Font Size: a A A

Research And Implementation Of Dance Motion Generation Model Based On Deep Learning Network

Posted on:2021-02-10Degree:MasterType:Thesis
Country:ChinaCandidate:Z W DangFull Text:PDF
GTID:2428330614970612Subject:Computer technology
Abstract/Summary:PDF Full Text Request
In the field of music-driven computer dance movement generation,the traditional music movement matching model and statistical mapping model have the following problems.First of all,the dance generated by the model does not fit the music well enough.Secondly,the completeness of the generated dance movement is insufficient.Thirdly,the smoothness and rationality of long time dance sequence need to be strengthened.Fourth,traditional models cannot generate new dance movements.How to generate smooth and complete dance posture sequence according to music is the problem of this paper.To solve these problems,we designed a deep learning-based dance generation algorithm to extract the mapping between sound and motion features.In the feature extraction stage,we used the prosodic features and audio beat features extracted from music as the music features,and the coordinates of key points of human bones extracted from the dance video as the movement features for training.In the stage of model construction,the basic mapping of music and dance movements is realized through the generator module of the model to generate smooth dance gestures.The consistency of dance and music is realized by the discriminator module.The audio features are more representative through the auto-encoder module.In the real humanization stage,the improved Pix2 Pix HD model was used to transform the dance posture sequence into the real dance.Finally,I got a live-action version of the dance in line with the music.The experiment obtained dance videos on the Internet as training data,analyzed the experimental results from five aspects: loss function value,comparison of different Baseline,evaluation of sequence generation effect,user study and quality evaluation of authentic dance videos.The experimental results show that our dance generation model can effectively extract the characteristics of music,generate the sequence of dance gestures that fit the music,and turn it into a realistic live dance video.
Keywords/Search Tags:Cross model generation, Dance motion generation, Attention Mechanism, Generative Adversary Network, Sequence to sequence
PDF Full Text Request
Related items