相机运动条件下的视频前景提取

提出一种基于非参数化运动估计和图像配准的方法来进行相机运动条件下的前景提取.通过对视频帧和接近的训练背景图像进行非参数化运动估计,动态地构造出一幅和视频帧的视角完全相同的背景图像, 再通过背景减除提取前景. 为了解决运动估计的计算效率问题, 又提出一种基于流形的改进算法: 在离线阶段, 预先对训练背景图像进行非参数化运动估计, 并利用流形学习对训练背景图像进行建模; 在在线阶段, 通过在背景流形上进行运动插值来快速地估计新视频帧和训练背景图像之间的运动. 实验表明,改进的方法在基本保持像素提取准确率的同时获得了很高的效率．

Video foreground segmentation with camera movement

A new approach based on non-parametric motion estimation and image registration was proposed for video foreground segmentation with camera movement.   First, non-parametric motion estimation was conducted between the video frame containing foreground and a similar training background image. Then, a new background image with exactly the same viewpoint as the video frame was constructed on the fly, and foreground pixels were segmented by classical background subtraction.  A manifold based extension was also proposed to alleviate the low efficiency of motion estimation. In the offline stage, motion estimation for training background images was pre-calculated, and a low-dimensional background manifold was discovered. In the online stage, motion estimation involving the new frame was conducted by fast interpolation on the manifold. Experimental results  show that the proposed method segments videos accurately and efficiently．