Increasing use of multimedia data makes it crucial to develop intelligent search mec:hanisms for retrieving multimedia data by content. Traditional text-based methods clearly do not suffice to describe the rich content of images, voice or video. Digital vidseo requires the incorporation of temporal information for any effective contentbased retrieval scheme. We present a novel technique which integrates object motion ancl temporal relationship information in order to characterize the events for subsequent search for "similar" clips. We propose a hybrid mechanism based on object motion trails similarity match and interval-based temporal modeling that leads to a unique framework for spatio-temporal content based access in digital video. We implemented the proposed methods and demonstrated that high-level query formulation can be achieved for the aforementioned purpose. Development of such technology will enable true multimedia search engines that will accomplish what current Internet search engines like Infoseek or Excite do today for textual data.

Date of this Version

September 1998