Adaptive video stabilisation with dominant motion layer estimation for home video and TV broadcast

Digital video stabilisation is useful to improve the viewing experience of shaky videos and eases further processing such as segmentation, encoding and restoration. In this paper we present an adaptive, low-latency video stabilisation technique targeting real-time applications for home video and broadcast content viewed on a mobile platform or TV. We stabilise videos by shifting each frame inversely to the measured 2D translational jitter which is computed coherently along the estimated dominant motion layer. This allows our technique to stabilise a wide variety of sequences, including the difficult case of zooming. Our smoothing filter preserves intentional motion and keeps a low latency throughout the processing. We propose an extension of our technique to remove rolling shutter artifacts and include a logo detection process to stabilise TV footage (e.g. news, sports) while preserving static areas. We compare our approach to existing methods and show its potential on several sequences captured casually with smartphone cameras and TV content.