Professional Documents
Culture Documents
discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/262202150
CITATIONS READS
15 104
2 authors:
All content following this page was uploaded by Christian Maas on 03 April 2017.
art ic l e i nf o a b s t r a c t
Article history: Ground Penetrating Radar (GPR) is used for the localization of supply lines, land mines, pipes and many
Received 5 November 2012 other buried objects. These objects can be recognized in the recorded data as reection hyperbolas with a
Received in revised form typical shape depending on depth and material of the object and the surrounding material. To obtain the
4 March 2013
parameters, the shape of the hyperbola has to be tted. In the last years several methods were developed
Accepted 15 April 2013
to automate this task during post-processing. In this paper we show another approach for the automated
localization of reection hyperbolas in GPR data by solving a pattern recognition problem in grayscale
Keywords: images. In contrast to other methods our detection program is also able to immediately mark potential
Reection hyperbola objects in real-time. For this task we use a version of the ViolaJones learning algorithm, which is part of
Haartraining
the open source library OpenCV. This algorithm was initially developed for face recognition, but can be
OpenCV
adapted to any other simple shape. In our program it is used to narrow down the location of reection
GPR
Adaboost hyperbolas to certain areas in the GPR data. In order to extract the exact location and the velocity of the
ViolaJones hyperbolas we apply a simple Hough Transform for hyperbolas. Because the ViolaJones Algorithm
reduces the input for the computational expensive Hough Transform dramatically the detection system
can also be implemented on normal eld computers, so on-site application is possible. The developed
detection system shows promising results and detection rates in unprocessed radargrams. In order
to improve the detection results and apply the program to noisy radar images more data of different
GPR systems as input for the learning algorithm is necessary.
& 2013 Elsevier Ltd. All rights reserved.
0098-3004/$ - see front matter & 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.cageo.2013.04.012
Please cite this article as: Maas, C., Schmalzl, J., Using pattern recognition to automatically localize reection hyperbolas in data from
ground penetrating radar. Computers & Geosciences (2013), http://dx.doi.org/10.1016/j.cageo.2013.04.012i
2 C. Maas, J. Schmalzl / Computers & Geosciences ()
reduced the data which has to be transformed to the Hough-space, independent and overcomplete basis of usable features for the
the computational demand is decreased dramatically and therefore analysis of sub-windows.
on-site application is possible. The computation of these features is based on the comparison
This paper is structured as follows: In the rst Sections 2 and 3 of pixel intensities. In the case of two rectangles the pixel sum of
the two parts of the developed detection program are presented, the black area is subtracted from the pixel sum of the white area
rst the ViolaJones implementation in OpenCV and after that the (Fig. 1). If the feature consists of three rectangles the sum of the
Hough Transform for hyperbolas. In Section 4 the training, the two white areas is added up and then subtracted from the sum of
utilized data and the detection program are introduced. Then in the inner black area.
Section 5 the results of the hyperbola localization are shown. At We assume that the top left corner of a rectangle r is located at
the end we nish with the discussion of the results (Section 6) and (x, y) and has the width b, the height h and the orientation f01; 451g,
conclusion and outlook in Section 7. so it is described by r x; y; b; h; . If the pixel sum of a rectangle is
denoted by RecSum, the features f have the following form:
f 1 RecSumr 1 2 RecSumr 2 ; 1
2. ViolaJones implementation in OpenCV where the weights 1 , 2 R consider differences in area size of the
different rectangles. To compute RecSum very quickly Viola and Jones
The Open Source Computer Vision Library (OpenCV) is writ- (2001) introduce the integral image.
ten in C/C++ and comes with many features for computer vision
and image processing. It was rst published by Intel in 1999, from 2.2. Integral image
2008 to 2012 it was further developed by Willow Garage, since
2012 it is supported by the company Itseez. In order to evaluate the features of Section (2.1) many pixel sums
The OpenCV-library contains a variant of the ViolaJones Face- have to be computed. This can be done very fast and in constant time
Detector (Viola and Jones, 2001) which was extended by Lienhart for any rectangle size by using the integral image approach, which is
and Maydt (2002) and is called Haartraining. This algorithm derived from the Summed Area Table (Crow, 1984). The integral
belongs to the supervised learning algorithms with strictly sepa- image intx; y of the original image therefore has to be calculated
rated learning and detection processes. In this section we want to only once. Then every feature can be determined by four table
present the basic concepts of the ViolaJones Algorithm. lookups, regardless of scaling and position. Intx; y is dened after
The basic idea of this algorithm is to learn during the so-called Viola and Jones (2001) as pixel sum of the rectangle with the top left
training to identify an object by analyzing many positive (containing corner at 0; 0 and the bottom right corner at (x, y)
object) and negative (not containing object) sample images. When
training is completed the output classier can be used for object intx; y ix; y: 2
x x;y y
recognition.
