International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 2

ISSN: 2321-8169
691 - 696

_______________________________________________________________________________________________

A New Approach for Stereo Matching Algorithm with Dynamic Programming

Miss. Priyanka M. Lohot

Mrs. Bhakti M. Sonawane

PG Student, Computer Engineering,

Shah & Anchor Kutchhi Engineering College,
Mumbai University.
[email protected]

Assistant Professor, Computer Engineering Department,

Shah & Anchor Kutchhi Engineering College,
Mumbai University.
[email protected]

Abstract Stereo matching algorithms are one of heavily researched topic in binocular stereo vision. Massive 3D information can be obtained
by finding correct correspondence of different points between images captured from different views. Development of stereo mat ching algorithm
is done for obtaining disparity maps i.e. depth information. When disparities computed for scan lines then dense reconstruction becomes time
consuming for vision navigation systems. So for pair of stereo images proposed method extracts features points those are at contours in images
and then a dynamic program is used to find the corresponding points from each image and calculates disparities. Also to reduce the noise which
may lead to incorrect results in stereo correspondence, a new stereo matching algorithm based on the dynamic programming is proposed.
Generally dynamic programming finds the global minimum for independent scan lines in polynomial time. While efficient, its performance is far
from desired one because vertical consistency between scan lines is not enforced. This method review the use of dynamic programming for
stereo correspondence by applying it to a contour instead to individual scan lines. Proposed methodology will obtain the global minimum for
contour array in linear time using Longest Common Subsequent (LCS) dynamic programming method with no disparity space image (DSI).
Keywords- Stereo Correspondence, Contour, Dynamic Programming, Disparity space Image, Longest common subsequent.
________________________________________________________*****_______________________________________________________

I.

INTRODUCTION

Stereo matching is an important branch of the research area

in computer vision. The mobile robot vision navigation system
suppose to detect the size and position of obstacle or object
from unknown environment in accurate manner and such
vision navigation system is carried by mobile robot are based
on fast and veracity stereo matching algorithm. In stereo
matching algorithm depth information can be obtained by
finding correct correspondence of different points between
stereo images captured from different views. Stereo
correspondence has traditionally been, and continues to be,
one of the most heavily investigated topics in stereo vision.
The known algorithms for stereo matching can be classified in
two basic categories: Area-based algorithms and Featurebased algorithms [2]. Feature-based algorithms (FBA) extract
features of interest from the images, such as corners, edge
segments or contours, and match them in two or more views.
These methods are fast because only a small subset of the
image pixels are used, but may fail if the chosen primitives
cannot be reliably found in the images; further more they
usually only yield Sparse depth map. FBA only calculate
points. In Area-based approaches, the system attempts to
correlate the grey levels of image patches in the views being
considered, assuming that they present some similarity. The
resulting depth map can then be interpolated. The underlying
assumption appears to be a valid one for relatively textured
areas; however, it may prove wrong at occlusion boundaries
and within featureless regions.
Alternatively, the map can be computed by directly fitting
a smooth surface that accounts for the disparities between the
two images. This is a more principled approach since the
problem can be phrased as one of optimization; how- ever, the
smoothness assumptions that are required may not always be
satisfied.

So the disparity map exported by area-based algorithms

can be dense enough. As a consequence, there are a large
number of spots in the disparity map and the algorithm is timeconsuming. Feature- based algorithms only calculate the
points which acquired by certain rules. So it produces a sparse
disparity map. The vision navigation of mobile robot just
needs to detect the outline of the obstacle without
reconstructing the surface of the obstacle precise. So, as above
considerations, we first extract the contour of the scene, then
calculate the disparity of the contour. When solving the
disparity using traditional matching algorithm (such as SAD,
SSD, etc.), the optimal disparity for each pixel is chosen
individually. A noisy disparity map which will generate the
wrong construction may be created, which will result in wrong
detecting of the obstacle. Aiming to solve this problem, Ohta
proposed a stereo correspondence method based on the
dynamic programming [5]. After initial matching costs were
computed by traditional matching algorithm, the dynamic
programming algorithm is applied to global optimization and
makes the quality of the disparity map obviously improved. In
order to eliminate stripes in disparity map, ground control
points (GCPs) are proposed by Stephen [8]. But all of these
algorithms are applied only to corresponding scan lines of
stereo pairs and not suitable for the contour. So a stereo
matching algorithm of contour based on dynamic
programming is proposed in this paper. Stereo vision is one of
the most actively researched area in computer vision. A large
number of stereo correspondence algorithms have been
developed. In order to gauge progress in the area, Schartein et
al. wrote a paper which provides an update on the state of the
art in dense two frame stereo correspondence algorithms under
known camera geometry [2]. The algorithms produce dense
disparity maps which are useful in a number of applications
including view synthesis, robot navigation, image-based
rendering and tele-presence. For researchers in stereo vision
one of the most useful outputs of the paper is a quantitative
691

