Research On Key Techniques For 3D Reconstruction Of Human Body Using RGB-D Camera

The human body three-dimensional model is widely used in animation,games,movies,manufacturing and other fields.How to quickly and effectively establish this model is a popular research direction in the field of computer vision.Traditional human modeling methods using laser scanning and stereoscopic vision constrain the extension of their application scope due to high cost,poor operability and other conditions.With the advent of Color-Depth(RGB-D)cameras,because of their low cost,simple operation,and the ability to obtain color depth information in real time,they have opened up new areas for human body modeling,laying the foundation of hardware for their widespread use,yet there is also the disadvantage of low resolution and high noise of the depth data.In addition,the human body inevitably fluctuates during the scanning process,which poses a challenge to the acquisition and registration of human body surface data.In order to solve the above problem,this article uses two RGB-D cameras to build a 3D human body reconstruction system.The system does not require tools such as dials and guide rails.The user can only obtain the complete and textured human 3D model for oneself by standing in the middle of rotation of two RGB-D cameras.The main research content of this article is as follows:First,the working principle,types and selection of RGB-D cameras are described.Then,by analyzing the sources of deep data noise,it lays a foundation for the establishment of a three-dimensional reconstruction system for human bodies,and introduces the relevant algorithm flow.Second,defects in scanned data need to preprocess.Noise from depth data is filtered.Unstable edges from human depth data are removed.Color data and depth data is converted into point cloud with color information.Background and ground point clouds are removed remaining human point cloud.Removing outline from human point cloud improves quality of human point cloud in order to guarantees the accuracy of the follow-up algorithm and further realize the down-sampling of human point cloud,which improves the efficiency of subsequent algorithms.Third,the human point cloud be slight deformation and located at different viewing angles is registered from different coordinate systems to the same coordinate system to achieve coordinate normalization,a registration algorithm combining rigid registration and non-rigid registration is proposed.The algorithm first performs global rigid registration and non-rigid registration on the human point cloud under different perspectives of the upper body to achieve the normalization of the upper body human point cloud coordinates,and the same way to achieve the normalization of the lower body human body point cloud coordinates.Then perform rigid registration and non-rigid registration on point cloud of the upper body and the point cloud of the lower body to realize the normalization of the entire human point cloud coordinatesFourth,according to the defects that lose data of the human body point cloud,a hole-filling algorithm is proposed that a Poisson surface reconstruction algorithm is used to merge different point cloud into a watertight mesh and the algorithm of texture reconstruction based on Poisson blending reconstructs texture of surface to yield a complete and textured merged model.Fifth,display 3D reconstruction results of different human bodies.The human 3D reconstruction system was compared with other similar human 3D reconstruction system in aspect of effect,time,and human body dimensions to verify the accuracy and efficiency of the reconstruction system in this article.