1. Human motion capture data recovery based on non-convex low-rank priors of temporal difference(非凸时序差分低秩约束的人体运动捕获数据恢复算法)
- Author
-
胡文玉(HU Wenyu), 彭绍婷(PENG Shaoting), 郭震宇(GUO Zhenyu), and 黄慧英(HUANG Huiying)
- Subjects
human motion capture(人体运动捕获) ,matrix completion(矩阵补全) ,temporal difference(时序差分) ,schatten-p norm(schatten-p范数) ,non-convex optimization(非凸优化) ,Electronic computers. Computer science ,QA75.5-76.95 ,Physics ,QC1-999 - Abstract
The objective of human motion capture data recovery is to recover the missing positions of motion markers and to remove noises. The existing low-rank matrix completion based methods that usually exploit the low-rank property underlying in human motion data matrix. However, as the frame number increases, the low-rank property may not be satisfied. To characterize the low-rank property of motion data well, a non-convex temporal difference low-rank constrained human motion capture data recovery algorithm is proposed by combining the Schatten-p norm and the q-norm. First, the data matrix is projected into the temporal difference space, and the temporal difference matrix is constructed. Then we introduce the non-convex Schatten-p norm to characterize the low-rank property of the data temporal difference matrix, and the non-convex q-norm to constrain the sparse noise terms. The alternating direction method of multipliers (ADMM) is used to solve the problem, and the Newton-Raphson iterative method is used to solve the sub-problems. Comparison is made between the proposed algorithm NTDLR and three classical algorithms, i.e., TrNN, CaNN and IRNNL Lp, on CMU dataset and HDM05 dataset. Both of the recovery error and visual effect show that our NTDLR has better recovery performance.(人体运动捕获数据恢复问题旨在恢复缺少的运动标记点位置信息,同时消除噪声。现有基于低秩矩阵填充的恢复方法大多利用人体运动捕获数据矩阵的低秩性。然而,随着运动数据帧数的不断增加,低秩性可能不再满足。为更好地刻画运动数据的低秩性,提出一种联合Schatten-p范数和lq范数的非凸时序差分低秩约束(NTDLR)的人体运动捕获数据恢复算法。首先,将数据矩阵投影至时序差分空间,构造时序差分矩阵。其次,引入非凸Schatten-p范数,刻画数据时序差分矩阵的低秩性,同时引入非凸lq范数约束稀疏噪声项。再次,利用交替方向乘子法求解模型,采用Newton-Raphson迭代法求解子问题。最后,在CMU数据集和HDM05数据集上,将NTDLR与经典的TrNN、CaNN和IRNNL Lp算法进行了比较,结果表明,NTDLR算法的视觉效果更优,具有更好的恢复性能。)
- Published
- 2025
- Full Text
- View/download PDF