英文摘要 | Over the last 3 years I spent in the Robot Vision Group of NLPR, my efforts have been primarily concentrated on Camera Self-Calibration and 3D Reconstruction. The main work can be summarized as follows: 1. A Plane Based Camera Self-Calibration Technique. The novelty of this technique is that it can determine LINEARLY all the FIVE intrinsic parameters of the camera. To our knowledge, techniques reported in the literature up to now can only deal with linearly four of the five ones, in other words, when the camera is of a complete perspective model, i.e., when the skew factor is non-zero, such techniques become invalid. The basic principle of our new calibration technique is to use the planar information in the scene and to control the camera to undergo several sets of orthogonal planar motions. Then, a set of linear constraints on the 5 intrinsic parameters is derived by means of homographies associated with scene planes in images. 2. An Epipole Based Camera Self-Calibration Technique. This technique is an extension of the above plane based technique. In this technique, epipoles, instead of homographies, are used to linearly determine camera's intrinsic parameters. In addition, since the fundamental matrix must be of an anti-symmetric one if the camera's motion between the two images is a pure translation, a 2-point algorithm, rather than the traditional 8-point algorithm, is used to estimate the fundamental matrix. The 2-point algorithm appears to be a great contributor to the substantial increases of the robustness and accuracy of the final calibration results. 3. A Projective Reconstruction Based Camera Self-Calibration Technique. With the camera undergoing a pure translation and two arbitrary motions (a rotation plus a translation), all the five intrinsic parameters can be obtained. This paper's main characteristics are two-fold: Firstly. since it is a linear one, it can avoid the local minima problem plagued in other nonlinear methods in the literature. Secondly, by means of a projective reconstruction, the information of all the available images is used, the method is inherently and fairly robust. 4. Realization of a Prototype of 3D Reconstruction System. At the present stage, the system can reconstruct a 3D scene surface from a pair of images, multiple image based reconstruction will be done later. The main steps of the system are: Firstly, the correspondences of image points are automatically established by using both gray level similarity and consistency of epipolar geometry; Secondly. the camera is calibrated separately; Thirdly, the 3D points are obtained with a standard SFM algorithm. Finally, triangulation and texture mapping are invoked. The system seems to perform nicely. |
修改评论