Research On Three-dimensional Human Body Reconstruction Method For Virtual Try-on With Multiple RGBD Cameras

With the advancement of the digital economy and the customization development in the clothing industry,virtual try-on technology has provided convenience for online clothing purchase and remote customization.Three-dimensional virtual try-on,which enables precise control of body and garment deformation,is gaining increasing attention.The accurate and efficient establishment of a three-dimensional human body model is a crucial prerequisite for achieving realistic three-dimensional virtual try-on,and it plays an important role in subsequent automatic measurement of body dimensions,garment pattern generation,and virtual clothing presentation.The emergence of consumer-grade depth(RGBD)cameras has provided a solution to the problems of expensive prices and complex operations associated with traditional three-dimensional body data acquisition methods such as laser scanners.However,RGBD cameras have limited visual range,resulting in low overlap of point cloud data captured from different angles,making it difficult to achieve precise registration.Furthermore,the acquired human body point clouds have complex surface morphology and are subject to noise interference,making it challenging to preserve fine geometric features of the human body surface during reconstruction.Therefore,this paper focuses on the following research regarding the three-dimensional reconstruction of virtual try-on using multi-angle RGBD cameras.1)To address the problem of partial overlap in multi-angle human body point cloud registration,a point cloud registration method based on overlap point filtering is proposed to improve registration accuracy.Firstly,in order to better extract point features and establish good point correspondence,a point feature descriptor is designed for point feature enhancement.The point features are integrated using self-attention modules based on all input points in the point cloud to better fuse global features.Then,to tackle the low overlap and noise issues in multi-angle human body point cloud data,an effective overlap point selection is performed through similarity matrix computation.This allows for rigid transformation,removal of noise points,and elimination of erroneous point correspondences caused by non-overlapping points,resulting in accurate rigid transformation and improved registration accuracy of partially overlapping human body point clouds.Experimental results demonstrate that the point cloud registration method based on overlap point filtering outperforms existing state-of-the-art algorithms significantly in datasets with no noise,low noise,and high noise,providing a complete human body point cloud model for subsequent surface reconstruction of human body point clouds.2)Considering the complex geometric features of the human body surface and the presence of noise interference in point cloud data,a three-dimensional human body surface reconstruction method based on adaptive filtering is proposed.Firstly,point cloud filtering is performed prior to mesh reconstruction using an adaptive point cloud filtering algorithm based on dynamic graph representation learning.This accurate denoising process recovers fine geometric details of the human body surface,enhancing the efficiency of subsequent point cloud mesh reconstruction.Then,to address the issues of low efficiency and poor quality in mesh reconstruction,a point cloud mesh reconstruction algorithm based on triangle growth criteria is proposed.It involves selecting growth points and constructing initial triangles based on criteria for departure points,improving the stability of subsequent triangle mesh growth.By projecting neighboring points onto a two-dimensional plane and designing triangle growth criteria,the best growth points that meet the conditions are selected,achieving higher-quality three-dimensional reconstruction of the human body and improving the efficiency of point cloud mesh reconstruction.Experimental results demonstrate that the proposed three-dimensional human body surface reconstruction method based on adaptive filtering accurately filters noise,more comprehensively describes the detailed features of the original human body data,and generates human body models that better reflect the surface structure of the human body.The generated number of triangular facets also meets expectations,exhibiting excellent performance in terms of quality and efficiency.3)A prototype system for three-dimensional human body point cloud reconstruction for virtual try-on is designed and developed.A data acquisition platform for threedimensional human body point cloud reconstruction for virtual try-on is built,and human body point cloud data is captured on-site using the established hardware platform.This enables three-dimensional human body point cloud reconstruction based on multiple RGBD cameras.The research work is validated through practical examples,providing an effective method for convenient implementation of three-dimensional human body point cloud reconstruction for virtual try-on.