Low-Complexity Super-Resolution Techniques for Object Detection in Frequency-Modulated Continuous-Wave Radars

Abstract

This application is titled "Low Complexity Super Resolution Techniques for Object Detection in Frequency Modulated Continuous Wave Radar". In the described example, the method for object detection employs a hierarchical approach of low complexity techniques. An initial FFT-based detection and range estimation (201) gives a rough range estimate of a set of objects within the Rayleigh limit or with different sizes caused by significantly varying reflection intensities. For each detected peak set, the method demodulates the input to approximate direct current (203), filters out other peaks (204), and decimates the signal (205) to reduce the data size. The method then applies a super resolution method (206) to this limited data size. The resulting fine range estimate (207) uses FFT processing to provide the distance relative to the rough estimate (202).

Description

[0001] This application is a divisional application of Chinese Patent Application No. 201680036357.2 (PCT / US2016 / 032631), titled "Low-Complexity Super-Resolution Techniques for Object Detection in Frequency-Modulated Continuous-Wave Radar", filed on May 16, 2016. Technical Field

[0002] This application relates to radar object detection and corresponding object position determination. Background Art

[0003] In traditional object detection techniques, the so-called Rayleigh distance limit is used to resolve the minimum distance between two nearby objects (radar reflections). These techniques often fail to detect smaller objects near larger objects. Moreover, there are several techniques known as super-resolution techniques, and these techniques can distinguish objects even below the traditional limits, but super-resolution techniques are computationally expensive and rarely implemented. Summary of the Invention

[0004] In the described example, a method for object detection employs a hierarchical approach of low-complexity techniques. Initial FFT-based detection and range estimation give a rough range estimate of a set of objects within the Rayleigh limit or having different sizes caused by significantly varying reflection intensities. For each detected peak, the method demodulates the input to approximate direct current, filters out other peaks, and decimates the signal to reduce the data size. The method then applies a super-resolution method to this limited data size. The resulting fine range estimate provides a distance relative to the rough estimate obtained using FFT processing. Description of the Drawings

[0005] FIG. 1 shows a conventional FMCW radar to which an example embodiment is applicable.

[0006] Figure 2 Shows signal data processing of an example embodiment.

[0007] Figure 3 Shows the steps involved in a multiple signal classification algorithm.

[0008] Figure 4 Shows the steps involved in a matrix pencil algorithm.

[0009] FIG. 5 shows the result of conventional processing of two objects at different ranges having the same reflectivity.

[0010] FIG. 6 shows the result of conventional processing of two objects at different ranges, where one object has a reflectivity 25 dB less.

[0011] Figure 7Shows the result of the processing according to an example embodiment of two objects as different ranges with the same reflectivity.

[0012] Figure 8 Shows the result of the processing according to an example embodiment of two objects as different ranges, where one object has a reflectivity 25 dB less. Detailed Description

[0013] In the described example, initial object detection is performed using a classical method. In the context of an FMCW (Frequency Modulated Continuous Wave) radar, this is done by a fast Fourier transform of the input data and then by searching for high value amplitudes. After potential objects are detected, a super-resolution algorithm is performed around each of the detected objects or reflections. To reduce the computational complexity of this search, the signal is demodulated (so that the detected objects are located around the approximate DC value), and then downsampled (so that the number of operation data points is reduced). Then the super-resolution technique acts on this reduced data set, thus reducing the computational complexity.

[0014] FMCW radars are often used to determine the position of an object and its moving speed. These radars are used in automotive, industrial measurement, and other applications. Figure 1 shows a typical FMCW technique.

[0015] A chirp signal (where the frequency is linearly changed) generated by a ramp oscillator 101 and a voltage controlled oscillator (VCO) 102 is transmitted by an antenna 103, and this chirp signal is reflected from (multiple) objects 104. The signal is received by an antenna 105, and in a mixer 106 the signal is mixed with the transmitted signal, and as given below, the resulting beat frequency 107 depends on the distance of the object:

[0016]

[0017] Thus, if the beat frequency or multiple beat frequencies for multiple objects can be estimated, the distances to these objects can be estimated. In the above equation, R is the range of the object, B is the bandwidth of the chirp signal, T r is the duration of the chirp, and c is the speed of light.