IJRITCC | February 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 2

ISSN: 2321-8169
691 - 696

_______________________________________________________________________________________________
test bed for stereo correspondence algorithms available, from
vision.middlebury.edu/stereo. This paper reviews stereo
correspondence algorithms and also proposes a new approach
for dynamic programming on contour.
II.

EXSITING SYSTEM

This section reviews already discussed stereo correspondence

algorithms mostly which have received attention in the
Middlebury evaluation [1].
Since the excellent taxonomy presented by Schartein and
Szeliski many new methods have been proposed [2]. Matching
methods can be grouped into those producing sparse output
and those giving a dense result. Feature based methods stem
from human vision studies and are based on matching
segments or edges between two images, thus resulting in a
sparse output. This disadvantage, dreadful for many purposes,
is counterbalanced by the accuracy and speed obtained. In
order to categorize and evaluate the stereo correspondence
algorithms that produced dense output, a context has been
proposed [2]. According to this, dense matching algorithms
are classified in local and global once. Local methods (areabased) trade accuracy for speed. The disparity computation at
a given point depends only on intensity values within a finite
window. Global methods (energy-based) are time consuming
but very accurate. Their goal is to minimize a global cost
function that combines data and smoothness terms.
A. Disparity Space Image
Stephen represents a data structure called as the Disparityspace image, or DSI. The data structure explores the occlusion
and stereo problem and it facilitated in development of a
dynamic programming algorithm that uses occlusion
constraints. The DSI is an explicit representation of matching
space [8].

consider as matching space, with x along the horizontal, and

disparity along the vertical.
The corresponding epipolar scan line from the left and right
images are used. The scan line from the left image is held still
as the scan line from the right image is shifted across. After
each pixel shift, the scan line is subtracted. The result from
overlapping pixels is placed in the resulting DSI.
III.

PROPOSED METHOD

Objective of proposed work is to compute disparity maps for

stereo image pair simultaneously and efficiently. The
motivation behind proposed algorithm is to use the powerful
and efficient optimization tool provided by dynamic
programming, and instead applying it on individual scan line
use it on contours which will more suited for vision navigation
system. Disparity computed at one contour pixel depends on
the disparity at all others contour pixels. Thus the proposed
dynamic programming is truly semi global optimization
algorithm and is not 1D optimization method because it
operates across both vertical and horizontal dimensions.
Dynamic programming is a classical global optimization
algorithm on stereo correspondence. The traditional dynamic
programming is applied to corresponding scanlines to find
lowest cost rout in the disparity space image (DSI) to optimize
result. Further, a contour based dynamic programming with
disparity space image (DSI) used to obtain optimal disparity
[7]. This paper introduces an improved stereo matching
algorithm of contour based with longest common subsequent
(LCS) dynamic programming. This will reduce computational
complexity comparing to DSI and more accurate result
comparing to DSI on contours.

Figure 2 The general block diagram for execution of algorithm.

Figure 1. This figure describes how disparity space image is generated [8].

