Relevance Ranking of Video Data using Hidden Markov Model Distances and Polygon Simplification

A video can be mapped into a multidimensional signal in a non-Euclidean space, in a way that translates the more predictable passages of the video into linear sections of the signal. These linear sections can be filtered out by techniques similar to those used for simplifying planar curves. Different degrees of simplification can be selected. We have refined such a technique so that it can make use of probabilistic distances between statistical image models of the video frames. These models are obtained by applying hidden Markov model techniques to random walks across the images. Using our techniques, a viewer can browse a video at the level of summarization that suits his patience level. Applications include the creation of a smart fast-forward function for digital VCRs, and the automatic creation of short summaries that can be used as previews before videos are downloaded from the Web.