Stereo Matching with Implicit Detection of Occlusions

Abstract

In this paper we introduce a new stereo matching algorithm, in which the matching of occluded areas is suppressed by a self-organizing process. In the first step the images are filtered by a set of oriented Gabor filters. A complex-valued correlation-based similarity measurement, which is applied to the responses of the Gabor filters, is used in the second step to initialize a self-organizing process. In this self-organizing network, which is described by coupled, non-linear evolution equations, the continuity and the uniqueness constraints are established. Occlusions are detected implicitly without a computationally intensive bidirectional matching strategy. Due to the special similarity measurement, dense disparity maps can be calculated with subpixel accuracy. Unlike phase-difference methods the disparity range is not limited to the modulation wavelength of the quadrature-filter. Therefore, there is no need for a hierachical coarse-to-fine control strategy in our approach.