A stereo disparity algorithm for 3D model construction

Stereo vision is an important passive method for extracting the 3D structure of a scene. It involves the analysis of at least two digital images with overlapping fields of view. We describe a hierarchical algorithm for the computation of a dense disparity field from a binocular view. Since our goal is the measurement of 3D scene structure with application to model-based image coding where a continuous surface description is required, we concentrate on area- rather than feature-based matching. The main features of the algorithm are: the synthesis of a variable disparity search range in a blockwise fashion; the active use of an occlusion detector; and the inclusion of a well-defined interpolation scheme that preserves discontinuities and avoids blurring near occlusion boundaries. These lead to a robust procedure for measuring continuous disparity fields