DSI representation for ith scanline in following way: Select

the ith scanline of the left and right images, sL sR respectively,
and slide them across one another one pixel at a time. At each
step, the scanlines are subtracted and the result is entered as
the next line in the DSI. The DSI representation stores the
result subtracting every pixel in sL and sR and maintains the
spatial relationship between the match points. As such, it may

The general block diagram for proposed system is shown in

Figure 2 which has three main sections as canny edge
detection, actual contour extraction and dynamic programming
with Longest Common Subsequent (LCS) method. This
method proposed new stereo matching technique based on
dynamic programming algorithm. Instead applying dynamic
programming on individual scan line this method uses dynamic
programming directly on contour array to find global
optimization and obtains optimal disparity. Traditional Scan
line based method requires polynomial time for execution as
well as disparity computed using search space technique on
692

IJRITCC | February 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 2

ISSN: 2321-8169
691 - 696

_______________________________________________________________________________________________
contour i.e. DSI on contour also takes polynomial time for
execution; this new method uses dynamic programming
directly on contour array which will execute in liner time. Sum
of points needed to be computed will be considerably reduced
when algorithm will compute stereo correspondence with high
speed and more accurate than dense reconstruction.
A. Canny edge Detection
The Canny edge detector is an edge detection operator that
uses a multi-stage algorithm to detect a wide range of edges in
images. Edges or contours in images are areas with strong
intensity contrasts. As the contour is the place where the great
changes have taken place on the image plane, we can remove
the area of near gray level effectively. Edge detecting an
image significantly reduces the amount of data and filters out
useless information, while preserving the important structural
properties in an image.
There are many traditional algorithms such as Sobel operator,
Prewitt operator, Laplace operator and so on. This entire
operator can perform at high speed. But detected edges are
always incomplete and uncontinuous and unsuited to be used
directly. By contrast, The canny operator is an optimization
algorithm for high signal to noise and precision of detection,
extensive used [12].
Apply canny edge detection algorithm for contour pixel
extraction of left and right image. The algorithm runs in 5
steps, apply each step on separate image of stereo pair ie. Left
and Right image.
Smoothing:
Blurring of the image to remove noise.
Finding gradients:
The edges should be marked where the gradients of the
image has large magnitudes.
Non-maximum suppression:
Only local maxima should be marked as edges.
Double thresholding:
Potential edges are determined by thresholding.
Edge tracking by hysteresis:
Final edges are determined by suppressing all edges that are
not connected to a very certain (strong) edge.
In proposed algorithm canny edge detection has been taken
directly. Which when applied on left and right images,
generated output gives detail edges with intensity 1 for all
contours and every non contour pixel appear with intensity 0.
This resultant canny left and right images pair used further for
getting stereo correspondence on contour.

Figure 3 Computing actual contour pixels from gray scale image.

Contour extraction mainly took place to obtain exact shape

and size information. In order to find correspondence in
reference image the actual intensity is needed. Therefore
actual contour intensities extraction took place.
C. Dynamic programming Principal
Stereo matching through dynamic programming (DP) an be
considered a semi-global method [8]. This algorithm gets the
global minimum for independent scanlines in polynomial time.
Problems with dynamic programming stereo include the
selection of the right cost for occluded pixels and the difficulty
of enforcing inter-scanline consistency, although several
methods propose ways of addressing the latter. Another
problem is that the dynamic programming approach requires
enforcing the monotonicity or ordering constraint. This
constraint requires that the relative ordering of pixels on a scan
line remain the same between the two views, which may not
be the case in scenes containing narrow foreground objects.
The stereo problem in this context becomes that of nding the
minimum cost path, or path of least resistance, through a cost
matrix such as the one shown in Figure 4. This matrix is
constructed by calculating the dierences of all the pixels in
the reference scan line and all the pixels of the corresponding
scan line over a range of disparity levels.
The name dynamic programming comes from a mathematical
process through which a problem is solved by breaking it into
smaller pieces. The smaller problems are solved rst and the
collection of all the smaller problems is equivalent to the
problem as a whole. The process is used to nd the minimum
cost path through a cost matrix derived from the image
scanlines. First the possible paths through the cost matrix are
dened using some constraints.

B. Actual Contour Intensity

Canny edge detection obtains same intensity for all contour
pixels. In solving problem for matching algorithm it is
necessary to acquire actual contour intensity. Here mapping of
canny edge with gray level image come to picture for stereo
pair is shown in following Figure 3.

Figure 4 Stereo matching using dynamic programming [8].

