Research On 3D Geometry,Appearance And Deformation Reconstruction

With the development of Virtual Reality(VR),Augmented Reality(AR),and 3D display technologies,3D content data,especially the 3D model data,has been widely used in human's work,life and production activities.The traditional 3D model data is typically used to describe the 3D shape of the real objects.But the complete 3D model data should also contain the information of the appearance as well as the physical properties(such as elasticity,density,etc.)of the real objects.In many applications based on 3D model data such as inverse engineering and scene reconstruction for autonomous driving,a key issue is how to quickly generate high-quality data.The conventional way is handcrafted by professionals using 3D software,but it is time-consuming and difficult to create high-fidelity data.Thus,it is of great research significance to quickly generate complex,high-fidelity 3D model data by reconstructing objects in the real world.Our research focuses on generating high-quality 3D model data,especially the shape data from geometry reconstruction,the appearance data from appearance modeling,and inner elastic parameter data from parameter optimization.The major innovations and contributions of this paper are described as follows:1.Typical surface geometry reconstruction systems are studied in this paper.In order to improve the accuracy of the reconstruction system,this paper proposes a supervised rectification method in terms of depth camera distortion and implements a weighted TSDF-based camera tracking algorithm,which yields a high accurate single depth sensor based reconstruction system.On the other hand,in order to improve the robustness of the reconstruction system,this paper develops a reconstruction system that integrates the localization sensor.To utilize the sensor,this paper proposes a multi-sensor fusion algorithm with adaptive weight adjustment,which improves the robustness of the system without accuracy loss.2.Conventional texture mapping algorithms often suffer from texture misalignment and other visual artifacts when the input geometry model or the image registration is inaccurate.For this problem,this paper proposes a robust and fast texture mapping correction algorithm.The algorithm innovatively uses alternative local/global texture mapping optimization framework,which effectively eliminates the influence of inaccurate input,and achieves significant speedup compared with the previous state-ofthe-art algorithm.Classical surface light field technology relies on high-precision acquisition equipment.If the input is inaccurate,the light field rendering results often appear blurring.Aiming at this problem,this paper proposes a robust surface light field reconstruction algorithm based on patch optimization.By optimizing the sampled patches,high-precision surface light fields can be reconstructed using lowprecision equipment such as mobile phones.3.With large parameter space and a large number of sampling data,the conventional elastic parameter estimation method suffers from computational complexity and long computational time.In terms of this issue,this paper introduces two improvements.The first one is a two-step elastic parameter estimation method with inexact iteration.The presented inexact convergence analysis proves that the relative error threshold exists in the inexact two-step method,which ensures the inexact solution can be solved using a fixed number iteration solver.The inexact two-step method reduces the number of iterations in a single step and shortens the overall calculation time.The second improvement is a fully automatic and adaptive online data acquisition method.This method produces the next optimal sampling position at while current sampling.So,this method can effectively reduce the number of sampled data while keeping the similar sample quality,which finally reduces the computational complexity of the optimization.