Radar signal direction of arrival super-resolution estimation

Abstract

An apparatus includes a radar processor that transmits a first radar signal at a first time, receives a received signal, and processes the received signal to generate a range-Doppler data frame. The radar processor determines a first snapshot including a first plurality of values associated with a first range-Doppler bin of the range-Doppler data frame, and processes the first plurality of values in the first snapshot to generate an autoregression model based on the first plurality of values. The radar processor extrapolates a second snapshot using the autoregression model, wherein the second snapshot includes the first plurality of values and a second plurality of values generated using the autoregression model; determining a full rank covariance matrix using the second snapshot; and identifying attributes of a plurality of objects using the full rank covariance matrix.

Description

Technical Field

[0001] The present disclosure generally relates to radar systems and associated methods of operation. In one aspect, the present disclosure relates to a radar system configured to determine the direction of arrival of an object using a super-resolution estimation method. Background Art

[0002] A radar system transmits an electromagnetic signal and receives the backscatter of the transmitted signal. The time delay and / or time delay variation between the transmitted signal and the received signal can be determined and used to calculate the distance and / or speed of the object causing the reflection, respectively. For example, in automotive applications, an automotive radar system can be used to determine the distance and / or speed of oncoming vehicles and other obstacles.

[0003] Automotive radar systems enable advanced driver assistance system (ADAS) functions, which may make driving increasingly safe and ultimately enable a fully autonomous driving platform. Such systems use a radar system as the primary sensor for ADAS operation.

[0004] The accuracy of an automotive radar system may be affected by the inaccuracies in the positions of the various transmit and receive antennas of the radar system. Therefore, techniques are needed to improve the effective accuracy of signals from real-world radar systems to increase the likelihood of object detection and the determination of the attributes of those objects (e.g., direction of arrival). Summary of the Invention

[0005] According to a first aspect of the present invention, there is provided a radar system comprising:

[0006] at least one transmitter and at least one receiver, wherein the at least one transmitter and the at least one receiver are configured to transmit and receive radar signals, and wherein the at least one transmitter and the at least one receiver are coupled to a vehicle; and

[0007] a radar processor configured to:

[0008] transmit a first radar signal at a first time,

[0009] receive the received signal using the at least one receiver,

[0010] process the received signal to generate a range-Doppler data frame,

[0011] determine a first snapshot comprising a first plurality of values associated with a first range-Doppler bin of the range-Doppler data frame,

[0012] Process the first plurality of values in the first snapshot to generate an autoregressive model based on the first plurality of values.

[0013] Use the autoregressive model to extrapolate a second snapshot, where the second snapshot includes the first plurality of values and a second plurality of values generated using the autoregressive model.

[0014] Use the second snapshot to determine a full-rank covariance matrix.

[0015] Use the full-rank covariance matrix to identify attributes of a plurality of objects, and

[0016] Send the attributes of the plurality of objects to a vehicle controller.

[0017] In one or more embodiments, the autoregressive model is configured to approximate the equation where a is a matrix containing the prediction coefficients of the autoregressive model, r x is a cross-correlation vector, and is the inverse matrix of the cross-correlation matrix.

[0018] In one or more embodiments, the radar processor is configured to process the first plurality of values in the first snapshot to generate the autoregressive model using the Burg method, the Yule Walker equation, or the Levinson method.

[0019] In one or more embodiments, the radar processor is configured to identify the attributes of the plurality of objects by:

[0020] Determine a set of eigenvalues of the full-rank covariance matrix; and

[0021] Use the set of eigenvalues to determine the attributes of the plurality of objects.

[0022] In one or more embodiments, the radar processor is configured to identify the attributes of the plurality of objects by:

[0023] Use at least one of the MUSIC spectral estimation algorithm and the CAPON spectral estimation algorithm to generate a pseudospectrum using the full-rank covariance matrix; and

[0024] Detect peaks in the pseudospectrum.

[0025] In one or more embodiments, the first snapshot includes a first number of values equal to the number of physical or virtual arrays in the radar system.

[0026] In one or more embodiments, the second snapshot includes a second number of values equal to twice the first number of values.

[0027] In one or more embodiments, the full-rank covariance matrix is a matrix having dimensions N×N, where the value of N is equal to the number of the second values divided by 1.5.

[0028] In one or more embodiments, the first snapshot includes four values, and the second snapshot includes eight values.

[0029] In one or more embodiments, the full-rank covariance matrix is a matrix having dimensions 5×5.

[0030] According to a second aspect of the present invention, there is provided a radar system, comprising:

[0031] at least one transmitter and at least one receiver; and

[0032] a radar processor configured to:

[0033] process the received signal to determine a first snapshot including a first plurality of values associated with a first range-Doppler bin of a range-Doppler data frame,

[0034] determine a second snapshot, wherein the second snapshot includes the first plurality of values and a second plurality of values generated using an autoregressive model,

[0035] use the second snapshot to determine a full-rank covariance matrix, and

[0036] use the full-rank covariance matrix to identify attributes of a plurality of objects.

[0037] In one or more embodiments, when determining the second snapshot, the radar processor performs the following steps:

[0038] process the first snapshot to generate an autoregressive model based on the first plurality of values; and

[0039] use the autoregressive model to extrapolate the second snapshot, wherein the second snapshot includes the first plurality of values and a second plurality of values generated using the autoregressive model.

[0040] In one or more embodiments, the autoregressive model is configured to approximate the equation where a is a matrix containing the prediction coefficients of the autoregressive model, r x is a cross-correlation vector, and is the inverse matrix of the cross-correlation matrix.

[0041] In one or more embodiments, the radar processor is configured to process the first plurality of values in the first snapshot to generate the autoregressive model using the Burg method, the Yule Walker equation, or the Levinson method.

[0042] In one or more embodiments, the first snapshot includes a first number of values equal to the number of physical or virtual arrays in the radar system.

[0043] In one or more embodiments, the second snapshot includes a second number of values equal to at least twice the first number of values.

