Progress In Electromagnetics Research M, Vol.

23, 153165, 2012

3D IMAGING METHOD FOR STEPPED FREQUENCY

GROUND PENETRATING RADAR BASED ON COMPRESSIVE SENSING
J.-L. Cai1, * , C.-M. Tong1, 2 , W.-J. Zhong1 , and W.-J. Ji1
1 Missile

Institute, Air Force Engineering University, Sanyuan, Shanxi

713800, China
2 State

Key Laboratory of Millimeter Waves, Nanjing, Jiangsu 210096,

China
AbstractLong data collecting time is one of the bottlenecks of the
stepped-frequency continuous-wave ground penetrating radar (SFCWGPR). We discuss the applicability of the Compressive Sensing (CS)
method to three dimensional buried point-like targets imaging for
SFCW-GPR. It is shown that the image of the sparse targets can be
reconstructed by solving a constrained convex optimization problem
based on l1 -norm minimization with only a small number of data
from randomly selected frequencies and antenna scan positions, which
will reduce the data collecting time. Target localization ability,
performance in noise, the effect of frequency bandwidth, and the effect
of the wave travel velocity in the soil are demonstrated by simulated
data. Numerical results show that the presented CS method can
reconstruct the point-like targets in the right position even with 10%
additive Gaussian white noise and some wave travel velocity estimation
error.

1. INTRODUCTION
Ground Penetrating Radar (GPR) is an important remote sensing
tool to detect the object buried in shallow ground and widely used in
such areas as civil engineering [1], landmine detection [2], archeological
investigations [3] and environmental applications [4], etc.
Generally speaking, there are two types of GPRs most commonly
seen in literatures. One is (Carrier-Free Pulse CFP-GPR), which
Received 12 December 2011, Accepted 29 January 2012, Scheduled 8 February 2012
* Corresponding author: Ji-Liang Cai (shitouji840716@126.com).

154

Cai et al.

images the subsurface by transmitting short electromagnetic (EM)

pulses and processing the reflections due to permittivity discontinuities
in the ground [5]. The other is the stepped-frequency continuous-wave
GPR (SFCW-GPR) [69]. When it works, a stepped frequency signal
probes the environment with a discrete set of frequencies. Compared
with CFP-GPR, SFCW-GPR has several advantages, such as greater
measurement accuracy, greater dynamic range, lower noise, easier
frequency band and frequency steps selecting [9]. Therefore, it is
becoming increasingly popular. In [68], SFCW-GPR is used to image
buried landmines and objects.
Although SFCW GPRs have very good properties, they are not
used widely in commercial systems. One important reason for this is
long data acquisition time [9]. There are tens of scan positions and
hundreds of measurements at each scan position, which means large
quantities of data and long measurement time and thus make it too
slow for some applications.
Compressive sensing (CS), proposed by Donoho [10], shows that
certain signals and images, which are sparse or compressible in some
domains such as time, space and frequency, can be recovered with
far fewer samples or measurements than traditional Nyquist sampling
theorem. Due to its compressed sampling and exact reconstruction
ability, CS has been widely used in radar field imaging [11], such
as Ground Penetrating Radar (GPR) [9, 12], through-wall radar
imaging [13], SAR [14], and ISAR imaging [15]. In recent years, CS has
received more and more attention in radar applications, for lightening
up sampling burden and improving the resolution of radar system.
In [12], a 3D GPR imaging method based on CS is proposed. The
target space is reconstructed from just a few compressive sensing data
obtained by random aperture measurements. However, it is an impulse
GPR. In [9], stepped-frequency GPR based on CS is proposed, but it
only deals with 2D targets. 3D imaging is the tendency of future GPR,
and SFCW-GPR has its own unique advantages as mentioned above.
As far as we know, CS has not been applied to 3D SFCW-GPR imaging
yet. And, in this paper, 3D imaging method for SFCW-GPR based on
CS is discussed.
The paper is organized as follows. Section 2 briefly describes the
theory of compressive sensing. The compressive sensing algorithm for
stepped frequency GPR and the selection of the related parameters
are explained in Section 3. Section 4 presents results for the simulated
stepped frequency GPR with a performance analysis. Conclusions are
drawn in Section 5.

