2. Proceedings of the 2nd International Conference on Current Trends in Engineering and Management ICCTEM -2014
17 – 19, July 2014, Mysore, Karnataka, India
has presented various local based stereo algorithms namely SAD, NCC, SSD and many more. In [7,
8, 9, 10, 11, 12], the author has presented various global based stereo algorithms and some of them
are global optimization and dynamic programming (DP). Local based stereo algorithms are also
similar to area based stereo algorithms, where only a smaller window is considered in order to avoid
unwanted smoothing. A larger window is considered in the areas of low texture because it contains
enough intensity variations so that a reliable stereo matching can be obtained. Area based stereo
algorithms mainly focuses on the aggregation of the matching cost. Alternatively, a global algorithm
makes a smoothness assumption first and then solves an optimization problem. In order to prevent
the problem of over smoothing, an energy function is necessary. Many methods are proposed past
from many decades in order to minimize the global cost. Graph cuts is an abstract representation of a
set of objects, where several pairs of the objects are connected by links. It is a mathematical structure
and is used to model Pair wise relations between objects from a certain collection. A different class
of global optimization algorithms is those based on dynamic programming. Dynamic programming
can find the global minimum for independent scan lines in polynomial time. Dynamic programming
was first used for stereo vision in sparse, edge-based methods (Baker and Binford, 1981; Ohta and
Kanade, 1985). Most of the global corresponding methods are very expensive sometimes which in
turn needs a huge set of parameters to determine. The motivation of this research is to develop an
algorithm for fast stereo matching that is able to produce smooth dense depth maps and preserve
enough depth discontinuity.
2. AREA BASED STEREO MATCHING ALGORITHMS
128
The main aim of the area based stereo matching algorithm is to estimate the similarities
between two or more images in order to obtain a dense disparity map from these stereo images.
Ideally, the block is very large enough to cover sufficient intensity variation so that the similarity
estimation is robust to noise. Similarity function plays a very important role in fast stereo matching.
Similarity function is also called as cost function. The cost function should be very robust to noise
and also illumination. Past from many decades most of the researchers have designed various cost
functions namely SAD, SSD, NCC, ZSAD, ZSSD, LSAD, and LSSD. Among these, SAD and SSD
are the popular cost functions and most widely used because of its simplicity in implementation. But
these two cost functions are very sensitive to illumination and camera gain. This is illustrated in
Figure.1.
Figure.1: Difference caused by various Camera Bias
The ZNCC stereo algorithm is used in order to deal with different camera bias. Even though
the ZNCC is very expensive, most of the researchers use this. As we all know that ZSAD is very
insensitive to differences in camera gain and computation is very less, we use this stereo matching
algorithm in our experiments. Further for the comparisons we use SAD, SSD, ZNCC, and ZSAD.
3. Proceedings of the 2nd International Conference on Current Trends in Engineering and Management ICCTEM -2014
17 – 19, July 2014, Mysore, Karnataka, India
129
7. %
3. ESTIMATION OF THE DISPARITY MAP
We consider a top layer of winner take all (WTA) in order to start disparity map estimation.
The result in terms of disparity map of winner take all contains an error when considering the flat
areas. Even though the window size is large we can see these errors. In Fig.2 we can see such errors.
As we can see in the figure that, the top layer is very small compare to the original image, so we are
applying dynamic programming in order to obtain a better quality estimation of the disparity map.
The cost measure can be defined as
'
' (
)*
Where C(x, y, d) is the cost measure at position (x, y) and d is the disparity and S is the
window size. I1 and I2 are the intensities of left and right stereo images. I1 and I 2 are the mean
values of intensities. We then apply interpolation on the disparity map of the top layer to obtain the
initial estimation of the disparity map for the second top layer.
If the disparity difference among neighboring positions exceeds a threshold value or the cost
measure at the position for the estimated disparity is too large, the disparity at that position will be
updated. The criteria for updating disparity are given as follows
|c(x, y)-c(x-1, y|μ, (6)
|c(x, y)-c(x, y-1)|μ, (7)
c(x, y, d)v, (8)
Where μ and v are thresholds and d (x, y) is the disparity at position (x, y). The updating is a
local process in which the continuity with two causal neighbors only is under the consideration.
