3D Motion Parameters Estimation Using MRF And Neural Network

Motion analysis is an important part in the field of computer vision research. Motionanalysis aims to distinguish an object from the image sequences, and at the same time itestimates the motion parameters of the object. Categorized by the analysis object, there are twokind motion analysis that is rigid and non-rigid motion analysis. In the early days, researches inmotion analysis were based on the rigid assumptions of an object, and a mature theoreticalframework has been formed. However, in the real world, there are so many motion objects thatare not rigid, then the rigid assumptions of motion description of a non-rigid object is inadequate.Therefore, it is necessary to do further research in non-rigid motion.To solve the problem of parameters estimation in non-rigid3D motion, the paper conductsthe following jobs:(1) Comprehensive summarizes the current research status of non-rigid motion analysis athome and abroad, and points out the challenges in non-rigid motion analysis, and presents anew algorithm. All of these lay a foundation for the following researches.(2) Introducing some basic problems about the3D motion analysis, while pointing out thedescription methods and common analysis methods of the3D motion and givingcomparisons at the same time.(3) Proposing a method basing on Markov Random Field for three-dimensional motionestimation, which establishes a MRF model corresponds to the motion estimation of featurepoints and constructs the energy function of joint probability distribution of motionparameters to reflect the constraint relations among three-dimensional motion parameters.Finally we proposes the simulated annealing algorithm for finding the optimal solution ofthe model,and then we get the motion parameters and the3D coordinates of all the featurepoints.(4) Proposing a algorithm using the neural network to do motion cluster. Firstly weintroduce the concept of the feature point and the regional triangle, and then we design theneural network according to the given three constraints. Then the detailed analysis of setting method of the weights and initial value of the neural network is presented, and theconvergence of the neural network is proved.(5) We combinates the MRF and neural network algorithm, and puts forward a concretemethod of fixing neighborhoods in time, so a complete theory framework ofthree-dimensional motion estimation of non-rigid has been constructed. The experimentalresults of simulation and cloth motion analysis prove the feasibility and accuracy of thisproposed algorithm.