Automatic Generation Of Labanotation Based On Deep Learning

In the context of the artificial intelligence era,the protection of traditional cultural resources through digital intelligence has become an important research topic.Laban dance notation is the most widely used dance symbol system in the world.Recording and preserving dance movements in the form of Laban dance notation is of great significance in protecting intangible cultural heritage.In order to eliminate the tedious work of manually recording the symbols of dance scores,a method of automatically generating Laban dance scores based on computer technology came into being.However,traditional methods do not make full use of timing information and spatial information.The current latest deep learning-based methods cannot achieve continuous action recognition,which limits the application scenarios.Under this background,this thesis conducts research on the automatic generation of Laban dance spectrum based on deep learning.Under the framework of recurrent neural network,this thesis proposes three dance spectrum generation algorithms based on recurrent neural network.main tasks as follows:(1)A dance spectrum generation algorithm based on spatial feature fusion is proposed.The existing skeletal node vector features lack spatial position information that can intuitively reflect human bone joint points,and ignore the influence of spatial position information on the accuracy of the generated dance spectrum,which limits the performance of the dance spectrum generation algorithm.In this thesis,based on the vector features of bone nodes,a feature representation method for the fusion of Joint and Line features for dance spectrum generation is proposed.The Line feature is a vector feature of bone nodes,which represents the topology of the human body.Joint is the threedimensional space position information of bone joints,which can intuitively reflect the movement trajectory of joints during the movement of the human body.This feature representation method more fully analyzes the changes in the position and direction of the human body movement process,and thus can better reflect the movement characteristics.(2)A dance spectrum generation algorithm based on multi-time series modeling is proposed.Existing dance spectrum generation algorithms based on long-term and shortterm memory networks use historical information to predict the state of the current moment,without considering the superimposed influence of future information on the current state,and because of the large number of network parameters and large amount of calculation,the performance of time series modeling is limited.On this basis,this thesis replaces the long-term and short-term memory unit with a bidirectional gating unit BiGRU for multi-sequence modeling.Fully consider the coherence and correlation of actions in time and space,make better use of spatiotemporal information recursively,reduce the computational complexity,and improve the prediction performance.(3)A continuous dance spectrum generation algorithm based on spatio-temporal cascade network is proposed.Existing dance spectrum generation algorithms based on long-term and short-term memory networks can only process and recognize a single action.Continuous motion needs to be manually divided into single motion fragments and then processed according to the existing dance spectrum generation algorithm,which limits the practical application ability.In this thesis,a continuous time classifier module is cascaded in the backend of the dance recognition network to realize continuous dance generation end-to-end.The Blank unit of the classifier effectively avoids the misjudgment of continuous repeated actions.Under the action of the continuous-time classifier module,the predicted probability distribution of actions at each moment forms a dynamic combination in sequence according to the time sequence.The group with the largest combination probability is the prediction result and is used as the recognition result of continuous action.It effectively avoids the tedious steps of manual segmentation actions,increases the possibility of practical application,and overall improves the recognition speed and accuracy.