3D Motion Analysis Of Non-rigid Object Based On Monocular Vision

3D motion analysis is one of active area in computer vision. Based on the progress of 2D image processing, it provides the motion and structure parameters of moving objects in scene by measuring and computing the coordinate of features. Now, it has become an important part of motion image processing.Compared with rigid motion, non-rigid motion is more general in real world. Because of the generality and variety, it has a wild application in many fields such as intelligent processing of agricultural information, intelligentization of agricultural machinery, agriculture and food industry, biomedical engineering, biometrics technology etc.Articulated and elastic non-rigid objects are our main research objects because they are more general than others in agriculture field. Based on the human, finger and simulated motion image sequences, some models and algorithms which respectively correspond to the motion of articulated and elastic non-rigid objects have been proposed in this dissertation. Our objective is to provide some useful theories or ways for the design of gait control of humanoid robot, the design of multi-fingered robot hand and other research works relevant to the non-rigid motion.The dissertation mainly consists of three topics, namely the basic concepts and methods of 3D motion analysis, motion-and-structure parameters estimation of articulated and elastic non-rigid object. The contents of this dissertation are arranged as follow.Chapter 1 demonstrates the significance of non-rigid motion vision, introduces the general classification of non-rigid motion forms, analyzes the existing methods of articulated and elastic non-rigid objects, discusses the limitations of these methods and outlines the main research contents of this dissertation.Chapter 2 introduces some basic concepts and methods concerned with this dissertation, explains the reason why the perspective projection model and the method based on feature correspondence is used to estimate motion of object in this dissertation.Chapter 3 proposes an approach in two parts to reconstruct 3D motion of human in monocular image sequence of human motion. In the approach's former part, Kalman filter is reasonably integrated into the basic framework of Genetic Algorithm(GA). Its main objective is to obtain the correct correspondence of 2D feature points between consecutive image frames. In order to solve this problem, several factors, such as the position estimators of feature point, pseudo feature points, the occlusion of feature point and the smoothness property of motion and so on, are considered during the design of GA's individual code and fitness function. The experiments result show that by means of the proposed approach the correct correspondence of feature points in image sequence can be realized after the pseudo feature points are eliminated and the occluded feature points is restored. In the approach's latter part, some constraints, for example the perspective-projection relationship between 2D image coordinate and 3D space coordinate, the local rigidity and smoothness property of human motion, the prior knowledge of human anatomy etc, are used to estimate 3D motion trajectory of joints based on the results of 2D feature-point tracking. The experiment results show that 3D motion trajectory of human's arm and leg can be effectively computed based on the constraints of rigidity, structure and motion smoothness under the perspective projection model.Chapter 4 mainly research the way to estimate 3D non-rigid motion based on the correspondence model of planar patch in monocular image sequence of finger motion. Through the motion-property analysis of finger's stick-figure model, three constraint equations, i.e. motion constraint, depth constraint and rigidity constraint, are formulated. Then, the objective function is created by using the penalty method. In order to establish the correspondence relationship of feature points on finger between consecutive frames, an improved block-matching algorithm is presented. The experiment results prove that the correct correspondence of feature points in image sequence can be attained by means of the improved block-matching algorithm and 3D motion trajectory of finger can be effectively reconstructed by optimizing the objective function.In chapter 5 and chapter 6, under the hypothesis that the correspondence relationship of feature points between consecutive frames have been established in motion image sequence of elastic non-rigid object, the problem to estimate 3D motion parameters of non-rigid object is transferred to the one to estimate 3D motion parameters of feature points on non-rigid object. Then, the local affine motion model is given under the perspective projection model. It needs to be mentioned that other appropriate models also can be used to substitute for the affine motion model.Chapter 5 presents a way to estimate local 3D motion of non-rigid object based on regularization technology. By means of this technology, the prior knowledge about non-rigid motion is integrated into the procedure of motion estimation. Then, the way how to formulate the objective function of motion estimation is explained and Levenberg-Marquart method is used to solve the objective function. The experiment results prove that the way with regularization technology is better than the way without regularization technology. In order to overcome the limitation of the regularization method, chapter 6 builds an irregular MRF model suitable for the motion estimation of feature point based on MAP-MRF frame. The energy function, which can reflect the joint probability distribution of motion parameters and the constraints relation of local 3D motion parameters, is established in irregular MRF. Then, a multi-level MRF is created so as to improve the efficiency of algorithm and solve the problem of solution uniqueness. In the multi-level MRF, two kinds of optimization algorithm, viz. least squared method and simulated algorithm, is used to minimize the energy function. Meanwhile, an improved SA is discussed according to the motion-estimation-property analysis of non-rigid object. The experiment results from two synthetic image sequences demonstrate the robustness of solutions and the feasibility of our algorithm.Chapter 7 firstly summarizes the research work and points out the originalities of this dissertation. Then, the recommendations for future work are also given.