4. EXPERIMENTAL RESULTS
We have evaluated the proposed fast stereo matching algorithm by considering several real
stereo image pairs. The first stereo pairs we have considered is Tsukuba stereo pair. This is one of
the widely used dataset to evaluate the stereo matching algorithms because this contains objects in
8. Proceedings of the 2nd International Conference on Current Trends in Engineering and Management ICCTEM
17 – 19, July 2014, Mysore, Karnataka, India
different depths. Our proposed stereo algorithm provides a very good quality in terms of disparity as
we can see in the Fig.2.
Fig.2: The estimation of the disparity map: (a) Left image of the Tsukuba. (b) The dispa
obtained by the proposed algorithm
The second example which we have considered in our evaluation is the images of road as we
can see in the fig.1. As we can see in the image that it contain trees and buildings. An illumination
difference in both the images play an important role as it is very significant. Except the cost function
both the disparity maps are evaluated and it is obtained by proposed algorithm called fast stereo
algorithm. The disparity map is shown if Fig.3. The results which
SAD based cross correlation algorithms (cost function). The results which is shown in fig.4. Is obtain
using ZSAD based cross correlation algorithms (cost function).
Fig.3: The estimation of the disparity
Fig.4: The estimation of the disparity map using ZSAD
5. CONCLUSION
This paper presents a novel fast stereo matching algorithm. In order to get the initial
estimation of disparity map, the dynamic
disparity map will be updated only in selected areas according to the local matching cost and the
depth difference among neighboring areas. The proposed algorithm is evaluated using rectified
stereo images. From the proposed algorithm we could obtain a better and very good result in terms of
high quality and also the speed. Experimental results shows that the ZSAD based stereo algorithms
are very robust to illumination and also robust to camera gain.
130
ained wn is shown in fig.3. Is obtain using
tion map using SAD-based cost function
ZSAD-based cost function
programming is applied on the top layer. The new dense
ages. -2014
disparity map
wn ion
9. Proceedings of the 2nd International Conference on Current Trends in Engineering and Management ICCTEM -2014
17 – 19, July 2014, Mysore, Karnataka, India
131
REFERENCE
[1] D. Scharstein and R. Szeliski, “A taxonomy and evaluation of dense two-frame stereo
correspondence algorithms,” International Journal of Computer Vision, vol. 47, no. 1/2/3,
pp. 7– 42, April-June 2002.
[2] L. D. Stefano, M. Marchionni, S. Mattoccia, and G. Neri, “A fast area-based stereo
matching algorithm,” 2002.
[3] C. Sun, “Fast stereo matching using rectangular subregioning and 3d maximum-surface
techniques,” International Journal of Computer Vision, vol. 47, no. 1/2/3, pp. 99–117,May
2002.
[4] J. Banks, M. Bennamoun, and P. Corke, “Non-Parametric Techniques for Fast and Robust
Stereo Matching,” in Proceedings of IEEE TENCON - Speech and Image Technologies for
Computing and Telecommunications, 1997, pp. 365–368.
[5] H. H. B. R. C. Bolles and M. J. Hannah, “The JISCT Stereo Evaluation,” in Proceedings of
DARPA Image Understanding Workshop, 1993, pp. 263–274.
[6] R. Zabih and J. Woodfill, “Non-Parametric Local Transforms for Computing Visual
Correspondence,”in Proceedings of ECCV, 1994, pp. 151–158.
[7] C. L. Zitnick and T. Kanade, “A cooperative algorithm for stereo matching and occlusion
detection,” IEEE Trans. on pattern analysis and machine intelligence, vol. 22, no. 7,
pp. 675–684, July 2000.
[8] O. V. Yuri Boykov and R. Zabih, “Fast approximate energy minimization via graph cuts,”
IEEE Trans. on pattern analysis and machine intelligence, vol. 23, no. 11, pp. 1222 – 1239,
November 2001.
[9] D. Terzopoulos, “Regularization of inverse visual problems involving discontinuities,” IEEE
Trans. on pattern analysis and machine intelligence, vol. 8, no. 4, pp. 413–424, April 1986.
[10] P. N. Belhumeur, “A bayesian approach to binocular stereopsis,” International Journal of
Computer Vision, vol. 19, no. 3, pp. 237– 260, 1996.
[11] S. Birchfield and C. Tomasi, “A pixel dissimilarity measure that is insensitive to image
sampling,” IEEE Trans. on pattern analysis and machine intelligence, vol. 20, no. 4,
pp. 401–406, April 1998.
[12] S. B. R. I. J. Cox, S. L. Hingorani and B. M. Maggs, “A maximum likelihood stereo
algorithm,” CVIU, vol. 63, no. 3, pp. 542–567, 1996.