Progress In Electromagnetics Research M, Vol. 23, 2012

155

2. BASIS THEORY OF COMPRESSIVE SENSING

When the object signal is sparse, compressive sensing provides a very
efficient sampling method which will significantly decrease the required
volume of the collected data. Sparsity means that the number of
nonzero coefficients representing a signal in a certain basis is much less
than its dimension. Considering a discrete signal vector x RN , we
say that it is K-sparse if K (K N ) of its coefficients are nonzero on
an orthonormal basis or over-complete dictionary C N N . Hence,
the true information is contained in x which lives at most K dimensions
rather than N . The sparse signal can be expressed as
x =

(1)

where vector RN 1 is the weighting coefficient. As the signal x

has a sparse representation in , x can be well approximated by the
best K term expression.
According to CS theory, the measured signal is acquired by linear
projections y = x. It makes sense that only M samples of signal
x need to be measured instead of N . Then, considering a linear
measurements matrix C M N with M < N , the measurements
signal y RM is descried as
y = x = =

(2)

where = is a M N matrix. This set of equations is

underdetermined, and (2) has infinitely many solutions. However, it is
indeed possible to recover the sparse signal via CS when the matrix
has the Restricted Isometry Property (RIP) of order K [16].
Knowing the observed vector y and the measured matrix , the
signal x can be recovered from the solution of a convex optimization
problem based on l1 norm
min kkl

st. y =

(3)

If noise is taken into account, the modified convex problem can be

described as
min kkl st. ky k2
(4)
where is weighted coefficient, and bounds the amount of noise in
measured data.
3. COMPRESSIVE SENSING FOR 3D SFCW-GPR
IMAGING
In the application of GPR, when detecting the ordnances, tubes or
metal steels, these small buried targets can be viewed as many discrete

156

Cai et al.

scatter centrals. If the object zone is subdivided, these scatter centers

usually take a small part of the space unite, i.e., the image to be
reconstructed is sparse, and the precondition of the CS is satisfied.
The proposed CS imaging method will make full use of the sparsity of
the targets to precisely image the buried targets.
3.1. Construction of the Dictionary Matrix
The two-layer scenario of 3D GPR imaging is shown in Fig. 1. The
target space T , which lies in the product space [xi xf ] [yi yf ] [zi zf ]
is discretized into a finite set of N points B= {1 , 2 , . . . ,N }. Here,
(xi , yi , zi ) and (xf , yf , zf ) denote the initial and final positions of the
target space to be imaged along each axis, and each j is a 3D vector
[xj , yj , zj ]. The transmit and receive antenna pair, situated at a known
height h from the ground/air interface, moves at the step of x and
y along x and y axes, respectively, to form a 2D measurement space.
This is a more general and realistic case for 3D imaging, but it is much
harder to figure out the received signal of each antenna. As the SFCWGPR scans a region, at each scan position, the transmitter transmits
M continuous sinusoidal signals, sequentially changing the frequencies
from the initial frequency f0 with frequency step f .
Suppose that the medium of the ground is uniform and
nondispersive, that there are P point-like targets underground,
illuminated in the main lobe of the antennas, and that the multiscattering among the targets do not happen. Then the frequency
response measurement at the ith scan position is:
XP
i () =
b (k) exp (ji (k ))
(5)
k=1

where is the vector of M signal frequencies i = 2 (f0 + mf ).

m = 0, 1, . . . , M 1. i (k ) denotes the time delay of the signal form

Figure 1. Scenario of 3D GPR

imagin.

Figure 2. GPR measurement

setup of antenna pair.

Progress In Electromagnetics Research M, Vol. 23, 2012

157

the transmitter to the target k and then reflected to the receiver at

ith scan position. b is a weighed indicator N 1 vector defining the
target space, i.e., if there is a target at j , the value of the jth element
of b should be the reflection coefficient j ; otherwise, it is zero. Our
goal is to find b, which is actually an image of the target space.
Equation (5) can be rewrite in the following matrix form
i () = i b

(6)

where i can be viewed as the dictionary matrix of M N , the jth

column of which, [i ]j , corresponds to the frequency response of the
target at j when the GPR is at the ith scan position
[i ]j = exp (ji (k ))

(7)

According to the ray theory view of electromagnetic wave

propagation, the transmitted signal follows the path in Fig. 2. At
the boundary between two different media (such as air and soil), the
direction changes according to Snells law, but exact calculation of the
refraction points requires the solution of a 4 degree polynomial. Several
approximations are available in [17]. After finding the reflection points,
the distances d1:4 can be calculated, and then the time delay i (k )
can be calculated as follow:
i (k ) = (d1 + d4 )/v1 + (d2 + d3 )/v2

