The USC Andrew and Erna Viterbi School of Engineering USC Signal and Image Processing Institute USC Ming Hsieh Department of Electrical Engineering University of Southern California

Technical Report USC-SIPI-223

“Computation of Dense Optical Flow with A Parametric Smoothness Model”

by Navid Haddadi and C.-C. Jay Kuo

October 1992

A new algorithm is presented for the computation of dense optical flow and motion boundaries from an image sequence using two or more frames. The algorithm is based on a novel parametric smoothness model by decomposing optical flow into irrotational and solenoidal fields, and imposing the smoothness constraint on each field separately. This model implies smooth translation and rotation of the underlying motion process. In contrast, the smoothness constraints used in all previous work do not distinguish the translational and rotational components but simply combine them as a whole. The derivation of the parametric smoothness model sheds new light on the interpretation of the conventional membrane model. The problem of over-smoothing across motion boundaries can be resolved to a high degree by successively improving the estimate of the parameters of the smoothness model. Significant improvements by the proposed new algorithm over classical gradient based methods have been obtained for a class of test problems.

To download the report in PDF format click here: USC-SIPI-223.pdf (1.2Mb)