US 8,285,079 B2Granted

Method for highly accurate estimation of motion using phase correlation

AssigneeSony Corporation

InventorsMark A Robertson, Ming-Chang Liu

Life of this patent

Estimating the term…

Loading patent…

Abstract

An image alignment method includes computationally efficient methods of achieving high-accuracy local motion estimates by using phase correlation. The method also estimates motion reliability that allows a generic robust model fitting algorithm to produce more accurate results while operating much more efficiently. One of three methods are used to determine sub-pel motion estimation with improved accuracy. Each of the sub-pel motion estimation methods uses phase correlation, and are based on fitting computationally efficient 2-D quadratic surfaces to a phase correlation surface. A pre-filter is applied which shapes the phase correlation surface to enable appropriate fitting to the quadratic surface. Bias is also compensated for prior to applying a sub-pel motion estimation method. The method also estimates the reliability of the sub-pel motion estimates determined using phase correlation.

Claims36

Dependent claims

An image alignment method comprising:

a. storing a reference image and a current image in a memory, wherein the reference image is defined by a first set of local pixel blocks and the current image is defined by a second set of local pixel blocks;
b. applying phase correlation to a first block in the reference image and a first block in the current image to generate a phase correlation surface between the first block of the reference image and the first block of the current image, wherein the phase correlation surface is defined by a set of correlation points each having a correlation value;
c. filtering the phase correlation surface using a filter that is fit by a first quadratic surface, thereby forming a filtered phase correlation surface having a filtered set of correlation points;
d. determining a sub-pel motion estimation according to a second quadratic surface and a sub-set of correlation values of the filtered set of correlation points and a largest peak correlation point, wherein the sub-set of correlation values correspond to a neighborhood of correlation points about the largest peak correlation point, thereby determining a motion vector corresponding to the first block in the current image;
and
e. determining a global motion between the reference image and the current image by fitting a set of motion vectors corresponding to the second set of blocks in the current image to a global motion model.

3 more elements

A system for aligning images, the system comprising:

a. a memory for storing a current image and a reference image, wherein the reference image is defined by a first set of local pixel blocks and the current image is defined by a second set of local pixel blocks;
and
b. a program for
i. applying phase correlation to a first block in the reference image and a first block in the current image to generate a phase correlation surface between the first block of the reference image and the first block of the current image, wherein the phase correlation surface is defined by a set of correlation points each having a correlation value;
ii. filtering the phase correlation surface using a filter that is fit by a first quadratic surface, thereby forming a filtered phase correlation surface having a filtered set of correlation points;
iii. determining a sub-pel motion estimation according to a second quadratic surface and a sub-set of correlation values of the filtered set of correlation points and a largest peak correlation point, wherein the sub-set of correlation values correspond to a neighborhood of correlation points about the largest peak correlation point, thereby determining a motion vector corresponding to the first block in the current image;
and
iv. determining a global motion between the reference image and the current image by fitting a set of motion vectors corresponding to the second set of blocks in the current image to a global motion model.

5 more elements

Dependencies are inferred from the claim text.