Dynamic Mode Decomposition Based Video Shot Detection

Abstract

Shot detection is widely used in video semantic analysis, video scene segmentation, and video retrieval. However, this is still a challenging task, due to the weak boundary and a sudden change in brightness or foreground objects. In this paper, we propose a new framework based on dynamic mode decomposition (DMD) for shot boundary detection. Because the DMD can extract several temporal foreground modes and one temporal background mode from video data, shot boundaries can be detected when the amplitude changes sharply. Here, the amplitude is the DMD coefficient to restore the video. The main idea behind the shot boundaries detection is finding the amplitude change of background mode. We can reduce error detection when the illumination changes sharply or the foreground object (or camera) moves very quickly. At the same time, our algorithm has a high detection accuracy, even the color changes are not obvious, the illumination changes slowly, or the foreground objects overlap. Meanwhile, a color space for DMD is selected for reducing false detection. Finally, the effectiveness of our method will be demonstrated through detecting the shot boundaries of the various content types of videos.