Research On Fore-Ground Fattening Problem In Local Stereo Matching

Stereo matching is one of key technologies in machine vision, and is a method to extract3D object information from2D images. Recently, stereo matching has made great process, and lots of algorithms will be proposed each year. Besides, many excellent algorithms and models have been used widely in medical image processing, robot vision, military, space explosion, etc.Stereo matching makes use of two rectified images to determine disparity of every pixel. In general, matching algorithms can be classified into two categories: local and global. Local and global algorithms are both based on the assumption that image has local smoothness. Local methods implicitly model the assumption into color or intensity values, and global methods explicitly model this assumption by performing a disparity optimization on the entire stereo pair. Global methods can get more accurate disparities but have higher complexity. Local methods are less accurate but faster and easier complemented.A new framework based on mean shift, geodesic-support weight and the proposed disparity calibration is developed in this paper. Mean shift approach is applied to the image clustering procedure to reduce the number of pixels to be matched and attain candidate pixels. Then geodesic-support matching algorithm is used on the candidate pixels for initial matching. Finally, a statistics based disparity calibration is proposed to take the place of weighted average based disparity calibration to calibrate the disparity of candidate pixels with the disparity of most pixels, so that foreground fattening can be eliminated.We present a range of comparative experiments on the platform of Middlebury benchmark demonstrating that our new method improves the matching accuracy and enhances the validity-efficiency in occlusion regions.