Study And Practice Of 3D Reconstruction Based On Image Sequence

The purpose to study 3D reconstruction in computer vision is to recover the 3D scene from the 2D image sequence. The theories and techniques of 3D reconstruction enjoy brilliant prospects of application from 3D terrain simulation, robot self-navigation, object reconnaissance, 3D communication and electronic commerce to protection of cultural relics. The key approaches of 3D reconstruction include automated rectification of the camera non-linear distortion based on image sequences, self-calibration of the internal camera parameters and photo-realistic reconstruction of the scene.According to the theory of multiple view geometry in computer vision, the procedure and key techniques of 3D reconstruction with image sequence taken by the common camera are studied and experimented in this paper, laying the emphises on the approaches to automated rectification of the camera radial distortion and camera self-calibration. The main study works and achievements include following aspects:1. Automated rectification of the camera radial distortionRadial distortion is the main factor to influence the geometric quality of image taken by common digital camera. In order to overcome radial distortion, this paper puts forward a new approach to automated rectification of the camera radial distortion based on fundamental matrix. With the proposed method, both the radial distortion coefficients k₁ , k₂ and rectified fundamental matrix F can be calculated based on the combination of RANSAC method with least square method of estimating fundamental matrix F. With this approach, specific calibration device required in traditional rectification process is no longer needed. Thus, the proposed rectification method is more flexible in application.2. Transforming the canonical projection matrice into projection matrice in a global frameThe pre-processing of camera automation calibration includes the estimation of the trifocaltensor and the transformation of canonical projection matrice into projection matrice in a global frame in a global frame. Trifocal tensor encapsulates all the projective geometric relations among three views that are independent of scene structure. The trifocal tensors of the image sequence in order are obtained with RANSAC method at the first, and then projection matrices computed by the trifocal tensors are transformed by a 3D homographic matrix into those in a global frame, which the first camera center is set as the origin. In this way, avoiding the troublesome process of bundle adjustments are avided.3. Camera auto-calibration and 3D reconstruction of the sceneCamera calibration is an important step in 3D Euclidian reconstruction. This paper improves auto-calibration method using the absolute dual quadric. The method linearises the polynomial equations, and then uses the Newton iterative method to find the resolution. Experiments show the proposed method is effective with both simulative data and the actual image data. On this basis, 3D homogeneous coodinates of the matched points on the images are calculated. The Delaunay triangle network is constructed and photo-realistic 3D model of the scene are reconstructed by mapping the texture to the corresponding 3D triangle areas.An experimental system is developed with VC++ and MATLAB to verify the effectiveness of the algotithms related in this paper.