Reliable Extraction of the Camera Motion

An accurate estimate of the epipole (direction of camera translation) is necessary if image motion is to be decomposed into ro-tational and translational components, which give the camera rotation and feature depths respectively. In this paper we introduce the Linearised Subspace Method to nd direct constraints on the epipole which are independent of camera rotation and scene structure. We present methods to compute reliable constraints and their uncertainties from image motion. We show how erroneous constraints due to errors in tracking can be rejected and how the valid constraints should be combined to form accurate estimates of the direction of translation. Experimental results show these methods lead to improvements in the recovery of camera motion and that the uncertainty estimates are accurate and useful in detecting degenerate scene structure or camera motions.