During detection the algorithm divides the input images into It can be derived with the following expression in one pass
many sub-images by moving a search window at multiple scales from left to right and top to bottom over all pixels of the original
over it. Then each sub-window is classied as object or no-object. image (Lienhart and Maydt, 2002):
For this task haar-like features are used. intx; y intx; y1 intx1; y ix; yintx1; y1; 3
with
2.1. Haar-like features int1; y intx; 1 int1; 1 0: 4
Based on Eqs. (2) and (3) the pixel sum of a rectangle with r (x, y,
Viola and Jones (2001) use a combination of simple haar-wavelet-
b, h, 01), for example rectangle D in Fig. 2, can be determined
like features shown in Fig. 1 to analyze each sub-image and classify it
with four table lookups
in two classes: object or no-object. Originally Viola and Jones
introduced the features 1a, 1b and 2a-c, Lienhart and Maydt (2002) RecSumr intx1; y1 intx b1; y h1
extended this set to 451 rotated features (1c, 1d, 2e-h, 3b). intx1; y h1intx b1; y1: 5
The advantage of using features instead of raw pixel values is
For rotated features the integral image can be computed
the possibility to compensate little variations in the appearance of
analogously. In this case two passes over the original image are
the object which makes classication easier. Furthermore they can
needed (see Lienhart and Maydt, 2002).
be calculated very fast by utilizing the so-called integral image
(see Section 2.2). In addition to that they can be scaled indepen-
2.3. Adaboost
dently in vertical or horizontal direction to build up a linear
Please cite this article as: Maas, C., Schmalzl, J., Using pattern recognition to automatically localize reection hyperbolas in data from
ground penetrating radar. Computers & Geosciences (2013), http://dx.doi.org/10.1016/j.cageo.2013.04.012i
C. Maas, J. Schmalzl / Computers & Geosciences () 3
Fig. 2. The principle of the integral image: The sum of all pixels in rectangle D can be calculated with four array references. At position 1 the value of the integral image is the
sum of the pixels in rectangle A. At location 2 it is the sum of rectangles A and B together, while at location 3 the value is the sum of the rectangles A and C. At 4 the value of
the integral image is A B C D. Therefore the pixel sum of rectangle D can be computed with 4 12 3 (Barreto et al., 2004).
Fig. 3. Schematic depiction of the detection cascade with N stages. In every stage one strong classier, which consists of several weak classiers, decides whether the sub-
window contains the object or not. If it does not contain the object, the sub-window is rejected for further detection (after Viola and Jones, 2004).
For the training process an enormous number of positive and The nal strong classier:
8
negative sample images is necessary. In every boosting round > T 1 T
<1 t ht x t
Adaboost chooses a low number of good classication functions hx 2t 1
t1
(weak classiers) that optimally describe the object and separate >
:
0 otherwise
the positive and negative sample images. To minimize the misclassi-
ed examples the weak classiers are repeatedly applied on different with t log1=t
input images. Misclassied samples get a higher, and correctly
classied examples a lower weight for the next boosting round. This
process is repeated several times. Therefore the boosting algorithm
concentrates on samples which are harder to learn.
Finally Adaboost forms a nal strong classier out of the 2.4. Cascade of classiers
trained weak classiers. In the following the Adaboost algorithm
after (Viola and Jones, 2001) is described. To increase the detection performance and reduce the computa-
tion time Viola and Jones use a cascade of classiers, which is
Given example images x1 ; y1 ; ; xN ; yN with yi f0; 1g, yi 1 comparable with a degenerated decision tree (Viola and Jones, 2001).
for positive sample images, yi 0 for negative sample images. Each stage of a cascade consists of a strong classier which was
Initialize weights: 1;i 1=2m; 1=2l for yi f0; 1g, m and l are trained with Adaboost and consists of a linear combination of weak
the number of negatives and positives respectively. classiers. During detection-process the strong classier of the rst
for t 1; ; T do stage makes a decision whether the sub-window contains the object.
