Content-and-disparity-aware stereoscopic video stabilization

Abstract

Filming stereoscopic videos has become easier with the development of science and technology, and such videos now proliferate on the Internet. Meanwhile, video stabilization is an important research topic. Thus, this study presents a method of stabilizing stereoscopic videos with preserving the disparities between objects in the frames. First, the feature points must be tracked and separated into many groups. We posit that the shaky motion is caused not only by translations but also by rotations. Thus, directly smoothing the path will not produce a similar trajectory so that we solve the shakiness of the turning before smoothing the path. To address such shakiness, we initially estimate the rotation angles between two adjacent frames. By determining the angle changes of all the frames, we can find out the preference of rotation in a video. Furthermore, the inconsistent angular velocity can be alleviated and the shakiness of the turning is solved by rotating the frame appropriately. Then, the Bezier curve is utilized to smooth the trajectories. We split a trajectory into a set of subtrajectories and subsequently smooth the latter independently. Unlike previous researches, we split the trajectory according to the feature tracking rate to obtain similar trajectories in the original video path. After making subtrajectories smooth, we merge them to attain a smoothed trajectory. The joint of the two subtrajectories is replaced by their interpolation. Finally, we optimize the smoothness and context preservation to stabilize videos without requiring extensive clipping.