Search

EP-4737940-A2 - RADAR SYSTEM FOR DIFFERENTIATING OBJECTS BY MEANS OF TWO DIFFERENTLY TRAINED NEURAL NETWORKS USING TWO DIFFERENT DATA PROCESSING ALGORITHMS FOR CALCULATING INPUT DATA FOR THE NEURAL NETWORKS ON THE BASIS OF RECEIVED RADAR SIGNALS

EP4737940A2EP 4737940 A2EP4737940 A2EP 4737940A2EP-4737940-A2

Abstract

Disclosed is a Radar system (3) for differentiating objects (61, 62) in an environment (140) of a vehicle (40), the radar system (3) comprising an evaluation unit (4), wherein the evaluation unit (4) comprises two differently trained neural networks (1, 2) and is configured to: - generate first input data (101) by means of a first data processing algorithm and dependent on receiving signals (100), - generate second input data (102) by means of a second data processing algorithm and dependent on the receiving signals (100), wherein the first data processing algorithm is different from the second data processing algorithm, - input the first input data (101) into a first one of the neural networks and in response receive from the first neural network (1) a first output (111), - input the second input data (102) into a second one of the neural networks and in response receive from the second neural network (2) a second output (112), generate an output (800) dependent on the first output (111) and the second output (112), the output (800) indicating a differentiation between the objects (61, 62).

Inventors

  • KUMAR, SANDEEP
  • SCHOLZ, Niko
  • POEPPERL, MAXIMILIAN
  • LUKIN, Artem
  • JENKINS, ALAN

Assignees

  • Valeo Schalter und Sensoren GmbH

Dates

Publication Date
20260506
Application Date
20251008

Claims (14)

