Motion estimation from three-dimensional data

In this thesis several new algorithms for the matching and motion estimation from 3-D data are introduced. First, various versions of the algorithms are classified into four levels based on the availability of the constraints such as object rigidity, point features, and matching correspondences. Then, at each level, several new algorithms are derived and their performance and applications are presented. The first level is the most primitive form of motion estimation in which rigidity, point features, and matching correspondences are required. To demonstrate such an algorithm the terrain matching problem is considered. Surface curvatures are used to obtain motion invariant feature points which are used for matching and motion estimation. Real range data is used to test algorithm performance. At the second level, the requirement of matching correspondences is removed for the purpose of computation reduction. First, algorithms for motion estimation from 3-D data without correspondences are introduced for both point and line feature points. Motion estimation is performed by fixing three orthonormal vectors to the 3-D set before and after the motion, and recovering motion parameters from positions of those vectors. Then the problem of matching of 3-D sets is considered and matching algorithms are derived. The stability of these algorithms is investigated and tested on the simulated noisy data. Finally, algorithms are used on real data of vehicle motion tracking and automated construction. At the third level, the point-feature requirement is removed in order to further reduce the computational complexity. This modification enables the algorithm to perform motion estimation from continuous image distributions. Both matching and motion estimation algorithms are presented and parameters which affect the stability of these algorithms are investigated. At level four, the rigidity constraint is removed while the requirement of point features and matching correspondences is retained. Algorithms in level four are intended for motion analysis of nonrigid or partially rigid objects and are based on tracking of surface curvatures. Algorithms are developed and applied, first, to the simulated piecewise-rigid and nonrigid object, and then to the human left ventrical data. The heart surface profiles and the surface stretching parameters are reconstructed from 3-D data obtained by the biplane angiography.