[0018] In the most commonly used object detection and distance estimation techniques, the Fourier transform is used to estimate the frequency. Usually, the FFT (Fast Fourier Transform) is used. The peaks in the FFT output shown at 108 correspond to the detected objects, and the frequency of the peaks corresponds to the distance. In this technique, the minimum distance for resolving two objects and determining their respective distances is called the Rayleigh limit, and the minimum distance is given by:

[0019]

[0020] In another problem with such detection, when the reflectivities of two closely spaced objects are different, the larger object tends to hide the smaller object.

[0021] To overcome the above limitations, super-resolution techniques have been proposed. Two such techniques are described herein.

[0022] The first technique is called MUSIC (Multiple Signal Classification), which divides the signal autocorrelation matrix 301 Rs into a signal subspace and a noise subspace 302. This is done by first using Singular Value Decomposition (SVD) 303:

[0023] R s = QΛQ H

[0024] And then the noise subspace is extracted from the eigenvectors with the lowest eigenvalues 304:

[0025] Q n = Q(:, N - M, N)

[0026] N: data dimension, M: signal dimension

[0027] The technique then creates a MUSIC pseudospectrum orthogonal to the noise subspace using the following equation 305:

[0028]

[0029] And finally, a search for peaks in the above spectrum is performed to determine the presence and location of the object in 306.

[0030] The second technique is called MPM (Matrix Pencil Method), which creates a Hankel matrix 401 of the signal vectors with delays:

[0031] S = [s0 s1 s2... s L-1 s L = [S0 s L = [s0 S1]

[0032] s n = [s(n) s(n + 1)... s(n + N - L - 1)] T

[0033] And then the generalized eigenvalue problem of the matrix pencil is solved in 402 (these eigenvalues encode the frequency estimates)

[0034] S1 - ξS0

[0035] The steps for solving the generalized eigenvalue problem are as follows: Perform a Singular Value Decomposition (SVD) 403, and select M highest eigenvalues in 404:

[0036] S H S = UΛU H ; U M = U(:,1:M)

[0037] And extract two eigenvector matrices in 405:

[0038] U 0M = U(1:L-1,:), U 1M = U(2:L,:)

[0039] And perform a second SVD in 406:

[0040]

[0041] And extract frequencies from the resulting eigenvalues (generalized eigenvalues) in 407.

[0042] Various variants of these techniques have been proposed. However, all these variants of the techniques have the common operation of performing eigenanalysis of signal vectors. For a data size of N, eigenanalysis has a computational requirement of the order of N 3 magnitude. For normal applications, N is of the order of 1000. Therefore, these techniques are not feasible for implementing embedded real-time applications.

[0043] In FMCW radar applications, additional signal dimensions of velocity, azimuth angle, and elevation angle can be used, and the effect is to increase the data size by several orders of magnitude.

[0044] In Figure 2 the proposed low-complexity techniques, super-resolution techniques are combined with FFT-based methods to create a hierarchical approach. First, perform FFT-based detection and range estimation in 201. This gives a rough range estimate 202 of a set of objects within the Rayleigh limit or with different sizes caused by significantly varying reflection intensities. For each detected peak group, the input is demodulated to approximate DC in 203, other peaks (or other object groups) are filtered out in 204, and then the signal is subsampled in 205 to reduce the data size. Then, a super-resolution method is performed on this limited data size in 206. The resulting fine range estimate 207 provides the distance relative to the rough estimate achieved using FFT processing.

[0045] The following study shows simulation results using a signal bandwidth = 4 GHz, chirp duration = 125 microseconds, and parameters of two objects at 5.9 m and 6 m. The two examples are: (a) the objects have the same reflectivity; and (b) the reflectivities of the objects differ by 25 dB. Reflectivity is measured based on RCS (radar cross section).

[0046] The outputs of conventional FFT-based processing are shown in FIGS. 5 and 6. FIG. 5 (corresponding to the same RCS of two objects) shows two peaks 501 and 502 corresponding to the two objects. In FIG. 6, in the case where the RCS of one object is 25 dB lower, the smaller object cannot be detected and is hidden as the peak of the larger object 601 spreads. The data size used is 512.