(8)

where v1 and v2 are the wave propagation velocities in the air and soil.
3.2. Construction of the Measurement Matrix
Standard SFCW-GPR measures at a fixed set of M frequencies for each
scan position, Hence the dimension of i () is M 1. By compressive
sensing sampling data acquisition method, a very small number of
random measurements carry enough information to reconstruct the
buried targets. Thus, a subset of L frequencies for each scan position
is measured. In the matrix form, the new measurements i can be
written as:
i = i i = i i b
(9)
where i is a L M measurement matrix constructed by randomly
selecting L rows of an M M identity matrix. This reduces the data
acquisition time by L/M .
For each scan position, the measurement matrix may be different
or keep the same. In [13], it is proved by simulation that they are the
same for the through-wall radar imaging, but the latter is easier for
hardware implementation.

158

Cai et al.

3.3. Compressive 3D SFCW-GPR Imaging

As mentioned above, SFCW-GPR needs measurements at different
scan positions. For 3D imaging, the transmitter and receiver antenna
pair moves along x and y axes to form a 2D grid measurement
space.
By the compressive sensing imaging method, we just
randomly select K scan positions to form a composite dictionary
T
matrix = [T1 , T2 , . . . , TK ] composite measurement matrix
= diag{1 , 2 , . . . , K } and the measurement vector =
T ]T .
[1T , 2T , . . . , K
Then the recovery of b is done by solving a
constrained l1 minimization problem
b = argmin kbk
s.t. = b
(10)
l1

The equality constraints in above equation are only valid for the
noiseless case. Generally, the data are always contaminated by the
noise in real measurements. Then constraints are as follow
= b + n

(11)

where n is the noise vector, following a normal distribution with mean

zero and standard deviation 2 , i.e., n N (0, 2 ). The reconstruction
is by solving a modified convex optimization problem,
of the image b
called Dantzig Selector [18].

= argmin kbk
b
s.t. AT ( Ab) <
(12)
l1
where A = , is the amount of noise in p
measured data. The
choice of is noise dependent, and selection = 2 2 log(KL) makes
feasible with high probability [18].
the true b
4. SIMULATION RESULTS
In this section, several simulated data results for CS method are
presented. It should be mentioned that this paper is mainly focused
on CS imaging, aimed at saving data processing time by sampling less
data. Before using CS, some data preprocessing methods in [7, 1921]
can be used to remove the ground reflecting effects.
First, to illustrate the random frequency sampling and random
antenna position selection idea, a 3D homogenous target space of
size 10 cm 10 cm 8 cm containing three randomly placed point like
targets (suppose P1 (8, 5, 6), P2 (5, 6, 3) and P3 (5, 8, 4)) with reflection
coefficients 1.0, 0.6 and 0.8 respectively are considered. The target
space is discredited into 10 10 8 points. Bistatic antenna pairs
with a 2 cm transmitter-receiver spacing at the height of 10 cm collect
frequency domain measurements at frequencies from 100 MHz to

Progress In Electromagnetics Research M, Vol. 23, 2012

159

10 GHz with 100 MHz frequency step. Thus, at each scan position,
100 frequency measurements are collected. The antenna pair moves
with the step of x = y = 1 cm along x and y axes, respectively, in
the measurement space with the size of 15 cm 15 cm. So there are
totally 225 positions. The permittivity of the soil r is supposed
to be

4, i.e., the signal transmitting velocity in the soil v2 = v1 / r = 0.5v1 .

For the numerical solution of the l1 -norm minimization problems in

2.2
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2

frquncy(100MHz)

10
20
30
40
50
60
70
80
90
100

(a)

100
150
antnna

(b)

200

(c)
0

-2

-2

-3

-2

-3

-4

-4

-6

-7

z(cm)

-5

-6

-8
6

-9

-10

-11

8
4

6
x (cm)

-8
-10

-12
2

(d) P 1y=5

6
x (cm)

-5

-6
5

-7

-8

-9
-10

-12
2

10

z(cm)

-4
z(cm)

50

10

(e) P2 y=6

6
x (cm)

10

(f) P 3 y=8
0

-10

-15

5
6
7

z(cm)

-5

-10

-15
5

-20

-25

6
x (cm)

(g)P 1y=5

10

-30

-10
-15

-20

-25

-20
-25

-5

4
z(cm)

-5

z(cm)

6
x (cm)

(h) P2 y=6

10

-30

6
x (cm)

10

-30

(i) P 3 y=8

Figure 3. (a) DSBF data acquisition. (b) CS data acquisition where

the black points indicates the randomly sampled data when about 7%
of the total measurements are used. (c) Noiseless space-frequency
domain target response for all the frequency and antenna positions.
(d)(f) DBSF imaging with the full set of data for P1 , P2 , P3 at y = 5,
6, 8 respectively. (g)(i) CS imaging with randomly selected subset of
data for P1 , P2 , P3 at y = 5, 6, 8 respectively.

