Fast and robust motion computations using direct methods

摘要

A respective intensity gradient of a first spatially aligned pixel in each frame of a sequence of images is multiplied by a respective intensity or function thereof of a second pixel in the new frame, this being done for all pixels in the set and the resulting products being summed. This is repeated so that in one of the sums the first and second pixels have identical locations in the respective frames and where in each additional sum the second pixel is shifted by a non-zero shift relative to its location in other sums. The transformation is computed using intensity and gradient values of the image frames in the sequence and the sums. Weighted averages of at least two of the sums are stored and used together with the intensity and gradient values of the frames in the sequence to re-compute the transformation.