| Abstract |
One of the most important problems in Computer Vision is the one of finding correspondences between two images. Having two images of a three dimensional scene and some features that have been extracted from them, we want to find for the features in the first view the corresponding features in the second view, if these exist. This function is a prerequisite for a large number of other functions, such as the computation of camera motion, three-dimensional reconstruction from a large number of views and object tracking in multiple views. Furthermore, the algorithms for finding correspondences can be categorised in the general framework of techniques for recognising three dimensional objects from different viewpoints, because they can answer if two views correspond to same 3D scene and if they do, report all the found correspondences between them. Additionaly, correspondence algorithms have a large number of applications, such as relative positioning of an observer, creation of 3D CAD models from multiple views, image registration applications, mosaic construction from multiple views and creation of artificial views. Because of all the above mentioned reasons, the correspondence problem has attracted numerous researchers. Though, the techniques developed so far limit the set of transformations the two views can be related. Most algorithms, assume a stereo pair, where assumptions are made like that the slopes of the lines or photometric information do not change significantly between the two views, or that corresponding points do not move significantly between the two views. Recently, there have been efforts for solving the corresponcence problem for pairs of views that differ significantly. A basic motivation for such a technique is the possibility of reducing the frame rate, and consequently the processing time required. The method presented in this thesis, allows for finding correspondences between features lying on a plane, where the two views of the plane may be related through general projective transformations. Furthermore, the presented method is robust in the presence of a large number of features that are not common in both views. The method belongs to the general category of uncalibrated vision techniques. It relies mainly on mathematical notions coming from projective geometry, like the projective invariants, the projective basis and the projectivity and is based on quantities and parameters that can be directly qualitatively explained. Shortly, the developed method relies on a procedure for predicting points in the second view. Having this procedure as a starting point, we can find subsets in the two views, for which there is strong evidence that they are corresponding. Next, we can compute the projectivity that relates the two views to verify that they depict the same three dimensional plane and if they do, find more correspondences through the process of alignment.
|
Χρήση γεωμετρικών αναλλοίωτων για την αντιστοίχιση χαρακτηριστικών σε όψεις επιπέδων σχετιζόμενων με γενικούς προβολικούς μετασχηματισμούς
