Algorithmic Research On Binocular Stereo Matching

Binocular stereo vision is one of the most classic research areas in computer vision. As a 3D modeling technique with low hardware dependency and strong environment adaptability, it maintains a great potential application in many fields. Meanwhile, stereo matching is the most challenging topic in binocular stereo vision. The development in this domain will largely promote the development of not only binocular stereo vision, but also the relative communities in computer vision.This thesis's work can be split into two parts. Firstly, to overcome the low efficiency of generalized belief propagation, three acceleration methods such as min-sum based caching method, direction set based complex reduction method, and hierarchical state space reduction method have been proposed. Experiments show these three speedup techniques can enormously improve the efficiency of the algorithm.Secondly, to improve the matching accuracy of single usage of current existing algorithms, a matching framework combining the strengths of both global based algorithms and local based algorithms has been proposed. In the basis of surface piecewise consistency assumption, the two categories of algorithms are successfully integrated, and more accurate matching results can be achieved through experiments.Overall, the research of this thesis is two-fold. One is the improvement of a current existing algorithm, and the other one is the integrity of two kinds of algorithms. Later research will focus on the algorithmic adjustment which should give a more detailed description of outer scenes.