Bayesian motion estimation without spatial and temporal gradients

Abstract

A Bayesian motion estimation technique is proposed which models the motion vector field with a discontinuity-preserving prior and the observation noise corrupting the video frames with a Gaussian density. The method is related to various optical flow techniques, except that it is not dependent on spatial and temporal gradients which are notoriously difficult to estimate from real image sequences. The objective function to be minimized contains a block matching likelihood term and an optical flow prior term, making the technique a hybrid of two popular motion estimation schemes. Simulations show that the proposed technique results in more accurate motion vector fields than those obtained through conventional block matching and Horn-Schunck optical flow estimation.