Accurate Body Shape Reconstruction Methods Based On A Single Image

The development of parametric model based on reconstruction promotes the research of human body reconstruction based on a single image.However,existing methods often ignore the affection of camera parameter estimation errors and the shape errors caused by the estimated sparse skeleton different from the real one.Therefore,in this thesis,we optimize the camera pose parameters and combine the sparse bone posture transformation,and consequently propose an accurate body reconstruction method based on a single image,get accurate shape appearance.The full text mainly includes the following two aspects:1.Body shape reconstruction with the iterative refinement of camera parametersTo tackle the estimation errors of camera parameter and pose,we combine camera parameter prediction and joint definition to design a layer-by-layer feedback structure optimization framework to refine camera parameters for human body reconstruction.The framework learns the initial shape and pose of the prediction model through deep convolutional neural network,uses the 3D-2D matching relationship to construct the least squares problem,and estimates the correct coordinate position of the camera.Each layer iteratively uses the original image and the result from the adjacent projection model as input to optimize the camera pose.Meanwhile,it introduces the transformation traction of the skeleton and the constraint of image contour,and uses the reprojection error to optimize the 3D model for an accurate modeling.Experiments demonstrate that this method can effectively improve the accuracy of 3D model reconstruction,thanks to the estimated accurate camera parameters.2.Body shape reconstruction based on the refinement of skeleton and poseTo overcome the limb penetration and pose errors caused by skeletal transformation of human model,we introduce a sparse skeletal node graph based on the linear skinning model with existing skeleton weights,and transfer the model estimation optimization into a skeleton transformation problem.The sparse skeletal node graph is a axes based pose model,and the temporal motion is carried out through dynamic time arrangement and spatial motion division.Meanwhile,the geodetic distance is taken to analyze and process the mesh topology of the model.A shape distance constraint system is created based on the minimum value of each voxel on the surface to the center skeletal point which limits the penetration of the intersecting parts during the overlapped movement of the limbs.Experiments show that this method improves the accuracy of human pose estimation in single images,fitting the 3D human model effectively.