Design And Implementation Of 3d Reconstruction System Based On RGB-D Camera

Compared with 2D images,3D reconstruction can more truly reflect the spatial location information and three-dimensional shape of objects.When collecting object depth information with RGB-D camera,there are a lot of noises and cavities in the output depth map due to the limitation of physical hardware,surface characteristics of objects and the influence of external illumination.Traditional filtering algorithm for denoising has been unable to meet the demand for prediction of large range of depth information,so people begin to turn their attention to the field of deep learning.Existing depth map completion methods can train the depth collected by RGB-D camera well,but cannot predict the depth not collected by the camera.Therefore,it is of great research significance to train a completion network that can calculate all pixel depth values and apply it to the general process of 3D reconstruction,so as to build a 3D reconstruction system with higher accuracy.The main contents of this thesis are as follows:First,the process of depth map restoration is subdivided into depth map denoising and depth map void completion,use the traditional denoising algorithm to remove most of the depth map noise,and using the deep learning approach,through high precision deep hole depth map figure true value data set training completion networks,in depth map are applied to solve the missing pixels achieve depth map hole completion.At last,the integrity of the reconstructed model is improved.Second,in the point cloud registration process,the 3D point cloud obtained by depth map transformation is first registered rough,and then the initial point cloud obtained by rough registration is used as the input of ICP algorithm for fine registration to improve the speed and accuracy of high point cloud registration and ultimately improve the accuracy of reconstruction model.Third,combined with PCL,OpenCV and other library functions,using the RGB-D camera Kinect v2 to obtain data,and then through the realization of depth map denoising,depth map void completion,point cloud processing and surface reconstruction,finally achieve a high-precision 3D reconstruction system for the indoor environment.In general,this thesis designed and implemented a 3D reconstruction system based on RGB-D camera,verified the functions of each module of the system through tests,improved the accuracy and integrity of the reconstruction model,and finally achieved the expected goal of this thesis.