Parallel Estimation of Translation for Video Stabilization

Translation motion estimation is the major aspect of video motion compensation, especially in two-dimensional turntable. This paper presents a robust parallel estimation algorithm of translation in video stabilization system. The equivalent model of translation that derived is accorded with small rotation angle comparatively. A parallel algorithm based on grey mapping is used in local motion estimation to shorten the estimation time and improve the computation efficiency. After local motion vectors are calculated, a fast filter can detect the spurious vectors avoiding possible jerks. The suggested translation estimation algorithm has been implemented on NVIDIA GTX 260, the experimental results show that , the average time of the estimation and compensation finished in 16 ms/frame, and stabilize 1k×1k pixels sequence gray video up to 60 frames/s.