Research On Profile Extraction And Key Technologies For 3D Reconstruction

Three-dimensional information obtained from two-dimensional images is one of the hot current studies and belongs to a multi-interdisciplinary area of research, which involves the projection geometry, digital image processing, computer graphics, computer vision and many other disciplines. Three-dimensional reconstruction is to restore the three-dimensional space information of objects through the basic elements (such as point, line, plane) of two-dimensional images, and it needs to study the relations between three-dimensional coordinates of points, lines and planes in three-dimensional space and the corresponding ones in two-dimensional images, for achieving quantitative analysis of the sizes and positions of objects. The three-dimensional information or three-dimensioal model, which is obtained by features extraction, features matching, reconstruction of key characteristics, triangulation, and data fusion, is widely used in many fields like visualization of the virtual plant, digital entertainment, appearance design of industrial products and virtual scene simulation.Reconstruction for three-dimensional objects based on two-dimensional images has made some achievements. The existing reconstruction methods can produce acceptable results for the regular profile curves (such as hyperbolic, parabolic, etc.) or man-made object profiles (housing, furniture, etc.). However, for the nature scenery or irregular profile reconstruction, the existing methods have many limitations.The thesis studies the methods of contour extraction and reconstruction of images obtained from different orientations by a single camera. The work mainly includes the scene contour extraction, feature detection, profile curve segmentation, feature matching, projective reconstruction, and quasi-Edclidean, and obtaining 3D spatial information from 2D image contour features.Finding out basic features in an image is the foundation of highlevel processing such as object recognition, reconstruction, analysis and description of shapes. After investigating the characteristics of scene images, we propose a novel algorithm for edge profiles extraction based on fractal and wavelet, and present the concept of fractal dimension standardization. The algorithm firstly computes fractal dimension values of R, G, B components of every pixel in images, and then synthesizes color image by normalized fractal dimension values, finally decomposes grey level histogram of synthetic image by wavelet transform and automatically determine the threshold of extraction edges according to wavelet decomposition coefficients. This not only makes full use of the adjacent pixel spatial relationships, but also improves the automation degree of edge extraction.For different high-level processing, we propose two methods for contour segmentation based on feature points, factal and curvature respectively. The first method segments profile curves according to the feature points which are detected according to relation between the eigenvalues of the covariance matrix and curvatures of points on curves. Feature points detection method firstly computes the smallest eigenvalues of the covariance matrix in different support regions, then takes the smallest eigenvalues of every point on curves multiplication and determines the feature points corresponding to the peaks of the product histogram. The proposed method not only avoids the problems coming from the definition of point curvature and the one in determining the threshold, but also improves the accuracy of the segmentation of edge contour and automation by suppressing noise and pseudo-feature points. The second one segments curves according to the relation between length of the curve and the errors produced in the conversion from grid maps to the vector maps, and merges similar curve segments according to fractal dimension of each curve segment.By analyzing the geometric relation between the points and the lines, a new matching method is proposed based on line feature. Each endpoint, obtained by segmented profile curves according to the feature points and cross points on curves, may be connected with another one or more, that is to say, each endpoint may be on one or more curve segments. After matching roughly endpoints by average of square difference (ASS), normalized cross correlation (NCC), endpoints sets of two view images are divided into many matching point subsets, among which the dynamic programming algorithm is used to match lines with the edge potential energy as a measure. The algorithm not only makes use of feature points, the geometric relation between feature points and line segments, as well as the gray spatial relations of pixels, but also reduces the research space and improves the running efficiency.On the profile reconstruction, after investigating the nature of space curve, spline curve fitting, and the relation of space curves and space surfaces, we proposes a new method to reconstruct curves in three dimensional space for the calibration image sequence with line feature as a basic element, and apply it to the profile reconstruction of space irregular objects.