Research On Stereo Matching Method Based On Graph Theory

Stereo matching technology is the key step in 3D-reconstruction system and non-contact measurement. And it is a hot research area in computer vision in recent years.The traditional stereo matching method based on global optimization has the complex algorithm framework and CPU cost. Having excellent mathematical properties of graph theory, the graph theory is applied to the stereo matching problem in computer vision, and can obtain good results and reduce the computation time.In this thesis, we first propose a stereo matching method based on the interactive image segmentation, which is based on the problem of measuring the disparity of some objects in most applications scenarios. In order to obtain dense disparity map, most of the stereo matching algorithm based on global optimization require iterative computation, which results to calculative cost up. Generally, to effectively reduce calculative cost, image segmentation and other means are introduced. In order to overcome the shortcomings of traditional stereo matching methods based on image segmentation which did not take full advantage of the segmentation information, this thesis introduces user interaction and set the seed point with graph cut algorithm to complete sense of the segmentation of region of interest extraction. A network graph is constructed by partition template. Then stereo matching is conducted. Due to the construction of network only for the segmentation region, comparing with the existing graph cut stereo methods which construct network graph on the whole image, our method has the advantage of matching accuracy and small computation.Secondly, although the graph cut algorithm is accurate, it has many parameters to be set and calculative cost. Furthermore, the algorithm has continuity in time which is not fit to parallel execute. So, it can not meet the requirements of real-time calculation. To solve this problem, a general-purpose parallel algorithm stereo matching method based on the minimum spanning tree is proposed. Based on the analysis we can get hot code, by using the OpenMP parallel computing, SIMD vector calculation is used to iteratively optimize the objective function. Finally, a general parallel stereo matching method which can compute in real-time is obtained in high accuracy.