Low complexity super-resolution technique for object detection in frequency modulation continuous wave radar

This application is a continuation of U.S. patent application Ser. No. 14/951,014, filed Nov. 24, 2015, which claims priority to U.S. Provisional Patent Application No. 62/162,405, filed May 15, 2015, each of which is incorporated by reference herein in its entirety.

The technical field of this invention is radar object detection and corresponding object location determination.

In classical object detection technique, the minimum distance to resolve two nearby objects (radar reflections) is limited by the so called Rayleigh distance. These techniques also often fail to find smaller objects in presence of close by larger objects. There exist several techniques known as super-resolution techniques to overcome these methods which can discriminate between objects even below the classical limits. However, these techniques are computationally expensive and rarely implemented in practice.

The solution to the computational problem is to perform an initial object detection using the classical method. In the context of FMCW (Frequency Modulated Continuous Wave) radar, this was done through Fast Fourier Transforms of the input data and then by searching for high valued amplitudes. Once potential objects are detected, super-resolution algorithms are performed around each of the detected objects or reflections. To reduce computational complexity of this search, the signal is demodulated so the detected object lies near DC values and then sub-sampled so the number of operating data points is reduced. The super-resolution technique then works on this reduced set of data thereby reducing computational complexity.

These and other aspects of this invention are illustrated in the drawings, in which:

FIG. 1 illustrates a prior art FMCW radar to which this invention is applicable;

FIG. 2 illustrates the signal data processing of this invention;

FIG. 3 shows the steps involved in the multiple signal classification algorithm;

FIG. 4 shows the steps involved in the matrix pencil algorithm;

FIG. 5 illustrates results of conventional processing for two objects as differing ranges with the same reflectivity;

FIG. 6 illustrates results of conventional processing for two objects as differing ranges with one object having 25 dB less reflectivity;

FIG. 7 illustrates results of processing according to this invention for two objects as differing ranges with the same reflectivity; and

FIG. 8 illustrates results of processing according to this invention for two objects as differing ranges with one object having 25 dB less reflectivity.

FMCW radars are often used to determine the location of an object and its speed of movement. These radars are used in automotive applications, industrial measurements, etc. A typical FMCW technique is shown in FIG. 1 .

A chirp signal generated by ramp oscillator 101 and Voltage Controlled Oscillator (VCO) 102 (where the frequency is changed linearly) is transmitted by antenna 103 and reflected from object(s) 104 . The reflected signal is received by antenna 105 , mixed with transmitted signal in mixer 106 and the resulting beat frequency 107 is dependent on the distance of the object as given by

beat ⁢ ⁢ frequency = B ⁡ ( 2 ⁢ R ) T r ⁢ c

Thus, if the beat frequency or frequencies for multiple objects can be estimated, the distances to those 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 time duration for the chirp and c is the speed of light.

In one commonly used object detection and distance estimation technique, the frequency is estimated using Fourier transforms. Usually an FFT (Fast Fourier Transform) is used. The peaks of the FFT output shown in graph 108 correspond to the objects detected and the frequencies of the peak correspond to the distances. In this technique, the minimum distance to resolve two objects and determine their respective distances are known as Rayleigh limit and is given by

c 2 ⁢ B

One issue with this detection method is when the reflectivities of the two closely spaced objects are different, the larger object tends to hide the smaller object.

In order to overcome the above limitations, super-resolution techniques have been proposed. Two such techniques are described here.

The first technique is called MUSIC (Multiple Signal Classification): it divides the signal auto-correlation matrix 301 R s , into signal subspace and noise subspace 302 . This is done by first using singular value decomposition (SVD) 303

R s =QΛQ H

and then extracting the noise subspace from the eigenvectors with lowest eigenvalues 304

Q n =Q (:, N−M,N )

N: data dimension, M: signal dimension

This technique then creates MUSIC pseudo spectrum orthogonal to noise subspace using the following equation 305

P MUSIC ⁡ ( ϕ ) = 1 S H ⁡ ( ϕ ) ⁢ Q n ⁢ Q n H ⁢ s ⁡ ( ϕ ) a

and finally a search for peaks in the above spectrum is carried out to determine the presence and the location of objects in 306 .

