Research On The Key Auto-matching Technique In Stereo Vision

Matches on two images that belong to the same scene are the foundation problem in computer vision, and also the important step in the process of the 3D reconstruction and stereo vision. Stereo vision is based on 3D reconstruction. The matching efficiency and performance deficiencies in 3D reconstruction are obvious, so research on this kind of matching method is a focus in the field of image, especially the research about feature based on matching method is a great practical significance.One difficult point of the feature based on matching method is feature extraction technique. In 3D vision reconstruction, whether the feature extracted from static scene or dynamic scene, how to find a accurate feature extraction algorithm to extract specific image feature is urgent need. SUSAN operator and Harris operator are widely used in the research of extracting a static scene feature. A new corner detector was also proposed based on the basic principles of the Forstnor operator. The advantages and disadvantages of these algorithms and some problems that needed to pay attention to in application are raised by experiment test and analysis. Besides, more bad matches appear in image due to noise, block and negative threshold selection. Epipolar constraint method is used to solve the problem. In the research of extracting a dynamic scene feature, in view of the matching efficiency and performance deficiencies of the SIFT key points descriptors, before construct octaves of scale space that repeatedly convolved with Gaussians, cubic interpolation is used to reduce original image efficiency, and then the octaves of scale space are decreased, which makes the final structure of the descriptors is distinctive and robust. The advantage of such improved method is that it gains much more key points information and takes the matching efficiency controlled descriptor account at the same time, which results a higher matching performance and a better matching efficiency. At first, images are deal with cubic interpolation and extracted by SIFT, and then matched by the nearest neighbor method, k-d tree and limit key points method to eliminate these bad matches. Experiment results shows the method is available.The above method provides a new method to effectively implement the matching technique in stereo vision. The advantages of them are on the use widely in many practical areas and obviously improved matching efficiency and performance, which is very important for many applications.