The main constraint used in DP is the ordering constraint. The

cost of a point is dened in a similar manner to the cost in
Equation 1.
693
IJRITCC | February 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 2

ISSN: 2321-8169
691 - 696

_______________________________________________________________________________________________
E(f) = Edata(f) + Esmooth(f) .........................(1)
Edata(f) =

Esmooth(f) =

Dp(fp),
p

qN(p)Vpq(fp, fq)

Where N(p) is the neighborhood pixels of p and Dp(fp) is the

data cost, which is a measure of the dissimilarity between
pixel p in the left image and pixel p & fp in the right image.
Vpq(fp,fq) is the smoothness cost that favors the piecewise
smooth objects. Taking these costs together, they arrive at the
optimal disparities by minimizing the total energy in formula
(1). The minimum cost to reach a particular point is then
calculated and used to calculate the minimum cost of reaching
the next point of any path that moves through the rst point.
One advantage of global and semi-global methods is that, to a
certain degree, they are insensitive to weakly-textured areas.
In DP the path will not stray far from the disparities at the
edges of the texture less area, because that would increase the
cost of the path.
D. Compute Disparity by Dynamic programming
Dynamic programming is recursion without repetition.
Dynamic programming example uses a multidimensional array
to store the result of recursive sub problems. To compute
disparity by dynamic programming between two pixel values
need to develop a recursive definition. Suppose we have two
pixel arrays which represent all pixels on contour positions
and to compute matching pixels values dynamic programming
approach will apply. Instead computing Disparity Space Image
for left and right image we will apply LCS dynamic
programming on contour arrays of left and right image and get
matching pixel value for respective contour array. This is done
by computing intensity difference of left and right contour
array. Here, the polynomial time require for computation of
DSI based dynamic programming is more complex than linear
time required LCS based dynamic programming. The new
approach of LCS based will ensures that computed value will
be optimal and more accurate than DSI.
In global optimization, the constraint on the disparity map d
are formulated into an equation 1 E(f) which is den minimize
over all image pixels. A typical objective function has the
following form:
E(d) = Edata(d) + Esmooth(d) .......................................(2)
The data term, Edata(d), measures how well the disparity
function d agrees with the input image pair. Using the
disparity space formulation,
Edata =

C(I,j,d(i,j)).......(3)

where, C is a (initial or aggregated) matching cost. The

smoothness term Edata(d) encodes the smoothness assumptions
made by the algorithm. Modeling is started on objective
function of the type in equation (2). The data term in equation
(3) can be written as,
Edata(d) = m(d(i,j)).....(4)

let i, j be the coordinates of the pixel in the left image and d(i,j)
be the value of disparity map d at pixel. Let m(d(i,j)) be the
matching penalty for assigning disparity d(i,j) to pixel, in our
frame work m(d(i,j)) can be sum of absolute differences (SAD)
to compare image contours,
m(d(i,j)) = SAD(i,j,d) = | Il(i+u, j+v) - Ir(i+u, j+v-d) |
.(5)
This means that for every contour pixel in the left image
search is performed along the same contour in right image that
best matches it, where d h. Note, if an image has j columns
and number of possible disparity values is h, then
straightforward dynamic programming take O(jh 2) time, but
running time can be reduced to O(jh). Let, s(dn) be the
smoothness penalty for assigning disparity dn and is disparity
penalty.
Thus the energy function required to be optimized is given by
equation (2) as follows,
E(d(i,j)) = m(d(i,j)) + s(dn)(6)
Then the minimum value of the energy in equation (6) can be
written as,
E(d(i,j)) = min(m(d(i,j)) + s(dn))..(7)
The optimal disparity assignment for d(i,j) in terms of LCS
dynamic programming can be written as follows, Let x = {x 1,
x2, x3xm} and y = {y1, y2, y3yn} be left and right
contour sequences and let z = {z1, z2, z3zk} be any LCS of
x and y.

If xm = yn, then zk = xm = yn implies that zk-1 is an LCS

of xm-1 and yn-1.

If xm yn, then zk xm implies that z is an LCS of x m-1

and y.

If xm yn, then zk yn implies that z is an LCS of x

and yn-1.

