Dual-mode video stabilization based on adaptive motion clustering

Many target tracking videos suffer from significant shake due to the undesirable camera motion or target vibration. These degraded video records have a negative impact on viewing experience as well as the follow-up analysis and application. Most existing digital video stabilization methods aim at removing camera shake using the techniques of global motion estimation and compensation between adjacent frames, while paying little attention to motion pattern and the relationship between the foreground and background contents. In this paper, we present a novel approach to stabilize the videos containing large dynamic object, which can realize the normal stabilization task and describe the target motion path at the meantime. We combine K-means and Gaussian mixture model to distinguish the motion modes of background and foreground without applying any prior knowledge about the moving object. The proposed approach gives a flexible choice in stabilizing background motion as traditional algorithms or foreground motion for salient target surveillance.