Optimal Algorithms for Stereo Correspondence Estimation

Abstract

Stereo correspondence has attained a position of overwhelming dominance in Computer Vision for long days for determining three-dimensional depth information of objects using a pair of left and right images from a stereo camera system. In this thesis we propose four novel ideas for improving the efficiency and accuracy of stereo correspondence estimation in stereo vision. First idea presents a “Real Time Approximation (RTA)” algorithm for computing the disparity of the stereo image sequences. The algorithm has been organized to make it dedicated for real timeapplications. To do this, the original image

is scaled down and obtained highest speed to compute the stereo correspondences. The second idea is a searching algorithm titled “Two Dimensional Real Time Spiral Search Algorithm (2DRTSSA)” to compute the stereo correspondence two dimensionally. The 2DRTSSA thus increases the speed and accuracy over the existing state-of-the-art methods of one dimensional and left-right searching strategy. The third idea is a new and significant searching method, is explored by the name “Self-Adaptive Algorithm (SAA)” for computing stereo correspondence or disparity of stereo image. According to the SAA method, stereo matching search range can be selected dynamically until finding the best match. The searching speed is almost doubled by reducing the search range half of its original, by dividing the searching range into two regions. First one is –d to 0 and second one is 0 to +d max max . To determine the correspondence of a pixel of the reference image (left image), the window costs of the right image are computed either for –d to 0 region or for 0 to +d region depending on the result of previous matching. The speed and accuracy are further improved by introducing the fourth idea entitled “Self-Guided Stereo Correspondence (SGSC) Estimation” algorithm. The SGSC algorithm is directed by photometric properties of the candidate-pixels. Searching performance is slightly improved by utilizing this photometric property of the candidate-pixels as well as by implanting the pioneer threshold technique. These two key techniques reduced the computational costs with further improvement of accuracy. The achievements of the SGSC method are testified on Middlebury standard stereo datasets of 2001, 2003, 2006 and Middlebury latest Optical Flow Datasets. Moreover, the newly invented algorithms RTA, 2DRTSSA, SAA and SGSC have been justified on real images which are acquisitioned in our laboratory in complex environment. The overall performances of max all algorithms are satisfactory in case of real stereo images. Finally, the proposed methods are compared with present state-of-the-art methods and our 2DRTSSA, SAA and SGSC outperforms the latest methods in terms of speed, visualization of hidden ground truth, 3D reconstruction and accuracy.