We present a scheme for simultaneous calibration for computing the 3-D position and focal length of a continuously moving and continuously zooming camera at each frame. This is motivated by virtual studio applications, in which zooming changes from frame to frame. Hence, we cannot pre-calibrate the camera in a controlled environment inadvance. The object we are observing also moves freely independently of the camera motion. Hence, we cannot self-calibrate the camera from the object images themselves. We resolve these difficulties by placing an easily distinguishable planar pattern in the scene: we detect an unoccluded portion of the pattern image in each frame, compute the position and focal length of the camera from it, and remove the pattern image by segmentation. Particular emphasis is placed on the following two aspects: Introducing a statistical model of image noise, we define a procedure for computing an optimal solution that attains the Cramer-Rao lower bound (CRLB) in the presence of noise. As a result, we can evaluate the reliability of the solution in quantitative term by simply computing the CRLB; no complicated step-wise covariance propagation analysis is necessary. When the camera optical axis is perpendicular to the pattern, the 3-D position and focal length of the camera are indeterminate. Also, the computed solution randomly fluctuates due to noise even when the camera motion is smooth or stationary. We avoid such degeneracy and fluctuations by model selection: we predict the 3-D position and focal length of the camera in multiple ways and select the best model by using the geometric AIC and the geometric MDL. The geometric MDL we use is a new concept distinct from the traditional MDL used in statistics and some vision applications in the past. We show the effectiveness of our method by simulations and real image experiments. Our technique can be applied to a variety of media applications, including virtual studio and 3-D analysis of sports broadcasting, for optimal 3-D computation and its stabilization.
© Copyright 2000 Computer Vision - ECCv 2000. 6th European Conference on Computer Vision Dublin, Ireland, June 26-July 1, 2000 Proceedings, Part II. Veröffentlicht von Springer. Alle Rechte vorbehalten.