Stereo Matching Based Estimation of Depth Map from Stereo Image Pair

Abstract

Stereo vision is a technique of depth perception, in which the information about the depth is inferred from the two (or more) images captured from different perspectives of a scene. These images is known as stereo image pair. Practical applications where stereo vision technology plays a role may include autonomous vehicle guidance, aerial photogrammetry, robotics vision, object tracking and industrial automation. Many automated vision systems could benefit substantially from depth maps. A depth map is a grayscale image that contains depth information for each pixel in an image. Traditionally depth maps were extracted using stereo camera approach. Depth estimation from stereo involves finding disparities along the same scanline, also called as epipolar lines. Such a search process typically requires a prior adjustment of the images known as rectification step to ensure that epipolar lines are well aligned. Still, the approaches for estimation of depth map suffer from either limited reliability and robustness when tested on stereo image pair or large time of computation. A novel approach for depth estimation that integrates the image filtering into a stereo matching framework is introduced. Experiment helps verifying the sustenance of high quality in depth maps, while reducing the average percent of bad pixels to 3.58%.