The invention discloses a multi-hypothesis prediction method for dynamically changing search window size. The method is characterized by comprising the following steps: performing block-based sampling on a reference frame and a non-reference frame of a video stream by use of a measurement matrix, and numbering different blocks; determining the search window size according to a block change degree, judging the change degree of each block in the non-reference frame relative to the block with the same number in the reference frame, wherein the non-reference frame obtains a non-reference frame residual error reconstruction signal by use of the search window size, and the reference frame performs the reconstruction by use of a fast reconstruction algorithm of a sampling value based on a compressed sensing projection smooth filtering algorithm of the block, thereby obtaining a reference frame hypothesis set; finally adding the reference frame hypothesis set and the non-reference frame residual error reconstruction signal to reconstruct a non-reference frame signal. By use of the prediction method disclosed by the invention, the algorithm time complexity in the multi-hypothesis prediction process is reduced, the reconstruction time is shortened, and the video reconstruction quality under a low sampling rate is improved.