Continuous Self-calibration and Ego-motion Determination of a Moving Camera by Observing a Plane

In many vision applications the depth relief of an observed scene is small compared with the extent of the image. Beside man-made environments these scenes may be approximated by a plane. Due to environmental influences camera parameters can gradually change which motivates the need for a continuous self-calibration. Based on the theory of recursive parameter estimation we present an update scheme for the parameters to handle endless video streams in real time. The geometric parametrization of the frame to frame homographies allows to incorporate information of other sensors. The application is the visual navigation of robots and unmanned aerial vehicles moving above nearly planar environments. The approach will be empirically evaluated with a synthetic data set and demonstrated with a real data set. A typical example is given by a real data set of an indoor robot observing the ground floor with one camera.