Music Style And Beat Recognition And Its Application On Performance Robots

With the development of science and technology,the application of robots is no longer confined to the fields of industrial production or home services,and more and more performing robots are continuously coming into people's field of vision.Most existing dance performance robots perform performances by embedding dance programs manually arranged by choreographers,lacking technological innovation and dance flexibility.Based on deep learning methods,this thesis studies music recognition from two aspects of style and beat,and applies the research results to the physical performance robot,so that the robot can automatically perform dances suitable for music style and beat according to the given music.The research results of this thesis include:(1)A music style recognition method using a combination of independent recurrent neural network and scattering transform is proposed.Firstly,the related characteristics of traditional audio processing methods are briefly analyzed,and their suitable application scenarios and their unsuitability in this task scenario are explained.Then,based on the principle of scattering transformation,the superiority and rationality of using scattering transformation in this task are explained.Then,based on the application scenario of this subject,the application effect of the recurrent neural network and its variants on this task is comparatively analyzed.Combined with experiments,it is shown that music style recognition based on independent recurrent neural networks can get better performance,and explained Its application is reasonable in this task.Finally,this thesis proposes a music style recognition method combining two strategies of scattering transform and independent recurrent neural network.Experiments show that this method improves the accuracy of music style recognition to a certain extent.(2)An adaptive model selection method combining music style is proposed to realize music beat recognition.Based on a comparative analysis of the short-time Fourier characteristics,Mel features,power spectrum features,and scattering transformations of audio signals in the task of music beat recognition,the Mel spectrum and power spectrum features were selected for music beat recognition tasks.In order to improve the accuracy of beat recognition,this thesis divides the music into three categories according to the music style and re-classifies the music beat data set,and then trains multiple music beat recognition models.Combining the results of previous music style recognition,the method of adaptive selection selects one of the most suitable models from multiple beat recognition models,thereby obtaining the model that is most suitable for the song to be tested for final beat recognition.The experimental results show that the method is practical.(3)Propose a method of intelligently building dance action library and complete dance performance.This thesis abandons the previous practice of using choreography software to arrange dance movements.Starting directly from dance videos with combining pose estimation technology extracts the pose stream of dancers and establishes a library of dance movements with various styles to increase the flexibility and aesthetics of dance performance.In the end,the real robot automatically selects appropriate actions from the established dance action library for performance according to the style and beat of the music.