Options
THREE-DIMENSIONAL VISION FOR STRUCTURE AND MOTION ESTIMATION
FUSIELLO, ANDREA
1999-02-12
Loading...
Contributor(s)
LONGO, O. GIUSEPPE
Abstract
This thesis addresses computer vision techniques estimating geometrie properties of the 3-D world /rom digital images. Such properties are essential for object recognition and classification, mobile robots navigation, reverse engineering and synthesis of virtual environments. In particular, this thesis describes the modules involved in the computation of the structure of a scene given some images, and offers original contributions in the following fields. Stereo pairs rectification. A novel rectification algorithm is presented, which transform a stereo pair in such a way that corresponding points in the two images lie on horizontal lines with the same index. Experimental tests prove the correct behavior of the method, as well as the negligible decrease oLthe accuracy of 3-D reconstruction if performed from the rectified images directly. Stereo matching. The problem of computational stereopsis is analyzed, and a new, efficient stereo matching algorithm addressing robust disparity estimation in the presence of occlusions is presented. The algorithm, called SMW, is an adaptive, multi-window scheme using left-right consistency to compute disparity and its associated uncertainty. Experiments with both synthetic and real stereo pairs show how SMW improves on closely related techniques for both accuracy and efficiency. Features tracking. The Shi-Tomasi-Kanade feature tracker is improved by introducing an automatic scheme for rejecting spurious features, based on robust outlier diagnostics. Experiments with real and synthetic images confirm the improvement over the original tracker, both qualitatively and quantitatively. 111 Uncalibrated vision. A review on techniques for computing a three-dimensional model of a scene from a single moving camera, with unconstrained motion and unknown parameters is presented. The contribution is to give a critical, unified view of some of the most promising techniques. Such review does not yet exist in the literature. 3-D motion. A robust algorithm for registering and finding correspondences in two sets of 3-D points with significant percentages of missing data is proposed. The method, called RICP, exploits LMedS robust estimation to withstand the effect of outliers. Experimental comparison with a closely related technique, ICP, shows RICP's superior robustness and reliability.
Insegnamento
Publisher
Università degli studi di Trieste
Languages
en
File(s)