Multiresolution Motion Estimation and Feature Extraction

Figure 2: A novel filter-based, multiresolution approach to velocity estimation allowing coarse-to-fine refinement of the velocity estimates. The algorithms are efficient and lead directly to precise descriptions of (affine) motion vector field. (Data sets from Jet Propulsion Laboratory at Caltech.)

We have made progress on two different methods for the extraction of features that have direct interpretations in terms of object composition or structure. In the first we combine so-called wavelet network approximations of images and signals with morphological smoothing in order to identify critical scales in the image of an object and to extract representations at these scales. In the other we have adapted the wavelet pursuit method of Mallat in order to extract representations based on prior models of the multiresolution structure of an image or signal. A set of novel metrics for measuring goodness of fit, which avoid the myopia of trying to fit one basis function at a time, have been developed and show considerable promise in modeling complex signals.

One application of our software algorithms is for autonomous hazard detection and motion estimation near the surface of Mars using a scanning laser radar (LIDAR) as shown in Fig. 2. These algorithms aim to enable safe and precise landing by spacecraft. (The P.I. of this project is Lachiver.) Another application our software strategy is for ``Multiresolution Morphodynamic (geometrical/topological) Imaging'', such as in tracking of the human heart, see Fig. 1.