Research On Human 3D Reconstruction System Based On Kinect Multi-view Point Cloud

With the rapid expansion of earth observing technology and the perpetual upgrading of artificial intelligence technology,the application of 3D point cloud data is becoming more and more diversified.As an important research direction of virtual reality technology and machine vision technology,point cloud registration and 3D reconstruction technology occupy an important position in major fields such as aerospace,automatic driving,cultural relics protection,simultaneous positioning and map construction,and anthropometry.Along with the fast development of Li DAR,Kinect and other high-precision sensors and 3D reconstruction technology,people pay more attention to the accuracy and authenticity of 3D models.This thesis designs and implements a 3D reconstruction system of human body based on Kinect and multi-view point cloud to solve this practical application problem.In addition,the following points are studied in depth:(1)Aiming at the problems of high cost,large volume and low accuracy of human body reconstruction model in the current non-contact anthropometric system,a 3D reconstruction system of human body based on Kinect and multi-view point cloud is designed.The system integrates hardware acquisition,server-side processing and terminal display.A hardware acquisition device is designed using a high-precision robotic arm and an Azure Kinect DK depth camera,and a multi-view acquisition scheme for the human body is given.The 3D reconstruction model of the human body is obtained by registering the multi-view point cloud of the human body.The test results show that the system is feasible and reliable,the 3D reconstruction of the human body has a high degree of restoration,has certain details of the human body surface,and has good research value and application prospects.(2)This paper introduces a novel point cloud registration approach,which overcomes the low registration accuracy associated with conventional methods.The proposed method is based on based on LIS(Locally Intrinsic Shape Signature of Histograms of Orientations)feature descriptor.To solve the problem that point cloud down-sampling will destroy the local structure of point cloud,a voxel down-sampling method based on spatial position is proposed by resegmenting voxels and positioning the spatial position.It is a technique that decreases the data size of a point cloud while maintaining the essential local features of the original data.In view of the fact that point cloud feature representation information is not enough to solve the problem of matching corresponding points,this paper introduces the LIS feature descriptor in detail,and verifies the local superiority of the descriptor through comparative experiments.Finally,through coarse registration experiments,fine registration experiments and comparative experiments,it is shown that the registration accuracy of this method is higher.(3)This paper suggests a new approach for improving the accuracy of point cloud registration that is based on deep learning.Specifically,it introduces a registration network that utilizes overlapping matching to enhance its performance.In view of the fact that the local information in the point cloud feature information extracted in the neural network is more conducive to point cloud processing,this paper proposes a local feature extraction based on clustering,which fuses the LIS feature descriptor with the basic point cloud features,and clusters.The point cloud is divided into multiple clusters,so that the obtained point cloud features contain more local geometric information.Aiming at the problem of wrong response estimation caused by using single-point feature distance as matching score,this paper proposes feature processing based on overlapping matching.By calculating the overlapping matching score and applying it to the feature distance matrix,the matching of corresponding points is more reliable.The experimental outcomes demonstrate that the suggested approach delivers superior accuracy,enhanced generalization capability,and greater noise resilience when performing registration tasks on point clouds.