Advanced Stereo Matching Algorithms

Nowadays, the improved computational power and the deeper understanding to the mechanism behind the human vision system promote the computer vision technology in all fields, which brings a greater application prospect for the technology in many fields such as autonomous driving, computer assisted monitoring and the shape/volume estimation. 3D/depth information is indispensable in computer vision applications. It not only depicts the shape of objects but also provides spatial relationship of objects in the scene. This information is widely used in a wide range of vision tasks.Stereo vision technology models the procedure of the human brain when taking information from two eyes and generating 3D shape about objects in the sense. Compared with some other methods such as Kinet and Lidar, stereo vision technology provides 3D information by only using conventional cameras without the need of extra equipments. Besides, this technique can be easily implemented in both indoor and outdoor applications. Therefore, it is a more flexible and less constrained manner with a lower cost. Despite of these advantages, stereo vision technology highly depends on the robustness and accuracy of algorithms being used, especially the matching algorithms. Unfortunately, current matching algorithms are vulnerable to the lack of texture and the illumination variation, which gives rise to the mismatching problem in matching results. The success of stereo vision technique requires algorithmic breakthroughs.This thesis mainly focuses on the matching algorithms in the stereo vision technology, including both sparse matching and dense matching. A cooperative strategy is proposed to address the former problem. The uniform distribution is beneficial to the camera geometry calculation which is essential in stereo vision and the cooperative strategy is specifically designed through encouraging uniform distribution. Three dense matching algorithms are proposed separately to address the second problem and a comparison between them over advantages and disadvantages is given. In addition, a filtering method is also proposed for addressing the post processing problems of dense matching, such as occlusion filling and depth map upsampling tasks.In summary, the contributions of this thesis are: (1) a new sparse matching algorithm for finding sparse correspondences in stereo image pairs (2) three new dense matching algorithms for calculating disparity maps, and (3) a new filtering algorithm