Modeling 3D Quantization Error For MotionEstimation

Abstract

Spatial anisotropic uncertainty of feature points must be taken into account to improve the precision in visual navigation. This paper includes two parts, which discuss error modeling and motion estimation respectively. In the first part we model the 3D reconstruction uncertainty in binocular stereo system as normal distribution and compute its propagation in stereo pair. Assume the uncertainty of image feature pixels geing normal distributed on the image plane, the reconstructed 3D error is analytically derived based on some error evaluation schemes. The closed-form solution of the 3D uncertainty is obtained for parallel camera setup. The second part of this paper proposes a novel iterative motion estimation algorithm that involves the anisotropic 3D uncertainty. We present a modified centroid coincidence theorem to divide the problem into two steps, rotation estimation and translation estimation. The translation estimation is straight-forward, and the latter can be obtained by a new iterative method as well. The LMS motion estimation criterion is linearized at 0th order and a motion estimation equation is proposed. The initial guess of the motion parameters is given by a SVD method. The iterative algorithm yields the optimal LMS motion estimation. Experimental data show that the iterative algorithm always converges under large and small point sets. It is much more robust and fast than previous quasi-Newton optimization based approaches. Transactions on Information and Communications Technologies vol 16, © 1996 WIT Press, www.witpress.com, ISSN 1743-3517