LIP6 2000/030
- Thesis
Classification de séquences audiovisuelles - J. Carrive
- 187 pages - 09/12/2000- document en - http://www.lip6.fr/lip6/reports/2000/lip6.2000.030.ps.gz - 21,489 Ko
- Contact : Jean.Carrive (at) nulllip6.fr
- Ancien Thème : OASIS
- Keywords : audiovisual indexing, program collection, template, sequence, description logics, temporal reasoning, constraint satisfaction
- Publisher : Valerie.Mangin (at) nulllip6.fr
In the context of automatic indexing of audio-visual documents, we are interested in collections of programs that respect a common general structure, such as broadcast news or some news or variety programs. We aim at representing and recognizing in scuch programs some characteristic sequences which can be annotated as a whole by a documentalist, such as a song in a variety program or a report in a broadcast news. We propose a language that represent such classes of sequences as templates. The definition of a template relies on the one hand on a vocabulary of directly observable primitive events, such as logos or jingles, and on the other hand on a set of temporal relations between these events. We propose to represent primitive events in a description logic formalism and we extend Allen's temporal interval algebra with an iteration operator « * » denoting a sequence of temporally adjacent events, and with specific constraints involving a negation. We present the template recognition problem as a constraint satisfaction problem and propose an efficient solving procedure, as well as an implementation of this procedure in an existing constraint satisfaction framework. We finally expose a set of experiments conducted at the French National Institute of Audiovisual (INA) on a corpus of broadcast news.