Video Stabilization for Sea Scenes via Low-Rank Alignment and effective Visual Cues

Abstract

Video stabilization(VS), as a fundamental video pre-processing task, benefits not only for human observers to have a good look at the scene, but also for many subsequent vision tasks to be properly tackled. However, numerous methods for video stabilization rely heavily on extracting enough features for matching. This requirement is difficult to meet in some adverse conditions, say sea-surface scene with dynamic and repetitive texture. In this paper, we propose a video stabilization method for sea scenes to deal with the above issue. Our method utilizes two visual cues, the sea-sky line and the Maximum Stable Extremal Region (MSER), to identify the sky and sea areas and use MSER features within the sea area to predict the camera path to obtain a coarse stabilization result. In further, a low-rank alignment model (LRA) is further employed to significantly boost the accuracy of the preceding stabilization result by enforcing a temporal low-rank restriction on image data. We also construct a sea-surface VS dataset for validation and deeper investigations. Experiments on the constructed dataset show that the proposed method outperforms mainstreaming VS methods under sea-surface scenes. Also, additional experiments on the public datasets manifest that our method can stabilize commonly natural scene videos and has a comparable performance with some conventional approaches.