Non-rigid Point Set Registration Algorithm And Its Application In Human Pose Estimation

Point set registration has always been a fundamental and critical problem in computer vision.Its goal is to find the corresponding relationship between two groups of related point sets.With the further development of science and technology,point set registration is widely used in face recognition,human body shape modeling,medical image analysis and other fields.Therefore,the research on point set registration is of great significance in many fields.At present,the robustness of the point set registration algorithm is still not high due to the influence of noise,and the uncertainty of the deformation of non-rigid transformation also leads to the challenge of point set registration.Therefore,the accuracy of human pose estimation based on point set registration is not high.In view of the above problems,the main research contents of this thesis are as follows:(1)A non-rigid point set registration method combined with clustering algorithm is proposed.In this thesis,the point set registration is regarded as a probability density estimation problem,and the Gaussian mixture model is constructed during registration.Furthermore,clustering,clustering corresponding projections and the local topological structure preservation are introduced.The local characteristic structure of the point set is maintained in the two stages of the registration.Then the parameter estimation is carried out by expectation maximization algorithm.A robust non-rigid point set registration algorithm is obtained,and the validity of the algorithm is verified on different data sets.(2)A method for human point set registration is proposed.In this thesis,Laplacian coordinates are added on the basis of non-rigid point set registration algorithm which maintains local topological structure,so that the local structure is maintained on the neighborhood scale when the human point set is registered.The two local structure retention constraints complement each other,so that the joint structure of human body can be maintained and the error of human point set registration can be reduced.This method does not involve the local rigid assumption,nor does it require any initial conditions,so it has certain flexibility and versatility when dealing with the deformation of non-rigid joints.(3)A human pose estimation framework based on point set registration is proposed.In this thesis,the point correspondence estimation obtained from the point set registration is used as initialization,and the human pose estimation is carried out with the improved articulated ICP algorithm,which can make up for the deficiency between the two at the same time.For the non-rigid probabilistic point set registration algorithm,the accuracy of joint deformation registration is not high enough because it violates the hypothesis of motion coherence.Articulated ICP algorithm is suitable for human pose estimation,but if there is no good corresponding initialization,it is easy to fall into local minimum.Therefore,the framework of this thesis combines the two algorithms,which has a good accuracy in human pose estimation,and the validity of the framework is verified by experiments.The non-rigid point set registration algorithm combined with the clustering algorithm proposed in this thesis has good robustness,and it also provides effective corresponding estimation in the application of human pose estimation.The human pose estimation results obtained by combining with the improved hinged ICP algorithm have good accuracy.This thesis provides theoretical and experimental references for the research of point set registration and human pose estimation.