Tracking of three-dimensional nonrigid motion and structure using points and curves

In this dissertation, we present a method of tracking 3D nonrigid motion and structure using a sequence of 2D point and/or curve features in perspective images. The points and curves are integrated by representing curves in Bezier form. The algorithm describes a general nonrigid motion with 3D affine motion model. We use Levenberg-Marquardt nonlinear minimization to recover the unknowns of motion and structure. The initial guess for the minimization function is obtained from the first frame range image of the sequence or aligning similar range image to the image sequence. The algorithm was tested with 625 synthetic and 22 real image sequences. We also demonstrate the application to the MPEG-4 FAP recovery.