, a drone, an airplane, a helicopter, a vertical take-off and landing aircraft, or the like; as a rail vehicle, for example, a locomotive, a wagon, or the like; or as a watercraft, in particular a ship, a boat, or the like. It is advantageous to provide a particularly safe automated vehicle.

[0025] Furthermore, a computer-implemented method for evaluating radar data from at least one MIMO radar sensor with a non-uniform antenna array is proposed. The received radar data is split into first radar data from a uniform portion of a virtual antenna array and second radar data from a non-uniform portion of the virtual antenna array. The first and second radar data are processed separately to obtain a uniform beam pattern and a non-uniform beam pattern, respectively. The uniform and non-uniform beam patterns are combined to obtain a complete beam pattern. This complete beam pattern is evaluated using a trained artificial neural network to obtain a unique angular position of at least one radar target. The artificial neural network is designed as a two-dimensional convolutional neural network.The artificial neural network comprises a convolutional arrangement that is trained block-wise for evaluation in at least two dimensions using a convolutional kernel, and in which another convolutional kernel reduces the depth of an output of the convolutional kernel. Advantageously, this can provide a computationally efficient method for the precise angular position determination of radar targets.

[0026] Furthermore, a computer program for evaluating radar data from at least one MIMO radar sensor with a non-uniform antenna array is proposed. The computer program comprises execution commands which, when executed by a computing device according to the invention, cause it to execute a method according to the invention. Advantageously, a computer program can be provided that enables computationally efficient and precise angular position determination of radar targets.

[0027] Furthermore, a method for training an artificial neural network, which is used by a computing device according to the invention to obtain a unique angular position of at least one radar target, is proposed. A plurality of radar data containing known angular positions of at least one radar target is fed into the artificial neural network. Depending on an output of the artificial neural network, parameters of the artificial neural network are set. In particular, the weights of the artificial neural network are set, especially such that the trained artificial neural network can recognize a correct angular position of at least one radar target in a plurality of ambiguous angular positions. Advantageously, the artificial neural network can be specifically trained for angular position detection in radar data.

[0028] The invention is illustrated by an exemplary embodiment in the following figures. They show: Fig. 1 a schematic representation of an automated vehicle according to the invention, Fig. 2 a computing device according to the invention of the automatically operated vehicle according to the invention Fig. 1 in a schematic representation, Fig. 3 an antenna array of a MIMO radar sensor of the automated vehicle according to the invention Fig. 1 in a schematic representation, Fig. 4 a virtual antenna array of the MIMO radar sensor Fig. 3 in a schematic representation, Fig. 5 beam patterns in a schematic representation, Fig. 6 an artificial neural network of the computing device according to the invention Fig. 2 in a schematic representation, Fig. 7 Output data of the artificial neural network from Fig. 6 in a schematic representation, Fig. 8 a flowchart of a computer-implemented method according to the invention in a schematic representation and Fig. 9 a flowchart of a method according to the invention in a schematic representation.

[0029] Fig. Figure 1 shows a schematic representation of an automated vehicle 16. In the present embodiment, the automated vehicle 16 is exemplary as a land vehicle, in particular as a passenger car. The automated vehicle 16 comprises at least one radar system 17. The radar system 17 comprises at least one MIMO radar sensor 2 with a non-uniform antenna array 3 (see Figure 1). Fig. 3) The radar system 17 comprises at least one control unit 15. The control unit 15 comprises at least one computing device 1. In an alternative embodiment, it is conceivable that the MIMO radar sensor 2 comprises the computing device 1. Furthermore, in Fig. 1. A radar target 13 is shown. By way of example, the radar target 13 is designed as a foreign vehicle.

[0030] Fig. Figure 2 shows the computing device 1 of the automated vehicle 16. Fig. Figure 1 is shown in a schematic representation. The computing device 1 is designed for evaluating radar data from the MIMO radar sensor 2. The computing device 1 includes at least one interface 4 for receiving the radar data. The computing device 1 includes at least one computing module 5.

