Research On Dense Matching Algorithm In Binocular Stereo Vision

The background of this project is to find a dense matching algorithm which is fast and accurate for cleaning robots automatic navigation system, so we research dense matching components and technical difficulties in stereo matching, through a combination of theoretical and practical ways, in order to find an accurate and real-time algorithm. Paper is divided into five chapters and each chapter is as follows:The first chapter introduces the basic theory of binocular vision - Marr visual framework and then reviews robot vision research home and abroad. After that we illustrate the idea of research projects and main difficulties during our work.The second chapter describes the linear camera projection model, and then describes the principle of Zhang Zhengyou calibration. We conduct experiments on camera calibration based on OpenCV and then rectify input images according to obtained parameters. Experimental results show that the accuracy of parameters obtained by calibration can meet real requirements.The third chapter introduces the basic theory of dense matching such as matching principles and matching constraints, and then analyzes the difficulties during dense matching such as matching errors and occlusions. After that we focus on stereo matching method based on Census non-parametric transformation and point out the limitations of traditional Census matching method. We use bilinear interpolation on input images in order to improve accuracy. Experimental results show that improved Census stereo matching algorithm has a better accuracy in non-occlusion area and depth of the discontinuity region than old one.In chapter IV, we present an improved global matching algorithm based on segmentation. First we divide image into segments with homogeneous color by Mean Shift and then calculate the coarse disparity map. A plane fitting method technique is applied to obtain the equation of disparity plane. After that we take Graph Cuts technique to accomplish global optimization. Final results show that not only it has a better accuracy than the Census stereo matching method, but also original and similar algorithms use Graph Cuts.In chapter V, we summarize our work and discuss the future research.