Research On The Key Technology For 3D Reconstruction Based On Optical Flow Of The Monocular Image Sequence

Reconstructing the 3D motion and structure of the moving object or scene from the monocular image sequence optical flow is a research focus of computer vision, since the optical flow contains not only the movement parameters of the object or scene, but also the abundant information of 3D structure, it has been widely applied in the robot vision, aerospace, military, medical research and other research fields. This thesis mainly studies the theory and technology for 3D motion estimation and structure reconstruction based on the monocular image sequence optical flow, it aims to improve the computing efficiency, develop the robustness and broaden the scope of the reconstruction algorithm. It proposes some new ideas and methods for 3D reconstruction based on the monocular image sequence optical flow from the view of projecting the variational optical flow model, reducing the neighborhood error influence, amending the pixel drift phenomenon caused by the brightness variation, reducing the constraint condition of the reconstruction model based on the straight-line optical flow and establishing the corresponding relation of point and straight-line optical flow. The main contributions of this thesis are as following:1). Start with the variational optical flow theory, it improves the data term of the variational optical flow energy function by adding local constraints and gradient constant assumption, and it designs an anisotropic diffusion smoothing term based on the image gradient to change the optical flow diffusion mode at the edge of the object or scene. For the computing efficiency, it employs cellular neural networks to optimize the variational optical flow estimation model, the anisotropic diffusion optical flow algorithm based on the adaptive cellular neural networks is proposed. Experiments show that the model can well deal with the noise of the image sequence, which has the high calculation accuracy and less time consumption.2). To reduce the terrible influence of the 3D reconstruction caused by the optical flow error of the neighborhood and to prevent the phenomenon of edge expansion, a linear method for 3D motion estimation based on the dense optical flow has been proposed according to the relationship of the 3D motion and image point optical flow, and the robustness of the proposed method is analyzed and proved by the Norm theorem. Experiments show that the proposed model can reduce the influence of the optical flow diffusion, keep the edge information of the object or scene.3). To avoid the error influence caused by the optical flow estimation, the 3D motion estimation and structure reconstruction model based on the variational optical flow theory has been proposed by the 3D motion constancy assumption based on the image brightness. In order to deal with the large displacement motion, the coarse-to-fine strategy is employed to optimize the motion estimation. For reconstructing the purely rotational motion, the 3D space displacement of the point is used to instead of its trajectory, and then the 3D reconstruction model for the purely rotational motion has been presented. Several experimental results prove the proposed method could well deal with the the brightness variation and large displacement motion in the image sequence, and it partly solved the robustness problem of motion occlusion.4). It researches the 3D reconstruction technology based on the straight-line optical flow, proposes the concept and calculation method of the straight-line optical flow. Then the linear 3D motion estimation model is projected according to the perspective projection relationship between the 3D straight-line and the corresponding image line. Experiment shows the proposed method just requests at least two straight-line optical flows to estimate the motion parameters of the moving object or scene.5). For the tracking and matching of the moving lines in the complex scene, a basic constraint formula between the point and straight-line optical flow is deduced and a robust moving lines tracking and matching method according to the relationship of the point and straight-line optical flow is proposed. According to the tracking result of the moving lines, the 3D motion could be estimated by the straight-line optical flow based method and the surface dense structure and sparse linear characteristics could be reconstructed by the 3D reconstruction model combination of the point and straight-line optical flow. Experimental results prove that the proposed method could accurately extract and track the moving lines of the image sequence and the reconstructed result could express both the surface details and the straight-line topological structures of the moving object or scene.