The output of z is used for disparity map computation further.

The proposed algorithm is simple to describe & implement but
also much more efficient. The obtained result of LCS on
contour is better than DSI results, which is shown in results
and discussions of this paper.
IV.

RESULT AND DISCUSSIONS

The algorithms have been tested on the benchmark

Middlebury Database; there are many stereo pairs which are
used in our experiments and results. Ground truth disparity
images present in Middlebury database are used for comparing
results. The error is computed as the percentage of pixels is far
from the true disparity by more than one. These statistics are
collected for non occluded contour pixels. Every stereo pair is
given as input to algorithm which will first extract the contour
of image, this is because algorithm mainly introduced for pick
694

IJRITCC | February 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 2

ISSN: 2321-8169
691 - 696

_______________________________________________________________________________________________
and place operations where size, shape and distance of object
is needed. Further stereo matching algorithm is applied on
actual contour extracted image.
Result is computed for different image pair from Middlebury
database to find depth information in which object near is
appear more brighter whereas object which is far is appear
darker.
The Figure below is the computed results for Midd 2 stereo
images pair from Middlebury database for DSI and LCS
mechanism on contour pixels. Detail evaluation is done in this
section.

comparing to object which is near to camera is appear more

brighter, refer highlighted area by red oval shape. Here,
accuracy is achieved by LCS method is more comparing to
DSI method of disparity computation.

Figure 7 Midd 2 disparity map by LCS.

(a)

(b)

The quality will be achieved from computing accuracy of

obtained results; this can be done by computing bad pixels
percentage and accuracy percentage. For each image pair, we
report percentage for all contour pixels. Only non occluded
pixels are considered in method using LCS. For bad pixels
percentage minimum is considered as best value. If minimum
bad pixels percentage is best, alternately maximum percentage
of accuracy become best value for given stereo pair.
TABLE I.

(c)
Figure 5 Midd 2 stereo pair and disparity map.

Figure 5(a) is left view image and Figure 5(b) is right view
image. Both images are given as input.. In left view object
which is near is more shifted towards right and in right view
that object shifted to left. On which canny edge detection and
actual contour extraction algorithm is applied to generate
disparity map. Figure 5(c) is ground truth disparity map
obtained by dense disparity map given in Middleburry
database for evaluation of accuracy. Cap is near so it more
brighter comparing to background in disparity map image
Figure 5(c).

Stereo
pair.

EVALUATION FOR 2 PAIRS D IFFERENT IMAGES.

RMS value

Bad Pixels
Percentage %

Accuracy
Percentage

DSI

LCS

DSI

LCS

DSI

LCS

Wood 1

7.65

3.22

6.56

6.85

93.44

93.15

Midd 1

1.04

1.06

9.87

1.008

90.13

98.99

Midd 2

1.06

1.07

9.92

1.002

90.08

98.99

Baby 1

5.46

5.37

4.81

4.51

95.19

95.49

Baby 2

4.72

4.61

3.89

3.83

96.11

96.17

Baby 3

8.47

8.57

7.46

7.80

92.54

92.2

Bowling
1

9.41

9.32

8.05

8.04

91.95

91.96

Flower

1.03

9.75

8.84

8.14

91.16

91.86

Wood 2

8.17

7.87

6.91

6.73

93.09

93.27

Figure 6 Midd 2 disparity map by DSI.

Figure 6 gives disparity map computed using disparity space

image (DSI) mechanism, in which we can see differentiating
between both images is difficult as we cannot easily indentify
in the image that which object is near or far, improved
mechanism with longest common subsequent (LCS) we can
see the object differs than one another.
Figure 7 gives disparity map by LCS where object behind cap
is far then computed disparity for that object is darker as

The above table shows different values computed for 2 pair

images from Middlebury database. We can see from above
table bad pixels percentage is decreases for LCS comparing to
DSI. RMS values for both DSI and LCS methods are also
calculated. Table above shows detail evaluation for all
different images from Middlebury database. The accuracy
percentage achieved by LCS method is better comparing to
DSI mechanism on contour.
695

IJRITCC | February 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________

International Journal on Recent and Innovation Trends in Computing and Communication