[0044] In one or more embodiments, the full rank covariance matrix is a matrix having dimensions N×N, where the value of N is at least equal to the second number of values divided by 1.5.

[0045] According to a third aspect of the present invention, there is provided a method comprising:

[0046] Processing the received signal to generate a range-Doppler data frame;

[0047] Determining a first snapshot including a first plurality of values associated with a first range-Doppler bin of the range-Doppler data frame;

[0048] Processing the first plurality of values in the first snapshot to generate an autoregressive model based on the first plurality of values;

[0049] Using the autoregressive model to extrapolate a second snapshot, wherein the second snapshot includes the first plurality of values and a second plurality of values generated using the autoregressive model;

[0050] Using the second snapshot to determine a full rank covariance matrix; and

[0051] Using the full rank covariance matrix to identify attributes of a plurality of objects.

[0052] In one or more embodiments, the method further includes processing the first plurality of values in the first snapshot to generate an approximate equation of the autoregressive model, where a is a matrix containing the prediction coefficients of the autoregressive model, r x is the cross-correlation vector, and is the inverse matrix of the cross-correlation matrix.

[0053] In one or more embodiments, the method further includes processing the first plurality of values in the first snapshot to generate the autoregressive model using the Burg method, the Yule Walker equation, or the Levinson method.

[0054] These and other aspects of the invention will be apparent from, and elucidated with reference to, the embodiments described hereinafter. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] A more complete understanding of the subject matter can be obtained by reference to the detailed description and claims considered in conjunction with the following drawings, in which like reference numerals refer to like elements throughout.

[0056] Figure 1 A simplified schematic block diagram depicting an automotive radar system.

[0057] Figure 2 Depicting an input signal that can be associated with a particular range-Doppler snapshot of received reflected signals captured by a 4-element receive array used by a vehicle radar system.

[0058] Figure 3 Depicting the original input signal that is increased in size by adding autoregressive prediction values at the start and end of the input signal. Figure 2 of the original input signal.

[0059] Figure 4 Depicting an algorithm that can be used to perform forward-backward spatial smoothing (FBSS) on the input signal.

[0060] Figure 5 Is a graph depicting the probability (vertical axis) of separating or identifying two and three objects when using conventional direction-of-arrival (DoA) estimation.

[0061] Figure 6 Depicting the performance of the MUSIC algorithm in a conventional signal processing chain when the phased array is perfectly calibrated.

[0062] Figure 7 Depicting the performance of the MUSIC algorithm with the super-resolution signal processing method of the present invention when the phased array is perfectly calibrated.

[0063] Figure 8 Depicting the probability of object separation of two objects when using the conventional CAPON algorithm when the phased array is perfectly calibrated.

[0064] Figure 9 Depicting the performance of the CAPON algorithm with the super-resolution signal processing method of the present invention when the phased array is perfectly calibrated.

[0065] Figure 10 Depicting the performance of the MUSIC algorithm in a conventional signal processing chain when the phased array has a residual calibration error.

[0066] Figure 11Describe the performance of the MUSIC algorithm of the super-resolution signal processing method of the present invention when the phased array has residual calibration errors.

[0067] Figure 12 Describe the probability of object separation of two objects using the conventional CAPON algorithm when the phased array has residual calibration errors.

[0068] Figure 13 Describe the performance of the CAPON algorithm of the super-resolution signal processing method of the present invention when the phased array has residual calibration errors. DETAILED DESCRIPTION

[0069] The following detailed description is merely illustrative in nature and is not intended to limit the application and use of such embodiments of the subject matter of such embodiments. As used herein, the words "exemplary" and "example" mean "serving as an example, instance, or illustration." Any embodiment or example described herein as "exemplary" or "example" is not necessarily to be construed as preferred or superior to other embodiments. Additionally, there is no intention to be bound by any theory, whether explicit or implicit, present in the foregoing technical field, background art, or the following detailed description.

[0070] Automotive radar systems are used to support advanced driver assistance system (ADAS) functions such as adaptive cruise control, emergency braking, and blind spot monitoring and warning. The spatial resolution of a phased array radar system depends on the size of the aperture, which can be physical, virtual (e.g., in a radar system with a multiple-input multiple-output (MIMO) configuration), or synthetic (e.g., as in a synthetic aperture radar (SAR)). Although the latter is outperformed by the former two in certain scenarios, due to the specific geometric shape constraints of the SAR system, radar systems with this configuration are generally not suitable for automotive applications. Therefore, automotive radars tend to rely strictly on the physical separation between the receiver (RX) and transmitter (TX) antennas of the radar system (e.g., arranged in a MIMO array). For practical considerations, this separation is typically very small (e.g., on the order of millimeters). Therefore, it is very difficult to position the antenna elements of the radar system with sufficient accuracy, which results in a reduced resolution of the actual operating radar system (i.e., compared to an ideal radar system with perfect RX and TX antenna placement).

[0071] To attempt to compensate for the reduced natural resolution of real-world radar systems, super-resolution algorithms have been proposed to improve the apparent resolution of such systems. However, in conventional super-resolution methods, such algorithms typically require certain conditions to operate properly, such as a relatively high signal-to-noise ratio (SNR), a known antenna separation, a narrowband plane wave approximation, and no antenna coupling or channel trace mismatch errors.

[0072] In conventional radar system design, super-resolution is only applied to range-Doppler detection, making the processing less exhaustive and the signal-to-noise ratio (SNR) high. However, these super-resolution methods may not be particularly successful because it is impossible to ensure perfect antenna spacing and completely no antenna coupling or trace mismatch errors in real-world radar systems, especially when considering the large temperature variations seen in automotive applications.

[0073] Accordingly, the present disclosure provides a method for processing radar system signals, which can improve the robustness of super-resolution algorithms to such inevitable system-level errors, such as those manifested in antenna placement errors. This is achieved by using an autoregressive model to extrapolate the values of the signal being processed before applying enhanced object estimation techniques. The autoregressive model can extrapolate harmonic signals, such that parasitic modulation of the spatial signal is ignored. The autoregressive model alone cannot provide sufficient resolution to be used alone, but can sufficiently enhance the received signal to enable the super-resolution algorithm to correctly extract the desired information.

