Long-term global motion estimation and its application for sprite coding, content description, and segmentation

摘要: We present a new technique for long-term global motion estimation of image objects. The estimated parameters describe the continuous and time-consistent over whole sequence relatively to fixed reference coordinate system. proposed method is suitable affine as well higher order models like parabolic model-combining advantages feature matching optical flow techniques. A hierarchical strategy applied estimation, first translation, motion, finally parameters, which robust computationally efficient. closed-loop prediction scheme avoid problem error accumulation in estimation. presented results indicate that very accurate approach can be used applications such MPEG-4 sprite coding or MPEG-7 description. also show efficiency significantly increased if model applied, we on-line applications. further demonstrate estimator serves powerful tool segmentation video sequences.