Ego-Motion Estimation Using Rectified Stereo and Bilateral Transfer Function

We describe an ego-motion algorithm based on dense spatio-temporal correspondences, using semi-global stereo matching (SGM) and bilateral image warping in time. The main contribution is an improvement in accuracy and robustness of such techniques, by taking care of speed and numerical stability, while employing twice the structure and data for the motion estimation task, in a symmetric way. In our approach we keep the tasks of structure and motion estimation separated, respectively solved by the SGM and by our pose estimation algorithm. Concerning the latter, we show the benefits introduced by our rectified, bilateral formulation, that provides at the same time more robustness to noise and disparity errors, at the price of a moderate increase in computational complexity, further reduced by an improved Gauss-Newton descent.