[0047] Then using the technique of the exemplary embodiment, the data is reduced to 32, resulting in a reduction of the computational complexity to 1 / 16 3 (reduction by a factor of 16 3 ) Figure 7 and Figure 8 show the outputs of the MUSIC method. Figure 7 Show steeper peaks 701 and 702 for the case of the same RCS. Figure 8 Show that the object is still missing for the case of a 25 dB RCS difference.

[0048] Providing the image output from the simulation of the MPM matrix pencil method from Figure 7 and Figure 8 is not practical. However, if the matrix pencil is run on this reduced data set, the matrix pencil provides two distance estimates for the following: the same RCS; and a 25 dB RCS difference. For the same RCS: the distance estimates are distance1 = 6.0012 m and distance2 = 5.8964 m. For a 25 dB RCS difference: the distance estimates are distance1 = 5.9990 m and distance2 = 5.8602 m. In this example, the objects are placed at 5.9 m and 6 m, so the MPM method accurately provides the distances with greatly reduced complexity.

Claims

1. A method for object detection, comprising: Transmitting a first signal using a transmitter; Receiving a reflected signal generated by a group of objects including a first object in response to the first signal; Mixing the reflected signal with the first signal to determine a beat frequency signal for the first object; Applying a Fourier transform to the beat frequency signal to generate a Fourier transform signal, the Fourier transform signal including a first peak corresponding to the first object and a second peak not corresponding to the first object; Based on the first peak in the Fourier transform signal, determining a first range estimate of the distance of the first object relative to the transmitter; Applying demodulation to a portion of the Fourier transform signal including the first peak and the second peak to generate a demodulated signal; Applying filtering to the demodulated signal to reduce the influence of the second peak, thereby generating a filtered signal; Performing subsampling on the filtered signal to generate a subsampled data set; And Based on the subsampled data set, determining a second range estimate of the distance of the first object relative to the transmitter.

2. The method according to claim 1, wherein the first signal has a linearly varying frequency.

3. The method according to claim 1, wherein the determination of the first range estimate is based on the frequency of the first peak in the Fourier transform signal corresponding to the distance of the first object relative to the transmitter.

4. The method according to claim 1, wherein the Fourier transform signal has a first data size and the subsampled data set has a second data size smaller than the first data size.

5. The method according to claim 1, wherein the second range estimate is more accurate than the first range estimate.

6. The method according to claim 1, wherein the determination of the second range estimate includes performing a feature analysis of the subsampled data set.

7. The method according to claim 1, wherein the determination of the second range estimate includes performing a multiple signal classification super-resolution processing technique, i.e., MUSIC super-resolution processing technique, on the subsampled data set by: Applying singular value decomposition (SVD) to the signal autocorrelation matrix corresponding to the subsampled data set to obtain an SVD result; Dividing the signal autocorrelation matrix into a signal subspace and a noise subspace by extracting eigenvectors from the SVD result, the eigenvectors with the lowest eigenvalues from the SVD result corresponding to the noise subspace, and the remaining eigenvectors from the SVD result corresponding to the signal subspace; Creating a MUSIC pseudospectrum orthogonal to the noise subspace; And Searching for peaks in the MUSIC pseudospectrum.

8. The method according to claim 1, wherein the determination of the second range estimate includes performing a matrix pencil method super-resolution processing technique, i.e., MPM super-resolution processing technique, on the subsampled data set by: Creating a Hankel matrix of signal vectors with delays; Applying a first singular value decomposition (SVD) to the Hankel matrix to obtain a first SVD result; Select a predetermined number of the highest eigenvalues from the first SVD result; Extract two eigenvector matrices; Apply a second SVD to the two extracted eigenvector matrices to obtain a second SVD result; and Search for peaks within the eigenvalues of the second SVD result.

9. The method according to claim 1, wherein applying the Fourier transform to the beat signal includes applying a fast Fourier transform to the beat signal.

