A New Stereo Matching
Algorithm for Binocular Vision
Tao Hu May Huang
International Journal of Hybrid Information Technology
Vol.3, No.1, January, 2010
Professor ： Tsai, Lian-Jou
Student：Tsai, Yu-Ming
PPT Production rate ：100%
Date ： 2012/02/29
South Taiwan University ESLab
Outline
• Introduction
• Gradient Vector
• Confidence Point
• Stereo Matching
• Stereo Matching Algorithm
• Experimental results
• Conclusion
South Taiwan University ESLab
1
Introduction (1/3)
• The research on binocular vision is a critical
issue in the field of computer vision.
• Binocular vision Can reconstruct a profile of the
object and define its location in 3D space.
• David Marr proposed a computational vision
theory in 80's and used it in binocular matching.
South Taiwan University ESLab
2
Introduction (2/3)
• Stereo matching for binocular vision :
Using two cameras to screen a scene.
Using two data in two-dimensional (2D) visions.
To implement a point in 3D vision.
• Three type’s stereo matching algorithms:
Local Matching
Feature Matching
Phase Matching
South Taiwan University ESLab
3
Introduction (3/3)
• New stereo matching algorithm:
(a) Images are taken by two cameras
(b) Set the left image as the original
(c) Compute the gradients
(d) Find out confident points based on the
gradients
(e) Compute the matching range
(f) Obtain the disparity range
(g) Find out the best matching points
(h) Generate a disparity map for the matching
South Taiwan University ESLab
4
Gradient Vector
• The difference of gradients between left and right images to
determine if the corresponding points are confidence points.
• Sobel operator
• The gradient of point （X,Y）
South Taiwan University ESLab
5
Confidence Point (1/3)
• Defining the difference of gradients in left and right
images as a threshold.
• Compare the calculated value with the threshold to
determine if the point is a confident point.
• Assuming a test line defined on the left image and
a point on the test line is (x, y).
• The disparity value tested is (dx, dy).
• Calculate the gradients of the points: GL(x, y) and
GR(x+dx, y+dy)
GL(x, y)
South Taiwan University ESLab
GR(x+dx, y+dy)
6
Confidence Point (2/3)
• Confident value of a point on the left image：
• Use Formula to determine a confident point：
• If the d_con >0, that the point (x, y) and the point
(x+dx, y+dy) are the pair of confident points.
• We can divide the test line on an image into pieces
ranges.
South Taiwan University ESLab
7
Confidence Point (3/3)
• Selecting the range which is closest to the camera.
• linear interpolation method can be used to confirm
a new disparity range.
• When the value of threshold is close to zero, we
call the state of the point as base case.
• In the base case, the disparity range is small, and
the accuracy of matching is high.
South Taiwan University ESLab
8
Stereo Matching (1/2)
• Fill the range with grey values using linear
interpolation method.
• Use the SSD (Sum of Squared Differences)
method to do stereo matching for this paragraph.
matching
South Taiwan University ESLab
9
Stereo Matching (2/2)
• Set a threshold d
• calculate the SSD(x, y) in different t in the disparity
range.
• If SSD(x, y)<d, set d equal SSD(x, y) and record t.
• When computed the whole disparity range, we can
obtain best disparity value T.
• The point (x+T, y) on right image which can
provide the best match with the point (x, y) on left
image.
South Taiwan University ESLab
10
Stereo Matching Algorithm
South Taiwan University ESLab
11
Experimental Results (1/4)
• Two Logitech QucikCamTM cameras were used for
capturing images.
• Two pairs of images taken by two cameras.
South Taiwan University ESLab
12
Experimental Results (2/4)
South Taiwan University ESLab
13
Experimental Results (3/4)
South Taiwan University ESLab
14
Experimental Results (4/4)
• Use different sizes of image pairs to test three
matching algorithms.
• SSD-based matching is much faster but the
matching results are not comparable to the new
algorithm.
South Taiwan University ESLab
15
Conclusion
• A gradient vector is used to determine confident
points and to estimate the disparity range.
• The new algorithm borrows the principle of SSD
matching method and reduces the matching range
to the limit with the range of confident points and the
range of disparity.
• The new algorithm is also shown the high
performance when finding proper matching and
disparity ranges.
South Taiwan University ESLab
16
Thanks for your listening!!
Any questions?
South Taiwan University ESLab