Local motion estimation with the Dynamic Pyramid

Abstract

The Dynamic Pyramid is a tool for analyzing monocular moving image sequences. Its process for achieving correspondence between successive frames is based on the physical model of the elastic membrane, where the cost function consisting of the similarity between the frames and the deformation of the transformation vector field is minimized.

A new concept is developped in order to determine the iniatial local displacements. It includes isolated points like corners but also makes use of the one given component of the displacement at edges. For the local correlation the sign of the Laplacian is used. This is a fast and reliable method to determine the initial displacements. A new approach of region oriented continuity constraints is developped in order to allow semantic discontinuities e.g. due to occlusion.

To make a scaleinvariant analysis of image sequences possible, the membrane model is extended to the Dynamic Pyramid. It makes use of the fact that at the coarser levels of the pyramid even large distances come into the realm of local operators. The results of one level are good starting values of the next. Thus also there local operators can be applied.

This achieves a reliable mapping in combination with an efficient calculation. The result of the dynamic pyramid is a local motion vector field. From there segmentation of moving objects is simple and the basic information for further steps like 3D-reconstruction is given.