Moving camera video stabilization using homography consistency

Abstract

Videos recorded on moving cameras are often known to be shaky due to unstable carrier motion and the video stabilization problem involves inferring the intended smooth motion to keep and the unintended shaky motion to remove. However, conventional methods typically require proper, scenario-specific parameter setting, which does not generalize well across different scenarios. Moreover, we observe that a stable video should satisfy two conditions: a smooth trajectory and consistent inter-frame transition. While conventional methods only target at the former condition, we address these two issues at the same time. In this paper, we propose a homography consistency based algorithm to directly extract the optimal smooth trajectory and evenly distribute the inter-frame transition. By optimizing in the homography domain, our method does not need further matrix decomposition and parameter adjustment, automatically adapting to all possible types of motion (eg. translational or rotational) and video properties (eg. frame rates). We test our algorithm on translational videos recorded from a car and rotational videos from a hovering aerial vehicle, both of high and low frame rates. Results show our method widely applicable to different scenarios without any need of additional parameter adjustment.