基于深度正則化的三維高斯人體重建算法

打開文本圖片集

中圖分類號(hào)：TP37 文獻(xiàn)標(biāo)志碼：A 文章編號(hào)：1001-3695（2025）07-041-2227-07

doi：10. 19734/j. issn. 1001-3695.2024.09.0368

Abstract：Duetothelackofaditionalgeometricconstraintsandpriorknowledge，thereconstructionresultsofexistingmultiview 3D humanbodyreconstruction methods are poor interms ofqualityandcompletenes.In response tothe above problems， this paper proposeda3DGausianreconstructionalgorithmDHGS forsparse views.Firstly，itimproved estimationmethod for humanbodymodelparametersbycombiningmulti-viewjointreprojectionandintersection-over-unionror，utilizedaccurate bodypriors toiitialize3DGaussianmodel.Secondlyitproposedanadaptivedepthadjustmentmodulethatincorpoateddepth estimationmodel，acieveddepthegularizationthroughdiferentiablerasterizationrendering，ndeancedthegeometriconsistencyofthereconstructionbyleveraginghumanbodyanddepthpriorknowledge.Finall，itgeneratedsyntheticpseudo-views duringtheoptimizationprocess toenforceadditional geometricconstraints.ExperimentalresultsontheZJU-MoCap，GeneBody， and DNA-Rendering datasetsshowthatthe DHGS algorithm achieves PSNRof 26.13dB，24.87d，and 25.25dBforimage reconstruction，represented improvements of 27.3% ， 32.6% ，and 17.4% over the original 3DGS algorithm.The experiments validateteeffectivenessofthealgorithm，withtheDHGSmodelbeingcapableoftraininginjust5minutestorenderhighquality 3D human body images in real time.

Keywords：3Dhuman reconstruction；3D Gaussiansplatting；depth regularization；diferentiablerasterizationrendering

0 引言

三維人體重建旨在生成具有真實(shí)外觀的人體三維模型，該技術(shù)在元宇宙、人機(jī)交互、虛擬和增強(qiáng)現(xiàn)實(shí)、游戲與電影制作等領(lǐng)域有著廣泛的應(yīng)用前景。（剩余17107字）