Analysis And Reuse Of Human Motion Capture Data

Human motion capture data is a new type of multimedia data and has been widely used in many areas,such as movies,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 motion capture system can obtain highly precise motion data and many applications can vividly reproduce human motion,which can bring wonderful visual feelings to users,so human motion capture data becomes more and more popular.However,professional motion capture devices are very expensive and ordinary users cannot afford.Only some big companies and research institutions can complete motion capture,so the motion capture technology cannot be widely used in human 's life.Via analysing and modeling existing motion capture data,motion reuse technology can produce new motion data according users' requirements.This effectively reduces the cost of time,money and labor.Therefore,motion reuse has attracted a lot of attentions and has been a hot research topic..Motion reuse contains a series of related technologies,such as motion segmentation,motion recovery,motion denoising,motion compression,motion retrieval,keyframe ex-traction and motion synthesis.Early related methods usually process motion data from kinematics and graphics perspectives,without modeling the special characteristics of mo-tion data.Then some machine learning based methods are proposed,but they usually directly apply classical machine learning models to motion data and the performances are not very good.Therefore,in this paper,we propose some motion data tailored machine learning methods on motion segmentation,motion recovery,keyframe extraction and motion synthesis.The main works and innovations of this paper include the following parts:(1)We propose a robust temporal sparse subspace clustering model to solve the segmentation problem of motion capture data which cotains non-Gaussian noise.Under the framework of sparse subspace clustering,we propose to use the geodesic exponential kernel to model the Riemannian manifold structure,use correntropy instead second-order statistics to measure the reconstruction error,use the triangle constraint to guarantee temporal continuity in each cluster and use multi-view reconstruction to extract the relations between different joints.Therefore,exploiting some special characteristics of motion capture data,we propose a new segmentation method which can not only complete temporal segmentation task but also suppress non-Gaussian noise.We also develop an efficient optimization algorithm with a linear complexity to solve the proposed model,while sparse subspace clustering is originally a quadratic problem.(2)We propose two motion recovery models exploiting the sparseness property of motion capture data.First,under the framework of sparse representation,we take the motion recovery process into dictionary learning methods,i.e.computing the sparse representations of incomplete frames and using them to update the complete dictionary.This makes the learned dictionary more suitable to motion recovery in theory.Then,to solve the out-of-sample problem of the above method,we propose a nonlinear low-rank matrix completion.Within the model,we use a multiple kernel learning process to find the feature space where human motion is linear and of low-rank.We can thereby use low-rank matrix completion to complete motion recovery in the learned space.In addition,two kinematic constraints are taken into these two recovery models as the prior knowledge of motion data and they can not only guarantee the kinematic properties of recovered motions,but also shrink the searching space during the optimization process.(3)We propose a joint kernel sparse representation model,which can exploit the sparseness property of human motion to complete keyframe extraction.Within the pro-posed model,we use the geodesic exponential kernel to embed human motion into Hilbert space where human motion is assumed linear and can be sparsely represented by itself.We use LP,2(0<p<1)norm instead of L1,2 norm to solve the redundancy problem of keyframes extracted by traditional sparse representation model.We assign each joint a reconstruction coefficient matrix and jointly represent them by these reconstruction coefficient matrices.Then these matrices can obtain much detailed information deeply embedded in motion data.And the triangle constraint can guarantee that each frame in the motion sequence is only represented by its neighbors,which can effectively solve the unreasonable distribution problem of periodic motions.(4)We propose a templated motion synthesis model,which simplifies the control-s of motion synthesis methods and improves the understandability of motion synthesis process.We use sparse principal component analysis(SPCA),group lasso and exclusive group lasso to model human motions and learn a set of low-dimensional parameters which respectively control an intrinsic degree of freedom(DOF)with intuitive meanings.Mean-while,our approach makes each joint controlled by as few low-dimensional parameters as possible to reduce the interferences between different DOFs.Users can control the motion features like amplitude of swing arm,kick height and jump distance by modify-ing the low-dimensional parameters intuitively in real time.This two-step approach of"template learning and template customization"can effectively reduce the complexity of synthesis control and the difficulty of using motion synthetic technology.