Research On 3D Reconstruction Technique Based On Multiple View Geometry

It is an important research issue in the field of computer vision to reconstruct the 3D model with 2D images taken by the digital cameras. 3D reconstruction technique based on multiple view geometry is widely applied not only in defense and military fields such as aerospace equipment, target detection, battlefield robot vision system, but also in the industrial circle and civil fields such as autonomous robot navigation, VR, industrial production automation, noncontact measurement, cultural relic protection, security, film special effects and 3D games.The main task of 3D reconstruction: Shoot two or multiple 2D digital images of the same 3D scene from different angles, obtain one to one mapping and position deviation between corresponding image points from two adjacent images by stereo matching algorithm, calculate internal camera parameters and restore the relative pose between cameras(or motion parameters of single camera), in order to obtain 3D geometric information of the scene, and realize the 3D reconstruction of the scene.The main work and research results are as follows:Methods of 3D reconstruction of different cases and external conditions are summarized. 3D reconstruction for binocular stereo vision imaging model of parallel alignment configuration is realized. Methods and Strategies of 3D reconstruction for general binocular stereo vision imaging model and uncalibrated camera vision model are summarized and analyzed in details. Varied estimation methods of fundamental matrix are analyzed and compared. Realized the 3D reconstruction under binocular stereo vision imaging model of parallel alignment configuration and the hand-held uncalibrated camera vision model respectively.2Complete 3D reconstruction under binocular stereo vision imaging model of parallel alignment configuration: improved traditional stereo matching algorithm. Evaluate my algorithm in the standard test platform established by Carnegie Mellon University Binocular Vision Laboratory. The assessment results prove that the algorithm has improved the effect, reconstruction results also confirm that this algorithm has a high matching accuracy, it can clearly reproduce the boundary information of object and recover depth information of weak texture regions and discontinuous disparity regions.Complete 3D reconstruction under the hand-held uncalibrated camera vision model. Specifically, take the standard testing image sequence downloaded from the website of computer vision as the experimental object, completed SIFT feature extraction and feature matching(initial matching) for image pair, realize robust estimation of fundamental matrix by improved RANSAC algorithm, eliminate the error matching produced in the previous process at the same time. The accuracy of slightly improved RANSAC algorithm can reach that average error(average distance between the inliers and respective epipolar lines) is less than 0.2 pixels, the maximum error is less than 0.8 pixels. Complete the camera self-calibration by the method of vanishing points estimation, then restore the essential matrix, achieve the restoration of camera motion parameters R and t, and then recover the camera matrix, calculate the 3D coordinate of the space point based on triangulation principle, automatically eliminate the point whose depth deviates from the reconstructed object seriously by a iterative method proposed in the paper, display 3D map of the reconstructed object after triangulation. Due to the roughness of the 3D map, complete texture mapping operation by openGL under the environment of VS2010 in order to improve the visual effect of the 3D map.In order to test popularity of algorithm, experiment with self-photograph image sequence shooted by rear camera of smartphone and digital camera respectively. To solve the problem that SIFT feature matching(initial matching) has a high rate of error matching for the self-photograph image, propose a slope consistent iterative method to eliminate a part of the error matching points before the improved RANSAC algorithm elimination, this operation can ensure that the RANSAC algorithm still has convergence although it is applied to the condition that initial matching has a high rate of error matching. The average error can be controlled within 0.6 pixels after two times of elimination. This method can ensure that accuracy of matching and accurate estimation of the fundamental matrix for the image sequence which has rich repetitive texture and serious occlusion of pedestrians, and lays the foundation for accurate solution of the follow-up camera motion parameters and 3D coordinate of the space points. The comparison between the slope consistent iterative method proposed by the paper and the similar methods confirms that effect of error matching elimination by the algorithm of this paper is as good as the similar methods, but the former’s real-time is much better than the latter’s.