Analysis And Research Of Human Motion Capture Data Based On Machine Learning

Human motion capture(mocap)data is a new type of multimedia data and has been widely used in many areas,such as movies7 computer games and robots.With a motion capture system,human motion is recorded as a frame sequence and each frame consists of the orientations or positions of all the joints in a human skeleton.The professional mocap system can obtain motion data with high precision and fidelity,but the high cost of mocap system prohibits its widespread use.Human motion data resuing,which aims to generate new motion data according to the requests of users using the data acquaired,has attracted a lot of attentions.After capturing,data may flow through a series of processing precedures such as mo-tion recovery,motion denoising,motion segmentation,motion compression and keyframe extraction.These indispensable precedures are of great significance to the final data syn-thesis and resuing.Early methods trackle these problems via kinematic and graphic stud-ies while ingoring the properties of mocap data.Recently,many researchers have turned to machine learning algorithms for help.However,applying these algorithms directly to use exhibits limited performance due to the nonlinear essence and complex structure of mocap data.Therefore,how to integrate the structural characteristics embedded in human motion within the algorithms is our main concern.This paper focuses on building tailored machine learning methods for mocap data,by exploiting the spatial-temporal properties and structural characteristics,we propose methods which achieve state-of-art performance motion recovery and segmentation respectively.The major contribution of this paper lies in two folds:(1)We propose a nonconvex LRMC based method for human motion recovery wherein we exploit both the state-of-art nonconvex truncated schatten-p norm and two impor-tant kinematic constraints.The proposed method utilize the low-rank prior of mocap data which solves the out-of-sample problem existed in many learning-based recovery methods.Based on the fact that existing LRMC based recovery methods all exploit the convex nuclear norm to approximate rank,we propose to use the state-of-art truncated schatten-p norm for approximation.Moreover,to preserve the kinematic characteris-tics and structral information of mocap data,we take two kinematic constraints(i.e.smoothness constraint and bone-length constraint)into the proposed model.Smooth-ness constraint is used to preserve the spatial-temporal stability of human motion and bone-length constraint can effectively avoid the unreasonable bone-length in the recov-ery results.To the best our knowledge,there is no existing LRMC based method have exploited nonconvex surrogate nor considered these two constraints simultaneously.We also develop a optimization framework based on alternating direction multiplier method(ADMM)to solve the resulting nonconvex problem.(2)We present a nonconvex low-rank kernel sparse subspace clustering method for human motion segmentation.Current existing segmentation methods all use predefined kernels to model the non-linear essence of mocap data,and conduct segmentation in the mapped high-dimensional feature space.However,due to the implicity of kernel functions,the data after mapping to feature space have no guarantee to be linear.Based on the above fact,we propose to learn a low-rank kernel fuction for mocap data,wherein we add a low-rank regularization term in the objective function to guarantee the multiple low-dimensional subspace structures of the mapped data in the feature space.To solve the resulting rank minimization problem,we exploit the weighted schatten-p norm to approximate rank.Since the related conclusion of solving the resulted nonconvex problem remains blank in the open literature,we also deduce the related lemma and provide the detailed proof in the appendix.Notably,the proposed lemma is general and can be fit to many other nonconvex rank surrogates.