160

Cai et al.

CS, a convex optimization package called CVX [22] is used. All the
simulations are done on a computer with 1.6 GHz Pentium 4 processor
and 1G memory.
For the CS method, we use the measurement strategy proposed
in [13], which is easy for hardware implementation, i.e., instead of
measuring all 100 frequencies at all 225 scan positions, a random subset
of 50 positions is firstly selected, and then for each position, the same
random subset of 30 frequencies is measured. To make a comparison,
the conventional delay-and-sum beam forming (DSBF) algorithm [13],
which performs coherent summation of full 100 225 sensor data, is
used. The target space slice images at y = 5, 6, 8 in Figs. 3(d)(f)
and (g)(i), the noiseless full space-frequency domain measured data in
Fig. 3(a), and the randomly selected space-frequency domain measured
data in Fig. 3(b) are shown respectively. It can be seen that when
applying CS, much less data are used.
For the DSBF imaging in Figs. 3(d)(f), it can be seen that
the three targets can be seen clearly with small blobs in the right

4
5
6
7
8
2

6
x(cm)

10

-2

-4

-6

5
6
7
8
4

6
x(cm)

5
6

-10

7
8
2

10

(b) P2 y=6

-15

z(cm)

-10

-10

-10

-15

-15

5
6

-25

-25

(d) P 1y=5

10

-30

-5

-20

10

6
x(cm)

-5

-20

6
x(cm)

(c) P 3 y=8
0

1
-5

-2
-3
-4
-5
-6
-7
-8
-9
-10

-8

-12
2

(a) P 1y=5

z(cm)

z(cm)

z(cm)

z(cm)

-2
-3
-4
-5
-6
-7
-8
-9
-10
-11
-12

z(cm)

-20
-25

8
2

6
x(cm)

(e) P2 y=6

10

-30

6
x(cm)

10

-30

(f) P 3 y=8

Figure 4. Imaging with 10% additive Gaussian white noise. (a)(c)

DBSF imaging with the full set of data for P1 , P2 , P3 at y = 5, 6,
8 respectively. (d)(f) CS imaging with randomly selected subset of
data for P1 , P2 , P3 at y = 5, 6, 8 respectively.

Progress In Electromagnetics Research M, Vol. 23, 2012

161

position, but for Fig. 3(d), a false point-like target at (5, 5, 2) is also
imaged. As seen in Figs. 3(g)(h), while using a much small subset of
data, the CS method can recover the targets with much less cluttered
image, which is better than DSBF method using the full data set.
Note that in real applications, the measured data are always
contaminated with noise. Next, the performance of the algorithm with
noise is disused.
4.1. Performance in Noise
To analyze the impact of noise data on the imaging, 10% Gaussian
white noise is added to the signal reflected by the above three point-like
targets. All other parameters are kept the same with above experiment.
The images are shown in Fig. 4.
From the simulation result shown in Fig. 4, with the additives
noise, though there is a subtle change for DSBF images the clutters and
false targets still exist. Whereas, for the CS method, there are some
changes, more clusters appear, but except the targets, the values of the
1

-5

-15

5
-20

6
7

-25

8
2

6
8
x (cm)

10

-30

z (cm)

-10

3
4

-15

5
6
7

-25

6
8
x (cm)

10

-30

-20
-25
2

-15

5
-20

6
7

-25

-5

-15

5
-20

6
7

-25

8
10

-30

10

-30

-5

-10

3
z (cm)

6
8
x (cm)

(c) P 3 y=8

-10

-10

3
z (cm)

-5

3
z (cm)

-15

(b) P2 y=6

(d) P 1y=5

6
8
x (cm)

-10

(a) P 1y=5

-20

-5

2
z (cm)

-10

3
z (cm)

-5

-15

5
-20

6
7

-25

8
2

6
8
x (cm)

(e) P2 y=6

10

-30

6
8
x (cm)

10

-30

(f) P 3 y=8

Figure 5. CS imaging with different bandwidth. (a)(c) CS imaging

with bandwidth of 2 [13] GHz for P1 , P2 , P3 at y = 5, 6, 8 respectively.
(d)(f) CS imaging with bandwidth of 5 [0.55.5] GHz for P1 , P2 , P3
at y = 5, 6, 8 respectively.

