| Abstract | To keep pace with the increased popularity of digital video as an archival medium, thedevelopment of techniques for fast and e cient analysis of video streams is essential. In
particular, solutions to the problems of storing, indexing, browsing, and retrieving video
data from large multimedia databases are necessary to allow access to these collections.
Given that video is often stored e ciently in a compressed format, the costly overhead
of decompression can be reduced by analyzing the compressed representation directly. In
earlier work, we presented compressed domain parsing techniques which identi ed shots,
subshots, and scenes.
In this paper, we present e cient key frame selection, feature extraction, indexing, and
retrieval techniques that are directly applicable to MPEG compressed video. We develop
a frame type independent representation which normalizes spatial and temporal features including frame type, frame size, macroblock encoding, and motion compensation vectors.
Features for indexing are derived directly from this representation and mapped to a low-
dimensional space where they can be accessed using standard database techniques. Spatial
information is used as primary index into the database and temporal information is used
to rank retrieved clips and enhance the robustness of the system. The techniques presented
enable e cient indexing, querying, and retrieval of compressed video as demonstrated by
our system which typically takes a fraction of a second to retrieve similar video scenes from
a database, with over 95% recall.
|
