Study On Optical Flow Computation Methods For Image Sequences

Optical flow algorithm is an important branch of computer vision, one of main methods of studying motion images, it describes the apparent motion of the moving object in the grayscale mode. Optical flow algorithm can extract the structure, depth,contour and other important information in the motion scene without being aware of position and characteristics,so it’s an advanced visual expression. This thesis mainly study on the optical flow algorithm, aimed at repaired the wrong flow in the occlusion area.In the first, we introduce classic Horn-Schunck algorithm and Lucas-Kanade algorithm,HS algorithm is proposed based on the conservation of gray level constraints and global smoothing,LK algorithm is based on local smoothing constraint;In the second,according to analysis the advantages and disadvantages of these two kinds of algorithms,we compare their performance;Finally we verified through many experiments.Classic+NL optical flow algorithm and TV-L1 optical flow algorithm, are the higher accuracy and robustness are better optical flow algorithm at present,without dealing with the wrong flow in the occlusion area.In the thesis,firstly,we deeply analyzes the effective optical flow detection method to detect occlusion area;Secondly use the monotone decreasing function to deal with occlusion area, suppressed the wrong flow to spread of the neighborhood; Finally proposed repair method based PSO,lateral repair method based on the neighborhood and longitudinal repair method based on flow of adjacent frames, to repair the wrong flow in the occlusion. Experiments show that proposed methods can effectively overcome wrong flow for the spread of the neighborhood, and effective to repair the wrong flow in the occlusion, accurately portrays the optical flow field, and improving the accuracy of optical flow.