| In the context of 5G and future 6G networks,holographic communication applications are considered to be one of the hottest applications.Holographic communication systems transmit local multidimensional human sensory information to be reproduced at a remote location to provide immersive communication experience.One of the key topics in this field is the 3D reconstruction of the human body on the local acquisition side.With the proliferation of lightweight depth camera devices,reconstruction applications with multiple views and simultaneous scanning of multiple devices are emerging.Alignment and data fusion between multi-view depth cameras are two major difficulties in this field.The traditional marker-based alignment process is cumbersome and requires manual intervention,making it a poor experience in practice.At the same time,the fusion algorithm of multi-view data still has room for improvement in terms of memory footprint and accuracy.This paper investigates and implements a point cloud alignment algorithm based on a combination of graph feature learning and optimisation as well as a multiview data fusion algorithm based on TSDF and hash tables,and implements a Kinect-based human 3D reconstruction system to address the limitations of existing human 3D reconstruction techniques.After all,the main work of this paper is as followed:(1)A point cloud alignment algorithm based on a combination of graph feature learning and optimisation is proposed to address the problem of the cumbersome process of existing marker-based alignment.The algorithm uses a deep learning network to extract point features and learn point cloud structures to construct graphs,and learns matching relationships between point clouds through graph matching to calculate the positional transformations between depth cameras.Finally,the positional transformations are accurately adjusted using optimisation.The proposed point cloud alignment algorithm based on a combination of graph feature learning and optimisation gives more advantageous results over both optimisation-based approaches and several representative feature learningbased algorithms on the ShapeNetSem dataset and real-world application scenarios.(2)A multi-view data fusion algorithm based on TSDF and hash tables is proposed to address the problems of poor real-time performance,high space occupation and low quality of implicit surface generation of existing multi-view data fusion algorithms.The proposed algorithm converts unordered point clouds into ordered voxels by introducing the voxel hash table technique for storing and manipulating voxel data,and improves the hash table function to achieve faster voxel data fusion operations.The algorithm achieves more accurate results in implicit surface extraction by improving the Marching Cube algorithm.(3)Based on the two algorithms mentioned above,this paper designs and implements a Kinect-based 3D reconstruction system for the human body.The system consists of a data acquisition module,a point cloud alignment module and a data fusion module,of which the two algorithms proposed in this paper are an important part.The system has been tested and validated in practical application scenarios to demonstrate the effectiveness and practicality of the two algorithms proposed in this paper,providing a new automatic fusion solution for the acquisition side of the holographic communication system. |
