Fast stereo matching for implementation in a 3-D vision sensor

An attempt is made to realize a 3-D vision sensor which can be applied to navigable robot and autonomous vehicles. These applications require fast, auto-adaptive algorithms which can be processed by parallel processors. The authors present a novel fast stereo matching algorithm ignoring epipolar geometry based on the prediction and verification of hypotheses. Experimental results obtained on several pairs of stereo images are given. Computational time is only 12 ms to match two lines. A complete stereo processing system including the proposed algorithm is currently being implemented on a specialized architecture, thanks to which the processing time will be reduced to 0.1 s.<>