162

Cai et al.

pixels are all less than 15 dB which can be viewed as the background,
i.e., the targets are all at the right position, and no false targets appear.
4.2. Effect of Bandwidth
The bandwidth of the measured frequency spectrum is an important
parameter in the SFCWGPR. To analyze the effect of frequency
bandwidth on the imaging, in this subsection, different bandwidths
of 2 GHz and 5 GHz are tested for the above three point-like targets.
In the simulation, 10% noise is added. The images are shown in Fig. 5.
It can be observed in Fig. 5 that the CS method is able to
reconstruct the true target points even for the low bandwidth of 2 GHz.
Besides, with smaller bandwidth, the number of false target points
becomes smaller, but with stronger intensity.
1

1
-5

-5

-15

5
-20

6
7
8
2

6
8
x (cm)

-10

3
z (cm)

-15

5
-20

-25

-30

10

(a) P 1y=5

6
8
x (cm)

-30

-25
-30
2

10

-15

5
-20

-25

7
8

-30
10

z (cm)

6
8
x (cm)

10

(c) P 3 y=8
1
-5

-10

(d) P 1y=5

-20

-5

-5

6
8
x (cm)

-15

(b) P2 y=6

-10

-25

-10

-15

5
-20

6
7
8
2

6
8
x (cm)

(e) P2 y=6

10

z (cm)

z (cm)

2
z (cm)

-10

z (cm)

-5

-10

-15

5
-20

-25

-30

-25
-30
2

6
8
x (cm)

10

(f) P 3 y=8

Figure 6. CS imaging with different wave travel velocity in soil and

10% additive Gaussian white noise. The true wave travel velocity
in soil is v = 1.5 108 m/s. (a)(c) Wave travel velocity in soil
v = 1.7321 108 m/s for P1 , P2 , P3 at y = 5, 6, 8 respectively. (d)(f)
wave travel velocity in soil v = 1.3416 108 m/s for P1 , P2 , P3 at
y = 5, 6, 8 respectively.

Progress In Electromagnetics Research M, Vol. 23, 2012

163

4.3. Effect of the Wave Travel Velocity Estimation Error

In above simulations, we view the wave travel velocity in soil as a known
parameter. However, in most of the real applications, it is unknown
or has a measurement error from the true value, which will directly
affect the imaging. In this subsection, the effect of the wave travel
velocity estimation error is discussed. The true velocity is 15 108 m/s
(r = 4). In the simulation, the measured data are sampled using
the true velocity, and the dictionary matrix is calculated using the
estimated velocity of 1.7321 108 m/s (r = 3) and 1.3416 108 m/s
(r = 5), respectively. The other parameters are kept the same in the
second simulation. The images are shown in Fig. 6.
Figure 6 shows that even if the wave travel velocity in soil is
wrongly estimated to some extent, the targets can still be correctly
imaged by the CS method, because in spite of certain difference
between the dictionary matrix constructed by the true velocity and
the one constructed by the estimated velocity, CS is based on the l1
norm optimization. By solving the convex optimization, the true image
can be reconstructed.
5. CONCLUSION
This work deals with the imaging of the three dimensional buried pointlike targets based on the compressive sensing method. By solving the
constrained convex optimization problem, the targets can be recovered
with much less sampled data, which will save the data collecting time.
Even with 10% additive noise, smaller bandwidth and some wave travel
velocity estimation error, the image of the targets can be reconstructed
with little degradation. Further work will be exerted on the application
of CS to small 3D block-like targets imaging of SFCWGPR.
ACKNOWLEDGMENT
This work was supported by the National Key Laboratory of Millimeter
Waves at Southeast University in China (No. K20121) and the Natural
Science Foundation of Shanxi Province, China (No. 2011JM8025).
REFERENCES
1. Grandjean, G., J. Gourry, and A. Bitri, Evaluation of GPR
techniques for civil-engineering applications: Study on a test site,
J. Appl. Geophys., Vol. 45, No. 3, 141156, 2000.

