A robust algorithm that estimates the motion parameters recursively from a sequence of noisy images is presented here. We propose the use of the least median of squares method in conjunction with a computationally efficient recursive scheme. The method works well even when nearly half of the features have been matched very poorly. A recursive constrained least squares method is developed while dealing with a range or stereo data sequence and a recursive total least squares method is proposed for the monocular data sequence. (C) 1996 .