The present invention provides an improved method for motion estimation. For this purpose, the prediction quality of a motion vector estimate is evaluated twice, namely by comparing the prediction quality with respect to the two previous fields individually and combining both calculated error values into a single error value for selecting the best motion vector estimate from a plurality of candidate vectors. In this manner, the determination of the real motion of an image object is improved by eliminating false estimates which are not detectable by only referencing a single previous image.