Polymorphic Statistics Based 3D Human Body Reconstruction And Its Application In Virtual Try-on

3D models of human body have been widely used in the field of 3D animation,garment design and virtual try-on.The reconstruction of 3D human body based on polymorphic statistics takes full advantages of a large number of sample data and statistical learning algorithms,which provides basic and universal material for these virtual reality scene.Currently,3D models of human body are acquired by interactive design using commercial softwares or high resolution 3D scanner,which are not suitable for ordinary commercial applications because of the price and expertise.Therefore,it is of vital importance for 3D animation,garment design,and especially virtual try-on to do research on how to reconstruct 3D models of human body by a convenient,fast and precise way.This dissertation has realized the reconstruction and deformation of individual 3D human body based on polymorphic statistical learning,which include the reconstruction of 3D human body based on shape statistics,optimized reconstruction based on skin color statistics and dynamic deformation of human body based on pose statistics.The shape statistics of 3D sample models makes the human body stay in human shape domain,the skin color statistics optimized the naked body shapes under loose clothes,the statistical deformation of human body based on poses expand the static models to dynamic motion sequences,making the reconstructed results to be applied in more fields.Furthermore,we have simulated clothes samples by physical method and map them to human body shape space.We also learned the correlation between shape parameters of human samples and their feature sizes,thus realizing fast virtual try-on according to human body feature sizes.The reconstruction of 3D human body based on body shape statistics.The statistical learning of 3D human body shape samples provides prior information for the reconstruction of 3D human body.In order to reconstruct the 3D naked model of dressed human body,we use Kinect to get depth data of human body.The dressed users are acquired to pose motionlessly in front of Kinect while shooting.An implicit surface based denoise algorithm has been proposed in order to reduce the noise of raw human depth image of Kinect while keeping the geometric information of body surface as large as possible.Even though the human body is non-rigid,each body part can be regarded as rigid approximately,thus part-wise method is adopted to register the human depth images.The template chosen from sample datasaet of 3D human models is morphed to the human registered point cloud by Laplacian deformation algorithm.Finally,the deformed template model is projected to 3D human shape space to obtain precise reconstrued result.Optimized reconstruction of 3D human body based on skin color statistics.Since Kinect acquires depth and RGB images at the same time,the optimized reconstruction of 3D human body reconstructed method has been proposed for more precise reconstructed results of human body in loose clothes.The main idea of the optimized method is to improve the probability statistical model of skin color detection,which resorts to the Grab Cut algorithm,helps to segment the skin and clothes region of human registered point cloud and finnaly puts different weights in the template fitting step.The efficient improved algorithm of skin color detection does not affected by clothing pattern.The dynamic deformation of human body based on pose statistics.The static 3D human models can reflect the true human shapes,but their applications are limited.We expand them to dynamic for more situations.Taking the mesh local surface transformation into account,the sample based dynamic deformation of human body could simulate realistic human motion.For pose deformation,based on the 3D human pose space,the models of 3D human body can be deformed freely by a learning deformation model which builds regression model between the transformation of triangles on meshes and skeleton poses of human body.Adding the shape into the deformation system,that is combining the shape parameters into the triangle transformation of model meshes.The pose and shape could deform simutaneouly and thus build 3D human animation sequences.The application of virtual try-on based on human body feature sizes.The foundation of virtual fitting room is the reconstruction of individual 3D human body,which could serve customs a large amount of garments to try on and build suitable clothes models for 3D human bodies,saving labor cost and time greatly.The shapes of users are diverse in practice,in order to realize fast individual virtual try-on we first construct the sample data of different body shapes,and then extract features of human bodies and garments.The garment sample dataset is built by position based cloth simulation and mapping relations are generated between human shape samples to garment samples and to human body feature sizes,and finally fast virtual try-on can be realized according to human body feature sizes.In conclusion,this paper realizes the reconstruction of 3D human body,pose deformation and its application in virtual Try-on based on polymorphic statistics.In the end,our work is summarized and the potential research directions are pointed out for the drawbacks of this dissertation.