3D reconstruction from uncalibrated-camera optical flow and its reliability evaluation

We present a scheme for reconstructing a 3D structure from optical flow observed by a camera with an unknown focal length in a statistically optimal way as well as evaluating the reliability of the computed shape. First, the flow fundamental matrices are optimally computed from the observed flow. They are then decomposed into the focal length, its rate of change, and the motion parameters. Next, the flow is optimally corrected so that it satisfies the epipolar equation exactly. Finally, the 3D positions are computed, and their covariance matrices are evaluated. By simulations and real-image experiments, we test the performance of our system and observe how the normalization (gauge) for removing indeterminacy affects the description of uncertainty. © 2002 Wiley Periodicals, Inc. Syst Comp Jpn, 33(9): 1–10, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.10133