[0074] Covariance matrix-based super-resolution methods typically require multiple snapshots of radar signal data to construct a full-rank covariance matrix for proper object detection. When only single-snapshot measurements are available, techniques such as spatial smoothing or Hankel matrix formation are typically required to construct a smaller but full-rank covariance matrix. However, in conventional radar systems, the reduced size of such covariance matrices has a significant adverse effect on both the resolution and the total number of objects that can be detected through such signal processing. To overcome these deficiencies, the present disclosure provides methods for constructing a full-rank covariance matrix that enables the super-resolution algorithm to detect a more comprehensive set of objects in a single-snapshot measurement environment.

[0075] Now refer to Figure 1 , Figure 1 FIG. 13 depicts a simplified schematic block diagram of an automotive radar system 100. The exemplary automotive radar system 100 includes a MIMO radar device 10 connected to a radar controller processor 20. In the selected embodiment, the MIMO radar device 10 can be embodied as a line-replaceable unit (LRU) or modular component designed to be quickly replaced in an operating position. Similarly, the radar controller processor 20 can be embodied as a line-replaceable unit (LRU) or modular component. Although a single or monostatic MIMO radar device 10 is shown, it should be understood that additional distributed radar devices can be used to form a distributed or multistatic radar. Additionally, the depicted radar system 100 can be implemented in integrated circuit form, where the LFM TD-MIMO radar device 10 and the radar controller processor 20 are formed by separate integrated circuits or chips or by a single chip, depending on the application.

[0076] Each radar device 10 includes one or more transmit antenna elements TXi and receive antenna elements RXj respectively connected to one or more radio frequency (RF) transmitter (TX) units 11 and receiver (RX) units 12. For example, each radar device (e.g., 10) is shown as including respective antenna elements (e.g., TX1,i, RX1,j) connected to three transmitter modules (e.g., 11) and four receiver modules (e.g., 12) respectively, but these numbers are not restrictive and other numbers are possible, such as four transmitter modules 11 and six receiver modules 12, or a single transmitter module 11 and / or a single receiver module 12.

[0077] Each radar device 10 further includes a chirp generator 112 configured and connected to supply a chirp input signal to the transmitter module 11. For this purpose, the chirp generator 112 is connected to receive a separate and independent local oscillator (LO) signal 110 and a chirp start trigger signal 111, but due to signal path differences and programmable digital delay elements in the signal path, the delays are likely to be different. A chirp signal 113 is generated and is typically sent to the plurality of transmitters 11 according to a predefined transmission schedule, where the chirp signal 113 is filtered at the RF conditioning module 114 and amplified at the power amplifier 115, and then fed to the corresponding transmit antenna TX1,i and radiated. By sequentially using each transmit antenna TX1,i to transmit consecutive pulses in the chirp signal 113, each transmitter element 11 operates in a time-division multiplexing manner relative to the other transmitter elements because these transmitter elements are programmed to transmit exactly the same waveform according to a temporally separated schedule.

[0078] The radar signals transmitted by the transmitter antenna units TX1,i and TX2,i can be reflected by an object, and a part of the reflected radar signal arrives at the receiver antenna unit RX1,i at the radar device 10. At each receiver module 12, the received (radio frequency) antenna signal is amplified by a low noise amplifier (LNA) 120 and then fed to a mixer 121, where the received (radio frequency) antenna signal is mixed with the transmitted chirp signal generated by the RF conditioning module 114. Specifically, the mixer 121 is configured to down-convert the RF signal to a baseband signal, which means that during the acquisition window, the object signal causes the superposition of constant-amplitude sine signals, while the interference signal generates a broadband chirp-like signal (i.e., time-limited after anti-aliasing filtering). The resulting intermediate frequency signal is fed to a first high-pass filter (HPF) 122. The resulting filtered signal is fed to a first variable gain amplifier 123, which amplifies the signal before feeding the filtered signal to a first low-pass filter (LPF) 124. This re-filtered signal is fed to an analog-to-digital converter (ADC) 125 and output as a digital signal D1 by each receiver module 12. In various embodiments, the receiver module compresses the object echoes of various delay signals into a plurality of sinusoidal frequency tones, the frequencies of the plurality of sinusoidal frequency tones corresponding to the round-trip delay of the echo.

[0079] The radar system 100 further includes a radar controller processing unit 20, which is connected to supply an input control signal to the radar device 10 and receive the digital output signal generated by the receiver module 12 therefrom.

[0080] In the selected embodiment, the radar controller processing unit 20 can be embodied as a microcontroller unit (MCU) or other processing unit configured and arranged for signal processing tasks such as, but not limited to, object recognition; calculation of object distance, object speed, and object direction; and generation of control signals. For example, the radar controller processing unit 20 can be configured to generate calibration signals, receive data signals, receive sensor signals, generate spectrum shaping signals (such as ramp generation in the case of a frequency-modulated continuous wave (FMCW) radar), and / or register programming or state machine signals for an RF (radio frequency) circuit enable sequence.

[0081] Additionally, the radar controller processor 20 may be configured to program module 11 to operate in a time-division manner by sequentially transmitting LFM chirps for coordinated communication between transmit antennas TX1,i and RX1,j. The result of digital processing at the radar controller processing unit 20 is to process the digital domain signal D1 through a fast-time range-frequency Fourier transform (FFT) 21, a slow-time FFT 22 for Doppler compression, and a constant false alarm rate (CFAR) object detection 23. For an object identified by the CFAR object detection 23, according to the present disclosure, a direction-of-arrival (DoA) analysis is performed using a DoA super-resolution processing algorithm 24 to determine the DoA of the object identified by the CFAR object detection 23. In the case where the DoA is determined, object tracking 25 processes the object range, Doppler, and DoA information to predict the path of the object, and the result is output 26 to other automotive computing or user interface devices for further processing or display. When performing object detection, the radar controller processor 20 is configured to determine the range image of the reflected radar signal, where the peak in the range image indicates a potential object at a known range.

