Three-dimensional Reconstruction Based On Images Of Monocular Vision

Three-dimensional reconstruction has always been a hot spot and one of the difficulties in the field of machine vision research. Three-dimensional reconstruction techniques is to study how to obtain two-dimensional information of the object by object of three-dimensional information. The outcome of the issue has been widely used in robot navigation, precision machining industry, non-contact measurement, virtual reality, field exploration and military. This paper presents a complete three-dimensional reconstruction algorithm based on the color image sequence of monocular vision, which has a wide range of advantages such as simple experimental equipment, low cost, high degree of automation, non-contact. In this paper, key technologies such as feature point extraction and matching, the camera matrix calculation,3D reconstruction techniques, have been discussed in detail, and some key difficulties are analyzed.First, This paper conducted in-depth research of camera model and epipolar geometry. Epipolar geometry is the basis of three-dimensional reconstruction. By use of pole geometry knowledge fundamental matrix between the two views can be obtained. By fundamental matrix between two images, we can get two camera matrices of two images.Second, in this paper a detailed analysis of the problem on how to strike the camera matrix is conducted in depth. For calibrated cameras, to strike the camera matrix problem is actually to strike the external parameters of the camera. It is a problem of solving highly nonlinear over-determined equations to strike the camera matrix. Such problems generally have these characteristics:a plurality of local minimum values, low accuracy, poor robustness. This paper describes three methods for solving these problems--DLT algorithm, pyramid algorithm, genetic algorithm. In this paper, difficult problems in approaches to obtain camera matrices are analyzed, and our own solutions are put forward. For different experimental conditions and different levels of error, you can select the best solution to improve the efficiency and scope of the algorithm.Third, in this paper common problems in three-dimensional reconstruction process are described in detail and discussed, and 3D reconstruction process of multiple views are described in detail. To improve the accuracy of three-dimensional reconstruction, bundle adjustment is put forward. In order to avoid excessive parameters, we use the alternate optimization method to reduce the computational complexity. Practice experiments show that this algorithm is necessary and effective. In the reconstruction process efficiency is greatly improved.Fourth, in this paper, a lot of simulation and actual experiments are done, comparing the similarities and differences between different algorithms. In the experiments the details of the genetic algorithm are improved to maintain high efficiency. In the simulation, calculation error is substantially the same as expected. In the actual experiment, the level of error is around 0.05 pixels.Fifth, a comprehensive application of these algorithms to complete the three-dimensional reconstruction of image sequences is applied. The level of error is around 0.75 pixels and the desired accuracy is achieved.