The goal of this research project is to develop algorithms to solve dual problem of structure and motion recovery of nonrigid objects from multiple 2D views. Previously, algorithms have been developed that can recover point correspondences in nonrigid motion from range data sequences. If only 2D projections (perspective or scaled orthographic) are available, not only motion but also structure (i.e. 3D coordinates) of the surface points need to be determined. There are several research questions related to nonrigid motion and structure recovery which are being addressed. .First, depending on the observed motion, which of the various nonrigid motion models should be utilized (that include determination of the number of parameters to be estimated)? Second, once the motion model is decided, how many feature points, over how many frames are required in order to estimate the motion and structure of the points under consideration? Further, the derived solutions need to be extended to line (and other) features. This local motion modeling approach is often noise sensitive and may have problems handling large deformations. Those drawbacks will be addressed bycombining local techniques with global models. Global models under investigation include time-varying superquadrics and hyperquadrics, spherical harmonics as well as FEM models.
