Displacement Estimation of a RC Beam Test based

on TSS algorithm

G. Almeida, Hugo Biscaia, Carlos Chastre, J. Fonseca F. Melcio

Faculdade de Cincias e Tecnologia Departamento de Engenharia Electrotcnica e Automao
Universidade Nova de Lisboa Instituto Superior de Engenharia de Lisboa
Lisboa, Portugal Lisboa, Portugal

AbstractThe traditional methodology used in civil engineering At first sight, the edge information is better because it
measurements requires a lot of equipment and a very complex occupies less data storage but it requires a careful choice of the
procedure especially if the number of target points increase. right threshold. Nevertheless, block-matching algorithms is a
Since the beginning of the current century, several studies have powerful technique but demands a lot of memory. Is it viable to
been conducted in the area of photogrametry using digital image use edge information for fast search on global areas and block
correlation associated with block motion algorithms to estimate matching for detailed processing on restricted areas?
displacements in reinforced concrete (RC) beams during a load
test. Using image processing techniques it is possible to measure What should the relationship between the entire image and
the whole area of interest and not only a few points of the tests the number of target points be? The information obtained by
materials. In this paper, block-matching algorithms are used in LVDT sensors will be useful to establish this relationship.
order to compare the results from photogrametry techniques and
the data obtained with linear voltage displacement transducer II. REVIEW OF RELATED WORK
(LVDT) sensors during the load tests of RC beams, which are Several works are under development in this area, namely
very common to find in civil engineering laboratories. at Dresden University of Technology [1], Waseda University
[2] and Lulea University of Technology [3] as discussed below.
Keywords- Photogrametry, Digital Image Correlation, Block
Motion Estimation The detection of early cracks with a fraction of a pixel can
be seen in Hampel & Maas [1]. The authors developed and
I. INTRODUCTION tested a cascade image analysis aiming at detecting fine cracks
Photogrammetry is more and more used as an additional in the micrometer range. Their paper started with a review of
power tool in several tests in civil engineering such as load related work in three categories of methods to automatically
tests, crack measurements, material tests inspection. With detect cracks in a non-cracked concrete surface: Edge detection
digital photography and image processing it is possible to techniques, Targeting and Full field displacement vector
automatically measure deformations, displacements, strain analysis. The present paper deals with cascade image analysis
fields and surface defects in many material tests and also because those techniques are applied subsequently in two steps.
analyze the entire area of interest, and not only a small area, Firstly, image-matching techniques are applied to consecutive
with simple procedures. images of an image sequence to generate sub pixel-accuracy
surface patch displacement vector fields. This displacement
The major challenge is to produce the most adequate
vector length image will show discrepancies in the vector
methodology with efficient algorithms in order to maintain
length at the position of the cracks when compared to the non-
good precision and measuring range.
cracked surface. In a second image-processing step, these
The digital image correlation is a method that examines discrepancies are detected and extracted by edge detection
consecutive images taken during the deformation period with a techniques applied to the displacement vector length image. As
mathematical correlation algorithm. Concrete is a common a result, an image containing information about the position
material used in civil engineering and has a good texture. and the width of cracks is obtained.
However, it is better to apply a speckle pattern to concrete
Yamaguchi & Hashimoto [2] introduce a new crack
surfaces (a random dot pattern) in order to get an easier image
detection method for a concrete surface image based on a
processing. Some experiments have been conducted, such as
percolation model. This model utilizes edge information to
the comparison of different block matching algorithms and the
reduce the number of starting points for percolation processing.
use of the edge techniques, but some questions remain without
Percolation is a physical model based on the natural
an answer.
phenomenon of liquid permeation. The authors use edge
What kind of correlation technique is more efficient? information for the choice of applicable seed pixels and
conducted tests with images of real concrete surfaces. In those
Is block matching better than edge information to obtain the tests, the methodology for crack detection was applied to 10
displacement information from an image? images of real concrete surfaces. The size of the images was
500 500 [pixels], corresponding to an area of approximately The match block chosen is based in the least cost. The cost
250 250 [mm] in the actual surface scale. function normally used is the mean square error (MSE) or the
mean absolute difference (MAD).
Carolin, Olofsson & Taljsten [3] present a non-touching
strain measurement method that covers a pre-defined area. N N
1
   Cij  Rij 
Several tests were carried out on RC beams with a span of 4.5
m and the results were compared to the traditional electrical  MAD  
strain gauges. In this study a speckle pattern correlation was
N i1 j 1
used and a photo was taken every 30kN until rupture. In their
tests they used a camera film and have to scan the film with a N N 2
1
high quality scanner. The picture was divided into sub-picture,
128x128, and then subjected to a threshold. After this they
 MSE     Cij  Rij 
N i1 j 1
    

