Motion estimation method using a 3D steerable filter

摘要

Scene motion is characterized by the orientation of its spectrum in the frequency domain. An image sequence can be divided into small regions which contain only translating information; thus, detecting the image motion corresponds to detecting the orientation of the spectrum of a moving image. In this paper, we use ‘steerable filters’ to estimate the motion of a scene. The steerable filters are synthesized by the interpolation of some ‘basis functions’ and ‘steered’ to arbitrary orientations. These filters measure the ‘oriented energy’ along certain orientations. When the oriented energy is minimized, the spectrum plane can be identified, for which the orientation indicates the velocity. Using steerable filters has two advantages: first, we have no orientation selection problem — the conventional ‘Gabor-type’ filters can have only certain directional pairs of filters; and second, it is computationally efficient because the output of steerable filters can be interpolated by the outputs of its basis filters. In our experiments, we show the accuracy and simplicity of our method using steerable filters.