[0082] During operation of the automotive radar system 100, the radar controller processor 20 sends control signals to various components of one or more radio frequency (RF) transmitter (TX) units 11 and receiver (RX) units 12 to control how the radar signal transmitted by the automotive radar system 100, which is the transmitted signal, is shaped and how the received signal is processed.

[0083] A phased array radar (e.g., Figure 1 the radar system 100 depicted in Figure 1 has a natural spatial resolution that may depend on the size of the aperture of the radar system, which may be physical, virtual (MIMO), or synthetic (SAR). Generally, in automotive applications, a radar system relies on the physical separation of the RX and TX antennas of the system (e.g.,

[0084] However, in conventional radar systems, super-resolution algorithms may require certain conditions to operate properly, such as high SNR, known antenna spacings, narrowband plane wave approximation, and no antenna coupling / trace mismatch errors. In most radar system designs, super-resolution is only applied to range-Doppler detection, such that the processing is less exhaustive and the SNR is high. However, known antenna spacings and no antenna coupling / trace mismatch errors may not be ensured, especially when considering the large temperature variations seen in automotive applications.

[0085] To compensate for these difficulties, the present disclosure provides a processing method that improves the robustness of super-resolution algorithms against these inevitable system-level deficiencies. Specifically, a signal processing pipeline (e.g., the pipeline implemented by the radar controller processor 20 of the radar system 100 of Figure 1 uses an autoregressive model to extrapolate and effectively increase the size of the input signal for the current snapshot, thereby increasing the size of the signal samples and the corresponding covariance matrix. A larger (compared to conventional methods) covariance matrix is then used to perform enhanced estimation. As described herein, the autoregressive model can extrapolate only harmonic signals, such that the model ignores the parasitic modulation of the spatial signals. Although the autoregressive model alone does not provide sufficient resolution to be used alone, this method sufficiently increases the size of the received signal snapshots such that the applied super-resolution algorithm can correctly extract the desired information.

[0086] In addition, covariance matrix-based super-resolution methods typically require multiple snapshots to construct a full-rank covariance matrix. Thus, when only single snapshot measurements are available, techniques such as spatial smoothing or Hankel matrix formation are conventionally required to construct a smaller but full-rank covariance matrix. The reduced size of the covariance matrix can have a significant detrimental effect on the resolution and the total number of objects that can be detected by such conventional signal processing methods, which in some cases can reduce the total number of objects that can be detected by two or more. In contrast, the signal processing method of the present disclosure enables the super-resolution signal processing algorithm to detect a larger number of objects that would otherwise be limited by the original array size.

[0087] In a particular embodiment of the present disclosure, pole-based prediction can be used to extrapolate the dominant spectral components of the input radar system snapshot signal, thereby reducing the effects of parasitic modulation, such as unresolved cross-coupling errors or phase / trace mismatches between phased array channels. Covariance spatial smoothing can then be used such that the size of the determined covariance matrix is equal to the number of the original covariance matrix or +1. This can enable the construction of a larger signal null space (noise subspace) and the identification of more objects than may be identified using conventional spatial smoothing covariance matrix methods.

[0088] Accordingly, the present disclosure provides an improved DoA estimation process that is robust to calibration errors and is also capable of detecting more objects than may typically be detected using conventional signal processing methods.

[0089] When using a super-resolution covariance matrix-based spectral estimation technique, a full-rank covariance matrix is required. In a radar system using single-snapshot estimation, a non-singular covariance matrix can be obtained via spatial smoothing or Hankel formation. This method can result in a reduction in aperture size by a factor of 1.5 to 2 (reflected in the reduction in the size of the covariance matrix) and additional signal degradation.

[0090] Thus, the effect of covariance spatial smoothing is a reduction in the spatial resolution of the radar system and a reduction in the number of maximum detectable objects (e.g., by approximately half).

[0091] To mitigate these detrimental effects of conventional super-resolution covariance matrix-based spectral estimation, in the present disclosure, a prediction filter (e.g., an autoregressive model) is used to effectively increase the size of the input signal, such that artificially larger snapshots are generated, thereby achieving improved system performance.

[0092] For illustration, Figure 2 An input signal 202 that may be associated with a particular range-Doppler snapshot of the received reflected signals captured by a vehicle radar system is depicted in block form. In this example, the individual blocks of the input signal 202 may represent samples of an analog input signal, where the samples are associated with particular range-Doppler bins. Once captured, the input signal 202 is processed to determine a corresponding covariance matrix 204. As described herein, the covariance matrix 204 is then used to perform object detection and estimation of object range, velocity, and direction of arrival.

[0093] As discussed herein, in the case where the size of the input signal 202 is limited (e.g., due to the finite size of the input signal 202), the overall accuracy and resolution of the radar system output and even the total number of objects that can be detected by the system can be significantly reduced.

[0094] To remedy these deficiencies, the present disclosure provides a method in which autoregressive estimation is used to increase the signal of the input signal being processed. The increase in the size of the input signal results in the generation of a larger full-rank covariance matrix, thereby increasing the resolution and accuracy of object estimation in the radar system.

[0095] For illustration, Figure 3 The original input signal 202 is depicted, the size of which is increased by adding autoregressive prediction values 302 at the start of the input signal 202 and prediction values 304 at the end of the input signal 202.

[0096] As described herein, in the case where the values of the samples of the input signal 202 are designed to be x 1:N , the value of the predicted value 302 added at the start of the input signal 202 can be specified as x -Nx+1:0 , where Nx is the total number of samples in the supplementary input signal 306 that combines the input signal 202 with the predicted values 302 and 304. Similarly, the predicted value 304 at the end of the input signal 202 can be specified as X N+1:Nx .

