Frequency analysis of contour orientation functions for shape representation and motion analysis

Abstract

The location and characterization of edges is a crucial factor for some image processing tasks, such as shape representation and motion analysis. In this paper, we present a set of filters which allow estimating edge orientation and whose output does not vary when the input signal is rotated or when a global illumination change occurs. The outputs of these filters are used to extract a one-dimensional representation of the contour in order to characterize shapes in an accurate way by using Fourier coefficients. An energy function is built to associate shapes and determine the similarities between object contours. The elements of this energy function are weighted so that those which provide the most relevant information have a larger contribution. By extracting the parameters of the transformations which relate every image in a sequence with the following, we can determine the temporal evolution of a moving object.