Motion analysis using orthogonal wavelets in the spatio-temporal filtering paradigm
Ronald L. Allen · 1993
Wavelet pyramid decompositions are applied for detecting and estimating the motion of rigid objects in image sequences. Motion in the image plane is found by measuring the orientation of surfaces or by measuring the energy of oriented textures in the space-time volume of the image sequence. This requires a three-dimensional representation. The computationally efficient two-dimensional orthogonal wavelet pyramid decomposition is extended to three dimensions so that a time coordinate is present. A direct derivation of coarse pyramid representations that immediately gives the desired coefficient sets is described; a parallel, rather than cascade, pyramid algorithm results. Algorithms for estimating motion are developed for two variations in the spatio-temporal filtering paradigm: orientation estimation and local spectral energy estimation. From the limited horizontal and vertical directional sensitivity given by orthogonal wavelet representations, a broader range of two- and three-dimensional orientations can be discerned. This becomes the basis for robust motion estimation using orientation. Filter responses across several pyramid resolution levels are used to detect localized, oriented textures. Motion estimation algorithms based on spatio-temporal frequency prove to be less reliable than the orientation estimation techniques. The direct derivation of the regions of the pyramid representation that are needed for motion analysis is shown to provide significant improvements in algorithm efficiency.