calculated the centre of gravity of the black & white picture.

With this methodology and the speckle correlation it was where N is the side of the block, and Cij e Rij are the pixels in
possible to find the same sub-picture in the second loading the current and reference block.
condition.
The general idea of TSS is to start the search location at the
One of the most popular algorithms for motion estimation middle and with a step size, S, and search the least cost
is the block-based search algorithm - Three Step Search (TSS). function of the eight locations around this location. The least
TSS is widely acceptable and is recommended for MEPEG 4 cost makes it the new search origin and the step size will be
[4]. There has been substantial improvement of TSS, for S=S/2. This procedure is repeated until S=1.
instance the Simple and Efficient Search (SES) [5, 6], which
reduce the computational complexity for motion estimation. The algorithm used in this paper is the Simple and Efficient
Search algorithm (SES) presented in Ming & Liou [5] that is
The digital image correlation is the technique applied in [7]. divided in two phases.
The images are divided into small subsets of gray level data.
Each subset of the original subset is compared with the subset The SES algorithm with these two phases promotes the
of the second image using the cross-correlation error function. reduction of time computation due to uniformity of the pattern.
The number of steps of the algorithm is dependent on the
When the image processing is used, it is important to know search window. The first phase consists of selecting the search
what the contribution of several factors is such as the resolution direction quadrant, as we can see in Fig.2. The goal of the
versus the specimen dimensions, focal length and distance second phase is to find the smaller error location in the
between camera and specimen, distortion effects and speckle previous selected quadrant.
pattern. All these considerations can be found in Cintrn &
Saouma [8]. Also in Bornet et al. [9] the general procedure to For the selection of the search direction quadrant it is
evaluate the measurement errors of the digital image necessary to compute the MAD function in three points, A, B
correlation (DIC) is discussed. The authors used synthetic and C, and with the following four rules the right quadrant is
images and evaluated the displacements with six DIC selected:
packages. If MAD(A)
MAD(B) and MAD(A)
MAD(C) quadrant
III. BLOCK MOTION ALGORITHMS I is selected;
The basic idea of block motion estimation is to divide the If MAD(A)
MAD(B) and MAD(A)  MAD(C) quadrant
current image into a matrix of blocks and then compare these II is selected;
blocks with the previous image in order to calculate the
If MAD(A)  MAD(B) and MAD(A)  MAD(C) quadrant
displacement vectors, as we can see in Fig. 1.
III is selected;
The current block is searching in the previous image in a
delimited search area, p pixels around the current block. If MAD(A)  MAD(B) and MAD(A)
MAD(C) quadrant
Usually a block takes a square side of 16 and the search IV is selected.
parameter p is 7. With the knowledge of the quadrant where the block is
Current image Reference image
moving it is necessary to calculate other points that will help to
decide the best block match. This is known as the second
ij Blo
ck
ij phase.
ck ing
Blo tch
Ma lock
B
rea
hA
Se
arc
III II
A B

