A new features-based fast algorithm for motion estimation: decimated integral projection (DIP)

Abstract

Block motion estimation using exhaustive search is computationally intensive. Most of the proposed strategies reduce the algorithm complexity by limiting the number of locations searched or by pixel and motion-field sub-sampling. This is done at the expense of the accuracy of the estimation and without a substantial reduction of the computational cost. A new algorithm based on the features representation of the luminance and chrominance pixels values using integral projection is proposed. The used search strategy has the same performance of the classical exhaustive search with a computation reduction of a factor of 30. In the presented implementation a six-step fast search procedure select a small set of blocks on which exhaustive search is applied to compute the final motion vector. The presented algorithm takes advantage of a statistical partitioning of the search set, in order to progressively reduce the number of matching. A direct comparison in terms of the performance/complexity is reported with respect to other proposed solutions.