Research On Optical Flow Computation For Motion Estimation

Optical flow field, as a motion parameter, contains not only information of theinterested moving objects, but also more hidden information, like the three-dimensionalstructure information of moving objects in the real world. We can easily achieve theidentification or positioning with the moving objects, and valid motion estimation bythese information. So how to improve the quality of optical flow field has always beenone of the hot issues in the field of computer vision.The main content of this Dissertation is in-depth study on which optimizationmethods will dramatically influence the quality of optical flow and how to improve thequality of optical flow based on a relatively fixed mathematical model by variationalmethods. A baseline method is proposed builded on HS model, some image processingand optimization methods, like Structure-Texture Decomposition, Coarse-to-FineEstimation, Warp, Graduated Non-Convexity and Median Filtering.The penalty function used in the mathematical model and optimization methods arestudied by changing one factor in the baseline method once each time. We find that thechoice of penalty function and different way using optimization methods willdramatically influence the quality of optical flow.An improved algorithm is proposed based on the conclusions of the study ofbaseline method. It use the optimal penalty function and optimization methods withrational parameters: the data term and prior term both use Charbonnier penalty function,three steps Graduated Non-Convexity and Warp five times at each image pyramid.The improved algorithm compares with I-TV-L1algorithm and A-Huber-L1algorithm, and we find that though it`s not the best results on each image sequences, theaverage endpoint error is minimum. And it proves that we use the original HS model,but we can still achieve high quality of optical flow based on the modern optimizationmethods.