An Event-driven Approach to Efficient Summarization, Visualization, and Browsing of Massive Combat Video Database

We propose a systematic solution to the problem of efficient summarization, visualization, and browsing of large combat video archives. Our approach is event-driven: spatio-temporal saliency is used to define significant visual events, which is in turn used in an event-based video model to partition the video into interleaved salient and less significant segments. Each of the two types of segments will then be summarized properly, with the salient segments been kept at higher spatio-temporal resolution while less significant segments being highly condensed. Such an approach naturally leads to a visualization and browsing scheme that can facilitate the consumption of the large amount of video data based on the relative importance of the raw video feeds, thus helping human analysts achieve efficient examination of the video without the drawbacks of conventional techniques that would discard some of the input data by thresholding or filtering. The innovation of the proposed approach include: 1) a unique method to define the significant segments (“events”) of a video based on saliency detection that can deal with low resolution videos; 2) high performance summarization schemes for different portions of the video based on their relative importance; 3) a novel visualization scheme to support event-driven non-linear browsing of the video feeds.