Method and system for determining direction of arrival of radar detections

Through computer hardware and machine learning modules processing complex-value beam vectors, the accuracy problem of radar signal incident angle detection is solved, more efficient radar signal processing and object positioning is achieved, and the positioning capability of autonomous driving vehicles is improved.

CN114442029BActive Publication Date: 2024-11-12APTIV TECHNOLOGIES AG
View PDF 11 Cites 0 Cited by

Patent Information

Application Number
CN202111276487.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2020-11-02
Filing Date
2021-10-29
Publication Date
2024-11-12
Estimated Expiration
2041-10-29

AI Technical Summary

Technical Problem

The prior art is difficult to accurately detect the incident angle of radar signals, especially in autonomous driving vehicles, which affects the effective processing of radar signals and object positioning.

Method used

Using computer hardware components and machine learning modules, by processing complex-value beam vectors, using multiple dense layers and complex activation functions, the complex domain to real domain maps, extracting phase relationships, calculating multiple phase differences, and through multi-layer perception The device extends the angle and combines the output to determine the arrival direction of the radar detection.

Benefits of technology

It improves the detection accuracy of radar signal incident angle, can effectively extract information in complex environments, improves the accuracy of radar signal processing and the positioning ability of autonomous vehicles.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114442029B_ABST
    Figure CN114442029B_ABST
Patent Text Reader

Abstract

The present disclosure relates to a method and system for determining a direction of arrival detected by a radar, and more particularly to a computer-implemented method for determining a direction of arrival detected by a radar, comprising the following steps performed by a computer hardware component: obtaining a complex-valued beam vector detected by the radar; processing the complex-valued beam vector in a complex domain by a machine learning module; and obtaining a direction of arrival as an output of the machine learning module.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to methods and systems for determining the direction of arrival of radar detections. Background Art

[0002] Detecting the angle of incidence of a received radar signal is an important task in various applications, especially for at least partially autonomously driven vehicles.

[0003] Therefore, there is a need to provide an improved method of detecting the angle of incidence of a received radar signal. Summary of the invention

[0004] The present disclosure provides a computer-implemented method, a computer system and a non-transitory computer-readable medium. Implementation embodiments are given in the dependent claims, the description and the drawings.

[0005] In one aspect, the present disclosure relates to a computer-implemented method for determining a direction of arrival of a radar detection, the method comprising the following steps performed (in other words: implemented) by a computer hardware component: obtaining a complex-valued beam vector of the radar detection; processing the complex-valued beam vector in the complex domain by a machine learning module; and obtaining the direction of arrival as an output of the machine learning module.

[0006] In other words, a network for processing complex-valued radar antenna array responses may be provided. In other words, a complex-valued neural network for direction of arrival estimation may be provided.

[0007] According to another aspect, the machine learning module includes a plurality of dense layers.

[0008] According to another aspect, each dense layer in the plurality of dense layers is followed by a corresponding complex activation function.

[0009] According to another aspect, the machine learning module further comprises another dense layer.

[0010] According to another aspect, the another dense layer is followed by at least one of a squaring layer, a convolutional layer, or a real activation function.

[0011] Using the methods described herein, a machine learning module (e.g., an (artificial) neural network) can optimize an objective function with respect to network learnable parameters. Since the minimum of a function or series in the complex domain may not be well defined, the objective function can be expressed in the real domain. Using the methods described herein, data can be processed in the complex domain to extract as much information as possible (especially phase relationships). In order to formulate an objective function, a mapping from the complex domain to the real domain (e.g., combining the amplitude and phase information of a complex number) can be provided. The mapping can be followed by one or more real-valued convolutions and real-valued activations, but a complex-to-real domain mapping (e.g., abs-square) can also be preceded by one or more complex-valued convolutions. It should be understood that operations in the complex domain are not limited to dense layers.

[0012] According to another aspect, the computer-implemented method further includes the following steps performed by computer hardware components: determining a plurality of phase differences based on the complex-valued beam vector.

[0013] According to another aspect, the plurality of phase differences are determined for antennas spaced apart a predetermined distance and for antennas spaced apart at least approximately 50% greater than the predetermined distance.

