Digital Image Stabilization Based on Log-Polar Transform

In this paper we present a novel log-polar transform based digital image stabilization algorithm to estimate large global transformations among consecutive frames. In our proposed algorithm, multi- resolution log-polar transform techniques in spatial domain are introduced as an initial motion estimation module to estimate arbitrary rotations, arbitrary translations and moderate scale changes. Then, a gradient-based nonlinear least squares optimization module is used to achieve sub-pixel accuracy. The experimental results show that the proposed algorithm can stabilize 15 frames per second in 320 X 240 image sequences with large interframe motions on a 3.0-GHZ Pentium 4 machine.