摘要
当场景中存在较大范围运动前景时,通常的电子稳像算法较难鲁棒地提取背景的全局运动。针对这种情况,提出了利用前景运动掩膜和对极几何关系抑制错误的全局运动估计,进而达到稳像的目的。首先,估计相邻帧间可视化密集光流图;将该光流图用颜色直方图统计各颜色比重,进而分割出运动前景并制作掩膜;利用掩膜删除掉分布在前景上的光流保留背景光流;之后进一步去除外点,利用剩余的光流进行全局运动估计;最后,采用卡尔曼滤波完成运动补偿。三组不同环境的实验结果均表明,该算法能够有效抑制运动前景对稳像结果的干扰,稳像前后图像的PSNR值提高了近36%,稳像效果明显。
When there is a scope of moving foreground,the common electronic image stabilization algorithm is difficult to be robust extraction of the global motion from background. It restrains the wrong global motion estimation with the use of the moving foreground mask and the epipolar geometry,to reach the goal of image stabilization. Firstly,the visualized dense optical flow between adjacent frames is estimated;then the proporation of this optical flow diagram with the color histogram is got. The moving foreground is detected and the mask is made;with the use of the mask the optical flow distributed in the foreground is removed and the background optical flow is retained. Then,further the exterior points are removed and the global motion estimation with the remaining optical flow is carried out. Finally,Kalman filtering is used to complete the motion compensation. The experiment results in three different environments all show that the interference of the moving foreground can be effectively restrained by this algorithm. The PSNR value has been increased nearly 36% after image stabilization and the effect of image stabilization is obvious.
出处
《长春理工大学学报(自然科学版)》
2015年第5期101-106,111,共7页
Journal of Changchun University of Science and Technology(Natural Science Edition)
基金
吉林省科技发展计划项目(20120333
20130101054JC
20140204047GX)
吉林省留学回国人员择优资助项目(RL201329)
关键词
电子稳像
密集光流
卡尔曼滤波
electronic image stabilization
dense optical flow
kalman filtering