Efficient Algorithms for Video Segmentation

Abstract

Describing video content without watching the entire video is a challenging matter. Textual descriptions are usually inaccurate and ambiguous, and if the amount of video is large this manual task is almost endless. If the textual description is replaced with pictures from the video, this is a much more adequate method. The main challenge will then involve which pictures to pick to make sure the entire video content is covered by the description. TV stations with an appurtenant video archive would prefer to have an effective and automated method to perform this task, with focus on holding the time consumption to an absolute minimum and simultaneously get the output results as precise as possible compared with the actual video content. In this thesis, three different methods for automatic shot detection in video files have been designed and tested. The goal was to build a picture storyline from input video files, where this storyline contained one picture from each shot in the video. This task should be done in a minimum of time. Since video files actually are one long series of consecutive pictures, various image properties have been used to detect the video shots. The final evaluation has been done based both on output quality and overall time consumption. The test results show that the best method detected video shots with an average accuracy of approximately 90%, and with an overall time consumption of 8.8% of the actual video length. Combined with some additional functionality, these results may be further improved. With the solutions designed and implemented in this thesis, it is possible to detect shots in any video file, and create a picture storyline to describe the video content. Possible areas of application are TV stations and private individuals that have a need to describe a collection of video files in an effective and automated way.