Robust estimation of 3-D motion parameters in presence of correspondence mismatches

It is noted that the accuracy of a motion analysis scheme based on feature point correspondence is quite poor when there are mismatches between points. A small amount of mismatch (i.e., outlier data) may degrade the performance of the least-squares estimator significantly. It is shown that the least median of squares (LMedS) estimator can provide the required robustness. This method works well even when almost half of the given data are outliers. A Monte Carlo sampling technique is used to reduce the computational load. Subsequent use of the total least-squares or the constrained least-squares method on the trimmed data set enhances efficiency. Simulation experiments are carried out to evaluate the performance of the proposed method. By virtue of having a very high breakdown point, the LMedS estimator provides the much desired robustness against the correspondence mismatches