Study And Implementation Of Image 3-D Reconstruction Based On Feature Matching Strategy By Geometric Constraints

Image-based 3-D reconstruction is a hot research topic in the field of computer vision.Making full use of the information of images and generating the real texture 3-D model by projection to images is the core of the research.The amount and accuracy of matching couples finally determine the completeness and quality of the 3-D model.On the basis of existing research,the strategy to improve the number and accuracy of matched couple points and to calculate space coordinates of foreground object using the image geometric constraints is studied to reconstructure the scenes which are photographed.Main works are listed bellow:(1)Fast feature extraction and preliminary matching between high resolution imagesIn view of the inefficient calculation and the huge amount of data produced in the process of SIFT feature extraction in high resolution images,parallel block strategy is improved to reduce the size of data and extract 10~4~10~5 feature points in a single image in 10~2 seconds.However,the computation of feature matching increases greatly along with the number of feature points.Thus,the scale distribution of feature points is analyzed to achieve the filtering method based on the scale of feature points,determining the reliable preliminary matching in 10 seconds.(2)The dense matching method based on scale distribution,epipolar constraint and parallactic restriction of feature pointsSince the amount of feature points in pictures are enlarged,similar features are more likely to exist,causing the limit of the quantity of matched couple points detected only through feature matching.According to the preliminary match,the stable basic-matrix and couples are detected by robust method,epipolar constraint and parallactic restriction are used to furtherly narrow the search scope.Matched couples are iterative improved by shrinking the parallactic estimation.Parallel computing is used to improve the matching process,matching 10~4 couples in 10~2 seconds.(3)The improvement of three-view unit 3-D reconstruction strategyAlthough the two-view matching relationship is established,wrong matches are still possible to happen.Existing schemes use three-view epipolar constraint to delete wrong matches.This paper furtherly use it to enlarge the amount of current matching couple points.Projection matrix in European space is calculated by attribute data of images.3-D point cloud is generated by reverse projection.For large scale terrain data,points are projected to the main plane and planar triangulation is made to simplify the estimation of the space triangulation of the 3-D point cloud.Eventually,a strategy is created to quickly form the structure and rich texture information of the 3-D model by three-view-unit.(4)The promotion of camera model in special situation such as specular reflectionThe interference of special situations such as specular reflection,uneven light,refraction and cover is discussed.Specular reflection,the situation which is still in line with the existing 3-D reconstruction calculation assumptions is studied to derive the related mathematical model,to get its detection method,to reduce the impact in reconstructure strategy and to gain new constrains to optimize the result.(5)Image 3-d reconstruction software with independent intellectual property rightsBased on the Exploration in the whole process of image 3-d reconstructure and the breakthrough of the key technology above,the fast and intensive feature extraction and matching in high resolution images is completed,the dense point cloud and real terrain surface model is formed,and the software with independent intellectual property rights is generated.