Video stabilization by estimation of similarity transformation from integral projections

The widespread use of hand held devices having video recording capabilities has made video capture prevalent. However, unwanted hand motion during capture introduces jerks which hampers the video viewing experience. In this paper, a robust method for global motion estimation between frames is presented. The algorithm estimates the rotation, scale and translation parameters between two adjacent frames. All the parameters of the similarity transformation are estimated by aligning the integral projections computed from the images. A novel framework is presented which uses angular integral projections to estimate rotation, scale and translation between adjacent frames. The curve warping technique, Derivative dynamic time warping (DDTW) is used for aligning the projection curves. Experimental results demonstrate the robustness of the proposed approach in stabilizing videos, including those captured under challenging conditions.