Motion Field Estimation for Temporal Textures

This paper presents a novel approach for estimating the flow fields of dynamic temporal textures whose motion differs radically from that of rigid bodies. Our approach is based on a local flow probability distribution function at each pixel using the STAR model and the data from a spatio-temporal neighborhood. The peak of this density function can be regarded as the estimated local flow vector. Our complete algorithm exploits a two-stage process. The first stage of the algorithm applies a simple tensor method to estimate the direction of the optical flow at each pixel in the texture. In the second stage, the flow probability function is used to perform a one-dimensional search along the flow direction to obtain the magnitude. Performance analysis and experiments with real video sequences show that our methods can successfully estimate flow fields.