3-D MOTION STRUCTURE RECOVERY FROM MOTION AND SHADING FLOWS AND STEREO DISPARITY

摘要

We address the recovery of both camera motion and scene structure given a sequence of images captured by a stereo vision system moving through a rigid environment. First, we simultaneously exploit the image motion and shading flow, a photometric cue, to develop a more robust 3-D motion estimation technique. This is implemented as a flow-based algorithm, without the need for image correspondences. The robustness is attributed to the use of shading flow in discriminating between the translational and rotational components of motion, without the need for a very wide field of view that usually comes hand in hand with severe lens distortion. Second, a novel Kalman filter formulation for structure recovery is proposed to integrate the motion and shading flows with the stereo disparity to achieve improved accuracy and robustness with respect to perturbations in any of the cues. The shading flow plays an important role by constraining the surface gradients, thus enabling a more accurate reconstruction of high-frequency details. The results of an experiment with a long sequence of underwater images demonstrate the performance of the proposed technique, including the construction of 3D maps of the entire scene by utilizing the estimated sensor trajectory from the motion cues.