Stereo Correspondence Algorithms For Stereo Vision

As an important part of stereo vision, stereo correspondence, which task is to calculate and identify the disparity map of stereo images, is the core technology to achieve three-dimensional information from two-dimensional image. After systematically introducing to the related theories, this thesis focuses on the research of stereo algorithms and proposes an improved method.Firstly, this thesis optimizes the computational efficiency of shape-adaptive support cost aggregation. For each anchor pixel an upright cross local support skeleton is adaptively constructed, with four varying arm lengths decided on color similarity and connectivity constraints. Then, we dynamically construct a shape-adaptive full support region. Our method can calculate the arm length quickly by changing the definition of the arm. In addition, our method uses the fast aggregation technology to speed up the cost aggregation step. Our method can optimize the computational efficiency without losing the quality.Secondly, this thesis proposes an improved dynamic programming method for calculating the final disparity map. First, the boundary information is detected by edge detection technology, and boundary pixels can select all disparity values instead of selecting similiar disparity to neighbor pixel. Smooth assumption is used for non-boundary pixels only. Our method can improve the disparity map without increasing the computational complexity.Finally, this thesis is almost a real-time method, which uses the improved shape-adaptive support cost aggregation and improved dynamic programming to achieve the high-quality disparity map.