Study On 3D Reconstruction Of Human And Extraction Of Skeleton Based On Multi-view Image Sequence

In recent years,3D reconstruction technology has become an important research issue in the field of computer vision,reverse engineering,virtual reality and so on,and it is also an important part of computer graphics.With the continuous development of science and technology,people are eager to find a kind of 3D reconstruction technology,which has the advantages of short operation time,less manual intervention and good effect.Therefore,the point cloud scanning technology based on 3D scanner and 3D reconstruction technology based on multi-view image are proposed.However,the 3D scanner is expensive,which limits the size of the reconstructed object,and it is difficult to scan the object in motion.The 3D reconstruction technology based on multi-view images can not only restore the scene structure and visual information,but also do not need to go through the complex modeling process and the global illumination calculation,and the limitation of the modeling object is relatively small.At present,the 3D reconstruction technology based on multi-view images has been applied to the 3D geometric model representation of large buildings.But for the movement of the human body,it is still a question of high research value to reconstruct the 3D structure accurately,and it is a great challenge to optimize the surface structure of the reconstruction model and extract the effective skeleton information.Therefore,this paper studies the 3D reconstruction of human bodies in motion.In this paper,we reconstruct the human models in consecutive frames,extract and optimize their skeletons based on the multi-view images collected by light field equipment.This paper focuses on the denoising method of the mesh models in continuous frame and the optimization method of the skeleton sequence.The concrete work is as follows:Firstly,based on the light field equipment,we collect multi-view images of the moving human body.We reconstruct the dense point cloud based on the multi-view image set of each frame.On this basis,this paper uses the poisson surface reconstruction method to reconstruct the surface of the point cloud.Secondly,based on the method of mesh denoising via L₀ minimization,we propose a denoising method considered the constraint of the corresponding points in the mesh models in continuous frame.In this method,the distance constraint of the matching points of the adjacent frame mesh model is added to the L₀ minimization model.For the abnormal data appearing in the mesh in individual frame,our method has a certain effect of denoising and optimization.Thirdly,we propose a skeleton optimization method via L₀ minimizing for the skeletons in continuous frame during human motion.Our method is for the skeleton sequence that is extracted from the human point cloud model of continuous motion,and the position of each skeleton point is optimized by considering the constraints of the skeleton points in the consecutive frames. L₀ norm directly represents a sparse solution,and we found that a large part of the skeletons in the adjacent frame of the continuous motion of the human models is coincident,and there is a strong correlation between the moving points and the corresponding points in adjacent frames.Therefore,this paper applies the L₀ minimization to optimize the skeleton sequence extracted from the moving human point clouds.Our method not only refers to the position of the original skeleton point,but also takes into account the position of the adjacent skeleton points,so as to get a more smooth and accurate human skeletons.