[0097] As Figure 3 shown, the supplementary input signal 306 generates a larger full-rank covariance matrix when being processed. The dimension of the larger full-rank covariance matrix 308 can depend on various factors, including the covariance smoothing technique used, whether the smoothing technique is forward, forward-backward or Hankel, the number of objects being estimated, and the input signal size. When the radar performs object detection, the larger full-rank covariance matrix 308 can achieve higher accuracy and resolution.

[0098] Figure 4 Depicts an algorithm that can be used to perform forward-backward spatial smoothing (FBSS) on the input signal. FBSS is typically used as a preprocessor during signal processing that can successfully generate a full-rank covariance matrix based on the input signal. Specifically, FBSS is used to ensure that the input signal is properly decorrelated. As shown, the FBSS algorithm receives the originally captured signal (e.g., Figure 2 the input signal 202) x and K as inputs, where K represents the size of the sliding window that will be used to generate the estimates. Then the variable N is defined as the dimension (e.g., size) of the input signal x.

[0099] Given those input variables, the algorithm is executed, where the value of n is incremented from an initial value of 1 to the value of N (the size of the input signal) minus K (the size of the sliding window plus 1).

[0100] As n is incremented, the values of R F,n and R B,n are determined.

[0101] Input: x, K

[0102] Output:

[0103] N = dim(x)

[0104] For n = 1, … N-K+1 do

[0105]

[0106] end loop

[0107]

[0108] Then the value R can be calculated according to the following expression F , R B and R S :

[0109]

[0110] denotes the forward smoothing covariance matrix constructed using x(n), is the backward smoothing covariance matrix, where the same operation is applied in reverse, and denotes and the average value of and The maximum rank of is N / 2, while the maximum rank of

[0111] Given a particular input signal snapshot, the present disclosure provides a method of increasing the size of the input signal to generate a larger covariance matrix. This in turn increases the total number of detectable objects of a particular radar system. Because a certain degree of signal estimation is involved, for a perfect signal (e.g., a signal captured by a radar system where the antennas of the system are positioned with perfect accuracy), there may be a slightly lower performance due to prediction errors, which slightly reduces the spectral estimation. However, if the phased array is properly calibrated, which is the case for almost all real-world radar systems, the input signal is affected by multiplicative noise, and the prediction errors involved in the present method become negligible, enabling the present method to significantly enhance the overall system performance. Specifically, current harmonic-based predictions (e.g., autoregressive (AR) predictions) reconstruct only a single frequency tone per pole and are unable to extrapolate broadband signals, such as the envelope / modulation resulting from calibration errors.

[0112] For example, in a particular embodiment of the present disclosure, the size of the input signal is increased by extrapolation / prediction involving the construction of an autoregressive model that fits the input signal data x, the input signal data x including a plurality of sampled values for a particular snapshot of a particular range-Doppler bin of a processed radar signal. The model is typically configured to approximate the following equation:

[0113]

[0114] where a is a matrix containing the prediction coefficients of the autoregressive model, r x is the cross-correlation vector, and is the inverse matrix of a cross-correlation matrix (which is Hermitian and Toeplitz), not to be confused with the covariance matrix denoted as R. In various embodiments, the prediction coefficients of an autoregressive model can be generated using established signal processing techniques such as the Burg method, the Yule-Walker equations, or the Levinson method. Generally, these methods take an input signal x as input and generate the coefficients of the autoregressive model as output. In an example, the model can then be represented using the Z-transform as follows:

[0115]

[0116] Predictions can then be generated using the following equation (e.g., Figure 3 predicted values 302 and 304 in the supplementary input signal 306 of), where the value n represents the index of the value being predicted positively. The function y() depends on the coefficient values a of the autoregressive model generated by the above methods.

[0117]

[0118] where the value p is a fine-tuning value that represents the number of objects for which an attribute (e.g., direction of arrival) is being determined. The value n represents the index of the current value being generated.

[0119] Thus, using this method, the given input signal x includes a snapshot of a processed range-Doppler frame of the received radar signal. The input signal has a number k of values. The input signal x is processed (e.g., via the Burg method, the Yule-Walker equations, or the Levinson method) to generate a matrix a containing the prediction coefficients of an autoregressive model, which can be used to predict additional values of the input signal x based on the existing values contained within the input signal x. Given the autoregressive model prediction coefficients, the autoregressive model can be used to predict additional values to be added to the input signal x to increase its size. The value p is selected, where the value p is a fine-tuning value that represents the number of objects for which an attribute (e.g., direction of arrival) is being determined.

[0120] In an embodiment, additional values are generated using the function y(n) (i.e., equation (1) above), where n is the index of the value being presented. Thus, to generate, for example, three additional values to append to the end of the input signal x, equation (1) is executed using the values n = k + 1, n = k + 2, and n = k + 3, where n is a value representing a discrete time interval. Positive values of n represent time values to the right, while negative values of n represent time intervals to the left. Similarly, to generate three additional values to insert at the start of the input signal x, equation (1) is executed using the values n = -1, n = -2, and n = -3.

[0121] By using the above autoregressive method to calculate the predicted values, a larger (effectively) input signal can be generated, and in turn, a larger full-rank covariance matrix can be calculated for the given input signal. Using the larger full-rank covariance matrix, various estimation methods (e.g., MUSIC or CAPON) can be used to perform spectral estimation, enabling the detection of objects in the reflected radar signal and determination of the distance, velocity, and direction-of-arrival information of the detected objects.

[0122] In still other embodiments of the present disclosure, other methods can be utilized to estimate or predict the generated values for increasing the dimension of the input signal, such as a neural network-based prediction network.

[0123] Once the input signal has been processed via an autoregressive model (or other method) to increase the size of the input signal and generate the corresponding covariance matrix, various spectral estimation methods can be used to analyze the covariance matrix to determine object properties.