(1) Normalize the weights: t;i t;i =nj 1 t;j If this is true, it is passed to the second stage of the cascade. When
(2) For each feature j, train a weak classier hj xf0; 1g. The the result is negative the sub-window is rejected immediately and is
error j is calculated with respect to t by not analyzed by the following stages (see Fig. 3).
j i i jhj xi yi j. The aim of this approach is to eliminate as much sub-windows
(3) Choose the classier ht with the lowest error t . as possible in the lower stages of the cascade. Therefore simpler
(4) Update the weights: i;t1 i;t 1e
t
i
; where ei 0 if exam- classiers are used, whereas in higher stages because of the
ple xi is classied correctly, otherwise ei 1, and increasing difculty of the problem the classiers are getting
t t =1t . complexer with each stage.
During training new stages are added to the cascade until given
end for detection and false-positive rates are achieved. These values can
Please cite this article as: Maas, C., Schmalzl, J., Using pattern recognition to automatically localize reection hyperbolas in data from
ground penetrating radar. Computers & Geosciences (2013), http://dx.doi.org/10.1016/j.cageo.2013.04.012i
4 C. Maas, J. Schmalzl / Computers & Geosciences ()
be determined for a cascade with K stages as follows: combination. Therefore the largest disadvantage of this method
K
is the high computational requirement: The more parameters are
F fpi ; 6 needed for the description, the more time is necessary for the
i1 calculations. Our approach is to rst limit the image area that has
with fpi false-positive rate of the i-th stage. The overall detection to be transformed by the ViolaJones Algorithm to decrease the
rate D of the cascade can be determined analogously computational demand dramatically and make an on-site use
possible.
K
D di : 7
i1
The second part of the program consists of a simple Hough For training and evaluation of the nal cascades four different
Transform for hyperbolas, which is based on the classical Hough sets of GPR data are used. One was recorded using Buttery-
Transform for nding lines in an image that was patented by antennas with a mean frequency of 50 MHz and a GSSI2 Terra-
Hough (1962). Duda and Hart (1972) extended it to the detection Sirch System SIRs -3000. The second data set was measured with
of simple geometrical shapes in binary images. The advantage of a 200 MHz GSSI antenna, type 5106 and Wu-King antennas with
this method is the reliability in noisy and disturbed binary images. a mean frequency of 30 MHz. The other two sets each contain data
It takes advantage of the fact that simple geometrical objects can of a GSSI 5103 antenna with 400 MHz and a antenna type
be described completely by a few parameters. For the classical 3101D with 900 MHz. All utilized GPR data is unprocessed, only
Hough Transform all points of the searched object are transformed an automatic gain function was applied which can also be
according to the appropriate mathematical equation from the calculated on-site.
binary image into the parameter or Hough space. Now the best For testing and evaluating the developed detection program all
tting parameters for the geometrical gure can be extracted by available data is divided into parts. Only the rst three sets are
locating the position of the local maximum in the parameter utilized for the learn process, out of them the input images for
space. training and also two test sets with 133 images each (test set 1
More details and applications of the Hough Transform can be and test set 2) are created. The test sets are not used for the
found for example in Illingworth and Kittler (1988). learning process and are chosen randomly, they are necessary to
For the implementation of the Hough Transform we utilize the evaluate and test the nal detection program. All images of the
concept of the accumulator array. This array has the dimension d fourth set form a third test set (test set 3) and are only used for
that corresponds to the number of parameters of the object, the evaluating the nal cascades. This data set consists of 76 images.