[0014] According to another aspect, the computer-implemented method further includes the following steps performed by computer hardware components: expanding the plurality of angles based on the plurality of phase differences, for example using a shared MLP (Multi-Layer Perceptron).

[0015] According to another aspect, the another dense layer is followed by a squaring layer, and the computer-implemented method further comprises the following steps performed by a computer hardware component: combining the output of the squaring layer and a plurality of expansion angles. The combining may, for example, include expanding the dimensions (to match the dimensions of the output of the squaring layer and the dimensions of the plurality of expansion angles) and then concatenating data of the same dimensions.

[0016] According to another aspect, the machine learning module further comprises at least one of a convolutional layer, a real activation function, and a dense combination layer.

[0017] According to another aspect, the machine learning module further comprises a class probability activation function.

[0018] In another aspect, the present disclosure is directed to a computer system comprising a plurality of computer hardware components configured to perform some or all of the steps of the computer-implemented methods described herein.

[0019] The computer system may include a plurality of computer hardware components (e.g., a processor, such as a processing unit or a processing network; at least one memory, such as a memory unit or a memory network; and at least one non-transitory data storage device). It should be understood that additional computer hardware components may be provided and used to perform the steps of the computer-implemented method in the computer system. The non-transitory data storage device and / or the memory unit may include a computer program for instructing a computer, such as using a processing unit and at least one memory unit, to perform some or all steps or aspects of the computer-implemented method described herein.

[0020] In another aspect, the present disclosure relates to a vehicle comprising a computer system and a radar sensor as described herein.

[0021] According to another aspect, the radar sensor includes a plurality of antennas, wherein at least some of the antennas are spaced apart by a predetermined distance, and wherein at least some of the antennas are spaced apart by a distance that is at least substantially 50% greater than the predetermined distance.

[0022] In another aspect, the present disclosure relates to a non-transitory computer-readable medium comprising instructions for performing some or all steps or aspects of the computer-implemented methods described herein. The computer-readable medium may be configured as: an optical medium, such as a compact disc (CD) or a digital versatile disc (DVD); a magnetic medium, such as a hard disk drive (HDD); a solid-state drive (SSD); a read-only memory (ROM), such as a flash memory; and the like. In addition, the computer-readable medium may be configured as a data storage device accessible via a data connection such as an Internet connection. The computer-readable medium may be, for example, an online data repository or cloud storage.

[0023] The present disclosure also relates to a computer program for instructing a computer to perform several or all steps or aspects of the computer-implemented method described herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Exemplary embodiments and functions of the present disclosure are described herein in conjunction with the following schematically illustrated drawings:

[0025] Figure 1 is the architecture of a machine learning module according to various embodiments;

[0026] Figure 2 is a diagram of the virtual antenna array layout of a short-range radar;

[0027] Figure 3 It is a diagram of the virtual antenna array layout of a medium-range radar;

[0028] Figure 4 is the architecture of a machine learning module according to various implementations.

[0029] Figure 5 is a flow chart illustrating a method of determining a direction of arrival of a radar detection according to various embodiments;

[0030] Figure 6 is a direction of arrival determination system according to various embodiments; and

[0031] Figure 7 A computer system having a plurality of computer hardware components configured to perform the steps of a computer-implemented method of determining a direction of arrival of a radar detection according to various embodiments.

[0032] Reference numerals list

[0033] 100 Architecture of a machine learning module according to various embodiments

[0034] 102 Beam vector

[0035] 104 First dense layer

[0036] 106 First complex activation function

[0037] 108 Dot Line

[0038] 110 Another dense layer

[0039] 112 Square absolute value

[0040] 114 Convolutional Layer

[0041] 116 Real Activation Function

[0042] 118 Dot Line

[0043] 120 Convolutional Layer

[0044] 122 Real Activation Function

[0045] 124 Dense combination features

[0046] 126 Final activation function

[0047] 128 corner grid

[0048] 200 Illustration of a virtual antenna array layout for a short-range radar

[0049] 300 Illustration of the virtual antenna array layout of the medium-range radar

[0050] 400 Architecture of Machine Learning Module According to Various Implementations

[0051] 402 Calculate Phase Difference

[0052] 404 Calculate Φ1

[0053] 406 Calculation of Φ2

[0054] 408 Extension Φ1 and Φ2

[0055] 410 Extended Dimensions and Connections

[0056] 500 is a flowchart showing a method of determining a direction of arrival of a radar detection according to various embodiments.

