Video Enhancement: Video Stabilization

Abstract

Video stabilization is required in digital video cameras as most of them are hand held or mounted on moving platforms (e.g. cars) undergo atmospheric vibrations and also the videos taken by the camera with CMOS sensors undergo rapid vibrations because of rolling shutter which scans an image horizontally or vertically instead of recording the whole image of the scene at same instant. And also the moving cameras cause geometric distortion of the objects in the video because of parallax. A method which handles parallax and rolling shutter effects along with removing the unwanted motion of frames caused by the movement of camera is proposed by modelling the camera motion with bundle of camera paths using bundled camera paths.
Bundled camera paths motion model involves dividing the video frames into mesh grid cells, extracting matched features for every two consecutive frames. Having these matched features and grid cell vertices, new vertices are obtained using Bilinear interpolation resulting in image warping of each frame in accordance with the camera motion. Homographies which represent the camera motion for each grid cell are obtained using original vertices and new vertices. For each grid cell, homographies of corresponding grid cell of all the frames of the video are multiplied to get camera path of each grid cell representing the bundle of camera paths. These camera paths are smoothed adaptively. Adaptive smoothing handles cropping. Smoothed homographies are applied to original video to get the stabilized video.