Motion-based Identiication of Deformable Templates

An algorithm is presented to compute the shape of moving objects in image sequences. We assume that locally the motion elds of both the object and its environment can be represented by aane approximations. Based on hypotheses about these elds, an optimal-ity principle is formulated to compute the shape of an object in terms of deformations of an arbitrary convex polygonal domain. As input data the algorithm merely requires rst order derivatives of the image function. Numerical experiments using synthesized and real image data are presented. The results suggest that our approach is suited for tracking tasks within an bottom-up/top-down framework.