Abstract
Because of the rapid development of computer technology, virtual reality has received much attention in recent years. For virtual reality, constructing a realistic three-dimensional (3D) virtual environment is of particular importance. Tree is very a common object in natural environment, thus its 3D model construction plays a quite important role for natural scene 3D reconstruction. In fact, modeling realistic trees has been a topic in computer graphics for many years and many approaches have been published in the literature. In computer graphics, trees are typically synthesized by some mathematical algorithms in conjunction with some botanic knowledge to generate the geometrical structures of the trees. Often some random variations are also imposed in the modeling process to avoid auto-similarity. Although these graphical methods can model quite realistic trees, the generated trees are, however, different from the real ones in our surrounding environment, i.e., they are grown themselves without any reference to the real trees in the nature. To create a 3D tree model similar to a real tree in the nature, we must model the tree according to the images of this tree. In this dissertation, a complete framework for constructing a realistic 3D tree model from uncalibrated image sequence (i.e., the images are captured with unknown camera internal and external parameters) is developed. We first construct the 3D trunk model via structure from motion technique, and then generate leaves on the created 3D trunk model. This manner can produce quite similar 3D trunk model since the 3D information of the trunk is directly computed from the images. Meanwhile, the generated tree crown will be also similar to the real tree as long as the leaves are appropriately generated. However, to achieve this goal several issues should be addressed. These issues, including correspondences searching, camera self-calibration, and tree segmentation, are also deeply investigated in this dissertation. To acquire the 3D information of the trunk from uncalibrated image sequence, camera self-calibration (i.e., calibrating the camera directly from the captured images) is a necessary step. However, to calibrate the camera, image correspondences should be identified first. Searching image correspondences is an elementary but ill-posed problem in computer vision. Nevertheless, it is convinced that optical flow can provide quite accurate image motion estimation, and hence image correspondences from an image sequence. In this dissertation, an accurate algorithm for computing optical flow under non-uniform brightness variations is first developed. Following this, the issue of camera self-calibration is deeply discussed. A camera self-calibration algorithm suitable for variant camera constraints is proposed. An algorithm for segmenting trunk and leaf regions from a single image is also discussed. The extracted trunk region is quite useful for the subsequent 3D trunk model construction. The skeleton of the segmented trunk region is first extracted to represent the 2D trunk structure of the tree. Subsequently, the 2D trunk skeleton is extended to 3D trunk skeleton by exploiting the calibrated camera and the trunk correspondences. After recovering the 3D trunk skeleton, a set of generalized cylinders is generated around the recovered 3D trunk skeleton to model the 3D trunk. Since the camera has been calibrated, trunk texture can be easily obtained by reprojecting the 3D trunk model onto the image plane. Mapping trunk texture can greatly improve the realism of 3D trunk model. Finally, the leaves are generated on the created 3D trunk model to produce a more realistic 3D tree model. Some experiments were conducted and the results demonstrate the feasibility of proposed system for creating a realistic 3D tree model from uncalibrated image sequence.