摘要: 结合两步法与传统梯度下降算法,提出一种改进的快速全局运动估计算法。采用稀疏抽样的MSEA快速块匹配算法估计局部运动矢量,使用迭代最小二乘法粗估计全局运动参数并排除外点(前景宏块),在排除外点的采样宏块集上选取特征像素,以上述两步法的全局运动估计参数为初始值,利用LM梯度下降算法对全局运动参数进行优化。实验结果表明,改进算法的估计速度达到11.42 ms/f,比FFRGMET算法快1.3倍,具有更高的全局运动估计精度。
关键词:
全局运动估计,
梯度下降算法,
块匹配算法,
迭代最小二乘
Abstract: This paper presents an improved fast Global Motion Estimation(GME) algorithm by combining with two-step method and traditional Gradient Descent(GD) algorithm. Sparsely sampling MSEA(Multilevel Successive Elimination Algorithm) fast Block Matching Algorithm(BMA) is used to get local motion vectors. Iterative Least Square(ILS) method is used to get rough estimation of the global motion parameters and excludes outliers(foreground macro-blocks). The rough global motion parameters is used as initial value and LM(Levengberg-Marquardt) GD optimization method is used on the feature pixels which are selected from the residual sampled blocks that have been excluded outliers with ILS. Experimental results validate that the estimation speed of improved algorithm reaches 11.42 ms/f, it is 1.3 times faster than FFRGMET algorithm, and it gets higher GME precision.
Key words:
Global Motion Estimation(GME),
Gradient Descent(GD) algorithm,
Block Matching Algorithm(BMA),
Iterative Least Square(ILS)
中图分类号:
刘磊, 王志良, 刘冀伟, 石志国. 一种改进的快速全局运动估计算法[J]. 计算机工程, 2010, 36(20): 28-31.
LIU Lei, WANG Zhi-Liang, LIU Ji-Wei, DAN Zhi-Guo. Improved Fast Global Motion Estimation Algorithm[J]. Computer Engineering, 2010, 36(20): 28-31.