Volume: 3 Issue: 2

ISSN: 2321-8169
691 - 696

_______________________________________________________________________________________________
V.

CONCLUSION

Proposed method is a new stereo matching algorithm based

on dynamic programming approach on contours. Instead
applying dynamic programming on individual scan lines this
method applies it directly on contour array to find global
optimization and to get optimal disparity. Traditional scan
lines based method require polynomial time for its execution
as well as disparity computed using search space based
technique on contour such as DSI on contour also requires
polynomial time for its execution; this proposed new
algorithm uses LCS dynamic program on contour array which
will executes each subsequence in linear time for length of
other contour sequence. Sum of points that needed to be
computed will be considerably reduced so this algorithm
computes stereo correspondence with high speed and more
accuracy than other.
Furthermore, this proposed method acts as a guideline for
dense stereo matching. Use of LCS dynamic programming for
stereo matching on contour is done as reference work for
dense matching.

[8] Stephen S Intille, and Aaron F Bobick, Disparity-space

images and large occlusion stereo, 3rd European Conf on
Computer Vision (ECCV94), 1994, pp.179-86.
[9] Elisabetta Binaghi, ignazio Gallo, Giuseppe Marino, Mario
Raspanti, Neural Adaptive stereo matching, Pattern
recognition Letters 25, pp. 1743-1758, 2004.
[10] L. Kotoulas, A. Gasteratos, G. Ch. Sirakoulis, C.
Georgaulas, and I. Andreadis, Enhancement of Fast
Acquired Disparity Maps using a 1-D Cellular Automation
Filter, Proceedings of the fifth LASTED International
Conference on Visualization, Imaging and Image
Processing, Benidorm, Spain, September 7-9, 2005.
[11] Fua, P., A Parallel Stereo Algorithm that Produces Dense
Depth Maps and Preserves Image Features, Machine
Vision and Applications, 1993, 6(1), 35-49.
[12] Canny J, A Computational Approach to Edge Detection,
IEEE Trans on PAMI, 1986, pp 679-698.

ACKNOWLEDGMENT
I thank to our college Shah & Anchor kutchhi Engineering
college, for providing its resources and other facilities;
helping us to come up with the topic in excellent manner. I
sincerely acknowledge my internal guide Mrs. Bhakti
Sonawane for her invaluable guidance in giving form and
substances to this paper. Her guidance and invaluable
suggestions has been major motivating factor. It is her
enthusiasm and sincere efforts that have given a sense of
direction and purpose to this.
I also thank the principal sir Prof. Vinit Kotak and the
Head of Department, Prof. Uday Bhave sir, for their able
guidance and inputs in implementing the presented topic. And
lastly our non-teaching staff for making our premises one
where working can be a pleasure.
REFERENCES
[1] http://vision.middlebury.edu/stereo/2010
[2] Scharstein D, and Szeliski R, A taxonomy and evaluation
of dense two- frame stereo correspondence algorithms,
International Journal of Computer Vision, 2002, 47(1-3):742.
[3] Fua, P., A Parallel Stereo Algorithm that Produces Dense
Depth Maps and Preserves Image Features, Machine
Vision and Applications, 1993, 6(1), 35-49.
[4] Hannah, M.J, A System for Digital Stereo Image
Matching, Phot. Eng. and Rem. Sens, 2001, 55(12), 17651770.
[5] Ohta, Y., and Kanade, T., Stereo by Intra- and InterScanline Search Using Dynamic Programming, IEEE
transactions on Pattern Analysis and Machine Intelligence,
1985, Vol. 7 No. 2, pp. 139-54.
[6] Y. Boykov, O. Veksler, and R. Zabih. Fast approximate energy minimization via graph cuts. In International Conference on Computer Vision, pages 377384, 1999.
[7] L.Junqin, D. Yan., X. Hu., Q. Xing., B. Yang., Dynamic
Programming algorithm for Stereo Correspondence of
Cntour, IEEE 5th International conference on Image and
Signal Processing, 2012.

696
IJRITCC | February 2015, Available @ http://www.ijritcc.org

_______________________________________________________________________________________