C
IV I
Figure 1. Current and reference image
Figure 2. First phase search direction
 In Fig.3, are indicated the initial points of phase 1, the black random speckle pattern manually painted using a large brush
circle, as the auxiliary points of phase 2 and the square point. with mate black ink (see Fig. 4).
The photos were taken during approximately 30 minutes
with intervals of 30 seconds, resulting on a sequence of about
A B A B 60 photos. In the last 10 minutes of the test, more images were
discarded for processing purposes because 20 minutes the test
C C the beam collapsed and its displacements were no longer object
of interest.
a) b) From the original image until the last one, the displacement
can be very large (see Figure 4 a) and b)) and the tests showed
that sometimes would be necessary to assume a tied interval
between consecutive images for a better correct speckle motion
A B A B
detection. A possible solution to this problem is to reduce the
time interval between frames on the faster parts of the tests.
C C
Alternatively, it could be sufficient to have some feedback
from the applied load to the acquisition system in order to
c) d) accelerate the image acquisition when the beam has more
Figure 3. Search points, phase 2, for each quadrant a) I, b) II, c) III, d) IV significant displacements. The prediction of the beams
behaviour by theoretical analysis should also improve the
IV. RESULTS image acquisition by determining, in advance, the expected
Several load tests have been carried out in partnership with reaction of the materials to the applied loads.
researchers from the Department of Civil Engineering from In these tests, LVDT sensors were also strategically
Universidade Nova de Lisboa (UNL). In Fig. 4 is possible to positioned in order to obtain reference data that could be
observe the undeformed RC beam (a) before the load test, the compared with the image processing results. These sensors
deformation of the RC beam near collapse (b) showing several require a lot of time to assemble and calibrate. Moreover,
cracks and one picture of the RC beam with the work area LVDT sensors can be easily damaged at the rupture instant of
analyzed in this study (c). the RC structure due to the indispensable proximity between
The image acquisition was done with a digital Cannon EOS them. This is one of the advantages of the photometric
400D camera with a resolution of 3888x2592 and two spots of approach since it doesnt require any setup physically under the
500W each guaranteed artificial light. The artificial lightning RC beam or even touching it. All the measurements are done at
was used in order to maintain a constant light environment. All a distance without any particular or long calibration, having a
the images were captured on RAW format in order to avoid low cost support, and they are easy to implement in the RC
image compression. beam test setup. In order to get a visual photometric reference,
an ordinary ruler was used on a sample photo as a reference for
all the other photos.
Another advantage of photogrammetry is that the RC beam
strains can be continuous measured over the entire area, while
strain gauges can only provide a measurement on a single
point.
In order to evaluate the RC block displacements between
consecutive images, several variations of the Three Step Search
(TSS) algorithm were tested: New Three Step Search (NTSS)
a) and Simple and Efficient Search (SES). The SES algorithm
revealed to be the most adequate to the study herein presented
and thus, it was the one used for the results shown in this paper
due to a less number of search points computations [6]. An
other variation of TSS which still under testing is the Diamond
Search (DS).
The digital image processing system compares two adjacent
c) images in order to evaluate the displacement. These results are
compared with the results acquired with one of the real sensors
b) used. In all tests more than 30 images were used and with the
Figure 4. Examples of the analyzed pictures a) First image b) Visible cracks successive displacements it was possible to make a graph of the
c) Detail of the beam painted with a crack displacement versus time (d vs t). In Fig. 5 the blue curve
indicates the information obtained by image processing and the
The RC beam was initially prepared with an underlying green one the signal of the local LVDT sensor in the same area.
painting of white mate ink and then with a superimposed
This graph was obtained by following the studied RC beam The final prototype developed will be used on the Civil
area in 38 consecutive images. Engineering Department of Universidade Nova de Lisboa for
