Three-dimensional Human Reconstruction Using A Single Depth Map

Three-dimensional human reconstruction refers to the use of computers to create virtual models of people in virtual scenes.In recent years,it has become a hot topic in computer vision and computer graphics research.Using various input methods and different technical routes,many existing research results have achieved good results.Among them,the use of deep neural networks to reconstruct the human body from RGB pictures is a common method.Through picture s or videos,we can obtain the reconstruction model of the target human body in real time,but there is still a lot of room for improvement in the accuracy of human body posture and posture.Traditional 3D human reconstruction technology decouples reconstr ucted human body into two sub-tasks: posture reconstruction and posture reconstruction.In order to better improve the accuracy of these two sub-tasks,this paper uses synthetic depth data to reconstruct the 3D human body.Unlike traditional RGB pictures,the data of each pixel of the depth image records the distance of the object from the camera.For tasks such as human body posture and pose estimation that are sensitive to spatial distance,depth pictures are more suitable as raw data.However,the use of deep learning for reconstruction learning requires a large amount of human deep picture data,and there is currently no public data set available.This paper uses computer graphics rendering technology to automatically synthesize massive human depth image data,and obtain a fast and accurate three-dimensional human body reconstruction model through the training of deep neural networks.In this work,the contents and difficulties of 3D human body reconstruction are explained firstly,and then several method s of 3D human body reconstruction are analyzed.Then,this paper uses Kinect to realize a traditional iterative human body reconstruction method,which can accurately reconstruct human body through depth image acquisition,point cloud preprocessing,point cloud alignment,solving fitting equations and other steps.Although the traditional iterative method has high reconstruction accuracy,the depth map preprocessing is complex,the number of iterative solutions is too many,and the time cost is large,which makes it unable to meet the needs of real-time application scenarios.In order to reconstruct 3D human body quickly and accurately,this paper proposes a 3D human body reconstruction method based on deep learning by using the strong learning ability and f ast inference ability of neural network.In this paper,a large number of depth map training data is synthesized to support the network learning,and a neural network model is designed to infer the human parameters of SMPL model.On the basis of joint point supervision,human parameter supervision and other common supervision information,a supervision method based on human mask is designed to improve the accuracy of network inference.Finally,the current 3D human body reconstruction technology is summarized and prospected.