Fast stereo vision algorithm for robotic applications

Abstract

Autonomous navigation applications demand sensors with a low sample time to be able to increase speed. We have developed a stereo vision algorithm, capable to deliver dense disparity maps for single, high-resolution scanlines at high speed (40 ms/line), even for wide disparity ranges. We have tested the algorithm with synthetic and real images. Our algorithm is based on a dynamic programming schema with a cost function based on a weighted sum of squared intensify errors. Weight factors are based on gradient values. The algorithm includes explicitly detection of occlusion. Occlusion cost changes dynamically depending on gradient values of matched points.