  1. Radar system (3) for differentiating objects (61, 62) in an environment (140) of a vehicle (40), the radar system (3) comprising an evaluation unit (4), wherein the evaluation unit (4) comprises two differently trained neural networks (1, 2) and is configured to: - generate first input data (101) by means of a first data processing algorithm and dependent on receiving signals (100) generated by means of received radar signals (501), - generate second input data (102) by means of a second data processing algorithm and dependent on the receiving signals (100), wherein the first data processing algorithm is different from the second data processing algorithm, - input the first input data (101) into a first one of the neural networks and in response receive from the first neural network (1) a first output (111), the first output (111) indicating a differentiation between the objects (61, 62), - input the second input data (102) into a second one of the neural networks and in response receive from the second neural network (2) a second output (112), the second output (112) indicating a differentiation between the objects (61, 62), - generate an output (800) dependent on the first output (111) and the second output (112), the output (800) indicating a differentiation between the objects (61, 62).
  2. The radar system (3) according to claim 1, wherein a type of the first input data (101) is different compared to a type of the second input data (102) and the first neural network (1) and the second neural network (2) comprise a different structure.
  3. The radar system (3) according to claim 1 or 2, wherein the evaluation unit (4) is configured to generate digitalized and filtered data dependent on the receiving signals (100) and to generate a 2-dimensional frequency spectrum dependent on the digitalized and filtered data, wherein a first dimension of the frequency spectrum represents first frequencies which relate to distances of the objects (61, 62) with respect to the radar system (3) and a second dimension of the frequency spectrum represents second frequencies which relate to relative velocities of the objects (61, 62) with respect to the radar system (3), and the type of the first input data (101) is designed such that the first input data (101) comprises data derived from the 2-dimensional frequency spectrum.
  4. The radar system (3) according to any of the previous claims, wherein the evaluation unit (4) is configured to - generate power spectra dependent on the receiving signals (100), the receiving signals (100) being received from a set of receiving antennas (11), wherein first frequencies of each power spectrum relate to distances of the objects (61, 62) with respect to the set of receiving antennas (11) and second frequencies of each power spectrum relate to relative velocities of the objects (61, 62) with respect to the set of receiving antennas (11), - select a pair of frequencies (300) for each power spectrum, wherein each pair of frequencies (300) comprises the same frequency of the first frequencies and the same frequency of the second frequencies of the respective power spectrum, wherein each power spectrum comprises a phase information of the selected pair of the respective power spectrum, - perform a spatial Fourier transform on the basis of the phase information of the selected pairs, the result of the spatial Fourier transform comprising first spatial frequencies and second spatial frequencies associated to each other in the form of pairs of spatial frequencies, wherein to each pair of spatial frequencies one of the first spatial frequencies, one of the second spatial frequencies and an intensity information and a phase information is associated, - generate the first input data (101) in the form of a first image comprising first pixels, each first pixel representing one respective pair of spatial frequencies of the pairs of spatial frequencies and comprising a set of pixel values representing different channels of the first pixel, wherein for the respective pair of spatial frequencies, the pixel values depend on the intensity information and the phase information of the pair of spatial frequencies.
  5. The radar system (3) according to claim 4, wherein the second input data (102) is provided in the form of an array and elements of the array comprise the intensity information and the phase information of the pairs of spatial frequencies.
  6. The radar system (3) according to claim 4 or 5, wherein the intensity information and the phase information of the respective pair of spatial frequencies is provided by means of a complex number, the complex number comprising a real part and an imaginary part, wherein the respective set of pixel values comprises a first pixel value representing the real part of the respective complex number and a second pixel value representing the imaginary part of the respective complex number and a third pixel value representing a phase value calculated dependent on the respective complex number.
  7. The radar system (3) according to claim 6, wherein the respective set of pixel values further comprises a fourth pixel value representing an absolute value of the respective complex number.
  8. The radar system (3) according to any of previous claims 6-7, wherein the evaluation unit (4) is configured to - generate the second input data (102) in the form of a second image comprising second pixels, each second pixel representing a pair of the spatial frequencies and comprising a second set of pixel values representing different channels of the pixel, the pixel values depending on the intensity information and the phase information of the respective pair of spatial frequencies, wherein the respective second set of pixel values comprises a first pixel value representing the real part of the respective complex number and a second pixel value representing the imaginary part of the respective complex number and a third pixel value representing a phase value of the respective complex number.
  9. The radar system (3) according to one of the preceding claims, wherein the first neural network (1) is a convolutional neural network and/or wherein the second neural network (2) is a transformer-based neural network.
  10. The radar system (3) of any of the previous claims 4-9, further comprising the receiving antennas.
  11. A vehicle (40) comprising a radar system (3) according to any of the previous claims 1 to 10.
  12. Method for operating a radar system (3) for differentiating objects (61, 62) in an environment (140) of a vehicle (40), the radar system (3) comprising an evaluation unit (4), wherein the evaluation unit (4) comprises two differently trained neural networks, the method comprising: - generating first input data (101) by means of a first data processing algorithm and dependent on receiving signals (100) generated by means of received radar signals, - generating second input data (102) by means of a second data processing algorithm and dependent on the receiving signals (100), wherein the first data processing algorithm is different from the second data processing algorithm, - inputting the first input data (101) into a first one of the neural networks and in response receiving from the first neural network (1) a first output (111), the first output (111) indicating a differentiation between the objects (61, 62), - inputting the second input data (102) into a second one of the neural networks and in response receiving from the second neural network (2) a second output (112), the second output (112) indicating a differentiation between the objects (61, 62), - generating an output (800) dependent on the first output (111) and the second output (112), the output (800) indicating a differentiation between the objects (61, 62).
  13. Method according to claim 12, the method further comprising training the first neural network (1) and the second neural network (2), the training comprising: - generating first training data (2001), the first training data (2001) comprising input datasets and target datasets, the generating of the first training data (2001) comprising generating the input datasets of the first training data (2001) dependent on training receiving signals using the first data processing algorithm, the training receiving signals being generated by means of training radar signals, - generating second training data (2002), the second training data (2002) comprising input datasets and target datasets, the generating of the second training data (2002) comprising generating the input datasets of the second training data (2002) dependent on the training receiving signals using the second data processing algorithm, - inputting the input datasets of the first training data (2001) into the first neural network (1) and in response receiving from the first neural network (1) first training output datasets (2011), - calculating a value of a first loss function dependent on the first training output datasets (2011) and the target datasets of the first training data (2001), - inputting the input datasets of the second training data (2002) into the second neural network (2) and in response receiving from the second neural network (2) second training output datasets (2012), - calculating a value of a second loss function dependent on the second training output datasets (2012) and target datasets of the second training data (2002), - calculating a value of a combined loss function dependent on a difference between the first training output datasets (2011) and the second training output datasets (2012), - adapting values of parameters of the first neural network (1) dependent on the value of the first loss function and the value of the combined loss function, - adapting values of parameters of the second neural network (2) dependent on the value of the second loss function and the value of the combined loss function.
  14. A computer program comprising machine executable instructions for execution by a computational system, wherein execution of the machine executable instructions causes the computational system to perform the method of claim 12 or 13.

