A number of important video domains like surgical
training, assembly manuals, surveillance, etc. require
highlevel access to the data. A video database
which could support these types of a query requires not
just "play" and "fast-forward" functions, but also semantic
video indexing.
Providing such concept level access to video data requires
video management systems tailored to the domain of
the data. That is, effective indexing and retrieval for highlevel
access mandates the understanding and use of significant amounts of
domain knowledge.
As an example, consider the following scenario. While surfing the web for basket ball related information a user hits the NBA site maintained by ESPN. Being a Miami Heat fan, the user looks at the statistics of different players. Impressed by Hardaway's scores, the user would like to access all game videos where:
Such a problem can be solved by examining a number of features that can be automatically extracted from a video. To do this there also needs to be a set of rules for interpreting these low level features along with a temporal model of how the game progresses to provide sufficient additional constraints.
Looking at the tables above, consider the segment of the game between minutes 1:30 and 1:40. There was a shot scored by the Hardaway of the Miami Heat. A possession change is indicated on the possession track (top panel). The event track shows that a shot scored event occurred and the player track indicates that it was scored by Hardaway. This conjunction satisfies the first part of the query which asks for segments in which Hardaway makes a shot. Now the speech track (left), shows the commentator (Mike) saying: "Hardaway off the dribble. Tim Hardaway creating his own shot, knocks it down, Hardaway now five of eight from the field". From this, the system can satisfy the restriction that the shot occurred "off a dribble". Finally, the shot table (right - shot numbers 10 and 11) show clips of Hardaway scoring the shot and a close-up of Hardaway. This would therefore be an appropriate portion of the video to show to the user in response to his query. As can be seen, a wide range of information needs to be integrated, and a fair amount of reasoning needs to be done, in order to answer such high-level queries. |
| Contact: Arun Hampapur | Last updated: 6/7/02 | ||
 |
 |
 |
 |