Translation and rotation invariant video stabilization for real time applications

Abstract

Use of camera has increased among professionals and nonprofessionals in recent years and videos are being captured widely and wildly for information, knowledge, surveillance, adventures and memories. So these videos are highly vulnerable to suffer from translational and rotational noises. These noise are caused by multiple factors and it is difficult to remove all those causes. So digital video stabilization is a process of acquiring and minimizing/removing the undesired motion from the video. In this paper we have presented a method which utilizes existing algorithms and techniques in a novel fashion for digital video stabilization. The quality of feature extraction is improved by using Speeded Up Robust Features (SURF) and the process for the selection of extracted features, for global motion acquisition, is also refined. Actual motion of the camera and the undesired motion are separated by applying the moving average filter. Finally, stable frames are obtained through affine transformation to produce an out of phase motion. We have also presented a way to use interpolation for improving the quality of video stabilization. Our system has been successfully tested on various videos including VIRAT dataset, disaster videos, rush hour videos, mountain cycling, street walking, TV reports etc.