ROBUST CORRESPONDENCE TECHNIQUES FOR STEREO VISION By

iii ACKNOWLEDGEMENTS I would like to thank my thesis director, Dr. Farag, for his support, patience and advice, all of which helped make this thesis possible. I am especially indebted to Moumen Ahmed, who provided invaluable guidance during my research for this thesis. I also wish to thank the other members of the CVIP Lab for their friendship. I would like to thank Dr. Brehob and Dr. Naber for serving on my thesis committee. Finally, I would like to thank my family for their support during my studies at the University of Louisville. iv ABSTRACT The purpose of this thesis is to develop a fast and accurate correspondence component for the CardEye active vision system. Correspondence is the first part of the stereo vision problem that must be solved and is also the most important, since the correspondence results affect the final 3-D reconstruction of a scene or object. Correlation has been commonly used in the past to produce dense disparity maps for correspondence. However, most classical linear correlation methods fail near depth discontinuities and in the presence of occlusions. Many robust methods have been proposed that claim to effectively deal with some or all of these issues. This thesis presents five requirements that should be met by a nonlinear robust correlation method. Several robust correspondence methods will be compared, and their utility on different data sets will be demonstrated. Based on these results, a hybrid solution to the correspondence problem is proposed that represents a compromise between the speed of classical correlation and the improved results obtained from a more robust correspondence method. The new hybrid correlation method can be used with the CardEye system. Since the CardEye system has a novel three-camera head, there exists more than just one method of performing the correspondence. Two of those methods are presented and evaluated in this thesis also. One method represents a slight modification of the correspondence method used with two images. The other one heavily exploits the v epipolar constraint, which can be used since the geometry of the cameras is known from camera calibration. Lastly, a new selective median filter is proposed to smooth the disparity maps produced by the different correspondence methods. The median filter is slightly modified to reduce the effect of removal of the ends of lines, one-pixel-wide lines and corners by the filtering.