New non-interpolative kernel for pel-recursion with integer estimate updates

Pel-recursive motion compensation, in its original form, is based on a continuous formulation. However, it has been shown previously that the implementation of pel-recursion in a sampled (digital) domain inevitably involves two distinct steps: estimation of integral part of the motion and fractional estimation. A class of four-dimensional LS-optimum operators have been introduced for fractional estimation. In this paper, we introduce another member of the family of higher dimensional operators as a kernel for fractional estimation. This five-dimensional linear operator works on a 5-pixel template in the past frame to yield a luminance estimate. Aside from improvements in the fractional part, the operator can be used, along with an appropriate classifier, to improve the integer estimate at each step. A constrained least squares approach is used to solve for this operator, and short cuts to reduce computational complexity are pointed out