Design And Implementation Of 3D Human Body Reconstruction System Based On A Single Photo

The increase in bandwidth and terminal computing power has inspired the transformation of the content form from text to pictures,pictures to videos on the Internet.With the continuous development of 5G,cloud computing,artificial intelligence and chip design,the next decade of the internet will belong to "three-dimensional".As for the 3D reconstruction of human body,full-body scanners and Kinect Sensor can speed the process of traditional manual reconstruction.The former is expensive,and the latter imposes requirements on the angle of view and hardware accuracy,which cannot be conducive to the promotion of 3D modeling technology as well.In this context,this paper chooses to use a single photo as input to complete the reduction of a two-dimensional image showing a human body to a three-dimensional model with the help of deep learning and SMPL.First,the extraction and alignment of human skeletal joint points in a photo was completed in three stages using Open Pose,namely,feature extraction by VGG-10,initialization and refinement of joint point confidence networks and PAFs.Next,the human body is divided into modules according to the joint information,and the modules are aligned by Platts analysis and then added to the capsule collision detector.Combined with the 3D key point fitting and posture prior distribution,the human body model parameter solution is converted into an unconstrained nonlinear optimization problem which can be solved by Powell's dogleg method.Finally,the 3D human body model is built and optimized using the solved model parameters and thin plate splines.Based on the algorithm above,this paper designs and implements a 3D Human body reconstruction System using B/S structure.The server side is responsible for the reconstruction module and database storage,the browser side is responsible for user interface interaction and model visualization,and the two sides communicate through Flask-Socket IO extension.The system test results show that the system can extract the skeletal joint points from a single photo,and build a 3D model of the human body with relative accuracy and stability.