Couple Dance Generation And Application Based On Transformer

Dance is one of the common and long-standing art forms.It is widely implemented with artistic expression,social interaction and emotional expression.Compared with traditional solo dance,couple dance represents its difference in the number of participators,various types of rhythm and styles of music echo with corresponding posture with dance rules.At the beginning of last century,a versatile of couple dances appeared all over the world,such as Tango Dance in Argentina,salsa dance in Cuba,fro dance in Brazil and so on.With the development of information technology,abundant dance learners tend to practice dancing via online courses.Generally,watching videos from the Internet is the limited way to learn how to dance,while more problems will be encountered under the process of couple dance learning compared to solo dance learning: first,additional dance partners should be taken into account in the process of couple dance learning;Secondly,the couple dance learning needs to follow the dance partners’ motion and movement to make the correct dance posture simultaneously;Third,the lack of interactive feedback and scientific dance movement quality evaluation when learning couple dance online.Therefore,the aim of paper is to design an improved transformer model structure,which can generate the dance partner’s action sequence according to the user’s action sequence.A somatosensory interactive environment is built in the virtual reality environment.Combined with the somatosensory equipment,the user’s actions are captured in real time,and the action sequence is input into the network to get the virtual avatar driven by the dance partner’s action sequence.The main research contents of this paper include:1.Based on Laban movement analysis method,this paper studies the semantic representation of couple dance,designs a set of evaluation indexes suitable for evaluating the action quality of couple dance,and visually verifies the effectiveness of the designed indexes through t-SNE manifold learning algorithm.2.The generation of dance partner sequence based on transformer model uses the mocap data set containing two dancers to complete the construction and training of the network under the pytoch deep learning framework to learn the structure of dance movement.By comparing with other common sequence generation model methods,the robustness and effectiveness of the designed algorithm are verified.3.Using the sequence generation model trained based on transformer model,combining the sequence generation technology in deep learning with motion capture and virtual reality technology,combined with the idea of natural human-computer interaction,and using Kinect to complete the interactive system of dance learning that can get the feedback of virtual dance partners in real time,and introduces Laban movement analysis method to evaluate the quality of users’ dance movements.