Description

FIELD OF THE INVENTION The invention relates to a radar system. Furthermore, the invention relates to a method for operating a radar system. BACKGROUND DE 10 2020 201 025 A1 describes a radar sensor with several receiving antennas and several transmitting antennas arranged on a printed circuit board. A large proportion of the transmitting antennas are arranged at different positions with respect to a first axis and at the same position with respect to a second axis orthogonal to the first axis. A proportion of the receiving antennas are arranged at different positions with respect to the first axis and at the same position with respect to the second axis. SUMMARY OF THE INVENTION It is an objective to provide an improved radar system for differentiating objects in an environment of a vehicle and an improved method for operating a radar system for detecting objects in an environment of a vehicle. The objectives underlying the invention are solved by the features of the independent claims. In one aspect a radar system for differentiating objects in an environment of a vehicle is disclosed. The radar system comprises an evaluation unit. The evaluation unit comprises two differently trained neural networks and is configured to generate first input data by means of a first data processing algorithm. In addition, the evaluation unit is configured to generate the first input data dependent on receiving signals generated by means of received radar signals. Furthermore, the evaluation unit is configured to generate second input data by means of a second data processing algorithm and dependent on the receiving signals. The first data processing algorithm is different from the second data processing algorithm. Furthermore, the evaluation unit is configured to input the first input data into a first one of the neural networks and in response receive from the first neural network a first output. The first output indicates a differentiation between the objects. Furthermore, the evaluation unit is configured to input the second input data into a second one of the neural networks and in response receive from the second neural network a second output. The second output indicates a differentiation between the objects. Furthermore, the evaluation unit is configured to generate an output dependent on the first output and the second output. The output indicates a differentiation between the objects. In another aspect a method for operating a radar system for detecting objects in an environment of a vehicle is disclosed. The radar system comprises an evaluation unit, wherein the evaluation unit comprises two differently trained neural networks. The method comprises: generating first input data by means of a first data processing algorithm and dependent on receiving signals generated by means of received radar signals,generating second input data by means of a second data processing algorithm and dependent on the receiving signals, wherein the first data processing algorithm is different from the second data processing algorithm,inputting the first input data into a first one of the neural networks and in response receiving from the first neural network a first output, the first output indicating a differentiation between the objects,inputting the second input data into a second one of the neural networks and in response receiving from the second neural network a second output, the second output indicating a differentiation between the objects,generating an output dependent on the first output and the second output, the output indicating a differentiation between the objects. In another aspect a computer program is disclosed. The computer program comprises machine executable instructions for execution by a computational system, wherein execution of the machine executable instructions causes the computational system to perform the method for operating the radar system for detecting objects in the environment of the vehicle. BRIEF DESCRIPTION OF THE DRAWINGS In the following, examples are described in greater detail making reference to the drawings in which: Fig. 1 is a schematic of an a radar system comprising an evaluation unit with a first and a second neural network,Fig. 2 is a flow chart for describing a data flow within the evaluation unit shown in Fig. 1,Fig. 3 depicts the radar system shown in Figure 1 with receiving antennas,Fig. 4 depicts an exemplary range-Doppler power spectrum generated dependent on receiving signals generated by means of one of the receiving antennas shown in Figure 3,Fig. 5 illustrates an application example of the radar system shown in Figure 1 in a traffic situation of a vehicle comprising two objects in an environment of the vehicle,Fig. 6 depicts a dataflow indicating a computation of a spatial Fourier transform of phase information provided by power spectra generated on the basis of a set of receiving signals generated by means of the receiving antennas shown in Figure 3,Fig. 7 shows an example