首页|基于单-多视图优化的足球球员三维姿态和体型估计

基于单-多视图优化的足球球员三维姿态和体型估计

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
足球比赛场景的三维重建有助于观众自由切换视角，增加了互动性和沉浸感。针对足球比赛场景中的足球球员，提出一种三维姿态和体型估计方法。对球员的多视图图像使用训练好的部分注意力回归的三维人体估计(PARE)模型生成初始的三维姿态和体型估计，并使用人工标注的二维关节点作为优化目标。单-多视图优化操作利用蒙皮多人线性模型(SMPL)和正交投影的可微性，将球员的三维姿态和体型参数映射到二维关节点，计算其与人工标注之间的差异，再使用神经网络的反向传播算法更新三维姿态和体型参数，持续这些过程直到差异最小化。在自建的足球球员多视图数据集上的实验结果表明，该方法能够有效估计足球球员的三维姿态和体型，与人体网格恢复、在循环中优化SMPL、PARE等方法相比，二维关节点精度在单视图上提高了9。2%～37。5%，在多视图交叉验证中提高了34。9%～54。1%。
Three-Dimensional Pose and Body Shape Estimation of Soccer Players Based on Single-and Multi-View Optimization
Reconstructing a Three-Dimensional(3D)scene of a soccer game allows the audience to switch their viewpoints freely,enhancing interactivity and immersion.A pipeline to estimate the 3D pose and shape of soccer players in a soccer scene is proposed.The proposed method obtains an initial 3D pose and shape estimation of soccer players with a trained Part Attention Regressor for a 3D human body Estimation(PARE)model,and utilizes annotated Two-Dimensional(2D)joints of the human body as the optimization target.Single-and multi-view optimization operations utilize the differentiability of a Skinned Multi-Person Linear(SMPL)model and orthogonal projection,map the 3D pose and shape parameters to predicted 2D joints,obtain the difference with annotations,and subsequently update these parameters with a backpropagation algorithm of neural networks.These processes continue until the difference is minimized.Experimental results on our soccer players multi-view dataset demonstrate that the proposed method is able to estimate the 3D pose and body shape of soccer players.Compared with Human Mesh Recovery(HMR),SMPL oPtimization IN the loop(SPIN),and PARE methods,the accuracy of 2D joints improved by 9.2%-37.5%on single-view,and by 34.9%-54.1%on multi-view cross validation.

Three-Dimensional(3D)pose and body shape estimationparametric human body modelsingle-and multi-view optimizationbackpropagationSkinned Multi-Person Linear(SMPL)model

谢欢、刘纯平、季怡

展开 >

苏州大学计算机科学与技术学院,江苏 苏州 215006

三维姿态和体型估计 参数化人体模型 单-多视图优化 反向传播 蒙皮多人线性模型

江苏省高等学校自然科学研究重大项目江苏高校优势学科建设工程资助项目

19KJA230001

2024

10.19678/j.issn.1000-3428.0067480

计算机工程
华东计算技术研究所　上海市计算机学会

计算机工程

CSTPCD北大核心
影响因子：0.581
ISSN：1000-3428
年,卷(期)：2024.50(3)
  • 37