Abstract
A feature-based stereo model designed to perform matching multi-resolution features is described. The multi-resolution features are zero-crossings (ZC's) of images convolved with different sized Laplacian-Gaussian operators. ZC's corresponding to small contrasts are removed according to gradient values of images convolved with Gaussian operators.
Candidate disparity intervals are determined using a disparity histogram of the total ZC's over the entire image. The image is then divided into small areas and a local disparity histogram is computed for the candidate intervals. Local disparity histograms in all resolution channels are searched for the most promising disparity for each area. If the disparity is found successfully, disparities for all ZC's in the area are determined by searching only the neighbor of the promising disparity. Once a disparity for a ZC is determined, the matching pair of ZC's are removed from a set of ZC's. This process is iterated until no more disparities are determined.
Experiments with sample scenes including objects with various shapes and brightnesses at different positions reveal that the model has advantages in efficiency and performance. Some discussions are also presented on the correspondence between the stereo model and the human binocular system.
