When using a gradient-basedmethod to determine the optical flow field for an image sequence, it is generally appreciated that some spatial pre-filtering of the images is usually needed, particularly for large motion values. However the characteristics of the filter are not generally given. In this paper we analyse the motion measurement from the point of view of sampling theory. We show how an aliasing problem can arise due to undersampling in the temporal domain, and how this problem can be alleviated by appropriate post-sampling spatial filtering. We also demonstrate a connection between this filtering and the methods used to generate the spatial and temporal image intensity gradients.
William J. Christmas