10. An apparatus for object detection, comprising: A voltage-controlled oscillator configured to generate a first signal; A first antenna coupled to the voltage-controlled oscillator and configured to transmit a radar signal based on the first signal; A second antenna configured to generate a reflected signal based on a reflected radar signal reflected from a group of objects including a first object; A mixer configured to mix the first signal and the reflected signal to generate a beat signal; A processor configured to: Apply a Fourier transform to the beat signal to generate a Fourier transform signal, the Fourier transform signal including a first peak corresponding to the first object and a second peak not corresponding to the first object; Demodulate the Fourier transform signal to generate a demodulated signal; Filter the demodulated signal to reduce the influence of the second peak on the demodulated signal, thereby generating a filtered signal; Subsample the filtered signal to generate a subsampled data set; And Determine a distance of the first object relative to the first antenna based on the subsampled data set.

11. The apparatus according to claim 10, further comprising a linear ramp generator coupled to the voltage-controlled oscillator, wherein the linear ramp generator and the voltage-controlled oscillator are configured to generate the first signal to have a linearly varying frequency.

12. The apparatus according to claim 10, wherein: The distance of the first object relative to the first antenna determined based on the subsampled data set is a first distance estimate; and The processor is configured to determine a second distance estimate of the first object relative to the first antenna based on the Fourier transform signal.

13. The apparatus according to claim 12, wherein the processor is configured to determine the second distance estimate of the first object based on a frequency of the first peak corresponding to the distance of the first object relative to the first antenna in the Fourier transform signal.

14. The apparatus according to claim 12, wherein the first distance estimate is more accurate than the second distance estimate.

15. The apparatus according to claim 10, wherein the Fourier transform signal has a first data size and the subsampled data set has a second data size smaller than the first data size.

16. The apparatus according to claim 10, wherein the processor is configured to determine the distance of the first object relative to the first antenna by performing a feature analysis of the subsampled data set.

17. The apparatus according to claim 10, wherein the processor is configured to perform a multiple signal classification super-resolution processing technique, i.e., MUSIC super-resolution processing technique, on the subsampled data set to determine the distance of the first object relative to the first antenna by: Applying singular value decomposition (SVD) to the signal autocorrelation matrix corresponding to the subsampled data set to obtain an SVD result; Dividing the signal autocorrelation matrix into a signal subspace and a noise subspace by extracting eigenvectors from the SVD result, wherein the eigenvector with the lowest eigenvalue from the SVD result corresponds to the noise subspace, and the remaining eigenvectors from the SVD result correspond to the signal subspace; Creating a MUSIC pseudospectrum orthogonal to the noise subspace; And Searching for peaks in the MUSIC pseudospectrum.

18. The apparatus according to claim 10, wherein the processor is configured to perform a matrix pencil method super-resolution processing technique, i.e., MPM super-resolution processing technique, on the subsampled data set to determine the distance of the first object relative to the first antenna by: Creating a Hankel matrix of signal vectors with delays; Applying a first singular value decomposition (SVD) to the Hankel matrix to obtain a first SVD result; Selecting a predetermined number of the highest eigenvalues from the first SVD result; Extracting two eigenvector matrices; Applying a second SVD to the two extracted eigenvector matrices to obtain a second SVD result; and Searching for peaks within the eigenvalues of the second SVD result.

19. The apparatus according to claim 10, wherein the processor is configured to apply the Fourier transform to the beat frequency signal by applying a fast Fourier transform to the beat frequency signal.

20. A method for object detection, comprising: Transmitting a first signal using a transmitter; Receiving a reflected signal generated by a group of objects including a first object in response to the first signal; Mixing the reflected signal with the first signal to determine a beat frequency signal for the first object; Applying a Fourier transform to the beat frequency signal to generate a Fourier transform signal, the Fourier transform signal including a peak corresponding to the first object; Applying demodulation to a portion of the Fourier transform signal including the peak to generate a demodulated signal; Applying filtering to the demodulated signal using a filter configured to retain the influence of the peak to generate a filtered signal; Subsampling the filtered signal to generate a subsampled data set; And Based on the subsampled data set, determining an estimate of the distance of the first object relative to the transmitter.

Images