Motion and structure from long stereo image sequences

Abstract

The treatment of uncertainty in 3D points determined by stereo triangulation affects significantly the reliability of the estimated 3D motion and structure. The authors first address this issue in the estimation of inter-frame motion. Two methods are discussed, a closed-form matrix-weighted least-squares solution and an iterative optimal solution. Using each of these two methods for two-view analyses, the authors proceed to deal with long image sequences. A recursive-batch approach is adopted to fuse multiple stereo views in order to achieve higher performance without suffering from excessive computational cost. Experiments with a real stereo image sequence have been presented to show the performance of the algorithms.<>