We propose a novel architecture for a video database system incorporating both spatio-temporal and semantic (keyword, event/activity and category-based) query facilities. The originality of our approach stems from the fact that we intend to provide full support for spatio-temporal, relative object-motion and similarity-based object-trajectory queries by a rule-based system utilizing a knowledge-base while using an object-relational database to answer semantic-based queries. Our method of extracting and modeling spatio-temporal relations is also a unique one such that we segment video clips into shots using spatial relationships between objects in video frames rather than applying a traditional scene detection algorithm. The technique we use is simple, yet novel and powerful in terms of effectiveness and user query satisfaction: video clips are segmented into shots whenever the current set of relations between objects changes and the video frames, where these changes occur, are chosen as keyframes. The directional, topological and third-dimension relations used for shots are those of the keyframes selected to represent the shots and this information is kept, along with frame numbers of the keyframes, in a knowledge-base as Prolog facts. The system has a comprehensive set of inference rules to reduce the number of facts stored in the knowledge-base because a considerable number of facts, which otherwise would have to be stored explicitly, can be derived by rules with some extra effort. (C) 2002 Elsevier Science Inc. All rights reserved. [References: 18]
展开▼
