A Segment and Fusion-Based Stereo Approach

Most algorithms in stereo vision work on rectified images and therefore find the point correspondences row by row. So especially for standard block-matching algorithms periodic patterns are a problem in determining corresponding features reliably.This contribution describes a segment-based approach that allows the detection and removal of single outliers in an arbitrary dense disparity map and so improves the data quality. The first step is a matching of vertical edge segments in the images in a coarse to fine strategy.Then the segment information is taken into account. Even more, when using segments there is only need to calculate feature correspondences for a fraction of the image rows, which considerably reduces computation time. By fusing this information with the disparity map of the standard block matching algorithm a significant improvement of the resulting disparity map in the presence of periodic patterns can be reached.