Over the last two years I spent in the Robot Vision Group of NLPR as a master student, my efforts have been primarily concentrated on camera self-calibration, matching and 3 D reconstruction. The main work of this thesis consists of the following three parts: 1. Camera Self-calibration Based on Circular Points. A new self-calibration technique is proposed which uses a new type of calibration pattern composed of a circle with lines passing through its center. The proposed technique only requires the camera to observe the calibration pattern at a few(at least three) different orientations, then all the intrinsic parameters can be determined linearly. The main points of our technique are. (1) It need not to establish the correspondence between points of the calibration pattern and its projected image;(2)It need not know the circle center and radius. The calibration process can be totally automatic. Hence it is especially convenient to those people who are not familiar with computer vision. Besides, we also list a catalogue of other possible calibration patterns. 2. Image Point Correspondence and Its Implementation. Based on recent literatures on point correspondence, we have developed a matching software. This software offers two matching facilities for a given pair of images. (1) When the baseline between the two viewpoints is short, it can automatically establish the point correspondence to sub-pixel accuracy. (2) In wide baseline cases, automatic correspondence becomes extremely difficult, the software permits manual intervention to manually select several seeding correspondences to recover the epipolar geometry, then the subsequent corresponding process proceeds automatically. Now the software has been successfully used in 3D reconstruction, IBR and other projects in our group. 3. Implementation of a 3D Reconstruction Demonstration System. First, the camera was calibrated separately, then the image point correspondence was done using the above matching software. After that, the sparse 3D points were obtained image-pair-wisely with a SFM algorithm, and followed by a data fusion process to transform the different reconstructed points into a same Euclidean coordinate system. And finally, by some manual intervention but a largely automatic process of dense matching, surface reconstruction and texture mapping were carried out. This demonstration system seems to perform nicely.
修改评论