基于3D高斯溅射的三维人体重建算法研究

Research on 3D Human Reconstruction Algorithms Based on 3D Gaussian Splatting

摘要: 三维人体重建在数字人、动作捕捉与沉浸式交互等应用中具有重要价值，但真实场景下衣物褶皱、软组织位移及快速运动带来的非刚性形变，使得仅依赖参数化人体模型的模板驱动方法难以获得稳定且细节贴合的重建结果。本文提出形变修正的高斯人体模型DCGaussianAvatar，结合3D高斯溅射的高效渲染能力与SMPL结构先验，实现可驱动的人体重建与新视角合成。该方法首先在SMPL网格表面初始化高斯基元，并通过姿态编码器提取全局姿态嵌入；在此基础上，设计姿态条件的非刚性形变矫正模块，以高斯中心位置嵌入与姿态嵌入构成的高级特征为输入，分别对蒙皮权重与外观参数进行残差式更新，从而补偿衣物与软组织等非刚性细节带来的残差误差。在ZJU-MoCap数据集上与多种代表性方法进行对比，结果表明DCGaussianAvatar在PSNR上取得最高值30.93，同时SSIM达到0.964、LPIPS为0.032，并在衣物褶皱与关节区域呈现更稳定的对齐效果与更清晰的细节纹理。

Abstract: 3D human reconstruction is of great importance for applications such as digital humans, motion capture, and immersive interaction. However, non-rigid deformations caused by clothing wrinkles, soft-tissue motion, and fast body movements in real-world scenarios make it difficult for template-driven methods that rely solely on parametric human models to achieve stable reconstructions with well-aligned fine details. To address this issue, we propose DCGaussianAvatar, a deformation-corrected Gaussian avatar model that combines the efficient rendering capability of 3D Gaussian Splatting with the structural prior of SMPL for animatable human reconstruction and novel-view synthesis. Our method initializes Gaussian primitives on the SMPL mesh surface and extracts a global pose embedding via a pose encoder. Building on this, we design a pose-conditioned non-rigid deformation correction module, which takes high-level features formed by concatenating the embedded Gaussian centers and the pose embedding, and performs residual updates for both skinning weights and appearance parameters to compensate for residual errors introduced by non-rigid details such as clothing and soft tissues. Experiments on the ZJU-MoCap dataset against multiple representative baselines show that DCGaussianAvatar achieves the best PSNR of 30.93, along with an SSIM of 0.964 and an LPIPS of 0.032, and delivers more stable alignment and sharper fine-grained details, particularly around clothing wrinkles and articulated joints.