Structure From Motion Based On Binocular Vision

Using multi-view images captured by moving camera to recover three-dimensional object structures is a hot research topic of 3D reconstruction. Multiple disciplines support its development, including computer vision, pattern recognition, image processing, computer graphics and so on. With the development of the basic theories, its applications for civil use is becoming wide, such as virtual reality, augmented reality, medical diagnosis, robot vision navigation.Focusing on the reconstruction of object structures from multi-view images, this article incorporates the characteristics of binocular stereo vision into the structure from motion(SFM) theory, to perform the 3D point cloud reconstruction for multi-view images. The main contents are as follows:(1) Some basic theories are introduced, such as the theories of single camera calibration, stereoscopic calibration and stereoscopic correction. Based on these theories, the binocular stereo camera calibration is implemented. Then, the basic procedures of SFM are elaborated in details.(2) A reconstruction algorithm is adopted to fuse the binocular point clouds which are reconstructed by the SFM algorithm. Firstly, two clouds corresponding to the two cameras are reconstructed respectively, based on the two-view cloud generation method of SFM and the epipolar constraints theory. Then, the two clouds are fused into one to obtain the final object structure, via the fusion partial cloud method of SFM and the characteristics of binocular image pairs whose corresponding pixel rows are aligned. The experimental results demonstrate the effectiveness of the proposed method.(3) An incremental structure reconstruction algorithm for binocular vision is presented. Firstly, the 3D cloud of the binocular image pair at every view is generated based on the parallax computation theory of binocular vision. Then, the public 3D cloud points are matched between two adjacent views, based on which the new binocular camera posture is estimated. Through the new binocular camera posture, the new binocular cloud is fused to the former incrementally. Finally, based on the fixed position characteristic of the binocular cameras, an improved binocular bundle adjustment algorithm is adopted to optimize the 3D point cloud and the binocular camera posture. The final 3D point cloud is the desired object structure. The generated point cloud models by the compared methods and the proposed method verify the feasibility and effectiveness of our proposed reconstruction algorithm.