Robust estimation of camera parameters from image sequence for video composition

In video composition systems, viewing points and image planes of multiple images to be composed have to be properly integrated for natural composed images, especially when some of the images undergo camera operations. Camera parameters are needed for the integration. In this paper, we propose a robust method for the estimation of camera parameters from image sequence. We first establish correspondence of feature points between consecutive image fields. After the establishment, we formulate a nonlinear least-square data fitting problem. When the image sequence contains moving objects, and/or when the correspondence establishment is not successful for some feature points, we get bad observations, outliers. They should be eliminated properly for a good estimation. Therefore, we propose an iterative algorithm for rejecting the outliers and fitting the camera parameters in turn. We attempt to decrease both the number of outliers and the energy of residual errors, given finite observations. Thus, we formulate a budget problem, outlier rejection formula. Then we construct a simple scheme for solving the problem. Finally, we show the validity of the proposed method using computer generated data sets and real image sequences.