Video Motion Compensation Prediction Technique Exploiting Zooming Characteristics

Motion compensated prediction (MCP) is one of the essencial techniques significantly improving the video compression efficiency. The block matching techniques based on a translational motion model are widely employed in inter-frame prediction. The translational motion model assumes that the objects are rigid and always moving in a fixed 2D plane perpendicular to the central axis of the video camera. But actually, objects may move in any direction in the 3D space, such as moving towards or away from the camera. Additionally, there may also be some camera operations such as panning and zooming. All these movements will cause the distances between the objects and the camera changing over time, which would be uniquely considered as zooming motions. Apparently, the zooming motions could not be effectively represented by only employing the translational motion model, which requires the zooming characteristics be utilized in MCP.To tackle the aforementioned problem, a new MCP technique considering both translational and zooming motions is investigated in this dissertation, which is named as translational and zoom integrated motion compensated prediction (TZIMCP). Besides translational motion, TZIMCP takes the characteristics of the zooming motion introduced by objects motions and camera operations into consideration in MCP, and provides an inter-frame predicted block with higher precision to the current block.The effectiveness of TZIMCP is verified by an observed significant reduction of the inter-frame prediction error or residue compared with the conventional one only using the translational model. Furthermore, to bring down the computational complexity introduced by TZIMCP, a pre-selection approach for the available reference frames is proposed based on intensive and extensive statistics analyses. Experimental results demonstrate that the TZIMCP scheme using pre-selection can achieve a precision of about 97% in searching the best inter prediction blocks, while its execution time is only about 19.9% of that of the full-search TZIMCP scheme, and can improve the prediction precision by 0.5% compared with the existed T style search TZIMCP scheme, while the computational complexity is notebly decreased by 25.8%.