Abstract. Velocity distributions are an enhanced representation of image velocity implying more velocity information than velocity vectors. Velocity distributions allow the representation of ambiguous motion information caused by the aperture problem or multiple motions at a given image region. Starting from a contrast- and brightness-invariant generative model for image formation a likelihood measure for local image velocities is proposed. These local velocities are combined into a coarseto-fine-strategy using a pyramidal image velocity representation. On each pyramid level, the strategy calculates predictions for image formation and combines velocity distributions over scales to get a hierarchically arranged motion information with different resolution levels in velocity space. The strategy helps to overcome ambiguous motion information present at fine scales by integrating information from coarser scales. In addition, it is able to combine motion information over scales to get ve...