Regularized Patch Motion Estimation

This paper presents a new formulation of the problem of motion estimation which attempts to give solutions to classical problems in the field, such as detection of motion discontinuities and insufficiency of the optical flow constraint in areas with low intensity variation. An initial intensity segmentation phase partitions each frame into patches so that areas with low intensity variation are guaranteed to belong to the same patch. A parametric model is assumed to describe the motion of each patch. Regularization in the motion parameter space provides the additional constraints for patches where the intensity variation is insufficient to constrain the estimation of the motion parameters and smooths the corresponding motion field. In order to preserve motion discontinuities we use robust functions as a regularization mean. Experimental results show that the proposed method deals successfully with motions large in magnitude, motion discontinuities and produces accurate piecewise smooth motion fields.