Progressive Model Refinement Global Motion Estimation Algorithm for Video Coding

摘要

This paper presents a Progressive Model Refinement (PMR) method for Global Motion Estimation (GME) in MPEG-4 video coding. Our contributions consist of two aspects. Firstly, a method of feature point selection is proposed based on the analysis of spatial distribution. It can effectively guarantee the number of feature point won't become too large and avoid most feature points congregated on a small region. Secondly, a PMR algorithm is proposed to select motion models progressively according to the complexity of the camera motion, which improves the convergence performance of GME and makes the PMR algorithm much more robust and faster than single-model based GME algorithms. Experiments show that the presented algorithm can always select the appropriate model to describe the camera motion.