[0031] Fig. Figure 3 shows the, in particular physical, antenna array 3 of the MIMO radar sensor 2 of the automated vehicle 16. Fig. Figure 1 is shown in a schematic representation. The antenna array 3 is formed by a plurality of antennas 18, 19, which are arranged at intervals from each other in a common plane. Some of the antennas 18, 19 serve as transmitters 18, which are designed to transmit radar signals. Another part of the antennas 18, 19 serves as receivers 19, which are designed to receive the radar signals transmitted by the transmitters 18 and reflected in the vicinity of the MIMO radar sensor 2, in particular by radar targets 13. The radar data are generated from the received radar signals. Fig. Figure 3 shows the positions of the phase centers 20, 21 of the antennas 18, 19. The phase centers 20 of the transmitters 18 are represented as crosses. The phase centers 21 of the receivers 19 are represented as plus signs. The antenna array 3 is configured as a non-uniform antenna array. In the present embodiment, the phase centers 20 of adjacent transmitters 18, viewed along a direction parallel to an abscissa 22 and along a direction parallel to an ordinate 23, are not equidistant from each other, particularly in the center of the antenna array 3. The phase centers 21 of the receivers 19 are equidistant from each other, viewed along the direction parallel to the abscissa 22 and along the direction parallel to the ordinate 23.

[0032] Fig. Figure 4 shows a virtual antenna array 7 of the MIMO radar sensor 2. Fig. Figure 3 in a schematic representation. The virtual antenna array 7 comprises a plurality of virtual antennas 24. The virtual antenna array 7 has a uniform component 6 and a non-uniform component 8. The uniform component 6 comprises virtual antennas 24 resulting from combinations of antennas 18, 19 arranged equidistantly in the antenna array 3. The non-uniform component 8 comprises virtual antennas 24 resulting from combinations with antennas 18, 19 arranged non-equidistantly in the antenna array 3. The uniform component 6 and the non-uniform component 8 are on the right side of the Fig. 4 shown separately. The computing module 5 is designed to divide the radar data into first radar data of the uniform component 6 of the virtual antenna array 7 and second radar data of the non-uniform component 8 of the virtual antenna array 7.

[0033] Fig. Figure 5 shows beam patterns 9-11 in a schematic representation. A uniform beam pattern 9, a non-uniform beam pattern 10, and a complete beam pattern 11 are depicted. The processing module 5 is designed to process the first and second radar data separately to obtain the uniform beam pattern 9 and the non-uniform beam pattern 10. The beam patterns 9-11 represent several hypotheses 25, 26 for the angular position of the radar target 13, of which a single hypothesis 25 corresponds to the correct angular position. The uniform beam pattern 9 results from the processing of the first radar data. The non-uniform beam pattern 10 results from the processing of the second radar data.

[0034] The computing module 5 is designed to further process the first radar data using a two-dimensional fast Fourier transform and / or the second radar data using digital beamforming. The further processing of the first radar data using the two-dimensional fast Fourier transform yields only a sub-area 27 of the uniform beam pattern 9, in particular corresponding to a sub-area of ​​the field of view of the MIMO radar sensor 2. The uniform beam pattern 9 is periodically structured due to the uniformity of the component 6 of the virtual antenna array 7. The computing module 5 is designed to multiply the sub-area 27 of the uniform beam pattern 9 obtained by the two-dimensional fast Fourier transform and to assemble it into a complete uniform beam pattern 9 with the size of the non-uniform beam pattern 10.The computing module 5 is designed to further process the second radar data using conventional digital beamforming.

[0035] Beam patterns 9-11 are two-dimensional power distributions, where hypotheses 25 and 26 correspond to power peaks in the distribution for a given angular position. One dimension 28 of beam patterns 9-11 is proportional to an azimuth dimension, and another dimension 29 of beam patterns 9-11 is proportional to an elevation dimension, shown here in an example UV representation. The computational module 5 is designed to combine, and in particular superimpose, the uniform and non-uniform beam patterns 9 and 10 to obtain the complete beam pattern 11.

[0036] Fig. Figure 6 shows an artificial neural network 12 of the computing device 1. Fig. Figure 2 is shown in a schematic representation. The computing module 5 is designed to evaluate the entire beam pattern 11 using the trained artificial neural network 12 in order to obtain a unique angular position of at least one radar target 13. The artificial neural network 12 is designed as a two-dimensional convolutional neural network. The artificial neural network 12 comprises a convolution arrangement that is trained block-wise for evaluation in at least two dimensions using a convolution kernel, and in which another convolution kernel reduces the depth of an output of the convolution kernel. The computing module 5 is designed to use the entire beam pattern 11 as input data for the artificial neural network 12.The computing module 5 is designed to divide the entire beam pattern 11, especially complex ones, into a real part 30 and an imaginary part 31, and to use the real part 30 and the imaginary part 31 together as input data for the artificial neural network 12.

