Fast digital image stabilization by motion vector prediction

In this paper, a fast algorithm for digital image stabilization is proposed that use standard division properties and motion prediction to reduce the computational cost of block matching algorithm (BMA) for motion estimation. In most cases global motions have the repeating behavior, such as car, ship, and etc. In proposed method previous global motion vectors as saved (all previous vectors before scene change) therefore it is possible to predict many motion vectors in the next frames. Proposed methods also has good performance for non periodic motions because if predicted vector has not good matching, this method can predict search direction (relative of the center point) and searching is continue in suitable direction to calculating the minimum sum of absolute difference (SAD) for other points and best matching is occurred faster than traditional search. Many calculations for motion estimation are eliminated also by mixing proposed method and partial distortion elimination (PDE), it is possible to estimate motion vectors very fast.