The Research On Rotational Motion Estimation Algorithm In Electronic Image Stabilization Systems

Electronic image stabilization is a new method to stabilize image sequences. It has high stabilizing precision, lightweight and cheap price, by contrast to traditional methods. Motion estimation and motion compensation techniques are used in electronic image stabilization systems to stabilize the image sequences. It is translation motion estimation algorithms that used in traditional electronic image stabilization systems. These algorithms are useful when there is only translation motion, but helpless to estimate the rotational motion. So, the rotational motion estimation algorithm with translations of local blocks is investigated in the thesis. This algorithm is able to estimate the central rotation.First the dissertation analyses the elements of rotational motion estimation algorithm using least squares method, and improves the full arranged method, called cross arranged method. The computation of cross arranged method is third of the computation of full arranged method, because of the decrease of local blocks. It is indicated from the experiment that cross arranged method approaches full arranged method finely, when the image is abundant of details.Then, for overcoming the disadvantage that the least squares method is complex and with big computation, a method of geometry is used to estimate the central rotational motion. First an alterable length searching method is used to estimate the translation of local matching blocks which are centrosymmetric, and then the translations of local matching blocks are put into affine model to calculate the rotation angle. From the data of experiment, it is concluded that the precision of geometry method is similar to the least squares method.To prevent the impact of motion of wisp in local, an average selection method is used. This method is good for reducing the error of rotational motion estimation. It is proved by the result of video stabilization experiment that the selection of accurate local translation estimations is useful and necessary.At last, through the simulation of local translation estimation algorithm and angel calculation algorithm in Quartus II, the thesis gets the result that local translation estimation needs 5 milliseconds. If the local translation estimation goes side-by-side, it is possible to estimate the rotation motion in real-time.