cells of the array match to discrete values of their corresponding
parameters. Therefore they have to be divided into discrete values, 4.2. Haartraining
the range has to be chosen before calculation. In the following the
general algorithm for the Hough Transform for different simple For the training of the nal cascade of classiers a huge number
geometrical shapes is illustrated (after Huang et al., 1985): of positive and negative sample images is necessary. The exact
position of each reection hyperbola has to be marked by hand in
Determine the minimum and maximum values for each para- the positive sample images. Therefore the GPR data is converted
meter. Through this the dimension of the accumulator array is into grayscale images and ipped vertically to double the number
dened. of available positive sample images. This seems to be reasonable
Initialize all accumulator cells with zero. because we assume the background noise in the data has no
Transform each non-zero binary element from the image plane vertical symmetry. Then every hyperbola is marked with a
to the parameter plane according to the mathematical equation rectangle, the exact position is saved and the resulting positive
and increment the corresponding accumulator cell by one. sample images (see Fig. 4) are scaled to a xed resolution for
Extract the matching parameter for the geometrical gure by training, for example to 20 20 pixel. The ratio of height to width
searching for a local maximum in the parameter space. of the marked hyperbolas is very important. It has to correspond
to the aspect ratio of the basic resolution which is xed for the
In our case the object that has to be located is a hyperbolic whole training process. In addition to that the hyperbola has to be
reection. After Ristic et al. (2009) the travel time of the reected centred in the rectangle which should not contain too much, but
wave can be expressed with the approximation of a point reector also not too little background. Otherwise the background would be
through recognized as part of the hyperbolas and they would not be
4 detected without this unique background. To obtain negative
t 2 t 20 xx0 2 : 8 sample images (Fig. 5) the grayscale images are cut, so that they
v2
do not contain any hyperbola. The resolution of the negatives can
The form of the resulting reection hyperbola depends on the
be random, however they should not be smaller than the basic
velocity v and the position x0 ; t 0 of the vertex, x is the horizontal
resolution, because the training algorithm automatically creates
displacement and t the corresponding travel time. For the Hough
negative sample images appropriate to the used resolution.
Transform we rearrange Eq. (8)
After preparations for the training process a total of 19 nal
r
4 cascades with different basic resolutions, sample images and
t 0 t 2 2 xx0 2 ; 9 learning parameters are created for this work in order to nd
v
the best parameter combination. All in all for training following
where (x, t) represent the coordinates of the reection hyperbola
settings are used:
and x0 ; t 0 and v are the parameters that are Hough-transformed.
Normally the whole image has to be transformed several times
for different parameter combinations to nd the best-tting 2
Geophysical Survey Systems, Inc.
Please cite this article as: Maas, C., Schmalzl, J., Using pattern recognition to automatically localize reection hyperbolas in data from
ground penetrating radar. Computers & Geosciences (2013), http://dx.doi.org/10.1016/j.cageo.2013.04.012i
C. Maas, J. Schmalzl / Computers & Geosciences () 5
Fig. 4. A selection of positive sample images for training, the resolution is 24 24 pixel.
Please cite this article as: Maas, C., Schmalzl, J., Using pattern recognition to automatically localize reection hyperbolas in data from
ground penetrating radar. Computers & Geosciences (2013), http://dx.doi.org/10.1016/j.cageo.2013.04.012i
6 C. Maas, J. Schmalzl / Computers & Geosciences ()
Fig. 6. Example for the output of the detection program, the image belongs test data set 3 (900 MHz antenna). On the left side no post-processing was done, that is why a lot
of multiple detections can be seen. On the right side post-processing was applied and all multiple detections disappear. In addition to that two correct detections are
removed (arrows), because they were only detected once.
Fig. 7. One result of the Hough Transform for hyperbolas. At rst the sub-window (at the top of the left) is smoothed and transferred into a binary image by a canny edge
detector (at the top in the middle). Afterwards a Hough Transform is applied on the sub-window, the calculation took 384 ms for a sub-window resolution of 47 62 pixel.
The calculated hyperbola equation is drawn in red (at the top of the right). At the bottom the accumulator array is illustrated, the cell with the maximum count can be seen.
vertex and the velocity v are extracted. These parameters are used cascades show the best results on test data set 3. The reason for
to calculate the travel time t with Eq. (8). At the end the detected that is, that this set contains the clearest hyperbolas and that the
shape of the reection hyperbola is drawn in the original input noise level of the data is low. The other two sets contain noisy data
image (see Fig. 7). and reections that are much more difcult to recognize.
