全局运动估计中特征点选取和鲁棒性分析

摘要

This paper studies good feature selection and robustness analysisof global motion estimation. At present the global motion estimation is applied widely, such as in sprite coding, virtual reality and panoramic image generation. But it is difficult to reach the balance of the speed and precision, so doing research on selection of good features and robust calculation to global motion estimation is necessary. The difficulty of global motion estimation is that it is apt to be influenced when local motion of foreground exists. Good feature used in global motion estimation can accelerate the calculation and robustness analysis can improve the accuracy of calculation. This paper proposes the criterion to select good features and a robust calculation for global motion estimation in global motion compensation coding. Motion feature is utilized for global motion estimation according to spatial gradient in the feature selection. There are two methods to exclude the noise in robustness calculation. One is based on residual histogram and the other is based on residual block. Adaptive weight function is added in the objective function in robustness analysis. The influence of noise can be restrained when motion features are used in calculation, and also the convergence of iterative calculation can be accelerated. Comparative experiments are performed to validate those proposals. Some sequences with global motion are utilized in the comparison experimentation. From the experimental results, conclusion can be made that the calculation of global motion estimation can be accelerated and the precision of results can be guaranteed using the good features and robust calculation. The effectiveness can be observed from the comparisons.%文中研究了全局运动估计中的关键问题，提出全局运动估计中有关特征点选取和计算中的鲁棒性分析等问题的解决方法，提高了全局运动估计计算的整体性能.全局运动估计用途非常广泛，如它可以应用于sprite编码、基于运动分割、虚拟现实和全景图生成等领域中.但是由于全局运动估计计算量大，噪声多等因素使得它很难在实时应用中使用.为了使计算速度和估计精度能够很好结合，文中提出了适合全局运动估计的特征点选取准则和鲁棒性分析方法，选取特征点进行计算可以减少计算量，提高运算速度；而在计算中引入鲁棒性分析方法可以抑制噪声影响，保证计算精度，因此这两者的结合就可以很好地解决快速全局运动估计问题.从实验结果来看该文方法是有效的，对全局运动估计的改进是明显的，速度提高而精度也可以得到保证.