Estimation of arbitrary camera motion in MPEG videos

Several algorithms have been proposed to solve the problem of camera motion estimation in digital videos. However, the distinction between translation along the x-axis (y-axis) and rotation around the y-axis (x-axis) has only rarely been considered, and no approach of this kind is known to us for the MPEG domain. In this paper, we present such an algorithm for camera motion estimation in MPEG videos. For performance reasons it is reasonable to extract motion vectors directly from the compressed stream. However, since motion vectors are optimal with respect to compression, they often do not model real motion adequately and can thus be considered as "outliers" with respect to camera motion estimation. Consequently, an outlier removal algorithm is incorporated into our approach to solve this problem. Furthermore, we have investigated the minimum number of motion vectors required to obtain satisfactory results. Comprehensive experiments with 32 video clips demonstrate the performance of the proposed approach.