基于运动信息与多尺度分水岭的视频目标分割

Video Object Segmentation Algorithm Based on Motion Information and Multi-resolution Watershed

从视频序列中提取视频目标是基于内容编码中的一项关键技术。提出了将高阶统计运动检测和多尺度分水岭相结合的视频目标分割算法。该算法首先利用高阶统计运动检测算法检测出运动区域,通过后处理得到运动目标的初始模板。然后,用小波变换对视频图像进行多分辨率分解。在最低分辨率上应用分水岭算法分割得到具有精确边缘的分割区域,通过将区域融合后的区域逐步投影到高分辨率图像上并结合高分辨率图像上的分水岭算法逐步提取出具有精确边缘的区域。最后,将运动目标的初始模板和多尺度分水岭分割得到的区域结合起来提取出具有精确边缘的视频对象。实验结果表明该算法能有效地分割和提取出视频序列中的视频对象。

Extracting the video object is a key technology in content-based video coding such as MPEG-4. An algorithm extracting video object is proposed by utilizing the HOS (higher order statistics) and multi-resolution watershed algorithm in video sequences. First, the HOS is used for motion detection. Then through post-processing, a rough motion mask can be obtained. After creating multi-resolution images using a wavelet transform, over-segmented regions with accurate boundaries are obtained by using watershed algorithm in low-resolution image. Second, the merged regions in low-resolution image are projected into a high-resolution image level by level. Third, the regions with accurate boundaries are extracted through combining the projection and the regions segmented using watershed in high-resolution images. Finally, the video object with accurate boundaries is extracted by integrated the motion mask and regions segmented by multi-resolution watershed. The experimental results show that the proposed algorithm improves the performance of detecting moving object and good segmentation results can be obtained.