Research Of Registration And Stitching For Heterogeneous 3D Human Body Model

Multimedia technology is developing rapidly,and the media is evolving from text,voice and image to 3D image.At present,in the movies,games and virtual scenes,the requirement for 3D human body models is getting higher and higher,and the demand for realistic 3D human body models is increasing.At present,there are many devices that can acquire 3D models,but each device or method has its own characteristics,resulting in different scale,quality,and processing speed of acquired 3D data.For the registration of the human body model data scanned by two or more devices,it is necessary to set the registration parameters manually according to the device parameters in advance.Some devices have large noise interference,which brings trouble to 3D reconstruction.During the process of scanning,the human body will be slightly deformed,which will also bring difficulties to the fusion and reconstruction.Aiming at these two problems,this paper presents a registration and splicing method for heterogeneous 3D human body model.Aiming at the initial rigid registration problem of heterogeneous 3D human body model,this paper proposed an initial rigid registration algorithm based on texture features and geometric features of heterogeneous model.In this paper,the initial rigid registration is realized by means of feature matching.First,the texture features and geometric features are calculated using point cloud data.Then,we calculate the point pairs by the feature matching,and the points that are optimally matched are selected.Finally,the scale parameters and the transformation matrix are calculated by using the matching point pairs,and finally the initial rigid registration is realized.Aiming at the deformation of human body,this paper proposes a non-rigid registration method based on motion vector and local rigid deformation.Based on the initial rigid registration,the method divides the model into three parts:the overlapping region,the boundary region and the non-overlapping region.Then,the motion vector is generated by using the point of the overlapping region of the model,and then the motion vector of the entire model is calculated.The model is then non-rigidly registered using a motion vector combined with local stiffness.We have experimented with scanned 3D human models and other models.According to the experimental results,it can be concluded that the proposed method can calculate correctly and select the feature point pairs and motion vectors of the 3D model,and achieve initial rigid registration and non-rigid registration.By comparing with other rigid registration methods,the rigid registration algorithm in this method can achieve rigid registration more accurately and efficiently.The non-rigid registration algorithm in this method can achieve non-rigid deformation quickly and smoothly.Moreover,this method can maintain the shape of the model mesh.