displacement measurements and strain fields mapping of
several RC structures in particularly, and among others, beams
and columns.
VI. CONCLUSION
The comparison between the measurements from LVDTs
sensors and the measurements obtained from image processing
techniques revealed to be very similar. Hence, the image
processing technique herein used seems to be a promising
technique of measurement displacements. The results obtained
in this preliminary study shows that it is possible to continue
and explore this subject deeply and test the whole area of a RC
Figure 5. Displacement calculated vs measured. structure in order to compare with all the results of the LVDTs
sensor. Moreover, a strain field of the RC area in study can be
It is possible to see that the result obtained by the image reproduced and compared to other techniques of modeling
processing have good accuracy when compared with the structures such as FEA, in order to calibrate and validate the
LVDT sensor. As mentioned above a solution to reduce the image processor technique. The research will eventually
slight discrepancy observed between the displacements at the provide software that enables a real time monitoring of the RC
end of the test is to reduce the time interval between frames in structure during the experimental test.
the fast part of the test. At the moment, only the vertical In these tests, each photo was taken every 30 seconds;
displacement was compared. Nevertheless, at this point, a however, in future tests, a different frequency can be adopted
precision of 1 pixel is achieved, which in case of the performed and it should be based on the RC beam applied load. With this,
tests is equivalent to 0.3 mm. This evidence revealed to be the number of photos per test could be reduced.
more than enough for the measurements of the RC beam
displacements. REFERENCES
[1] H. Hampel, H.-G Maas, Cascade image analysis for dynamic crack
V. FUTURE WORK detection in material testing, ISPRS journal of Photogrammetry and
The next step of this work will be the study of the Remote Sensing, 64(2009)345-350, 2009.
importance of the pattern on the efficiency of the detection in [2] T. Yamaguchi, S. Hashimoto, Automated Crack Detection for concrete
order to achieve better resolution and more reliable results. Surface Image Using Percolation Model and Edge Information, IEEE,
2006.
Different patterns and pattern densities will be tested in order
[3] A. Carolin, T. Olofsson, B.Taljsten, Photographic strain monitoring for
to evaluate their influence on the final results. civil engineering, FRP Composites in Civil Engineering - CICE 2004,
Once determined the best pattern, our main goal will be to Seracino, 2004.
achieve the analysis of all RC beam area under testing in order [4] M. Al-Mualla, C.S. Canagarajah, David R. Bull, (2002). Video Coding
for Mobile Communications, Elsevier Science.
to determine strain maps of the material. The strain maps of a
RC element have the advantage of monitoring a substantial [5] J Lu, Ming L. Liou, A Simple and Efficient Search Algorithm for
Block-Matching Motion Estimation, IEEE TRANSACTIONS ON
area of the element, dispensing strain gages. Moreover, this CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 7,
technique can be compared with non-linear software based on NO. 2, 1997.
the finite element analysis (FEA) in order to corroborate the [6] A. Barjatya, Block Matching Algorithms for motion Estimation, DIP
strain fields in the RC structure. Is expected a good accuracy 6620 Spring 2004 Final Project Paper, 2004.
between this technique and FEA, so the next step will be the [7] S. Kurtz, P. Balaguru, J. Helm, Experimental Study of interfacial shear
monitoring of strain fields of RC structures in real time. stresses in FRP-strengthened RC beams, Journal of composites for
However, the image processing will be initially done off-line construction, ASCE, 2008.
but, as mentioned, the final objective is to achieve real time [8] R. Cintrn, V. Saouma, Strain Measurements with Digital Image
Correlation System Vic-2D, NEES at CU Boulder, 2008.
performance in order to provide feedback to the operator
during the experimental tests. Since the amount of computation [9] M. Bornet, F. Brmand, P. Doumalin, J.C. Dupr, M. Frazzini, M.
Grdiac, F. Hild, S. Mistou, J. Molimard, J. J. Orteu, L. Robert, Y.
required for the motion estimation is considerable, the selection Surrel, P. Vacher, B. Wattrisse, Assessment of Digital Image
and optimization of the image processing algorithms will be a Correlation Measurement Errors: Methodology and Results,
central point of the future work. Experimental Mechanics, HAL-CCSD, 2008.