Video summarization with key frames

Video summarization is an important tool for managing and browsing video content. The increasing amount of consumer level video recording devices combined with the availability of cheap high bandwidth internet connections have enabled ordinary people to become video content producers and publishers. This has resulted in massive increase in online video content. Tools are needed for efficiently finding relevant content devoid traditional viewing. Video summaries provide a condensed view of the actual video. They are most commonly presented as static still images in the form of storyboards or dynamic video skims, which are shorter versions of the actual videos. Although methods for creating summaries with the assistance of computers have been long studied, practical implementations of the summarization methods are only a few. In this thesis, a semi-supervised workflow and a tool set for creating summaries is implemented. At first, the implemented tool creates a static storyboard summary of an input video automatically. Users are able to use the storyboard summaries to select the most important content and the selected content is then used to create a video skim. Major part of the thesis work consists of evaluating and finding the best methods to detect single key frames that would best depict the contents of a video. The evaluation process is focused mainly on motion analysis based optical flow histograms. In the experimental part, the performance of the implemented workflow is compared to state of the art automatic video summarization method. Based on the experiment results, even a rather simple method can produce good results and keeping the human in the loop for key frame selection is beneficial for generating meaningful video summaries