A new parallel feature-based stereo-matching algorithm with figural continuity preservation, based on hybrid symbiotic genetic algorithms

The calculation of shape from stereo requires matching of points from the left image with points from the right image. Many classes of stereo-matching algorithms have been proposed [1}3] with di!erences in the type of tokens they match and the optimisation method and the constraints they use. They are mainly categorised as area-based or feature-based. Feature-based methods can be either low level, which try to match individual edgels or high level, which try to match edge structures, such as lines or curves. While high-level methods enable comparisons of more complex features and incorporate naturally the "gural continuity (FC) [1,4] constraint, they can su!er from problems of imprecise disparity estimation as well as fragmentation and fragility [3]. In this work we present a novel low-level edge-based algorithm. Its main advantages are explicit optimisation of the FC across bands of epipolars, elimination of the ordering constraint (OC) and highly parallel distributed execution. Each epipolar is searched in parallel for a high con"dence intra-row matching decision (at the local level). Simultaneously, temporary matching information is propagated to the adjoining epipolars for minimisation of FC violations (at the global level).