Comparative Analysis of Video Stabilization using SIFT Flow and Optical Flow

Abstract

In any domain of computer vision it is good to have a high quality stabilized video but in many practical applications such as UAVs and Mobile Videos, it is difficult to get such videos. To work in these scenarios, it is best to stabilize the video for further processing. Video stabilization is the post processing method to eliminate the undesirable camera motion in a video sequence. In this paper, we are proposing a useful strategy to eliminate the accidental jitter and produce the high quality stabilized video. In this context we are using Scale Invariant Feature transform (SIFT) Flow to calculate the inter- frame motion. For better estimation of model between two frames, we use Random Sample Consensus (RANSAC) to reject the outlier and only use inlier for model estimation. We use projective transformation to estimate the model between two frames. In the end, we compare our results with conventional optical flow based video stabilization. It is demonstrated that our method effectively produce better stabilized video. The effectiveness of proposed frame work has been verified over a widely variety of videos.