基于新视角合成的视频摘要交互式浏览

摘要

现有视频摘要方法大多仅在固定视角下对视频内容进行压缩，限制了视频内容的浏览空间，导致压缩后的视频内容相互碰撞和遮挡。最近的方法改变视频内容的时空关系，但导致摘要结果与原始视频的相似度降低。本文研究基于新视角合成技术的交互式摘要视频浏览算法，通过为视频内容构建新的观察视角，增加视频浏览空间，从本质上减少视频内容碰撞和遮挡的可能性。为完全避免碰撞现象，提出了特定视角下投影遮挡消除技术，仅需较小地改变视频内容的时空关系，保证摘要视频与原始视频内容的一致性。最后，为方便用户交互式浏览摘要视频，构建了观察视角自动生成算法，根据当前摘要内容实时推荐并调整到最优观察视角。实验结果和用户调查表明本文提出的交互式视频摘要算法能有效压缩视频冗余信息，提高了摘要视频准确度、易理解程度。%Most of the previous video synopsis methods compact the video under a fixed viewpoint,which limits the brows-ing space of the video.Recent method changes the spatiotemporal relationships between video objects with the side-effect of reduc-ing the accuracy of synopsized video.This paper presents a novel-view synthesis based interactive video synopsis browsing method. Our method constructs wider browsing space that can inherently reduce the probability of collision and occlusion occurrences.To completely remove collision and occlusion artifacts,we propose a projective occlusion removing algorithm for a specified viewpoint, slightly changing the relationships between video objects.Finally,to facilitate user interactive browsing the video,we further propose an automatic viewpoint suggestion method,to compute and adjust to the optimal viewpoint in real-time according to current synop-sized contents.The experiments and user study demonstrate the effectiveness of our method.