164

Cai et al.

2. Feng, X. and M. Sato, Pre-stack migration applied to GPR for

landmine detection, Inverse Prob., Vol. 20, 99115, 2004.
3. Groenenboom, J. and A. Yarovoy, Data processing and imaging
in GPR system dedicated for landmine detection, Subsurf. Sens.
Technol. Appl., Vol. 3, No. 4, 387402, 2002.
4. Hubbard, S., C. Jinsong, K. Williams, Y. Rubin, and J. Peterson,
Environmental and agricultural applications of GPR, Proc. 3rd
Int. Workshop on Adv. Ground Penetrating Radar, 4549, 2005.
5. Daniels, D., Ground Penetrating Radar, 2nd edition, London, UK,
2004.
6. Counts, T., A. C. Gurbuz, W. R. Scott, Jr., J. H. McClellan,
and K. Kangwook, Multistatic ground-penetrating radar experiments, IEEE Trans. Geosci. Remote Sens., Vol. 45, No. 8, 2544
2553, Aug. 2007.
7. Lopera, O., E. C. Slob, N. Milisavljevic, and S. Lambot, Filtering
soil surface and antenna effects from GPR data to enhance
landmine detection, IEEE Trans. Geosci. Remote Sens., Vol. 45,
No. 3, 707717, 2007.
8. Fang, G.-Y. and M. Sato, Stepped frequency ground penetrating
radar and its application for landmine detection, Acta Electronica
Sinica, Vol. 33, No. 3, 436439, 2005.
9. Gurbuz, A. C., J. H. McClellan, and W. R. Scott, A compressive
sensing data acquisition and imaging method for stepped
frequency GPRs, IEEE Transactions on Signal Processing,
Vol. 57, No. 7, 26402650, 2009.
10. Donoho, D. L., Compressive sensing, IEEE Trans. on
Information. Theory, Vol. 52, No. 4, 12891306, 2006.
11. Baraniuk, R. and P. Steeghs, Compressive radar imaging, Proc.
IEEE Radar Conf., 128133, 2007.
12. Yu. H.-M. and Y. Fang, Research on compressive sensing based
3D imaging method applied to ground penetrating radar, Journal
of Electronics & Information Technology, Vol. 32, No. 1, 1216,
2010.
13. Huang, Q., L. Qu, B. Wu, and G. Fang, UWB through-wall
imaging based on compressive sensing, IEEE Trans. Geosci.
Remote Sens., Vol. 48, No. 3, 14081415, 2010.
14. Wei, S.-J., X.-L. Zhang, J. Shi, and G. Xiang, Sparse
reconstruction for SAR imaging based on compressed sensing,
Progress In Electromagnetics Research, Vol. 109, 6381, 2010.
15. Zhang, L., M. Xing, C. Qiu, et al., Achieving higher resolution
ISAR imaging with limited pulses via compressed sampling,

Progress In Electromagnetics Research M, Vol. 23, 2012

16.
17.

18.
19.
20.
21.
22.

165

IEEE Geoscience and Remote Sensing Letters, Vol. 6, No. 3, 567

571, 2009.
Candes, E. J. and M. Wakin, An introduction to compressive
sampling, IEEE Signal Processing Magazine, 2130, Mar. 2008.
Johansson, E. M. and J. E. Mast, Three dimensional ground
penetrating radar imaging using a synthetic aperture time-domain
focusing, Proc. SPIE Conf. Adv. Microw. Millimeter Wave
Detectors, Vol. 2275, 205214, 1994.
Candes, E. and T. Tao, The Dantzig selector: Statistical
estimation when p is much larger than n, Ann. Statist., Vol. 35,
No. 6, 23132351, 2007.
Tuncer, M. A. C. and A. C. Gurbuz, Ground reflection removal in
compressive sensing ground penetrating radars, IEEE Geoscience
and Remote Sensing Letters, 2011.
Picardi, M., Background subtraction techniques A review,
Proc. IEEE Int. Conf. Syst. Man. Cybern., 30993104, Oct. 10
13, 2004.
Mayordomo, A. M. and A. Yarovoy, Optimal background
subtraction in GPR for humanitarian demining, Proc. 5th Eur.
Radar Conf., 4851, Oct. 2008.
Grant, M. and S. Boyd, CVX: Matlab Software for Disciplined
Convex Programming (Web Page and Software), 2011, Available:
http://stanford.edu/ boyd/cvx.