Research On 3D Human Registration Based On Spectrum Embedding And Feature Points

In recent years,with the rapid development of computer graphics,3D modeling technology has been widely used in virtual reality,film animation special effects production and other fields.3D registration technology is a basic problem of 3D reconstruction.The design of point cloud registration algorithm has a direct impact on matching accuracy,speed,model acquisition and other aspects.When acquiring 3D human body model data,it is difficult for imaging sensor equipment to obtain human body point cloud data in a short time.With the extension of model acquisition time,human body must have different degrees of deformation.For example,tilt of body parts,changes of body movements,etc.Therefore,it is necessary to register the 3D models acquired from different perspectives through non rigid registration algorithm,and then get a more complete 3D human model.In order to obtain the real and natural human registration model after registration,this paper presents a processing flow of 3D human non rigid registration based on the combination of spectral embedding and feature point extraction.First of all,this paper uses the depth acquisition equipment to obtain the real different human body movements,and carries on the denoising and the simplification processing to the collected point cloud data.Then,the curvature extremum of all point cloud data is searched iteratively by the method of fitting the surface,and all feature points of the model are obtained by comparing the curvature value with the threshold value.Then,the spectral embedding algorithm is used to calculate the corresponding relationship between models.Finally,the feature points with higher curvature value are selected as the control points,and the corresponding relationship between the data matching points is calculated according to the spectrum embedding algorithm.The control points and the thin plate spline function are used to drive the change of the whole 3D model,so as to complete the registration of the 3D human body model.In view of the difficulties caused by the body deformation during the registration process,this paper presents a non rigid registration method based on the spectrum embedding algorithm for the human body model,which is used to register the large-scale deformation model.The main difficulty is to find the precise correspondence between point cloud data.Therefore,this paper uses the spectrum embedding algorithm to calculate the similarity between point clouds,The affinity matrix of the model is constructed,the eigenvalues and eigenvectors of the matrix are obtained and visualized,that is,the three-dimensional model in spatial coordinates is transformed into the two-dimensional spectrum domain.This method retains the main feature information of the model and eliminates the influence of rigid body transformation when collecting the model.The experimental results show that the non rigid registration method proposed in this paper can correctly calculate the geometric feature point pairs and their corresponding relations of the three-dimensional model,and achieve the final non rigid registration.Compared with other registration methods,the proposed method can complete registration more accurately and efficiently.