[0124] One such spectral estimation method, MUSIC, requires decomposing the full-rank covariance matrix R into its eigenvalue components: R = R S +R W = VΛV H . The space generated by the eigenvector V can be divided into two subspaces V = [S G]. In this case, S is a matrix with the eigenvectors corresponding to the largest eigenvalues as the columns of the matrix and represents the signal subspace, and G is formed by the eigenvectors corresponding to the smallest eigenvalues and represents the noise subspace.

[0125] By multiplying the right-side eigenvalue decomposition equation by V H , the equation becomes:

[0126]

[0127] The matrix equation can then be rewritten in a system form as:

[0128] Given that the last system equation can be rewritten as:

[0129] Since AP cannot be 0, then A H G = 0, which means that the noise subspace is orthogonal to the subspace generated by the steering vector at the desired angle of arrival, e H (ω i )G = 0.

[0130] The MUSIC algorithm is based on the orthogonality between the signal subspace and the noise subspace. It is known that: for ω = ω i , e H(ω)G = 0, where ω i is the angular frequency of the desired signal. For the real-valued representation of the pseudo-spectrum, the Euclidean norm can be defined as follows:

[0131] For ω = ω i , ‖e H (ω)G‖ 2 = e H (ω)GG H e(ω) = ~0.

[0132] In a conventional signal processing system, the estimation error of the covariance matrix is caused by a relatively small number of signal samples (given by the number of antennas), and the norm is not exactly 0 at the angles of interest. This problem can be alleviated by using an autoregressive method to increase the size of the processed input signal and thus increase the dimension of the covariance matrix. In addition, in this algorithm, e H (ω)G can be calculated as iFFT(G) using an FFT accelerator. The MUSIC pseudo-spectrum that can have a significant maximum at the angles of interest (i.e., the DoA angles of the identified objects) is then defined as:

[0133]

[0134] Another spectral estimation method, CAPON, which can be used to process the covariance matrix generated by this signal processing method, is not based on the signal subspace but on filtering the input signal spectrum of s[k]. Using this algorithm, the conditions imposed are: for the filter, the coefficient h must have unit gain within the passband: h H e(ω) = 1 and the output power must be at a minimum:

[0135]

[0136] This optimization problem can be solved as follows using the first-order Lagrangian optimization method:

[0137] L(h) = h H Rh + λ(e(ω)h H - 1) + λ * (e H (ω)h - 1)

[0138] Minimizing the Lagrangian corresponds to In addition,

[0139] The last two equations can be combined to yield: And the CAPON pseudo-spectrum is defined as the output power of h:

[0140]

[0141] By using the present method to increase the effective length of the input signal to the object estimation signal processing chain of a vehicle radar system, a larger covariance matrix can be generated, and the total number of objects detectable by the radar system can be increased. For a perfect signal (i.e., generated by a perfectly implemented and unrealistic radar system), there is a slightly lower performance caused by the prediction error of the autoregressive model, which slightly reduces the spectral estimation. However, if the phased array is not properly calibrated (i.e., most real-world radar systems), resulting in the input signal being affected by multiplicative noise, the prediction error of the present autoregressive method becomes negligible, enabling the present method to significantly enhance the overall system performance of real-world applications. This is because the harmonic-based prediction (e.g., the prediction provided by the present autoregressive estimation method) reconstructs only a single frequency tone per pole and cannot extrapolate broadband signals, such as the envelope / modulation generated by calibration errors.

[0142] In the simulation of the super-resolution signal processing method of the present invention, two sets of simulated DoA estimation results are generated, one for a perfectly calibrated phased array (i.e., an 'ideal' radar system that does not exist in the real world), and the other for a non-ideally calibrated (i.e., 'real-world') phased array.

[0143] In the simulation, two spectral estimators MUSIC and CAPON are utilized, but it should be understood that other spectral estimators can be combined with the present disclosure to achieve similar benefits. The present disclosure proposes using an autoregressive model for forward-backward harmonic extrapolation before applying covariance-based spectral estimation to an extended signal. Since the autoregressive model is a discrete harmonic, the model does not reproduce noise and errors, and thus can be used as a filter even for multiplicative noise and parasitic modulation.

[0144] In this simulation, the phased array consists of 4 identical receiving antenna elements spaced at half-wavelength intervals (e.g., Figure 1 the receiving antenna units RX of the receiver unit 12 of the radar system 100 1,1 -RX 1,4 ). The antenna element array is linear and uniformly spaced (ULA). The simulation results are obtained in 10,000 Monte Carlo runs, where the radar system has a field of view (FoV) between -90 degrees and 90 degrees and an SNR randomly selected for each object from a uniform distribution between 15 dB and 30 dB.

[0145] To further illustrate the benefits of the super-resolution signal processing method of the present invention, Figure 5 is a graph depicting the probability (vertical axis) of separating or identifying two and three objects when using a conventional FFT-based DoA estimation when using a 30 dB Chebychev window. The horizontal axis represents the separation between the objects in degrees. InFigure 5 Within, trace 502 represents the probability of distinguishing between two objects, while trace 504 represents the probability of distinguishing between three objects. As Figure 5 depicted, using conventional techniques, the likelihood of the radar system successfully separating or identifying objects is no more than 75% until two objects (trace 502) are separated by at least 30 degrees. The conventional method cannot achieve a 75% probability of separating three objects (trace 504).

[0146] Figure 6 Depicts the performance of the MUSIC algorithm in a conventional signal processing chain. In Figure 6 within, trace 602 represents the probability of distinguishing between two objects. As Figure 6 shown, the MUSIC algorithm is unable to detect three objects using the 3×3 spatial smoothing covariance matrix obtained from 4 antenna elements in a conventional signal processing method. In contrast, Figure 7 depicts the performance of the MUSIC algorithm utilizing the super-resolution signal processing method of the present invention. In Figure 7 within, trace 702 represents the probability of distinguishing between two objects, while trace 704 represents the probability of distinguishing between three objects. Referring to Figure 7 , it can be seen that for two objects, the separation probability of the present invention is lower when there is no calibration error. However, as described herein, this signal processing method enables the MUSIC algorithm to detect three objects (as indicated by trace 704) by extrapolating the original input signal to 8 samples and constructing a 5×5 spatial smoothing covariance matrix.

