Research On Feature Representation And Synthesis Of Human Motion Dat

As a new type of multimedia data,motion capture data has been widely used in many fields,such as movie production,computer games and medical rehabilitation.This kind of data records the position and orientation of human joint points at each time and constitutes the whole motion frame sequence.Because the position information of human body captured by it is very accurate and it can display the movement of character model in various complex scenes,motion capture data gives consumers strong visual impact,it has won the favor of a large number of users.However,professional motion capture equipment is very expensive and ordinary users cannot afford it,so only large institutions or companies can complete the task of motion capture.Motion synthesis technology can use the existing motion data to synthesize the user-wanted motion sequence through analysis and modeling,so it has attracted a lot of attentions.However,due to the high degree of freedoms and complex structure of motion capture data,motion feature learning methods are needed to obtain temporal and spatial information in motion data to complete motion segmentation,keyframe extraction,motion synthesis and so on.Some early motion feature learning and motion synthesis methods are mostly based on the theory of statistics,and the kinematic characteristics of motion data are ignored.Later,some researchers proposed relevant models based on machine learning,but only applied some classical machine learning methods to motion data,and did not achieve good results.Therefore,this thesis utilizes the important characteristics of motion data to design a depth model which is suitable for human motion capture data for motion feature representation and motion synthesis tasks.The specific work includes two aspects:(1)Motion feature representation learning based on local self-representation subspace learning: this thesis propose a novel temporal self-supervised learning model for motion capture data.Within the model,this thesis transforms the individualities of motion data and segmentation task as supervisions to guide the subspace learning.Specifically,considering the temporality of motion data,this thesis uses the temporal convolution module to extract temporal features.To implement the local validity of self-expression in temporal tasks,this thesis designs the local self-expression layer which only maintains the representation relations between temporally adjacent motion frames.To simulate the interpolatability of motion data in the feature space,this thesis imposes a group sparseness constraint on the weights of the local selfexpression layer to select keyframes to represent the whole motion sequence.Besides,based on the subspace assumption,this thesis proposes the subspace projection loss,which is induced from distances of each frame projected to the fitted subspaces,to penalize the potential clustering errors.In the experiments of synthetic data,through visualizing the learned explicit features we verify the linear feature learning ability of our model.In the experiments of real motion capture data,the superior performances on both motion segmentation and keyframe extraction demonstrate the effectiveness of the proposed model.(2)Motion synthesis based on start-end frame pairs and motion pattern: In order to improve the controllability and comprehensibility of motion synthesis,this thesis propose to relate the start-end frame pair and motion pattern by learning the nonlinear mapping relationship between them.The proposed model outputs the corresponding motion sequence according to the given Start-end frame pairs.The model is composed of motion pattern extraction network and motion synthesis network.The function of network based on LSTM is to extract the motion pattern efficiently to solve the problem of high time cost of the optimization based algorithm.The motion synthesis network extracts the feature information contained in the start-end frame pairs by the synthesis module,and learns the nonlinear relationship between the start-end frame pairs and the motion patterns to solve the theoretical deficiency of existing methods using joint feature space modeling.The accurate and efficient motion pattern extraction results and high-quality synthesized motion sequences in the experiments verify the effectiveness of the proposed model.