Research On Optical Flow Computation Based On Generalized Dynamic Image Model

Visual motion analysis is a major field in computer vision. Because the optical flow field provides abundant information of motion and 3D structure research on optical flow field is always regarded as one of the effective way of solving the problems in visual motion analysis. The optical flow approach has emerged as a major technique to estimate object's motion in image sequences.The optical flow field is widely applied in industry and military, such as vision system of robot, resource prospecting, satellite tracking system, etc.This thesis is engaged in research on optical flow. Firstly, The traditional and often used methods of optical flow estimation have been introduced. The application range and problems in each kind of algorithms are analyzed in depth, which provides the basis of knowing optical flow. Secondly, Horn and Schunck's notable method is introduced in detail on the basis of introduction of optical flow estimation. The problem of bad effect of optical flow under situations of motion discontinuity and uneven illumination change is found by study. Traditional methods of optical flow estimation are poor in uneven brightness variation environment. This thesis solves this problem by generalized dynamic image model ( GDIM) and robust statistical framework. This method combines Gennert and Negahdaripour's GDIM formulation with the robust statistical framework of Black and Anandan. This model of illumination variation is general. The robust statistical framework recovers motion structures by main pixels in regions and eliminates inconsistent pixels to main pixels at same time to improve the ability in anti-jamming. GDIM includes radiometric information and works better under uneven brightness variation. The method in this thesis possesses advantages of both methods and gets better effect under situations of motion discontinuity and uneven illumination change.The method is simulatively implemented in this thesis. Comparisons between the experimental results of this method and the traditional method are presented to demonstrate its better performance. The estimation results are improved especially under situations of motion discontinuities and uneven illumination variations.