Stereo Matching Using Reliable Points And Its Applications

Stereo matching is the process of taking two or more images and estimating a 3D model of the scene by finding matching pixels in the images and converting their 2D positions into 3D depths.It is a critical process of computer vision.The current global methods can achieve high quality result at the cost of mass computing,while the local methods can not get satisfied results because of its weak constraints. This thesis is aimed at soloving these problems.Firstly,this thesis looks into the mass computation problems in traditional graphcut based global methods.The number of pixels and range of disparity values together leads to a big graph model,which leads to unnecessary expensive computing cost.In this thesis,we put forward a two-step refined vision of graph cut.Firstly,the map are divided into overlapping blocks in which region-based graph cut is applied,combined with estimations of the disparity values in a unique block,we can get an optimized solution when we use the graph-cut model.Experiments results proves the e?cient of the method.Secondly, we focus on the applications of reliable points in those local methods.In the new methods,reliable points and initial disparity maps are used as strong constraints to update the matching cost and the cost aggregation, final disparity maps prove the e?ciency of the optimization work.As to the current popular tree-model based methods,the construction process of its edges leads to failures in many areas.This thesis introduce an optimized version.Except reliable points and initial disparity maps,the segmentation process and local gratitude also strengthen the constraints.During the whole process, the matching cost and weights of edges are updated and thus we can achieve a new minimal spanning tree.Experiments show that each optimization step can upgrade the results to a higher level.Finally, this thesis give a introduction of my specialized practice in the stereoscopic microscope system.During the course, a method aimed at achieving real-time disparity maps is proposed and the disparity information is used in both segmentation and the three-dimensional tracking of the surgical knife.