Research On Bilocular Stereo Matching Algorithms For Blind Guider System

Research groups use their own hardware and software for various demands, which lead to the performance vary widely. Therefore, considering the performance of the practical application, modeling the stereo vision calculation methods, paying focus on the system design, analyzing the stereo vision theory will be helpful to promote the application of stereo vision in practical applications.Due to the uncertainty of match result, there is no uniform approach for the stereo matching problem. Bi-eButton as a wearable device will be used as a blind guider. The research work on the binocular system of Bi-eButton is the background of the thesis. The key research of the thesis is area-based stereo matching method. It emphasized more on the property of real-time than the precision on local stereo matching method. Contrary with the local method, the global method emphasized more on the property of precision than the real-time. When it processed the real-world scene image, the property of robust is more important. The main contributions of the thesis are as follows:1. The main models in binocular stereo matching are camera calibration and epipolar rectification, stereo matching, depth acquirement, whose theories are analyzed. According to the basic theory of the binocular stereo vision system, we built a simple experimental system with the help of OpenCV and OpenGL. The experiment results are also presented.2. After analyzing the advantages and disadvantages of several classic matching algorithms, the thesis find that the research on the local stereo matching algorithm is a research on local filter essentially. This filter can smooth the disparity space image at the same time save discontinuity. In fact, the research on the filter is an old field. And there are also many excellent filters in the existing applications (such as, image denoising, image dehaze). Whether these excellent filters can be used into the matching problem? A new method for constructing an accurate disparity space image and performing an efficient cost aggregation in stereo matching based on local affine model is proposed. This model has been used successfully in image mapping and image denoises. The results show that the local stereo matching algorithm ensures the requirements of real-time processing at the same time the accuracy is improved.3. After giving detail studies in dynamic programming and Markov random field, analysis the importance of edges, the thesis present a Bayesian stereo matching method based on edge constraints. The Bayesian approach is utilized for dense stereo matching problem formulated as a maximum a posteriori Markov Random Field (MAP-MRF) problem. In order to improve stereo matching performance, edges are incorporated into the Bayesian model as a soft constraint. Accelerated belief propagation is applied to obtain the maximum a posteriori estimates in the Markov random field. The proposed algorithm is evaluated using the Middlebury stereo benchmark. Comparing with some state-of-the-art stereo matching methods, the experimental results demonstrate that the proposed method provides a disparity map with subpixel precision.4. Due to the mismatch always appears on uniform regions, the thesis adds a segment constrain when deals with these regions. Assuming the difference of disparity value in the same segment region is small. First, the local affine aggregation and color images over-segmentation are used to get initial disparity map and segmentation result, respectively. Then the disparity plane fitting are used to solve the mismatch problem in uniform region.5. The "staircase" phenomenon appears on slant regions. To solve these problems, we present a local method based on sub-pixel support window, which uses a rotatable plane to find the support window. The experimental results show the algorithm pro forms very well in the slant region. Finally, a more reasonable and complex method based on second-order priori assumption global stereo matching methods is introduced to solve these phenomenon.6. Blind guider system based on Bi-eButton will be used both indoor and outdoor. Due to the lack of Bi-eButton picture, we use outdoor pictures in the KITTI lib to do the experiments. And an effectiveness cross-based Census semi-global stereo matching algorithm is presented. Considering the robustness is the more important part when processing real world images, the thesis compared several similarity measurements, the Census method is chosen. Then, use the cross aggregation method to improve the robustness. Finally, the thesis uses the semi-global method to do the energy minimization.