Content-based Retrieval System for Video Databases, Based on Event-Activity Model

针对视频数据库的构建问题，提出了事件-活动模型，将事件定义为从帧间差异中观察到的对象特征，将活动表示为连续有序的事件序列。此外，我们实现了原型系统的视频数据库管理，基于这样一个模型。在许多传统的方法中，检索设备依赖于具有多个帧的场景的索引，并且仅识别预设场景。另一方面，在我们的方法中的索引被分配给个别的特征帧作为事件和检索设施可以选择各种场景匹配的事件序列自适应。也就是说，检索设备将检索条件视为事件序列的模式的活动，以及与来自视频帧的事件模式一致的事件序列。支持检索功能的自动机基本上按照输入事件的顺序将其自身的内部状态转换为可接受的状态，并有效地识别与活动相对应的场景。我们的研究课题计划在2000年之前的3年期间进行，并为管理视频数据库贡献了以下机制和功能：-通过事件序列进行场景检索;- 由自动机机制控制的基于内容的检索;-对视频帧的自动索引;-基于事件/对象之间的语义角色的面向用户的检索;-相似场景检索;-针对主题场景的检索表的自动合成;-用于有效浏览的超视频。

Concerning the subject about the construction of video databases, we developed the event-activity model.The events are defined as the features of objects observed from the difference among frames and also the activities are composed as a sequence of continuously ordered events. Additionally, we implemented the prototyping system STRIKE for managing video databases, based on such a model. In many of the traditional approaches, the retrieval facilities are dependent on the indexes for scenes with several frames, and identify only the preset scenes. On the other hand, the indexes in our approach are assigned to individual characteristic frames as events and the retrieval facilities can select various scenes which match with a sequence of events adaptively. Namely, the retrieval facility looks upon the retrieval conditions as activities for patterns of a sequence of events, and a sequence of events to be consistent with the event pattern from video frames. The automaton which supports the retrieval facility basically transfers its own internal state to the acceptable state in accordance with the order of input events, and identifies scenes, corresponded by activities effectively.Our research subject was scheduled during 3 years until 2000, and contributed the following mechanism and functions for managing video databases :- scene retrieval through a sequence of events ;- content-based retrieval controlled by automaton mechanism ;- automatic indexing to video frames ;- user-oriented retrieval, based on semantic roles among events/objects ;- similar scene retrieval ;- automatic composition of digests for topic scenes ;- hyper-video for effective browsing.