[0037] The computing module 5 is designed to execute the artificial neural network 12. The artificial neural network 12 is trained, in particular by means of supervised training, to select the correct hypothesis 25 from the hypotheses 25, 26 contained in the entire beam pattern 11 for an angular position of the radar target 13 and to output this angular position as the unique angular position of the radar target 13.

[0038] The artificial neural network 12 has a plurality of intermediate layers 14, which differ from each other at least with respect to their parameters. The parameters by which the intermediate layers 14 differ from each other are the weights of neurons in the intermediate layers 14. The parameters are set by training the artificial neural network 12.

[0039] Fig. Figure 7 shows output data from the artificial neural network 12. Fig. Figure 6 is shown in a schematic representation. The artificial neural network 12 is trained using a multitude of radar data containing known angular positions of at least one radar target 13 to recognize a correct angular position of the at least one radar target 13 in a plurality of ambiguous angular positions. The artificial neural network 12 is trained using supervised training. The radar data contains a multitude of different angular positions of at least one radar target 13, in particular different combinations of angular positions in the azimuth and elevation dimensions. The artificial neural network 12 is designed to output a beam pattern 32 as output data, which comprises only a single power peak corresponding to the unique angular position for each radar target 13.

[0040] The convolution array is an array of convolution kernels. The convolution array is trained block-wise for evaluation in the azimuth and elevation dimensions. In the convolution array, points that represent the same position of the uniform portion 6 of the virtual antenna array 7, whose ambiguity, in particular, has not yet been resolved, are processed together.

[0041] The artificial neural network 12 includes at least one incomplete kernel designed to jointly process points that represent the same position of the uniform portion 6 of the virtual antenna array 7.

[0042] The artificial neural network 12 is designed to process points representing the same position of the uniform portion 6 of the virtual antenna array 7 during the depth reduction of the convolution kernel output, particularly jointly, especially by means of the incomplete kernel. The computation module 5 is designed to rearrange the uniform portion 6 and the non-uniform portion 8 of the virtual antenna array 7 such that the virtual antenna array 7 is extended by one dimension in which points of the uniform portion 6 of the virtual antenna array 7 are arranged sequentially, particularly to enable processing during the depth reduction of the convolution kernel output.

[0043] Fig. Figure 8 shows a schematic flowchart of a computer-implemented method for evaluating radar data from the MIMO radar sensor 2 with a non-uniform antenna array 3. In a first process step 33, the received radar data are split into first radar data from the uniform component 6 of the virtual antenna array 7 and second radar data from the non-uniform component 8 of the virtual antenna array 7. In a second process step 34, the first and second radar data are processed separately to obtain a uniform beam pattern 9 and a non-uniform beam pattern 10. In a third process step 35, the uniform and non-uniform beam patterns 9, 10 are combined to obtain a complete beam pattern 11.In a fourth process step 36, the entire beam pattern 11 is evaluated using the trained artificial neural network 12 in order to obtain a unique angular position of at least one radar target 13.

[0044] A computer program product for evaluating radar data from at least one MIMO radar sensor 2 with the non-uniform antenna array 3 includes execution instructions which, when the program is executed by the computing device 1, cause it to execute the computer-implemented method.

[0045] Fig.Figure 9 shows a schematic flowchart of a procedure for training the artificial neural network 12, which is used by the computing device 1 to obtain a unique angular position of the at least one radar target 13. In a first procedure step 37, radar data containing a known angular position of the at least one radar target 13 are input into the artificial neural network 12. The radar data can contain several radar targets 13, in particular with different angular positions. In a second procedure step 38, parameters of the artificial neural network 12 are set depending on an output of the artificial neural network 12.Through training, the weights of the artificial neural network 12 are set, in particular such that the trained artificial neural network 12 can recognize a correct angular position of at least one radar target 13 in a plurality of ambiguous angular positions.

[0046] The process steps 37 and 38 can be repeated multiple times, so that a large number of radar data containing known, in particular different, angular positions of the at least one radar target 13 are input into the artificial neural network 12. The radar data can contain a large number of different angular positions in combination with different and / or identical velocities and / or distances of the at least one radar target 13 to the MIMO radar sensor 2. The MIMO radar sensor 2 is exemplified as a 4D MIMO radar sensor that can detect an angular position in the azimuth dimension, an angular position in the elevation dimension, a velocity, and a distance of radar targets 13. The radar data can be simulated radar data. The radar data can be simulated depending on a configuration of the MIMO radar sensor 2, in particular the antenna array 3.Alternatively or additionally, it is conceivable that the artificial neural network 12 is trained using real radar data. Reference sign 1 calculating device 2 MIMO radar sensors 3 antenna array 4 Interface 5 Calculation module 6 equal proportion 7 virtual antenna array 8 unequal proportion 9 uniform beam pattern 10 uneven beam pattern 11 entire beam pattern 12 artificial neural network 13 Radar target 14 Intermediate shift 15 Control unit 16 vehicles 17 radar systems 18 Transmitter 19 recipients 20 Phase Center 21 Phase Center 22 Abscissa axis 23 ordinate axis 24 virtual antennas 25 Hypothesis 26 Hypothesis 27 sub-area 28 Dimension 29 Dimension 30 Real part 31 Imaginary part 32 Output beam patterns 33. Procedure step 34th process step 35th process step 36th process step 37th process step 38th process step

