A new composite multi-constrained differential-radon warping approach for digital video affine motion stabilization

- Tilt estimation using correlation of multiple differential-Radon angular slices.
- Novel multi-constraint warping is proposed for error prone boundary-vector removal.
- Warping singularities are removed using interest vectors based on optima matching.
- Slope of extracted interest vector is utilized for inter-frame zoom estimation.
- A new affine video stabilizer framework based on derivative-projection is proposed.

This paper presents a new projection based affine motion stabilizer framework for video stabilization using differential-Radon (DRadon) curve warping. Extending the translational domain of classical projection based algorithms towards affine stabilization, multiple angular curves obtained with Radon or rotated images have recently been explored for combined rotation and zoom estimation. Radon provides efficient projection extraction, but use of integral intensity under local variation degrades the desired motion accuracy. DRadon works on derivative of each angular slice to incorporate shape based matching for better projection alignment. Based on human perception of inter-frame tilt in unsteady camera recordings, the proposed angular DRadon curve estimation is confined to angular search space of [−20ο, 20ο] with the angular increment of 0.1ο. Out of complete set of angular DRadon curves of reference frame, five key-angular slices are selected and correlated with their corresponding neighbourhood in target DRadon for inter-frame tilt estimation. Best matched key slices of reference and target DRadon are warped using a novel multi-constrained approach and the extracted warping vectors are further processed for translation and zoom estimation. A vector-slope algorithm based on relative stretching/contraction between the DRadon-projections is used for camera zoom estimation. Combining the estimated motion parameters, an affine transformation is developed for inter-frame stabilization. Comparative performance using motion accuracy and frame stability is evaluated over different categories of real-world videos.