Structure and motion from road-driving stereo sequences

Abstract

We introduce a unified framework for scene structure and motion estimation on road-driving stereo sequences. This framework is based on the slanted-plane scene model that has become widely popular in the stereo vision community. Our algorithm iteratively and alternately solves for scene structure and motion. Surface estimation is done using our own slanted-plane stereo algorithm. Motion estimation is achieved by solving a MRF labeling problem using Loopy Belief Propagation. We show that with some specific assumptions about the motion of the camera and the scene, the motion estimation problem can be reduced to a 1D search problem along the epipolar lines. We also propose a novel evaluation metrics, based on the notion of view prediction error. This metrics can be used to evaluate the performance of structure and motion estimation algorithms on stereo sequences without ground truth data. Experimental results on road-driving stereo sequences demonstrate that our algorithm successfully improves the view prediction error although it was not designed to directly optimize this quantity.