Research On Image Selection For Multi-view Stereo Matching

Multi-view matching can integrate information of many overlap images, and to some extent, improve the reliability of matching. However, the existing multi-view matching method simply reduce all overlap images and ignore the difference of image quality, which results into bad matching performance in many condition. This paper summarizes the traditional multi-view matching constraints and common strategies. The effection to image matching from obstacles and baseline ratio is analysied. And the relationship between baseline ratio and matching precision, as well as relationship among the terrain slope, baseline ratio and image distortion, is derived under a certain condition. After that analysis, the dissertation proposes an image selection with the reference information of terrain and images. With the help of the method, multi-view mixed matching with real-time optimum is proposed. Both of the methods perform very well. The main content of this paper includes:(1) After the thorough analysis of the multi-view stereo matching model and constraint strategy, a multi-view feature matching method with the pyramid strategy is completed. The method provides an initial object information for image selection and real-time optimum.(2) The effection of obstacle and baseline ratio on the existing multi-view matching is discussed in detail. The discipline of impaction from baseline ratio to image distoration is reducted under the view of height accuracy and terrain slope. When matching precision is integer pixels, the maximum value of base to height ratio provides a theoretical basis for image screening. Besides, when the terrain exist relief, relative deformation in images of objects is connect ed with base to height ratio and feature vector location of the space station.(3) An image selection with the reference information of terrain and images. is proposed. The method select optimum images from overlapping datas in each of feature point with the accout of space geometic and image information. Then k-d tree strategy is applied to the work of interpolation for selection information of each non-feature point from feature points. It is demonstrated by experiments that this method can remove the image of bad matching, and improves the matching success rate while reducing the calculation.(4) When cross correlation coefficient adopt as a matching measure, the feasibility of transformation between single stereo matching model and multi-stereo matching model is proved. Then a real-time optimum and multi-view mixing matching method is raised. The method can get target points through deleting "negative information" image with matching information and achieving optimum combination with the rest image. The method using image preferential strategy aimed at the difficult matching area with traditional methods, this method changes the reference image and the search image getting an improved matching success rate. The experiments show that the method is superior to the traditional method over the matching success and matching efficiency.