Video Stabilization Based on Multi-scale Local Color Invariants

Abstract

Feature extraction and matching is the key process of motion estimation, and determines the performance of video stabilization to a great extent. A novel approach of video stabilization was proposed based on multi-scale colored local invariant features. The proposed approach transformed the image from RGB color model to color invariant model, and built up multi-scale color invariant space based on Gaussian pyramids, then extracted FAST feature points in the multiscale space and matched the feature points by building Fast Retina Key-point (FREAK) descriptors, finally estimated interframe motions in the video by M-estimator Sample Consensus (MSAC) algorithm, and processed image compensation and smoothing. Experiments demonstrated that the approach was efficient and more robust than general methods especial in harsh imaging conditions.