I have developed a new optical flow algorithm which overcomes many of the problems associated with the standard algorithms, giving excellent results on noisy image sequences. It works by estimating the probability that an image patch (receptive field) informs about any of a number of possible displacements. The most probable displacement is selected as the flow value; this is repeated for each image patch in the image. This method gives accurate results near motion shear boundaries, unlike other methods. It is intrinsically simple, and does not rely on temporal or spatial difference information, nor is it a correlation-based method.
The algorithm has been tested on imagery from surface strain (using photo-micrographs in collaboration with Prof Philip Withers and British Aerospace), microsurface structure (using SEM imagery courtesy of Dr Rob Yates, University of Sheffield) and image registration (using MRI imagery by kind permission of the Department of Radiology, Addenbrooke's Hospital). The algorithm is being used in a research project in the Materials Science department funded by British Aerospace, and its further application to materials testing is the subject of a EPSRC funded project by myself and Phil Withers (Surface strain mapping as a tool for the study of materials behaviour across a range of scales).
The most recent development is that the algorithm has been found to be equivalent to a robust estimator using a Welsch influence function, and has been reformulated in a Bayesian framework by Phil Torr of Microsoft Research Cambridge.