In this paper we consider motion parameter estimation in dynamic images with the constraint of using as little computation time as possible. We propose using a two-stage divide-and-conquer strategy to achieve this goal. At the first stage, moving objects are located with modest computation time. At the second stage, motion parameter estimation is performed on the localized region. This two-stage strategy reduces the extensive computation task and provides a flexible mechanism to allow termination between the two stages. It is designed to mimic certain features of human visual perception. The derived computation formulas have simplicity and nice regularity, as well as a good local structure that enables the computation to be performed locally.