A multiresolution block matching technique for image motion estimation

To better utilise the information in the high-resolution small-blocks of an image it is necessary to generate accurate estimates of the motion. We present a bottom-up multiresolution block matching technique. We first compute a set of candidate motion vectors for each comparatively small block. We then take an intersection of these sets of neighbouring blocks. The resulting set is the set of candidate motion vectors for the larger block that encompasses the neighbouring blocks. This way, it preserves the motion information extracted at high resolution and uses the information at low resolutions only selectively to improve the motion estimates