[0057] 502 Steps for obtaining complex valued beam vector detected by radar

[0058] 504 Steps for processing complex-valued beam vectors in the complex domain by machine learning modules

[0059] 506 Step of obtaining the direction of arrival as the output of the machine learning module

[0060] 600 Arrival direction determination system according to various embodiments

[0061] 602 Beam vector acquisition circuit

[0062] 604 Machine Learning Circuit

[0063] 606 Arrival direction acquisition circuit

[0064] 608 connections

[0065] 700 Computer system according to various embodiments

[0066] 702 processor

[0067] 704 Memory

[0068] 706 Non-transitory data storage device

[0069] 708 Radar Sensor

[0070] 710 connections DETAILED DESCRIPTION

[0071] In order to locate objects that are radar reflective, the angle of incidence of the received EM (electromagnetic) waves needs to be determined.

[0072] According to various implementations, neural networks may be provided in a manner that respects the underlying complex domain and outperforms existing real-valued methods.

[0073] Figure 1 An architecture 100 of a machine learning module (eg, an artificial neural network) is shown having M densely connected complex-valued layers followed by K convolutional layers 120 whose kernels may be set to identity, according to various embodiments.

[0074] The network input may be a complex valued antenna array response (further referred to as beam vector) 102, which may be multiplied by a calibration matrix to correct for production defects.Although a neural network is also capable of learning calibration, calibration may be applied to generalize better since calibration is unique to each unit.

[0075] The beam vector 102 may be processed by one or more densely connected layers computing inner products of complex-valued inputs and a complex-valued weight matrix that may be optimized during training. A complex-valued nonlinear activation function is applied to the result of each inner product operation.

[0076] For example, the complex-valued beam vector 102 may be provided to a first dense layer 104, which may be followed by a first complex activation function 106. Further dense layers may be provided, each followed by a further complex activation, as indicated by dotted line 108.

[0077] Defining an activation function for complex-valued inputs is not a trivial task, since it should take into account both the phase and the absolute value of the input. According to various embodiments, a model with a complex-centered form (Equation 1) may be used as the activation function:

[0078]

[0079] Complex center type with its phase Attenuation Complex Value After the one or more densely connected layers, a last densely connected layer without an activation function (in other words: another dense layer) 110 may be applied. The squared absolute value of the output of this layer may be calculated 112, which may be considered as an activation function that transfers data from the complex domain to the real domain. The output may then be processed by operations in the real domain (e.g., one or more convolutional layers 114, 120) to refine the output and achieve position-invariant encoding. Each convolutional layer 114, 120 may be followed by a real activation function 116, 122. Further convolutional layers followed by real activation functions may be provided, as indicated by dotted lines 118.

[0080] The box “Densely combine features” 124 is used to combine features learned by convolution (in case they are not set as identities).

[0081] The final activation function 126 (e.g., softmax or sigmoid) can generate a class probability value for each output neuron, which corresponds to the angle of arrival, which can be provided in the angle grid 128. The angle grid 128 can be the output of the network (in other words: the angle grid 128 can be the classification result). The angle grid 128 can be interpreted as the probability of the target being reflected from the different angles represented by each cell of the angle grid 128.

[0082] According to various embodiments, a complex-valued neural network for direction of arrival (DoA) estimation for phased arrays utilizing sub-Nyquist sampling may be provided.

[0083] In order to apply the classical method to DoA estimation, a uniformly sampled linear array with respect to the Nyquist criterion is required. This means that the array elements need to be spaced at least λ / 2 (where λ is the wavelength of the signal) to sample the signal unambiguously. For larger spacings, the determination of DoA may become ambiguous.

[0084] Figure 2 Diagram 200 of a short range radar virtual antenna array layout is shown.

[0085] Figure 3 Diagram 300 of a mid-range radar virtual antenna array layout is shown.

[0086] Figure 4 An architecture 400 of a machine learning module is shown in accordance with various embodiments. Figure 4 The various parts can be Figure 1 The illustrated architecture 100 is similar or identical, so the same reference numerals may be used and repeated descriptions may be omitted.

