Research on the Evaluation Model of Dance Movement Recognition and Automatic Generation Based on Long Short-Term Memory

Table 1

Dance generation algorithm based on deep learning.

Algorithm	Content
Yi wang and other researchers proposed NPHHMM in 2005.	This hierarchical hidden Markov model can capture action data. NPHHMM is trained to include a wide variety of ballet and disco movements.
In 2006, Bai niao takagi created an action synthesis method based on motion capture system [21], which can input music synchronously.	When the system inputs music, this method can calculate the similarity between motion and music features. The system can use correlation tracking motion graph to generate new motion.
In 2012, a new many-to-many statistical mapping framework emerged. The framework proposed by ofli et al. can learn musical features and dance figures.	The framework is also a discrete hidden Markov model. In addition, it introduces an improved Viterbi algorithm to synthesize dance sequences. It can use MFCC features to classify music and generate corresponding dance postures.
Chor-RNN’s system was proposed in 2016.	RNN of LSTM type was used for modern dance training. This is the first time that deep recurrent neural network is used to automatically generate dances and generate new dance sequences.
Omid alemi et al. designed an application called GrooveNet in 2017.	This application can learn and generate dance patterns on a very small training set. It uses FCRBM and RNN to generate dance movements.
A deep recurrent neural network with good performance is proposed.	Its encoder adopts one-dimensional convolution layer and multilayer LSTM, which can process the audio power spectrum. The decoder part uses LSTM layer to generate dance movements.
Vondrick et al. designed a deep learning model [22].	Sound and motion features are used to extract mapping. In order to get better performance, time index and masking methods are also used.