New area-edge-based stereo algorithm and its experimental demonstration using natural objects

We present a new stereo algorithm which includes some features of area-based and edge-based approaches, eliminating their main drawbacks. An addition to usual pre-and post-processing steps (edge points and contours extracting, segmentation, least squares plane fitting and interpolation), the algorithm includes a nem: solution of correspondence problem. For each left edge point within the some support area (typically 20x20 pixels) we search at first the possible corresponding points (candidates) in the right image (Located at the same scan line within some disparity range and having the same local characte-ristics) selecting than only the single candidate having the maximum support sum. To find a support sum for each disparity candidate, vee use a disparity gistogram taking the magnitude of the nearest gistogram peak as a support sum to be found. To reject false matches, each support area is classified depending on fraction F of edge points connected with the major gistogram peak, If F is more than some threshfold, say 80%, all points not connected with the major peak are simply rejected marking the rest points as a true. In opposite case, more complicated procedure is used taking into account the edge contours information by testing it for some global consistency constrains (uniqueness, continuity and ordering). The algorithm have been successfully tested using specially developed computer program operating with different natural objects including some hard features (noise and edge detecting defects repetitions, sharp depth variation and discontinuities, occlusions, etc.).