Research On Optical Flow Computation Of Color Image Sequences

Motion vision analysis is an important research area of computer vision. It hastwo main methods: feature-based and optical flow-based. Therefore the opticalflow computation is a key technique for motion vision analysis, it has emerged as amajor technique to estimate object's motion in image sequences. Optical flowanalysis researches on the relationship of image brightness varying in time and theobject's structure and motion in the background. Many important information ofmotion and structure of object are included in optical flow, and optical flow canhelp us to know about many important motion and structure characteristics ofmotion bodies. This thesis uses GDIM mold to estimate optical flow for colorimage sequences, and the research on optical flow mainly based on two points: thelimitations of brightness conservation assumption and the under-determinedproblem of the optical flow fundamental equation. First, the optical flow techniqueusual used is based on the assumption that the image brightness remainsunchanged under the image motion. According to this assumption then the opticalflow fundamental equation is received. When additional constraints are added, theoptical flow is obtained. But in non-uniform brightness variation environment, theoptical flow fundamental equation which is base on brightness conservationassumption is unsuited, and the result computed from these methods is unreliable.We solve this problem by using generalized dynamic image model (GDIM). GDIMrevises the optical flow fundamental equation, it includes radiometric informationin the optical flow computation, and works better under non-uniform brightnessvariation. Second, the optical flow fundamental equation which is based on thebrightness conservation assumption only has one equation, and the optical flowcould not uniquely determined if there is no additional constraint added. In colorimages, there is abundant color information, according to the three-basic-colorprinciple and optical flow fundamental equation, we can obtain three optical flow constraint equations corresponding to three-color components. This gives us anover-determined linear system with three equations and two unknowns, and theoptical flow of color images that what we find is the solution of the linearequations' system.For the optical flow from color images sequence, the color systemsnormalized rgb are used, and the optical flow field of color image sequences iscomputed by using GDIM mold. This thesis targets motion image sequences thatusually used. Comparisons between the experimental results of this method and thetraditional method are presented. The results demonstrate the better performance ofvelocity and simpler computation and more motion pattern suited. For theincreasing of the accuracy and reliability of optical flow computation, the result ofthis research has certain significance.