Motion Segmentation: A Robust Approach

A motion segmentation method, based upon robust least K-th order statistical model fitting (LKS), is proposed. Similar LKS based segmentation algorithms have been proposed for range data segmentation but non have been applied to motion segmentation. Moreover, the algorithm we propose here differs from other contemporary approaches using versions of LKS in a number of important ways. Firstly, the value of K is not determined by a complex (and heuristically justified) optimization routine. Secondly, having chosen a fit, the method of determination of the scale (the value used to define the “inlier” threshold) is not based upon the K-th order statistic of the residuals, but solely on the (ordered) unbiased scale estimates from the sorted residuals. Other aspects of the full segmentation scheme include the use of segment contiguity to: a) reduce the number of random sample fits used in the LKS stage, and b) to “fill-in” holes caused by isolated miss-classified data.