Three-dimensional Reconstruction Of Non-rigid Target Based On RGBD Camera

3D reconstruction technology has a wide range of application prospects in modeling and navigation,human-computer interaction,visual monitoring,virtual and reality,computeraided design and many other related fields.With the rapid development and popularization of RGBD(color image + depth image)camera in recent years,3D reconstruction technology based on RGBD camera has gradually become one of the main research directions.At present,there are still many problems in 3D reconstruction technology for non-rigid target.Therefore,this thesis focuses on the research of non-rigid target 3D reconstruction,designs and completes a non-rigid target 3D reconstruction system.The main works are as follows:(1)In this thesis,a RGBD multi-cameras optical acquisition system is built.The optical equipment uses the Azure Kinect DK camera,released by Microsoft in 2019 year.Then,the synchronous non-rigid target database of RGBD multi-cameras is constructed,which mainly includes calibration image,non-rigid target image(head,arm motion,etc).The total number of multi-frame image in this database is 3000.In addition,in view of the depth image noise problem,which is caused by the RGBD camera.This thesis researches four typical filtering algorithms and analyzes their running time and denoising effect in the above database.(2)In order to reduce the complexity of initial RGBD multi-cameras pose calibration,this thesis investigates a method for RGBD multi-cameras initial pose calibration and joint optimization based on QR code and bundle adjustment.This method consists of two parts:one is the study of QR code corner detection algorithm and DLT(Direct Linear Transform)algorithm by using the diffuse QR code calibration board for fast and robust pose estimation between multiple RGBD cameras.The other is global optimization of initial RGBD multi-cameras pose on the basis of bundle adjustment approach to further reduce the pose error.(3)Aimed at the problem of RGBD multi-cameras accurate registration,this thesis investigates a RGBD multi-cameras calibration method from the 3D point cloud perspective.We propose a ICP(Iterative Closest Point)registration algorithm based on multi-dimensional semantic mapping.This method solve the camera accurate registration problem with large angle deviation.In addition,the performance of point-to-point ICP algorithm,point-to-plane ICP algorithm and colored ICP algorithm are analyzed qualitatively and quantitatively on the database established by this thesis.According to the experimental results,the proposed method achieves high accuracy of RGBD multi-cameras pose estimation based on the colored ICP algorithm.(4)3D point cloud reconstruction and surface texture reconstruction for non-rigid target based on screened poisson and CUDA(Compute Unified Device Architecture)are researched.In the aspect of non-rigid target point cloud reconstruction,this thesis reconstructs non-rigid target point cloud model in real time and switches visual angle display by using computational imaging theory,ray tracing method and CUDA parallel computing technology.As for surface texture reconstruction,screened poisson surface reconstruction and texture mapping.Algorithm are utilized for model reconstruction and texture mapping.