[0147] Presents the simulation results of the CAPON algorithm. Similar to the MUSIC algorithm, CAPON also requires a full-rank covariance matrix to generate a pseudospectrum, which means that for a single snapshot, a 4-element signal is transformed into a 3×3 FBSS covariance matrix. Using the super-resolution signal processing method of the present invention, the 4-element signal is extrapolated to 8 samples and transformed into a 5×5 FBSS covariance matrix. Figure 8 Depicts the probability of object separation for two objects (depicted by trace 802) using the conventional CAPON algorithm. Figure 9 Depicts the performance of the CAPON algorithm utilizing the super-resolution signal processing method of the present invention. In Figure 9 within, trace 902 represents the probability of distinguishing between two objects, while trace 904 represents the probability of distinguishing between three objects. It has been shown that the results are consistent with those of MUSIC.

[0148] The simulation results presented above are provided in conjunction with an ideal signal use case (i.e., generated by a radar signal with idealized perfect antenna placement). The following are simulation results generated for a more realistic radar system phased array that is subject to coupling and channel gain / phase mismatch errors that have not been properly calibrated, which may be caused by improper antenna placement, among other real-world imperfections.

[0149] The calibration error is modeled as an amplitude error for each antenna element drawn from a zero-mean normal distribution with a standard deviation of 1.5 dB, and an additional phase error drawn from a zero-mean normal distribution with a standard deviation of 2.5 degrees.

[0150] By combining these simulated real-world conditions, the performance of the conventional MUSIC algorithm is degraded by a factor of 2, while the performance of the MUSIC algorithm combined with the super-resolution signal processing method of the present invention is only slightly degraded and exceeds the performance of MUSIC for both object use cases.

[0151] For the three object use cases, the super-resolution signal processing method of the present invention enables the MUSIC algorithm to provide robust results.

[0152] Specifically, for real-world simulations, Figure 10 Characterize the performance of the MUSIC algorithm in a conventional signal processing chain. Figure 10 In FIG. 1 , trace 1002 represents the probability of distinguishing two objects. Figure 10 As shown in Figure 1, the MUSIC algorithm is unable to detect three objects using the 3×3 spatial smoothed covariance matrix obtained from 4 antenna elements in conventional signal processing methods. In contrast, Figure 11 Depict the performance of the MUSIC algorithm using the super-resolution signal processing method of the present invention. Figure 11 , trace 1102 represents the probability of distinguishing two objects, while trace 1104 represents the probability of distinguishing three objects. Figure 11 , it can be seen that for two objects, the separation probability of the present invention is lower when there is no calibration error. However, as described herein, the present signal processing method enables the MUSIC algorithm to detect three objects (as indicated by trace 1104) by extrapolating the original input signal to 8 samples and constructing a 5×5 spatial smoothing covariance matrix.

[0153] Simulation results for 'real world' simulations are presented for the CAPON algorithm. Similar to the MUSIC algorithm, CAPON also requires a full rank covariance matrix to generate the pseudo-spectrum, which means that for a single snapshot, the 4-element signal is converted to a 3×3 FBSS covariance matrix. Using the super-resolution signal processing method of the present invention, the 4-element signal is extrapolated to 8 samples and converted to a 5×5 FBSS covariance matrix. Figure 12Depicts the probability of object separation for two objects (depicted by trace 1202) using the conventional CAPON algorithm. Figure 13 Depicts the performance of the CAPON algorithm using the super-resolution signal processing method of the present invention. Inside Figure 13 , trace 1302 represents the probability of distinguishing two objects, while trace 1304 represents the probability of distinguishing three objects. It has been shown that the results are consistent with those of MUSIC.

[0154] In summary, when the system is affected by antenna coupling / mismatch, the super-resolution signal processing method of the present invention can improve the radar system performance by enhancing the resolution and separation ability. Additionally, regardless of calibration errors, the super-resolution signal processing method of the present invention can estimate more objects than conventional methods.

[0155] In some aspects, the techniques described herein relate to a radar system including: at least one transmitter and at least one receiver, wherein the at least one transmitter and the at least one receiver are configured to transmit and receive radar signals, wherein the at least one transmitter and the at least one receiver are coupled to a vehicle; and a radar processor configured to: transmit a first radar signal at a first time; receive the received signal using the at least one receiver; process the received signal to generate a range-Doppler data frame; determine a first snapshot including a first plurality of values associated with a first range-Doppler bin of the range-Doppler data frame; process the first plurality of values in the first snapshot to generate an autoregressive model based on the first plurality of values; use the autoregressive model to extrapolate a second snapshot, wherein the second snapshot includes the first plurality of values and a second plurality of values generated using the autoregressive model; use the second snapshot to determine a full-rank covariance matrix; use the full-rank covariance matrix to identify attributes of a plurality of objects; and send the attributes of the plurality of objects to a vehicle controller.

[0156] In some aspects, the techniques described herein relate to a radar system including: at least one transmitter and at least one receiver; and a radar processor configured to: process the received signal to determine a first snapshot including a first plurality of values associated with a first range-Doppler bin of a range-Doppler data frame; determine a second snapshot, wherein the second snapshot includes the first plurality of values and a second plurality of values generated using an autoregressive model; use the second snapshot to determine a full-rank covariance matrix; and use the full-rank covariance matrix to identify attributes of a plurality of objects.

[0157] In some aspects, the techniques described herein relate to a method including: processing a received signal to generate a range-Doppler data frame; determining a first snapshot including a first plurality of values associated with a first range-Doppler bin of the range-Doppler data frame; processing the first plurality of values in the first snapshot to generate an autoregressive model based on the first plurality of values; extrapolating a second snapshot using the autoregressive model, where the second snapshot includes the first plurality of values and a second plurality of values generated using the autoregressive model; determining a full-rank covariance matrix using the second snapshot; and identifying attributes of a plurality of objects using the full-rank covariance matrix.

