A method and an apparatus for elevating compression efficiency of a motion vector by effectively predicting a motion vector of an enhanced layer by means of a motion vector of a base layer in a video coding method employing a multi-layer structure are disclosed. A motion vector compression apparatus includes: a down-sampling module for down-sampling an original frame to have a size of a frame in each layer; a motion vector search module for obtaining a motion vector in which an error or a cost function is minimized with respect to the down-sampled frame; a reference vector generation module for generating a reference motion vector in a predetermined enhanced layer by means of a block of a lower layer corresponding to a predetermined block in the predetermined enhanced layer, and motion vectors in blocks around the block; and a motion difference module for calculating a difference between the obtained motion vector and the reference motion vector.