Analysis And Generation Of Dance Movements Based On Deep Learning

Dance is an art about the movements of the human body.It takes professional movements as the main means of expression,and focuses on the inner and spiritual world that is difficult to express in language or other artistic forms.Generally speaking,it is characterized by exquisite emotion,profound thought,clear relationship,standard action,organization and beauty.As a choreographer,it is a difficult skill to master the characteristics of dance and have the ability to choreograph,because it requires a lot of learning time,deep professional knowledge background and corresponding economic foundation.As a hot technology in the field of computer vision,deep learning can effectively model related problems.Therefore,it is of great significance to realize the automatic generation of dance by using the method of deep learning and the internal connection of dance movements.Based on the above considerations,this paper divides the task into two parts according to the analysis and generation of dance movements.Part of the research is based on the audio-visual multi-modal framework of music and dance,mainly through the generation of music corresponding dance segments.This part explores the multimodal dance generation network by constructing the corresponding relationship between visual information and audio information.Specifically,a 2D prediction module is proposed to predict future frames by combining visual and audio features,and a 3D conversion module is proposed to realize the transformation from 2D skeleton to 3D.In addition,this part also puts forward some new evaluation indexes of dance generation to evaluate the generation results.The experimental results show that the proposed musical dance framework can meet the requirements of authenticity and diversity of dance.The other part is to consider how to control the direction of dance generation.A typical task is to predict the next dance sequence from a given starting frame,or to convert the music to the corresponding dance movement using audio information.However,these tasks did not control the direction in which the dance was produced,and there was a lack of appropriate controls to evaluate the results produced.This part explores a multi-channel based dance trend frame by giving the starting frame and the ending frame.According to the tips of the first and last frames,there can be a specific generation direction,so as to complete the dance generation task under the requirements of standardization.The experimental results show that a weighted multi-channel convolutional neural network can accomplish this new task well.In addition,the whole task will also evaluate the experimental results on a controlled basis.