摘要
从观测到的人体运动捕获数据中恢复缺失数据是一个重要的研究问题.传统方法利用矩阵核范数和矩阵l1范数分别对运动捕获数据潜在的低秩性和噪声稀疏性进行约束,并额外添加时序光滑项,构建的目标函数包含3个正则化项.为更好刻画运动数据的低秩性并简化模型,提出一种利用时序差分低秩先验(Temporal Difference Low-rank Prior,TDLRP)且目标函数只含2个正则化项的人体运动捕获数据恢复模型;在保证收敛的前提下,采用交替方向乘子法求解模型.在公开的CMU数据集和HDM05数据集上与现有算法进行比较,实验结果验证了TDLRP算法具有较好的恢复性能.
Recovering missing data from observed human motion capture data is an important research issue.The traditional method uses the matrix nuclear norm and the matrix l1 norm to constrain the potential low-rank property and noise sparsity of the motion capture data,and adds an additional time series smoothing term.The constructed objective function contains three regularization terms.In order to better describe the low-rank property of motion data and simplify the model,a human motion capture data recovery model using temporal difference low rank prior(TDLRP)and the objective function only contains 2 regularization terms is proposed;under the premise of ensuring convergence,using the alternating direction multiplier method to solve the model.Compared with the existing algorithms on the open CMU dataset and HDM05 dataset,the experimental results show that the TDLRP algorithm has better recovery performance.
作者
郭震宇
邱熙雯
赖烨辉
胡文玉
GUO Zhenyu;QIU Xiwen;LAI Yehui;HU Wenyu(School of Mathematics and Computer Science,Gannan Normal University,Ganzhou 341000,China)
出处
《赣南师范大学学报》
2023年第3期41-49,共9页
Journal of Gannan Normal University
基金
国家自然科学基金项目(62266002)
江西省自然科学基金项目(20224BAB202004)
赣南师范大学研究生创新基金项目(YCX21A002)
赣南师范大学学位与研究生教育教改研究项目(YJG-2021-13)
关键词
运动捕获
低秩矩阵补全
低秩稀疏分解
时序差分
交替方向乘子法
motion capture
low-rank matrix completion
low-rank and sparse decomposition
temporal difference
alternating direction method of multipliers
