3D Non-rigid Human Body Reconstruction Based On Kinect Depth Sequence

3D reconstruction is an hot research topic in the field of computer vision and graphics.With the advent of Microsoft's Kinect depth camera,the study of three-dimensional reconstruction based on depth images has culminated.The 3D reconstruction based on Kinect depth data is generally divided into the following steps:depth image acquisition,depth image processing,depth image conversion,point cloud registration and point cloud fusion.Among them,non-rigid human body reconstruction requires non-rigid point cloud registration and non-rigid surface fusion.In this paper,we used the depth image sequence obtained by Kinect to reconstruct the realistic non-rigid 3D human model.We designed a human body depth data scanning system and made the following work in the point cloud correspondence estimation,point cloud non-rigid registration and non-rigid surface fusion and so on:(1)Human body depth data scanning system based on multiple KinectsThis paper proposed a human body depth data acquisition scheme based on multiple Kinects,and designed a human body scanning system.This system could scan quickly the human body and obtain human body depth data.(2)Correspondence estimation of point cloud estimation based on vector field consistencyCorrespondence estimation is the key step of point cloud registration,and the corresponding points that are non-rigidly registered are more difficult to estimate.The nearest neighbor algorithm is a common method for corresponding point estimation,however,this method will get a lot of wrong matching points.Vector field consistency algorithm based on vector field "smoothness" prior was used to filter image feature's match.In this paper,the vector field consistency algorithm was applied to the correspondence estimation of 3D point cloud,which accelerated the convergence speed of point cloud registration.(2)Constructing a non-rigid registration model based on sparse nodesThe non-rigid registration of point cloud requires estimating a large number of transformation parameters and the computation is complexity.The non-rigid registration based on sparse nodes constitutes a “graph” by uniformly collecting points at the source point cloud,and indicates the deformation capability of the entire point cloud through the transformation matrix of the“graph”nodes.It improves the performance of the registration greatly.Due to the mismatch,the normal of the corresponding points will be greatly different,which will affect the performance of non-rigid registration.In order to reduce the influence of mismatch on registration,this paper introduced a normal regularization term to improve the registration'sperformance of the stration model based on sparse ndoes.(3)The acquisition of point cloud non-rigid fusion's warped-filedDepth data non-rigid fusion based on truncated signed distance field(TSDF)model requires a warped-field.The traditional methods perform generally linear interpolation on matrix set to obtain the warped-field.This kind of linear interpolation can't obtain a smooth warped-field.In this paper,the 3D thin-plate spline interpolation algorithm was applied to transformation matrix interpolation to obtain a smooth warped-field.