The Study On Object 3D Reconstruction Theory And Key Technology Based On Image Sequences

With the rapid development of technologies in recent years, digital camera has become an important tool of image acquisition because of its advantages of low price, convenience and flexibility. It is an important research topic in the field of computer vision to reconstruct the three-dimensional model with planar images taken by uncalibrated camera. The technique of object 3D reconstruction based on image sequence is widely applied not only in the civil domain such as 3D model display, virtual reality, industrial measurement, archaeology, cultural relic protection, real estate management, electronic business, movie stunt and computer game, but also in the military scope of target reconnaissance, vision system of battlefield robot as well as precise evaluation of strike effect. After analyzing and summarizing the theoretical and technological achievements in object 3D reconstruction based on image sequence, the dissertation lays the research emphasizes on such key topics related to 3D reconstruction as uncalibrated image matching, fundamental matrix estimation, trifocal tensor estimation, camera self-calibration, scene 3D reconstruction and error model analysis. The theories and viewpoints on this topic in both computer vision and photogrammetry are also analyzed and discussed. The principal works and research achievements include following aspects.(1) Fundamental matrix expresses the intrinsic relations between two images with overlapped image area. By studying and analyzing a few popular estimation algorithms of fundamental matrix, the thesis puts forward a robust estimation method of fundamental matrix based on Hough transformation. The redundant feature points near the epipolar lines are deleted in line with the results of Hough transformation. The accuracy and stability of fundamental matrix are evidently improved with the proposed algorithm.(2) Radial distortion brings errors to the coordinates of image point, destroys the geometrical relation of camera model and produces negative impacts on image matching, fundamental matrix computation as well as point measurement. This paper puts forward a method of automated correcting the camera radial distortion based on the fundamental matrix. The algorithm introduces radial distortion parameters in the process of fundamental matrix computation. The elements of fundamental matrix together with the radial distortion parameters, as the unknown quantities, can be estimated with least square method. The advantage of the proposed algorithm is that only the two images of the same scene are needed to resolve the radial distortion parameters, while no knowledge about the camera is needed in advance. In this way, the flexibility of radial distortion correction is improved.(3) Image matching is a classical topic in computer vision. The thesis studies the matching algorithm for uncalibrated images,which makes use of the exclusive epipolar geometry relation between two uncalibrated images as constrained conditions. A strategy of image matching is proposed by eliminating those collinear redundant points with Hough transformation. By using the fundamental matrix estimated with the matching points with sub-pixel precision, the accuracy and robustness of image matching results are improved evidently.(4) Camera auto-calibration is an indispensable step of 3D reconstruction based on uncalibrated images sequence. The thesis proposes a stratified camera self-calibration algorithm, in which the infinite plane is firstly determined,followed by the determination of the IAC ( image of absolute conic), and finally the calibration matrix of the camera is calculated. The practical formulae of the stratified camera self-calibration are deducted. On this basis, experiments are carried out with actual image sequence, and the results show that the proposed algorithm is correct and effective.(5) As to the theory and application of image sequences 3D reconstruction, the basic algorithm and procedure of 3D reconstruction with image sequence are investigated in the thesis. The error analysis models of both canonical stereo camera model and convergence stereo camera model are also proposed. The methodologies of calculation of 3D coordinates of the matched points, the triangularization 3D space and texture mapping are deeply discussed. An experimental system of 3D reconstruction of image sequence is designed and developed. With the support of OpenGL software, the mapping of texture and the display of 3D reconstructed model are realized, and a satisfactory effect of 3D visualization is acquired.