结合Kinect深度图的快速视频抠图算法

Quick matting for videos based on depth images of the Kinect

现有视频抠图算法主要存在人机交互繁琐、计算复杂度高的问题,为此,该文提出了一种利用Kinect深度图的新的快速视频抠图算法。首先结合彩色图信息改进区域生长算法,估计出三色图(原始图像被3种颜色标记出前景、背景和未知区域)以避免深度图中遮挡区域的影响。其次,提出前景和背景样本点集二次筛选机制,保证估计精度的同时大幅降低计算复杂度。最后,采用深度、彩色和置信度图对抠图结果进行加权滤波,减少不透明度图像中低置信度的像素点和不平滑区域。实验结果证明了该算法精度高、速度快且交互简单。

Video matting can be computationally expensive, This paper presents a quick matting algorithm for videos based on depth images of the Kineet. First, the color image information is used to improve the region-growth process to estimate the trimap （marked as the foreground, the background and unknown regions）. This scheme avoids the effect of occlusion regions. Secondly, samples refinement of the foreground and background regions is used to preserve the accuracy of the matting results while reducing the computational cost. Finally, the depth, color and confidence image are combined into a weighting filter to smooth the matting results and reduce the number of low confidence pixels. Tests verify the accuracy and speed of this algorithm.