A New Method for Camera Motion Estimation in Video

Camera motion estimation is important for video semantic analysis and video classification. Camera motion can be classified into several classes: camera rotation, camera translation, and camera zoom, but ordinarily we consider only camera translation and camera zoom. Traditional approaches for camera motion parameters estimation is features-based, and they are dependent on the image quality and image features' characteristic. Traditional approaches are not valid in the low-density texture images and deformed images. In this paper we consider image's algebraic character, and introduce a robust approach for video camera motion parameters estimation. Scale parameter is extracted based on edge images' singular value decomposition, and the translation parameters are extracted based on their phase difference of cross-power spectral in frequency domain. Experiments show that this method is simple and efficient, and it is independent of image features and robust in most cases.