Abstract
This dissertation presents techniques for generating static representations of video content,which can be used for video compression, video generation, and for allowing users to more
efficiently locate information in videos and video databases. These static representations
create a correspondence between large video files and concise image representations of their
content. In this work I focus on two domains of video, blackboard lecture videos and videos
containing background content which exhibits complex dynamic motion. In both instances,
there are regularities in the videos which are inherent to the video category, and which can
be exploited to create efficient static representations which succinctly describe the content
of the video.
For blackboard lecture videos, videos are processed to remove the speaker from the fore-
ground by detecting regions of the frame where the content is rapidly changing. Once the
background video feed has been generated, it is further processed to locate frames before the
information is erased from the board, and reduced to a series of slides.
In the case of videos containing dynamic background motion, these motions can be de-
tected by processing pairs of adjacent frames with optical flow techniques. Using frequency
domain analysis, these motions are summarized into static oscillation modes, which can then
be used to reanimate still frames from the video in a manner that captures the oscillatory
quality of the motion in the original video.