[0087] The beam vector 102 may be provided to (eg Figure 1 ) first dense layer 104 in the network shown, and is used to Figure 4 Processing in the left branch is shown.

[0088] In addition, a beam vector 102 may be provided for Figure 4 Processing in the right branch (402, 404, 406, 408).

[0089] As a first preprocessing step, the virtual array needs to be collapsed so that all elements lie on one line ( Figure 3 The three vertically deviated elements in need to be moved downward). Afterwards, all λ-spaced pairs may be collected and all 1.5λ-spaced pairs may be collected, and the phase differences between these pairs may be calculated (402).

[0090] For the complex numbers containing the phase differences of λ and 1.5λ interval pairs, respectively, the complex numbers containing the phase differences of λ interval pairs will be further referred to as z1 and z2. In 404, Φ1 can be determined as the independent variable of the complex number in z1, and can contain the real-valued phase difference of the λ interval pair. In 406, Φ2 can be determined as the independent variable of the modified complex number derived from both z1 and z2, thereby deriving the real-valued phase difference from the (virtual) λ / 2 interval pair. This can be achieved by dividing z2 by z1. As a last step, in 408, Φ1 and Φ2 can be expanded to the size of the target angular grid. Therefore, a multi-layer perceptron (MLP) can be applied, which shares the weights of Φ1 and Φ2 to ensure that the two vectors will be expanded in the same way.

[0091] The expanded vectors from the similar "phase comparison" paths (402, 404, 406, 408) can enable the network to learn to pay attention to specific regions in the power spectrum-like vectors from the left path (104, 106, 108, 110, 112) of the network. The results from the left path and the results from the right path can be expanded in their dimensions and concatenated (410).

[0092] One or more convolutions 114, 120 may be applied to the three vectors stacked together, whose weights may be shared along the angular dimension. A final real-valued dense layer 124 combines the features generated by the convolutions, followed by an activation function 126, which produces the desired class probability value.

[0093] Figure 5 A flow chart 500 illustrating a method of determining a direction of arrival of a radar detection according to various embodiments is shown. At 502, a complex-valued beam vector of a radar detection can be obtained. At 504, the complex-valued beam vector can be processed in the complex domain by a machine learning module. At 506, a direction of arrival can be obtained as an output of the machine learning module.

[0094] According to various implementations, the machine learning module may include multiple dense layers.

[0095] According to various embodiments, each dense layer of the plurality of dense layers is followed by a corresponding complex activation function.

[0096] According to various embodiments, the machine learning module may further include another dense layer.

[0097] According to various embodiments, the further dense layer may be followed by at least one of a squaring layer, a convolutional layer, or a real activation function.

[0098] According to various embodiments, the method may further include determining a plurality of phase differences based on the complex-valued beam vector.

[0099] According to various embodiments, the plurality of phase differences are determined for antennas that are spaced apart by a predetermined distance and for antennas that are spaced apart by at least approximately 50% greater than the predetermined distance.

[0100] According to various embodiments, the method may further include expanding the plurality of angles based on the plurality of phase differences.

[0101] According to various embodiments, the further dense layer is followed by a squaring layer, and the method may further include combining an output of the squaring layer with the plurality of expansion angles.

[0102] According to various embodiments, the machine learning module may further include at least one of a convolutional layer, a real activation function, and a dense combination layer.

[0103] According to various embodiments, the machine learning module may further include a class probability activation function.

[0104] Each of steps 502, 504, 506, as well as the other steps described above, may be performed by computer hardware components.

[0105] Figure 6 A direction of arrival determination system 600 is shown according to various embodiments. The direction of arrival determination system 600 may include a beam vector acquisition circuit 602, a machine learning circuit 604, and a direction of arrival acquisition circuit 606.

[0106] The beam vector acquisition circuit 602 may be configured to acquire a complex-valued beam vector detected by the radar. The machine learning circuit 604 may be configured to process the complex-valued beam vector in the complex domain. The arrival direction acquisition circuit 606 may be configured to obtain an arrival direction as an output of the machine learning module.

