A new stereo matching paradigm is introduced as an integrated process of highly discriminating steps, adopting congruously all the fundamental steps of the stereo vision problem. The central objective is the extraction of a disparity map from which the depth map will be derived. A unique representation of the two dimensional (2-D) stereo images into linear, orthogonal, and spatially-varying attributes serve as the mathematical foundation from which the proposed stereo matching method has evolved. The devised attributes contribute equally to the decision making process and provide information on the characterization of a potential match and its validation through a consistency check.
A fundamental contribution of this thesis is in creating the possibility for the design of a dimensionally-augmented vision system (2½ -D representation) based on an effective stereo paradigm with realistic computational requirements. In this design configuration, the geometrical mappings between the 3-D real-world measurements with the measurements obtained using the proposed 2½ -D-D representation are established. Computer results for the intended objective of creating highly accurate disparity maps for various scenes with varying complexities clearly demonstrate the soundness of the proposed method both in terms of its matching effectiveness and its realistic computational power requirements. Future objectives point to the development of enhanced algorithms for scene interpretation and understanding based on this augmented representation.
