Video stabilization: A comprehensive survey

Abstract

Video stabilization is one of the most fundamental and challenging tasks in video processing, which can be widely applied in many areas, such as video surveillance, robotics, unmanned aerial vehicles and smartphones. It is a video enhancement technology that aims to improve original video quality by removing potentially shaky camera motion. Generally, conventional methods can be divided into three types: 2D, 3D and 2.5D models, which mainly have boundedness. In the past decade, deep learning has emerged as a powerful technique for learning feature representations directly from data, leading to significant progress in video stabilization. However, previous surveys mainly focus on conventional methods and lack performance comparison. In this paper, we present a comprehensive survey of video stabilization. Firstly, we briefly review three stages of video stabilization. Then, we deliver the conventional methods and the deep learning-based methods in detail. Furthermore, we pay special attention to video stabilization quality assessment, benchmark datasets, and state-of-the-art performance. Finally, we provide discussions on current challenges and future directions to overcome the limitations of the existing methods and better meet the needs of researchers in this active area.