[0107] The beam vector acquisition circuit 602, the machine learning circuit 604, and the direction of arrival acquisition circuit 606 may be coupled to one another, for example, by exchanging electrical signals via an electrical connection 608 such as a cable or a computer bus, or via any other suitable electrical connection.

[0108] A "circuit" may be understood as any type of logic-implemented entity, which may be a dedicated circuit or a processor executing a program stored in a memory, firmware, or any combination thereof.

[0109] Figure 7 A computer system 700 having a plurality of computer hardware components configured to perform the steps of a computer-implemented method for determining a direction of arrival of a radar detection according to various embodiments is shown. The computer system 700 may include a processor 702, a memory 704, and a non-transitory data storage device 706. A radar sensor 708 may be provided as part of the computer system 700 (e.g., Figure 7 as shown), or may be provided external to the computer system 700.

[0110] Processor 702 may execute instructions provided in memory 704. Non-transitory data storage 706 may store computer programs including instructions that may be transferred to memory 704 and then executed by processor 702. Radar sensor 708 may acquire radar measurements based on which a complex-valued beam vector may be determined.

[0111] The processor 702, the memory 704 and the non-transitory data storage device 706 may be connected to each other to exchange electrical signals, for example, via an electrical connection 710 (e.g., a cable or a computer bus) or via any other suitable electrical connection. The radar sensor 708 may be connected to the computer system 700, for example, via an external interface, or may be provided as part of the computer system (in other words: inside the computer system, for example, connected via the electrical connection 710).

[0112] The terms “coupled” or “connected” are intended to include a direct “coupled” (eg, via a physical link) or direct “connected” as well as an indirect “coupled” or indirect “connected” (eg, via a logical link), respectively.

[0113] It should be understood that what is described for one of the above methods may be similarly applicable to the direction of arrival determination system 600 and / or the computer system 700 .

Claims

1. A computer-implemented method for determining a direction of arrival of a radar detection, The computer-implemented method comprises the following steps performed by computer hardware components: - obtaining a complex-valued beam vector detected by the radar (502); - Processing the complex-valued beam vector in the complex domain by a machine learning module (504), wherein The machine learning module includes a plurality of dense layers and a square layer; - determining a plurality of phase differences based on the complex-valued beam vector, wherein the plurality of phase differences are determined for antennas spaced apart by a predetermined distance and for antennas spaced apart by at least 50% greater than the predetermined distance; - expanding a plurality of angles based on the plurality of phase differences; - combining the output of the squaring layer with a plurality of expansion angles; and - obtaining the direction of arrival as an output of the machine learning module (506).

2. The computer-implemented method according to claim 1, in, Each dense layer in the plurality of dense layers is followed by a corresponding complex activation function.

3. The computer-implemented method according to claim 1 or 2, in, The machine learning module further includes another dense layer after the plurality of dense layers.

4. The computer-implemented method according to claim 3, in, The another dense layer is followed by at least one of a square layer, a convolutional layer, or a real activation function.

5. The computer-implemented method of claim 4, in, The further dense layer is followed by the square layer.

6. The computer-implemented method of claim 1, in, The machine learning module also includes at least one of a convolutional layer, a real activation function, and a dense combination layer.

7. The computer-implemented method of claim 1, in, The machine learning module also includes a class probability activation function.

8. A computer system (700) comprising a plurality of computer hardware components configured to perform the steps of the computer-implemented method according to any one of claims 1 to 7.

9. A vehicle comprising the computer system (700) according to claim 8 and a radar sensor (708).

10. The vehicle according to claim 9, in, The radar sensor (708) includes a plurality of antennas, wherein at least some of the antennas are spaced apart by a predetermined distance, and at least some of the antennas are spaced apart by a distance that is 50% greater than the predetermined distance.

11. A non-transitory computer readable medium comprising instructions for executing the computer-implemented method according to any one of claims 1 to 7.

Citation Information

Patent Citations

  • Massive MIMO DOA system based on neural network and implementation method

    CN111366892A

  • Smart-device-based radar system performing angular estimation using machine learning

    CN111433627A

  • Deep learning direction finding method based on phase optimization

    CN111580042A

  • Removal of false alarms during the club formation phase for radar detection using a deep neural network

    DE102020102712A1

  • Neural network-based radar system

    US20160097853A1