Research On Automatic Generation Of Labanotation Based On 3D Human Motion-captured Data

In recent years,based on the modern information technology,the digital recording and storage of excellent traditional cultural resources has become an important research direction.Labanotation,as a scientific symbolic motion recording system,is a powerful tool to record and protect the dynamic arts such as folk dance,opera and martial art.However,the main way to obtain the Labanotation is to draw manually by the professionals,which requires a high cost of time and labor.It is of great significance to realize the automatic generation of Labanotation through computer technology,which can greatly improve the efficiency.Therefore,this thesis studies the automatic generation method of Labanotation based on 3D human motion data.According to the theory of motion segmentation and unit movement recognition in Labanotation,the generation process is divided into two steps.Firstly,the Laban segmentation algorithm of time series motion data is studied to segment the continuous limb motion and obtain the unit movements which can be represented by Laban symbols.Then,the recognition algorithm of upper limb and lower limb movements is studied to determine the corresponding symbol,so as to generate the Labanotation.The main researches and contributions of this thesis are summarized as follows:(1)A Laban segmentation algorithm for time series limb motion is proposed.The movement of upper and lower limbs of human body has coordination.At the same time,the movement of upper and lower limbs in Laban analysis is different because of its different relationship with the center of gravity of human body.Therefore,based on the synergy,the limb movement is divided into different behavior segments as a whole,and then the upper and lower limb movements in each behavior segment are divided based on the difference.Firstly,in the process of behavior segment segmentation,subspace clustering algorithm based on elastic network regularization constraint is used to segment limb behavior segments by association between adjacent frames of time series data.And then,in order to enhance the robustness of upper limb motion segmentation,the temporal analysis of the change of velocity and the spatial analysis of the change of Laban orientation are used to segment the upper limb time series data.In order to improve the quality of lower limb motion segmentation,the lower limb time series data are segmented at two levels,from analyzing the change of motion trend to modeling the unit movements in the trend using Gaussian mixture model.Experimental results show that the proposed segmentation algorithm can segment the time series limb motion data according to Laban theory,and has higher segmentation accuracy than the existing methods.(2)A decision fusion Labanotation generation algorithm for upper limb movement is proposed.In order to solve the problem that the feature representation is sensitive to scale and angle changes,this chapter proposes to use normalized node feature and Lie group feature to represent the motion data.In order to solve the problem of insufficient attention to the ending posture in the recognition of upper limb movement,this chapter proposes to analyze the ending posture from two perspectives of artificial rule making and extreme learning network,and then fuse the results with strategies to improve the recognition ability.In addition,based on the strategy fusion,using the synergy of lower limb and upper limb movements,this thesis further proposes to use the upper limb and lower limb movement data to model the relationship between upper and lower limbs.Experimental results show that the normalized node feature and Lie group feature can represent 3D motion data robustly.Compared with the existing methods,this algorithm achieves higher recognition accuracy and improves the quality of upper limb Labanotation generation.(3)A spatiotemporal network Labanotation generation algorithm for lower limb movement is proposed.In order to solve the problem of insufficient attention to the temporal and spatial information in the movement process,this chapter proposes to analyze the temporal and spatial information by using the double branch network structure based on the bidirectional gated recurrent unit neural network and based on the Lie group network.The branch of bidirectional gated recurrent unit can model the long-term dependence of time series data,and the branch of Lie group network can model the spatial relationship of data.In order to improve the recognition ability of network model,the time analysis ability and spatial analysis ability of the two networks are combined through network combination.In addition,based on the synergy of upper and lower limb movements,the association relationship of upper and lower limbs is further proposed by using the upper limb classification results and upper limb movement data on the basis of spatio-temporal network.Experimental results show that,compared with the existing methods,this algorithm achieves higher recognition accuracy and improves the quality of lower limb Labanotation generation.(4)Design and implement a platform of automatic Labanotation generation and multi-dimensional display.The functions of the platform include: the recording and saving function based on the automatic Labanotation generation,and the display function based on multi-channel videos,motion capture data and Labanotation.It realizes the combination of computer technology and Laban theory based on 3D motion capture data.Labanotation becomes a new way of digital recording of dynamic arts.It makes contributions to the protection and inheritance of national folk dance,opera and martial arts.This thesis focuses on several key technologies of automatic generation of Labanotation based on 3D human motion data.Through motion segmentation and unit movement recognition,this work proposes the corresponding algorithms.Finally,we complete a solution of automatic generation of Labanotation based on the motion capture data.