Advanced Technology Research On Multi-view Real-time 3D Reconstruction Of Human Body

The era of full 3D reconstruction of objects has arrived in recent years with colordepth(RGB-D)sensors becoming more common.The accurate and full 3D reconstruction of objects has become a research topic that has received extensive attention.The majority of methods in full 3D reconstruction still focus on static objects.However,the accurate and full 3D reconstruction of moving objects is a challenging task.In this paper,in order to achieve real-time,high-quality and full 3D reconstruction of the moving objects,we research on the accuracy,robustness,and time complexity of reconstruction algorithms.The main work is as follows:1.The existing multiple RGB-D sensor calibration methods are studied and compared.Since the existing calibration methods need to be operated by multiple times,the captured calibration images may be not synchronized and the structure of the calibration objects is complex and bulky,a novel method is proposed that a calibration object is based on the triangular pyramid structure with multiple fixed markers.The result of calibration is accomplished by accurately positioning multiple fixed markers,and is optimized by the ICP registration method.The method overcomes the problem of accumulated errors due to multiple operations in conventional calibration methods,calibrates multiple RGB-D sensors accurately and simultaneously and can be done in one step.2.The existing global registration methods are studied and compared.Because the depth images of RGB-D sensor have lower accuracy and are affected by environmental changes,a novel global registration,NWGPA-ICP(New Weight on Global Procrustes Analysis and Iterative Closest Point),is proposed.The method uses grid-based motion statistics and feature matching of color to set the weights of GPA-ICP(Global Procrustes Analysis and Iterative Closest Point)global registration.It can solve the problem of poor accuracy and stability of registration led by inaccurate depth images through controlling the confidence of matching pairs.3.The existing color processing methods are studied and compared.Because the existing color processing methods can not accurately and stably process the color difference problem of the adjacent view of RGB images,a novel color correction and blending method is respectively proposed in order to achieve color balance between adjacent views.The method of color correction adjusts the light to fit the color point pairs non-linearly between adjacent point clouds.It can better maintain the consistency of color between adjacent RGB views from a global region.The method of color blending uses distance transform of foreground area to set the weights of each point of point cloud.It can more stably blend color differences of all the same local regions.In order to evaluate the accuracy,robustness and time complexity of the reconstruction algorithm proposed in this paper,a system of multi-view real-time full 3D reconstruction of human body based on low-precision sensors and limited computing resources is built in a natural environment.This system combines with triangular pyramid calibration method,NWGPA-ICP global registration algorithm,color correction and blending algorithm.Experimental results show that the reconstruction system of this paper achieves high-quality,robust and real-time full 3D reconstruction of human body.