Human motion recovery utilizing truncated schatten p-norm and kinematic constraints

Human motion capture (mocap) data, which records the movements of joints of a human body, has been widely used in many areas. However, the raw captured data inevitably contains missing data due to the limitations of capture systems. Recently, by exploiting the low-rank prior embedded in mocap data, several approaches have resorted to the low-rank matrix completion (LRMC) to fill in the missing data and obtained encouraging results. In order to solve the resulting rank minimization problem which is known as NP-hard, all existing methods use the convex nuclear norm as the surrogate of rank to pursue the convexity of the objective function. However, the nuclear norm, which ignores physical interpretations of singular values and has over-shrinking problem, obtains less accurate approximation than its nonconvex counterpart. Therefore, this paper presents a nonconvex LRMC based method wherein we exploit the state-of-art nonconvex truncated schatten-p norm to approximate rank. Moreover, we add two significant constraints to the low-rank based model to preserve the spatial-temporal properties and structural characteristics embedded in human motion. We also develop a framework based on alternating direction multiplier method (ADMM) to solve the resulting nonconvex problem. Extensive experiment results demonstrate that the proposed method significantly outperforms the existing state-of-art methods in terms of both recovery error and agreement with human intuition. (C) 2018 Elsevier Inc. All rights reserved.