Claims

[1] Computing device for evaluating radar data of at least one MIMO radar sensor (2) with a non-uniform antenna array (3), comprising at least one interface (4) for receiving the radar data, and at least one computing module (5) designed to split the radar data into first radar data of a uniform portion (6) of a virtual antenna array (7) and second radar data of a non-uniform portion (8) of the virtual antenna array (7), to process the first and second radar data separately to obtain a uniform beam pattern (9) and a non-uniform beam pattern (10), to combine the uniform and the non-uniform beam patterns (9, 10) to obtain an overall beam pattern (11), and to evaluate the overall beam pattern (11) using a trained artificial neural network (12) to obtain a unique angular position of at least one radar target (13),wherein the artificial neural network (12) is designed as a two-dimensional convolutional neural network comprising a convolution arrangement which is block-wise trained for evaluation in at least two dimensions by means of a convolution kernel and in which another convolution kernel reduces a depth of an output of the convolution kernel. [2] Computing device according to claim 1, wherein the artificial neural network (12) has a plurality of intermediate layers (14) which differ from each other at least with respect to their parameters. [3] Computing device according to claim 1 or 2, wherein the artificial neural network (12) is trained by means of a plurality of radar data containing known angular positions of at least one radar target (13) to recognize a correct angular position of the at least one radar target (13) in a plurality of ambiguous angular positions. [4] Computing device according to one of the preceding claims, wherein the computing module (5) is provided to further process the first radar data by means of a two-dimensional fast Fourier transform and / or the second radar data by means of a digital beamforming. [5] Computing device according to one of the preceding claims, wherein the artificial neural network (12) comprises at least one incomplete kernel designed to jointly process points that represent the same position of the uniform portion (6) of the virtual antenna array (7). [6] Computing device according to one of the preceding claims, wherein the artificial neural network (12) is provided to process points representing the same position of the uniform portion (6) of the virtual antenna array (7) during the depth reduction of the convolution kernel output. [7] Control unit for a vehicle (16), comprising at least one computing device (1) according to any of the preceding claims. [8] Radar system comprising at least one MIMO radar sensor (2) with a non-uniform antenna array (3) and at least one computing device (1) according to any one of claims 1 to 6 or at least one control unit (15) according to claim 7. [9] Automated vehicle comprising at least one radar system (17) according to claim 8. [10] Computer-implemented method for evaluating radar data from at least one MIMO radar sensor (2) with a non-uniform antenna array (3), wherein the received radar data are split into first radar data of a uniform portion (6) of a virtual antenna array (7) and second radar data of a non-uniform portion (8) of the virtual antenna array (7), wherein the first and second radar data are processed separately to obtain a uniform beam pattern (9) and a non-uniform beam pattern (10), wherein the uniform and the non-uniform beam patterns (9, 10) are combined to obtain an overall beam pattern (11), and wherein the overall beam pattern (11) is evaluated by means of a trained artificial neural network (12) to obtain a unique angular position of at least one radar target (13),wherein the artificial neural network (12) is designed as a two-dimensional convolutional neural network comprising a convolution arrangement which is block-wise trained for evaluation in at least two dimensions by means of a convolution kernel and in which another convolution kernel reduces a depth of an output of the convolution kernel. [11] Computer program product for evaluating radar data from at least one MIMO radar sensor (2) with a non-uniform antenna array (3), comprising execution instructions which, when the program is executed by a computing device (1) according to one of claims 1 to 6, cause it to execute a method according to claim 10. [12] Method for training an artificial neural network (12) used by a computing device (1) according to one of claims 1 to 6 to obtain a unique angular position of at least one radar target (13), wherein a plurality of radar data containing known angular positions of at least one radar target (13) is input into the artificial neural network (12) and parameters of the artificial neural network (12) are set depending on an output of the artificial neural network (12).

Images