l 1/2 Regularized RPCA Technique for 3D Human Action Recovery
M. S. Subodh Raj, Sudhish N. George · 2020
Even though human action data is widely used for various applications, the inevitable occlusions and limitations of motion capture systems lead to loss of information. Different methods have been used for recovering the lost information in motion data, among which low rank approximation gives promising results. However, the convex approximation of low rank minimization problems which uses nuclear norm minimization completely ignores the importance of larger singular values compared to the smaller ones. In this paper, we propose a new method for the recovery of lost 3D human action data using weighted nuclear norm approximation, which is a non-convex low rank approximation technique. The recovery of human action data is made possible by minimizing the weighted nuclear norm of the clean motion matrix and the l1/2norm of the error matrix by 2 effectively utilizing the aspects of Robust Principle Component Analysis (RPCA). Experimental results show that our method is efficient in recovering clean human action data from the corrupted skeletal information.