合理地组织视频数据对于基于内容的视频分析和检索有着重要的意义。提出了一种基于运动注意力模型的电影视频摘要生成方法。首先给出了一种基于滑动镜头窗的聚类算法将相似的镜头组织成为镜头类；然后根据电影视频场景内容的发展模式，在定义两个镜头类的3种时序关系的基础上，提出了一种基于镜头类之间的时空约束关系的场景检测方法；最后利用运动注意力模型选择场景中的重要镜头和代表帧，由选择的代表帧集合和重要镜头的关键帧集合建立层次视频摘要（场景级和镜头级）。该方法较全面地涵盖了视频内容，又突出了视频中的重要内容，能够很好地应用于电影视频的快速浏览和检索。

It is important to properly organize the unstructured video data for content-based video analysis and retrieval. In this paper, we propose a unified approach for film video summarization based on the analysis of video structure and motion attention model. Video shots are firstly grouped into shot clusters. Afterwards, according to the characterization of film video scene, a temporally and spatially integrated strategy is presented to parse shot clusters into semantic scenes in terms of the definition of temporal relationships between two shot clusters. Finally, representative frames and highlight shots are selected from scenes by using motion attention model. The scheme offers an efficient mean for browsing and effectively retrieving film video.