In order to utilize the performance-program sample images
with labelled hyperbolas are needed as input. For this we use the
5. Results images we created for testing purposes during training prepara-
tion (Section 4.2). The performance-program utilizes the trained
In this section the cascade of classiers with the best results is cascade of classiers, applies it on all input images and compares
introduced. For evaluating the detection results and comparing the detected and manually marked positions of hyperbolas. If a
them to other cascades we use the performance-program (see detection matches the manual labelling it is rated as correct
Intel and Seo, 2006) that is delivered with the Haartraining detection. On the other hand, if the position of the detection
package of OpenCV. Every nal cascade of classiers was applied differs from the marking, or the detection is not manually marked
on the three test data sets using this program (see Table A1). All as hyperbola, it is regarded as false-positive detection. In addition
Please cite this article as: Maas, C., Schmalzl, J., Using pattern recognition to automatically localize reection hyperbolas in data from
ground penetrating radar. Computers & Geosciences (2013), http://dx.doi.org/10.1016/j.cageo.2013.04.012i
C. Maas, J. Schmalzl / Computers & Geosciences () 7
Fig. 8. Two examples for the output of the performance program. In image (a) two false detections can be seen, apart from that all manually labelled hyperbolas (blue)
were detected (green). At the bottom ve hyperbolas were identied correctly, furthermore one unmarked hyperbola was found (red) whereas the hyperbola at the top right
corner was missed. (For interpretation of the references to color in this gure caption, the reader is referred to the web version of this paper.)
Please cite this article as: Maas, C., Schmalzl, J., Using pattern recognition to automatically localize reection hyperbolas in data from
ground penetrating radar. Computers & Geosciences (2013), http://dx.doi.org/10.1016/j.cageo.2013.04.012i
8 C. Maas, J. Schmalzl / Computers & Geosciences ()
Fig. 10. Two results of the nal detection program. Radargram (a) belongs to test data set 2 and was recorded with a 400 MHz antenna. Radargram (b) belongs to test data
set 3 and was measured with a 900 MHz antenna. The detection results of the ViolaJones detector are drawn in red, the hyperbola equations that are calculated by the
Hough Transform are drawn in blue. In the left image all (clear) hyperbolas were found, on the right at least two ones were missed (arrows). (For interpretation of the
references to color in this gure caption, the reader is referred to the web version of this paper.)
Please cite this article as: Maas, C., Schmalzl, J., Using pattern recognition to automatically localize reection hyperbolas in data from
ground penetrating radar. Computers & Geosciences (2013), http://dx.doi.org/10.1016/j.cageo.2013.04.012i
C. Maas, J. Schmalzl / Computers & Geosciences () 9
Fig. 12. Three more results of the nal detection program. All radargrams were measured with a 200 MHz antenna and belong to test data set 2.
Al-Nuaimy, W., Huang, Y., Nakhkash, M., Fang, M., Nguyen, V., Eriksen, A., 2000.
Automatic detection of buried utilities and solid objects with GPR using neural
networks and pattern recognition. Journal of Applied Geophysics 43, 157165.
A comparison with other available methods is difcult, because Barczak, A., Dadgostar, F., 2005. Real-time hand tracking using a set of cooperative
the most studies use processed data. For example Simi et al. (2008) classiers based on haar-like features. In: Research Letters in the Information
achieve slightly higher detection and lower false-positive rates but and Mathematical Sciences. Citeseer.
these results are not comparable to our since they process and Barreto, J., Menezes, P., Dias, J., 2004. Humanrobot interaction based on haar-like
features and eigenfaces. In: Proceedings of the IEEE International Conference
optimize the data for hyperbola extraction. All in all our results on Robotics and Automation, 2004. ICRA'04. IEEE Comput. Soc. pp. 18881893.
promise a detection system with much better results than could be Bradski, G., Kaehler, A., 2008. Learning OpenCV: Computer Vision with the OpenCV
achieved with the few available data for this work. As can be seen Library. O'Reilly Media.
in several other studies the achieved results strongly depend on Canny, J., 1986. A computational approach to edge detection. IEEE Transactions on
Pattern Analysis and Machine Intelligence 8, 679698.
the data quality available for training. Chen, H., Cohn, A., 2010. Probabilistic robust hyperbola mixture model for
interpreting ground penetrating radar data. In: The 2010 International Joint
Conference on Neural Networks (IJCNN). IEEE Computer Society, pp. 18.
