The Research On Stereo Matching Algorithm Based On Region-growth

Computer vision is a fast developing field which is full of challenges and a research direction combined with many subjects. Binocular and stereo matching is a very important step and also a very difficult problem involved in a lot of factors. Although dense matching algorithm performs very well among several matching algorithms, there are still some problems on quantity and precision of matching points. The thesis has done some researches on feature points detection based on Marr's theory of computer vision and has made some improvement on region-growth matching algorithm according to the problems which happened in dense matching algorithm.Through the researches on the theory of classical edge-detection operators, the thesis compared with their detection results and got some conclusions. The influence of bigger adjacent area was considered in Prewitt operator which detected clear and continuous edge. In addition, the positioning accuracy was high and the course of detection was not easy to be affected by noise.The theories and implemental methods of SUSAN operator, MIC operator and Harris operator were compared and Harris corner detection was finished by the tool of VC6.0 in the thesis. The detection results showed that the speed of Harris corner detection method was fast and this method can get even and accurate corners. So this algorithm was stable.At present, these common matching algorithms are very hard to get dense and high-precision information of image with high speed. The dense matching algorithm has shown its advantage in this aspect. The thesis mainly discussed the problems of region-growth algorithm and made some improvements. It adopted Harris corner detection to finish the function of achieving seeds automatically and it provided manual way to guarantee the accuracy and enhance the speed. The matching relation was diffused to the surround area of seeds according to the continuity of pixels and seeds until the matching points in the whole image were found. At last parallax of matching points was calculated according to the theory of parallax and the parallax image was drawn by the tool Visual C++6.0.Theoretical analysis and experimental results showed that the method of picking out the seeds automatically and manually assured the accuracy of seed. The course of region-growth achieved plentiful matching points according to the continuity of pixels as well as the clear and dense parallax graph was gotten.