This work presents a new approach to calculate the optical flow, which uses the decentralized information filter to accomplish a data fusion step. The proposal consists in taking some least mean square estimates of the optical flow vector associated to a certain region of the image considered and to fuse them in order to get a more accurate final estimate of such a vector. Besides this, the way the initial least mean square estimates are obtained is also modified, by reducing the number of observations used to get each one of them, in order to make the overall algorithm faster. The idea is that such initial estimates can be rougher, once the fusion engine improves the accuracy of the final estimate, in the sense that the decentralized information filter generates a final estimate whose variance is lower than the smallest variance associated to the initial estimates. ©2005 IEEE.