Crow, F., 1984. Summed-area tables for texture mapping. ACM SIGGRAPH Computer
7. Conclusions and outlook
Graphics 18, 207212.
Cui, Y., Wang, L., Xiao, J., 2010. Automatic feature recognition for GPR image processing.
Altogether the results of the developed detection program is World Academy of Science, Engineering and Technology 61, 176179.
capable of the automatic localization and tting of reection hyper- Dailey, M., Bo, N.,. Towards real-time hand tracking in crowded scenes DOI:
10.1.1.59.6410, in press.
bolas in unprocessed GPR data in real-time. By using this method the Duda, R., Hart, P., 1972. Use of the hough transformation to detect lines and curves
expenditure of time when searching for objects that produce a in pictures. Communications of the ACM 15, 1115.
Please cite this article as: Maas, C., Schmalzl, J., Using pattern recognition to automatically localize reection hyperbolas in data from
ground penetrating radar. Computers & Geosciences (2013), http://dx.doi.org/10.1016/j.cageo.2013.04.012i
10 C. Maas, J. Schmalzl / Computers & Geosciences ()
Freund, Y., 1996. A decision-theoretic generalization of on-line learning and an Peters, V., 2006. Efzientes Training ansichtsbasierter Gesichtsdetektoren. Master's
application to boosting. Journal of Computer and System Sciences 55, 119139. Thesis. Universitt Bielefeld.
Freund, Y., Schapire, R., 1996. Experiments with a new boosting algorithm. In: Ristic, A., Petrovacki, D., Govedarica, M., 2009. A new method to simultaneously
Proceedings of the 13th International Conference on Machine Learning. Morgan estimate the radius of a cylindrical object and the wave propagation velocity
Kaufmann Publishers Inc., pp. 148156. from GPR data. Computers & Geosciences 35, 16201630.
Hough, P., 1962. Method and Means for Recognizing Complex Patterns. Technical Shihab, S., Al-Nuaimy, W., Eriksen, A., 2002. Image processing and neural network
Report. techniques for automatic detection and interpretation of ground penetrating
Huang, K., Fu, K., Sheen, T., Cheng, S., 1985. Image processing of seismograms: radar data. In: Proceedings of Sixth WSEAS International Multi-Conference on
(a) hough transformation for the detection of seismic patterns; (b) thinning Circuits, Systems, Communications and Computers.
processing in the seismogram. Pattern Recognition 18, 429440. Simi, A., Bracciali, S., Manacorda, G., 2008. Hough transform based automatic pipe
Illingworth, J., Kittler, J., 1988. A survey of the hough transform. Computer Vision, detection for array gpr: algorithm development and on-site tests. In: Radar
Graphics, and Image Processing 44, 87116. Conference, 2008. Radar'08. IEEE. IEEE Computer Society, pp. 16.
Intel, C., Seo, N., 2006. Haartraining documentation. visited 11.05.2011. Treptow, A., Zell, A., 2004. Real-time object tracking for soccer-robots without color
Kaneko, T., 1991. Radar image processing for locating underground linear objects. information. Robotics and Autonomous Systems 48, 4148.
IEICE Transactions 74, 34513458. Viola, P., Jones, M., 2001. Rapid object detection using a boosted cascade of simple
Lienhart, R., Kuranov, A., Pisarevsky, V., 2003. Empirical analysis of detection cascades of features. In: IEEE Computer Society Conference on Computer Vision and Pattern
boosted classiers for rapid object detection. Pattern Recognition, 297304.
Recognition. IEEE Computer Society, pp. I511.
Lienhart, R., Maydt, J., 2002. An extended set of haar-like features for rapid object Viola, P., Jones, M., 2004. Robust real-time face detection. International Journal of
detection, in: Proceedings of the International Conference on Image Processing, Computer Vision 57, 137154.
2002. IEEE Computer Society, pp. I900.
Please cite this article as: Maas, C., Schmalzl, J., Using pattern recognition to automatically localize reection hyperbolas in data from
ground penetrating radar. Computers & Geosciences (2013), http://dx.doi.org/10.1016/j.cageo.2013.04.012i
View publication stats