Stereo Matching Algorithms In3D Reconstruction

In recent years, more and more computer vision researchers pay attention to3D reconstruction based on images or videos, especially images. Stereo matching isone of the most important problems in the3D reconstruction. This thesis aims tofind better stereo matching algorithms. We focused on template based stereomatching algorithm using Graph cut (TGC), including the implementation andanalysis of the typical stereo matching algorithms and some improvements aboutTGC based on the basic framework of stereo matching algorithms.In the first part, four typical stereo matching algorithms are implemented andthe results are analyzed, then we improve them. The algorithms include WTA andreal-time which are local stereo matching algorithms, and two global ones, they areDynamic Programming and Simulated Annealing. Then we find BT algorithm is auseful technique in stereo matching algorithms, and the global ones take too moretimes, so efficient global stereo matching algorithms are needed, especially forpractical applications.Based on the idea of graph cut, we proposed a global stereo matchingalgorithm--template based stereo matching algorithm using Graph cut (TGC) whichneeds three steps to implement: the segmentation of reference image, disparityplanar estimation, templates assigning to disparities using graph cut. Then we givethe PBM and running time of several algorithms on the Middlebury test set, theproposed (TGC), graph cut, dynamic programming, simulated annealing,winner-take-all, and real-time. Experimental results show TGC can obtainsatisfactory stereo matching results.In the last part, we do four experiments to improve TGC based on the basicframework of stereo matching algorithms. The first improvement, BT algorithm isadded to global stereo algorithms, according to the above results, we confirmed it isalso helpful for the global ones. The second one, four algorithms based on graph cutare compared to find the two critical parameters that make the performance ofstereo matching algorithms different. We also find the best range of parameters withwhich the stereo matching algorithms can achieve better results that are competitivewith the state-of-the-art algorithms.