[0158] Although examples have been described with reference to automotive radar systems, the systems and methods described herein can be implemented in conjunction with other types of radar systems.

[0159] The foregoing detailed description is, to a large extent, merely illustrative in nature and is not intended to limit the embodiments of the subject matter or the application and use of such embodiments.

[0160] As used herein, the word "exemplary" means "serving as an example, instance, or illustration." Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments. Additionally, there is no intention to be bound by any theory expressed or implied in the foregoing technical field, background, or detailed description.

[0161] The connecting lines shown in the various figures contained herein are intended to represent exemplary functional relationships and / or physical couplings between various elements. It should be noted that many alternative or additional functional relationships or physical connections may exist in embodiments of the subject matter. Additionally, specific terms may be used herein solely for purposes of reference and thus are not intended to be limiting, and unless the context clearly indicates otherwise, the numerical terms "first," "second," and other such terms referring to structures do not imply a sequence or order.

[0162] As used herein, "node" means any internal or external reference point, connection point, junction point, signal line, conductive element, etc., at which a given signal, logic level, voltage, data pattern, current, or quantity exists. Additionally, two or more nodes can be implemented by one physical element (and although received or output at a common node, two or more signals can still be multiplexed, modulated, or otherwise distinguished).

[0163] The foregoing description refers to elements or nodes or features being "connected" or "coupled" together. As used herein, unless stated otherwise explicitly, "connected" means that an element is directly joined to another element (or directly communicates with another element), and not necessarily joined mechanically. Similarly, unless stated otherwise explicitly, "coupled" means that an element is directly or indirectly joined to another element (or directly or indirectly communicates with another element electrically or otherwise), and not necessarily joined mechanically. Thus, although the schematic illustrations shown in the figures depict an exemplary arrangement of elements, additional intervening elements, devices, features or components may be present in embodiments of the subject matter depicted.

[0164] Although at least one exemplary embodiment has been presented in the foregoing detailed description, it should be understood that a vast number of variations exist. It should also be understood that one or more of the exemplary embodiments described herein are not intended to limit in any way the scope, applicability or configuration of the claimed subject matter. Indeed, the foregoing detailed description will provide those skilled in the art with a convenient guide for implementing one or more of the described embodiments. It should be understood that various changes may be made to the functionality and arrangement of the elements without departing from the scope defined by the claims, which includes known equivalents and foreseeable equivalents at the time of filing this patent application.

Claims

1. A radar system, characterized in that: include: at least one transmitter and at least one receiver, wherein the at least one transmitter and the at least one receiver are configured to transmit and receive radar signals, wherein the at least one transmitter and the at least one receiver are coupled to a vehicle; as well as A radar processor, the radar processor being configured to: Send the first radar signal at the first time, receiving the received signal using the at least one receiver, processing the received signal to generate a range-Doppler data frame, determining a first snapshot comprising a first plurality of values ​​associated with a first range-Doppler bin of the range-Doppler data frame, processing the first plurality of values ​​in the first snapshot to generate an autoregressive model based on the first plurality of values, extrapolating a second snapshot using the autoregressive model, wherein the second snapshot includes the first plurality of values ​​and a second plurality of values ​​generated using the autoregressive model, Use the second snapshot to determine the full rank covariance matrix, identifying properties of multiple objects using the full rank covariance matrix, and The properties of the plurality of objects are sent to a vehicle controller.

2. The radar system according to claim 1, characterized in that The autoregressive model is configured to approximate the equation where a is the matrix containing the prediction coefficients of the autoregressive model, r x is the cross-correlation vector, and is the inverse of the cross-correlation matrix.

3. The radar system according to claim 2, characterized in that The radar processor is configured to process the first plurality of values ​​in the first snapshot to generate the autoregressive model using Burg's method, Yule Walker equations, or Levinson's method.

4. The radar system according to claim 1, characterized in that The radar processor is configured to identify the attributes of the plurality of objects by: determining a set of eigenvalues ​​of the full rank covariance matrix; and The property of the plurality of objects is determined using the set of feature values.

5. The radar system according to claim 1, characterized in that The radar processor is configured to identify the attributes of the plurality of objects by: Using at least one of a MUSIC spectrum estimation algorithm and a CAPON spectrum estimation algorithm to generate a pseudo spectrum using the full rank covariance matrix; as well as Peaks in the pseudo-spectrum are detected.

6. The radar system according to claim 1, characterized in that The first snapshot includes a first number of values ​​equal to a number of physical or virtual arrays in the radar system.

7. The radar system according to claim 6, characterized in that The second snapshot includes a second number of values ​​equal to twice the number of the first values.

8. The radar system according to claim 1, characterized in that The first snapshot includes four values ​​and the second snapshot includes eight values.

9. A radar system, characterized in that: include: at least one transmitter and at least one receiver; as well as A radar processor, the radar processor being configured to: processing the received signal to determine a first snapshot comprising a first plurality of values ​​associated with a first range-Doppler bin of the range-Doppler data frame, determining a second snapshot, wherein the second snapshot includes the first plurality of values ​​and a second plurality of values ​​generated using an autoregressive model, Determine the full rank covariance matrix using the second snapshot, and Attributes of a plurality of objects are identified using the full rank covariance matrix.

10. A method, characterized in that include: processing the received signal to generate a range-Doppler data frame; determining a first snapshot comprising a first plurality of values ​​associated with a first range-Doppler bin of the range-Doppler data frame; processing the first plurality of values ​​in the first snapshot to generate an autoregressive model based on the first plurality of values; extrapolating a second snapshot using the autoregressive model, wherein the second snapshot includes the first plurality of values ​​and a second plurality of values ​​generated using the autoregressive model; determining a full rank covariance matrix using the second snapshot; and Attributes of a plurality of objects are identified using the full rank covariance matrix.

Images