[Japanese | Thesis | Researches in Minoh Lab | Minoh Lab]
Video images taken by automatic filming system with multiple cameras are required to be sophisticated. There are three criteria to evaluate video images, one is based on visual psychology, another is framing of a shot and the other is shot composition of adjacent shots. These criteria form a hierarchy that visual psychology criterion is the lowest and shot composition is the highest. Lower criterion should be fulfilled to satisfy higher criterion.
In this study, we compose a video sequence with fixed camera shot segments to eliminate unstable background motion in video image, which depress the evaluation of the lowest criterion. Since pan-tilt-zoom cameras except for the one which is selected as an output are allowed to change their direction and zoom, it is important to control cameras before they are selected.
We propose a method in which filming rules are made automatically before online filming process is conducted. Since the best strategy to control cameras depends on target motion, we make use of motion history of each target object that is to be shot. By using a graph that is generated based on a motion history, we make filming rules that guarantee cameras are controlled before they are selected and that fulfill all criteria as much as possible. Our experiment shows the validity of our method.