Digital video stabilization

Abstract

The development of multimedia equipments has allowed a significant growth in the production of videos through professional and amateur cameras, smartphones and other mobile devices. However, videos captured by these devices are subject to unwanted vibrations due to camera shaking. To overcome such problem, digital stabilization aims to remove undesired motion from videos through software techniques, without the use of specific hardware, to enhance visual quality either with the intention of enhancing human perception or improving final applications, such as detection and tracking of objects. The two-dimensional digital video stabilization process is usually divided into three main steps: camera motion estimation, removal of unwanted motion, and generation of the corrected video. In this work, we investigate and evaluate digital video stabilization methods for correcting disturbances and instabilities that occur during the process of video acquisition. In the motion estimation step, we develop and analyzed a consensual method for combining a set of local feature techniques for global motion estimation. We also introduce and test a novel approach that identifies failures in the global motion estimation of the camera through optimization and computes a new estimate of the corrected motion. In the removal of unwanted motion step, we propose and evaluate a novel approach to video stabilization based on an adaptive Gaussian filter to smooth the camera path. Due to the incoherence of assessment measures available in the literature regarding human perception, two novel representations are proposed for qualitative evaluation of video stabilization methods: the first is based on the visual rhythms and represents the behavior of the video motion, whereas the second is based on the motion energy image and represents the amount of motion present in the video. Experiments are conducted on three video databases. The first consists of eleven videos available from the GaTech VideoStab database, and three other videos collected separately. The second, proposed by Liu et al., consists of 139 videos divided into different categories. Finally, we propose a database that is complementary to the others, composed from four videos collected separately, which are excerpts from the original videos with moving objects in the foreground and with little representative background extracted, resulting in eight final videos. Experimental results demonstrated the effectiveness of the visual representations as qualitative measure for evaluating video stability, as well as the combination method over individual local feature approaches. The proposed method based on optimization was able to detect and correct the motion estimation failures, achieving considerably superior results compared to when this correction is not applied. The adaptive Gaussian filter allowed to generate videos with adequate trade-off between stabilization rate and amount of frame pixels. The results reached with our optimization method for the videos of the proposed database w...