Research And Application Of Image-oriented 3D Reconstruction Of Non-rigid Body

In recent years,most of the research on 3D reconstruction of objects has been on rigid objects whose shape and size will not change.Non-rigid bodies also are the important part of the real world,the early methods of 3D reconstruction are based on rigid body 3D reconstruction.These methods try to reconstruct the model as a whole,which is prone to model reconstruction deviation,which makes it difficult to restore the local characteristics and physiological form of non-rigid bodies.Based on this,this thesis takes the study of non-rigid birds as an example,analyzes and studies existing data sets,redefines bird skeleton and key point information,and designs a non-rigid 3D reconstruction algorithm based on local stiffness energy optimization.Comparing the existing algorithms and conducting experiments,this algorithm has achieved the effect of non-rigid body 3D reconstruction with better detail restoration.The main research content of this thesis is divided into the following parts:(1)A deformation algorithm for the local stiffness of non-rigid bodies is proposed.Through the analysis of the bird's skeleton,this thesis redefined the key points of the bird,and highlighted the key parts that are prone to deformation.At the same time,based on the definition of the rotation matrix and the local stiffness,combined with the shell energy function optimization method,a deformation algorithm for the local stiffness of the nonrigid body is designed.This algorithm minimizes the stiffness energy of each part according to the weight distribution of the deformation,so that the small deformation of the local part of the model better fits the local characteristics of the image,and does not affect the deformation of the rest of the model.Comparative experiments show that the Percentage of Correct Key-point(PCK)is on average 2.55% higher than the comparison algorithm model,which verifies the promotion of local stiffness to the local deformation of the non-rigid model.(2)An image-oriented non-rigid 3D reconstruction algorithm is proposed.The deformation algorithm based on the local stiffness of non-rigid bodies aims to reconstruct a complete non-rigid body.This algorithm combines the non-rigid deformation module,the camera pose prediction module and the texture prediction module,and preprocesses the optimized data set by minimizing the joint loss function,execute the average model deformation of birds,reconstruct the three-dimensional model of birds under the constraints of key point annotation and local non-rigid deformation,and then fit the physiological characteristics of the non-rigid species of birds.This algorithm is better than the average model in the Intersection over Union(Io U),and the Percentage of Correct Key-point(PCK)is higher than the local algorithm by 3.01%,which proves that the algorithm is achieved in the image-oriented non-rigid 3D overall reconstruction a positive effect.(3)An image-oriented non-rigid 3D reconstruction system is designed and implemented.Based on the algorithm framework of this thesis,the specific requirements and significance of non-rigid body reconstruction are analyzed,and the functions of each module under the system as a whole are designed,and a Django-based image-oriented non-rigid body 3D reconstruction web application system is realized.This system can not only show users the existing model reconstruction results,but also provide the function of realizing three-dimensional reconstruction of custom uploaded images.In addition,the system provides users with a rich interactive interface and a good interactive experience.This thesis focuses on the local deformation of non-rigid bodies,so that the reconstruction model can achieve the effect of non-rigid deformation in certain indexes.And we designed and implemented an image-oriented non-rigid 3D reconstruction system and forms an application.