Method for fast multiple reference frame motion estimation

A method for fast multiple reference frame motion estimation, which is used to code an image frame sequence including a current frame Fn and multiple previous reference frames Fn−1, Fn−2, . . . , Fn−(k−1) and Fn−k positioned rearward in time with respective to the current frame Fn, where n and k are an integer, the frames being divided into a plurality of blocks, the method finding motion vectors for a block in the current frame referring to the multiple reference frames, the method comprising:

• a block-matching step, which performs a special block matching to find a plurality of motion vectors for blocks of each frame referring to a previous one frame;
• a motion vector composing step, which obtains composed motion vectors of the current frame referring to the reference frame Fn−k for the block, wherein
• MVn(i)=MVn(i−1)+MVn−(i−1)(1)
• where MVn(i) is the composed motion vectors of the current frame referring to the reference frame Fn−i, MVn(i−1) is the motion vector of the current frame referring to the reference frame Fn−(i−1) and MVn−(i−1)(1) is the motion vectors of the reference frame Fn−(i−1) referring to the reference frame Fn−i, i is an integer and 2≦i≦k;
• a selecting step, which selects one with a minimum of cost function from the composed motion vectors MVn(i) produced in the motion vector composing step;
• a fine adjusting step, which finely adjusts the composed motion vector selected, thereby obtaining a final motion vector of the current frame referring to the reference frame Fn−I;
• and
• a global selecting step, which selects a final motion vector with a minimum of cost function from final motion vectors produced when the block in the current frame refers respectively to the reference frames Fn−1, Fn−2, . . . , Fn−(k−1) and Fn−k as a best motion vector of the block in the current frame,
• wherein in the global selecting step, the cost function is represented by J(m, ref, λmotion)=SAD(s, c(m, ref))+λmotion•(R(m−p)+R′(ref)), where m=(mx, my)T is a motion vector, p=(px, py)T is a predictive motion vector, R(m−p) is a motion vector information, λmotion=(0.85·2QP/3)1/2 is a corrective factor, ref is a reference frame and R′(ref) is information of the reference frame.