Stereoscopic scene flow estimation with global motion prior

Abstract

Scene flow estimation jointly recovers dense scene structure and motion from at least two pairs of stereo images, thus generalizing classical disparity and optical flow estimation. Such a complete description of the scene has many uses in the field of automated driving such as dynamic traffic object detection or infrastructure element detection. Estimation of the structure and motion of each scene element is a difficult problem because of the large number of unknowns that need to be assessed. In order to increase the accuracy and the robustness of the estimation, we propose to extend the piecewise rigid scene model used in modern state of the art scene flow algorithms with a global motion prior that presumes that a large number of objects in the scene are static. For obtaining the scene flow result, we proposed a two-step iterative approach: A Nelder-Mead nonlinear minimization accompanied by a spatial propagation of current best estimation to neighboring image regions.