Research On Realistic Reconstruction Technology Based On 3D Point Cloud

The rapid development of 3D vision technology has promoted the revolutionary transformation of artificial intelligence.As an indispensable technical support for VR/AR,intelligent robot SLAM and other fields,3D reconstruction technology is more and more affecting people's daily life.It can help people digitize objects in daily life quickly and accurately,and display them in the form of three-dimensional models on the screen,giving people a real feeling.However,the traditional image-based 3D reconstruction method has defects in reconstruction speed,model accuracy and robustness,which can not meet the requirements of high-speed,high-precision and robust realistic 3D model reconstruction.In response to the urgent need,this paper proposes a realistic reconstruction method based on 3D point cloud with Kinect developed by Microsoft Corporation,and builds a corresponding 3D point cloud hardware acquisition system.The main contents are as follows:1.Acquisition and preprocessing of target 3D point cloud data.This part firstly built a 3D point cloud hardware acquisition system using Kinect depth sensor,rotating booth,background frame and PC processor.Then,with the help of the rotating booth,the point cloud data of the three targets at 16 different viewing angles are respectively collected by means of the simultaneous Kinect sensor shooting of the target object by the 22.5 degree.Finally,a redundant background segmentation algorithm based on a straight-through filter is studied in the turntable coordinate system to extract the region of interest of each target.2.3D point cloud registration in multiple viewing angles.This part firstly improves the shortcomings of the classical ICP registration algorithm,and proposes a matching optimization ICP registration algorithm based on multi-dimensional features of key points.The algorithm optimizes the corresponding point estimation process,enhances the robustness of the global registration process,improves the matching accuracy by increasing the fourth-dimensional curvature feature,using voxel filtering to reduce the number of point clouds and slow down computer computing pressure,utilizing KD-Tree neighbor search mechanism to improve the matching speed.Then,the traditional sequential incremental registration process is abandoned in the registration and splicing process of the multi-view point cloud image.instead of adopting a layered and pairwise registration strategy,which greatly reduces the cumulative number of registration errors and effectively solves the head-to-tail offset.Finally,through the comparative experimental analysis,the superiority of the registration algorithm and the registration strategy in registration accuracy,registration speed and robustness is demonstrated.3.Surface reconstruction of target 3D point cloud model.In this part,a Poisson surface reconstruction method based on point cloud geometric model estimation is proposed.The method of moving least squares and Poisson equation surface reconstruction is combined to effectively solve the existence of holes in local regions and "double wall" in overlapping regions during surface reconstruction.Then,compared with the experimental results obtained by the greedy projection triangulation method,the algorithm in this paper is better than the greedy projection triangulation method in terms of model reconstruction accuracy,local detail features,time efficiency and model file capacity.4.Realistic light rendering of the target 3D surface model.This part firstly derived a pixellevel illumination rendering model suitable for the research goal based on the ray tracing principle of human eye vision.Then based on the mathematical model,the lighting renderer is designed independently to paint the 3D surface model of Bear,Star and David through the steps of importing and identifying models,building the light models,calculating vertex normal vector,controlling target material,displaying the window,which make the 3D models exhibit significant bright shadow changes.Finally,the multiple results of contrast experiments show that the illumination renderer designed in this paper has improved the realism of the target 3D model,satisfies the human sensory requirements,and achieves the ultimate realistic 3D reconstruction visual effect.