This paper addresses the problem of LMS motion estimation

in the presence of feature point uncertainties. Optimal motion estimation is substantially a nonlinear problem with 6 depees of freedom. In this paper a modified centroid coincidence theorem and a motion parameter decomposition theerem are introduced

which degenerate it into a pure rotation problem with only 3 degrees of freedom. We propose a linear algorithm as well as an iterative algorithm

both providing optimal estimations. The linear one can be implemented in realtime

however the iterative one has better convergence abilities.