Motion estimation algorithms based on motion models

Motion estimation is a critical component of any video compression system aimed at achieving efficient storage and transmission of video data. Knowledge of the motion is not readily available from video data. It has to be extracted from the data with various techniques, which are usually very computationally intensive. Block-based motion estimation has been adopted in international standards for video coding to exploit the temporal redundancy that exists between succeeding frames. Conventional motion estimation algorithms in video coding that rely on displacement motion vectors have limitations in capturing potential motions such as scaling, rotations and deformations in a video scene other than the translation. Models characterizing non-translation motions will thus be beneficial as they offer more accurate motion estimation, which can lead to a higher compression ratio. In this thesis, we introduce linear motion models based on the Lie operators, which have the advantage of being linear operators with low complexity. We develop transformation estimation techniques to detect non-translational motion of an object in video scene. Finally, we embed the new transformation estimation system in MPEG-2 codec and evaluate its performance.