Research On Optical Flow Computation For Motion Image Analysis

Motion image analysis is an important task of Computer Vision, and the optical flow computation is a key technique for motion image analysis. The optical flow is a convenient and useful image motion description, it contains rich 2D motion cues and is often used to recover 3D scene structure and 3D motion parameters of visual sensors. It also contains some important cues of elastic deformation of non-rigid bodies, and flow structure of fluid motion images. It can help us to know about many important motion characteristics of motion bodies by optical flow, It is undoubted that optical flow plays an important role in motion image computation and analysis.This thesis addresses some problems of optical flow computation and analysis including: increase the accuracy of optical flow computation, cope with the problems of motion boundary, propose a optical flow method for fluid motion image computation, give a vector filter based on temporal correlation constraints for the more accurate and reliable description of fluid motion characteristics, and analyze the motion patterns of video sequences by optical flow statistic. This research is significant for increasing the accuracy and reliability of optical flow computation, and it also has significant academic and practical value for motion images computation and analysis.The first chapter systematically introduces the background , the significance and the state of the art of optical flow computation, and indicates the limitations of traditional methods. This chapter also gives the structure and the contributions of this dissertation.In chapter 2, we propose a new optical flow method based on Local Structure Constancy Model (LSCM) instead of the traditional BCM for reducing the model error and increasing the robustness. Local image structure is less sensitive to illumination variation than intensity, which can increase the robustness of optical flow computation in real applications. Here we use two order structure tensors of an image to describe its local structure. A bi-directional computation scheme based on LSCM is also adopted for reducing the model error resulted from the linear Taylor's expansion, which allows two order Taylor's expansion of optical flow equation for a more accurate approximation and avoids nonlinear equation calculation.Chapter 3 addresses the issue of motion boundary problems to which two ideas are introduced. First, a probability-control selecting scheme is adopted, which divides the local neighbor into different regions with consistent motion vector. By a local refining step, the new method can preserve motion boundaries efficiently; Second, we view optical flow computation as a nonlinear diffusion process instead of an energy minimization process. It avoids the restrictions of convexity and differentiability required by normal regularization methods. We use a Coherence Vector Enhancing Diffusion (CVED) scheme and a bilateral filtering diffusion scheme to improve the efficiency of preserving motion boundaries, respectively.In chapter 4, an optical flow method based on nonlinear filtering scheme is proposed which is the extended version of the method based on bilateral filtering scheme proposed in chapter 3 and is applied to fluid motion images computation (PIV). In addition, a Least-squares (LS) vector-filtering scheme based on temporal correlation constraint is proposed for offline computation & analysis, which can help to suppress the computation noise and increase the reliability of fluid motion characteristics' description. In view of the div and curl fields as the main objects of computation & analysis, we extend the proposed vector-filtering scheme to multivector space, which takes into account div and curl along with motion vectors. With the help of Clifford algebra, we construct multivectors by vectors, div and curl to characterize fluid motion. By a multivector filtering computation, we obtain the reliable description of fluid motion characteristics.The fifth chapter of the thesis is concerned with video motion analysis based on optical flow computation. From optical flow fields we can obtain the motion vector of each pixel per frame, which provides richly redundant information for statistic methods acting on optical flow fields. In this chapter we develop a statistics method for motion recognition based on motion accelerate fields. A spatio-temporal multiresolution histogram (STMRH) acting on motion accelerate fields as the motion character's description of dynamic texture videos is proposed. In order to reduce computational consumption, a simplified multigrid computation scheme for optical flow is adopted.The final chapter gives the conclusion and the contributions of this dissertation, and indicates some related works that are yet remained and attractive for future research.