The second technique called MPM (Matrix Pencil Method):

create a Hankel matrix 401 with delayed signal vector

S =[ s 0 s 1 s 2 . . . s L-1 s L ]=[ S 0 s L ]=[ s 0 S 1 ]

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

and then solve a generalized eigenvalue problem of the matrix pencil in 402 (these eigenvalues encode the frequency estimates)

S 1 −ξS 0

The steps to solve the generalized eigenvalues problems are as follows: perform Singular Value Decomposition (SVD) 403 and choose M highest eigenvalues in 404

S H S=UΛU H ;U M =U (:,1: M )

extract two eigenvector matrices in 405

U 0M =U (1: L− 1,:), U 1M =U (2: L ,:)

perform a second SVD in 406

U 1M H U 0M

and extract frequencies from the resulting eigenvalues in 407 (the generalized eigenvalues).

Various variations of these techniques have been proposed. But they all have the common operations of performing eigen-analysis of signal vectors. For a data size of N, the eigen-analysis has computational requirement on the order of N 3 . For typical applications N is of the order of 1000. This makes implementation of these techniques unfeasible for embedded real time applications.

Note that in FMCW radar applications, additional signal dimensions of speed, azimuth and elevation angle can be used whose impact is to increase the data size by several orders.

In the proposed low complexity technique illustrated in FIG. 2 , the super resolution techniques are combined with the FFT based method to create a hierarchical approach. First an FFT based detection and range estimation is performed in 201 . This gives a coarse range estimate 202 of a group of objects within the Rayleigh limit or with varying sizes resulting from widely varying reflection strengths. For each group of detected peaks, the input is demodulated to near DC in 203 , other peaks (or other object groups) are filtered out in 204 and the signal is then sub-sampled in 205 to reduce the data size. Super-resolution methods are then performed on this limited data size in 206 . The resulting fine range estimations 207 provide distance relative to the coarse estimation done using FFT processing.

The following study shows simulation results using the following parameters: signal bandwidth of 4 GHz; chirp time duration of 125 microseconds; 2 objects at 5.9 meters and 6 meters in two examples (1) the objects have the same reflectivity and (2) the objects differ in reflectivity by 25 dB. The reflectivities are measured in terms of RCS (radar cross section).

The output of the prior art FFT based processing are shown in FIGS. 5 and 6 . FIG. 5 (corresponding to the same RCS of two objects) shows the two peaks 501 and 502 corresponding to the two objects. In FIG. 6 where the RCS of one object is 25 dB lower, the smaller object cannot be detected and is hidden with the spread of the peak of the larger object 601 . The data size used is 512 data points.

The data size is then reduced to 32 data points using the technique of this invention leading a computation complexity reduction by a factor of 16 3 . The output of the MUSIC method (described in FIG. 3 ) is shown in FIGS. 7 and 8 . FIG. 7 shows much sharper peaks 701 and 702 for the case of same RCS. FIG. 8 shows that an object is still missed for the case of 25 dB RCS difference.

It is not possible to provide pictorial output from the simulation of the MPM matrix pencil method like shown in FIGS. 7 and 8 . However, if we run matrix pencil on this reduced data set, it provides two distance estimates for both the same RCS, and 25 dB RCS difference. The results are noted below. For the same RCS: the distance estimates are distance1=6.0012 m and distance2=5.8964 m. For 25 dB difference RCS: the distance estimates are distance1=5.9990 m and distance2=5.8602 m. Comparing with the fact that the objects are placed at 5.9 meters and 6 meters, the MPM method provided the distances accurately with much reduced complexity.