Feature Constraint Based Symmetric Optical Flow For Motion Estimation And Image Registration

Optical flow representing image apparent movement contains the observed object’s motion information and three-dimensional structure information. Optical flow is an important entry in higher-level compute vision algorithm, playing an important role in target tracking, motion estimation, image registration and shape information recovery. Optical flow contains object’s motion information and structure information which have the role of guidance in motion estimation and image registration. Based on the research of optical flow computation, this paper propose a feature descriptor matching based symmetric optical flow. The contributions are:1) A feature descriptor matching algorithm is proposed to improve the accuracy of feature matching in motion image, which adopt multi-scale Zernike moments descriptor and a new feature matching strategy. Multi-scale Zernike moments descriptor represents neighbor information from different scale, which can describe local information well. In addition, a descriptor fuzzy matching strategy is proposed to improve corresponding point matching result.2) To improve detail preservation of motion image and reduce inaccuracy matching of detail, the descriptor obtained by Zernike moments are defined on the driving points in an image is proposed. The driving points are obtained by a union overlap of the boundary points by the Canny detector and Superpixel method. So driving points contain real boundary points with less noise. Furthermore, we propose a delaunary based fuzzy matching model. In this model, we redefine neighbor relationship which contributes to improve descriptor matching result. 3) A feature constraint based optical flow model is proposed to improve estimation accuracy of motion detail. This model use feature matching to correct motion detail in every image resolution (from coarser to fine). In addition, we propose an efficient symmetric optical flow to keep consistency of optical flow.Finally, this work integrates feature descriptor constraint into symmetric optical flow to improve the accuracy of motion estimation and image registration. Compared to the traditional optical flow methods, this method achieves better result on Middlebury beach dataset, MIT dataset, KITTI dataset, UCL dataset and magnetic resonance brain images, which demonstrates the effectiveness of the proposed method.