Research On Variational Optical Flow Computing Technology For Large Displacement And Complex Scene Image Sequences

Optical flow is the two-dimensional instantaneous velocity of the pixel on the visual surface of the moving object or scene in the image sequence.It not only contains the motion parameter information of the moving object and scene in the image,but also includes the three-dimensional structure information.Thus,the research of optical flow computing from image sequence is a key task in the fields of image processing and computer vision.The purpose of studying on optical flow computation is to approximately acquire the motion field which can't be obtained from the image sequence directly.Since the 21st century,with the great developments of the computer software and hardware,optical flow estimation and the related technologies have been widely applied to UAV positioning and navigation,automatic testing equipment,foreground obstacle detection,human gesture recognition and many other areas.With the continuous improvement of optical flow computing technology,the accuracy and reliability of optical flow estimation in simple scene have been greatly developed.However,the accuracy and robustness of the existing optical flow models still need to be further enhanced when the image sequence contains difficult motion scenes such as large displacement and complex texture structure.This thesis mainly studies on the optical flow computing technology in large displacement motion and complex scene,which is devoted to deal with the issues of motion discontinuities and edge-blurring.The main contents of this thesis are summarized as following:1.In section I,the basic concept,research significance and background of optical flow estimation are introduced firstly.Then,the research progress,typically related work and some remained key issues of optical flow computing technology are detailed discussed.Finally,the research contributions and chapter arrangement of this thesis are briefly concluded.2.Section II introduces the basic knowledge of variational optical flow computation.Specifically,the gray constant assumption between consecutive frames,the classical Horn-Schunck optical flow model,the energy function of the variational optical flow and the commonly optimization strategies in optical flow calculation are elaborately discussed and analyzed.3.To improve the robustness of large displacement optical flow computing,a layered nearest neighbor field based variational optical flow model is presented in Section III.First,the relationship and difference between the nearest neighbor field and optical flow are discussed.Then,a layered nearest neighbor field based motion segmenting optical flow model is proposed to obtain the prior optical flow which includes the abundant motion information.Afterwards,the prior optical flow is substituted into the classical TV-L~1 optical flow model to compensate the large displacement motion.The experimental results indicate that the proposed method has high estimation accuracy and robustness for large displacement optical flow estimation.4.In order to address the issue of edge-blurring caused by the complex scene,section IV explores a non-local joint filtering based TV-L~1 variational optical flow estimation model.First,the input image is divided into three types:mutual structural regions,inconsistent regions and smooth regions according to the different local feature structures.Then,the mutual structure guiding filtering model of optical flow is presented by utilizing the mutual structural region as the guiding information.Furthermore,the non-local joint filtering based TV-L~1 optical flow model is constructed to remove the outliers by jointing mutual structure guided filtering and weighted median filtering,which can preserve the image and motion boundaries and improve the robustness of optical flow computing.The experimental results demonstrate that the proposed approach can produce accurate and robust optical flow,especially with significant capacity of edge-preserving.5.Section V firstly summarizes the research content of this thesis,and then the potential limitations and possible solutions of the presented models are discussed.