Research On Action Recognition Using LSTM Method Based On Extended Data Set

With the development of information technology,pattern recognition technology and artificial intelligence technology have gradually get people's attention,especially the continuous development of deep learning technology and human behavior recognition technology.which has finally gained the attention of the academic community and has become one of the most popular research areas.Deep learning technology and human behavior recognition technology are increasingly welcomed by people from all fields of life and have extensively been applicated in various fields.Therefore,how to apply deep learning to accurately and efficiently identify human behavior becomes a challenging problem.This paper,based on the three-dimensional data of human skeletal points collected and extracted by Kinect equipment and other equipment,has extensive research on two aspects of human data set expansion and human action regonition.First of all,this paper analyzes the requirements of deep learning for data,and finds that the collection of human data sets is a very time consuming and labor-intensive project,and it is difficult to meet the needs of deep learning moderns.In order to solve this problem,this paper proposes a semi-supervised deep learning structure that can generate a large number of reliable data sets based on the small-scale data that has been collected.In this paper,by combining the Recurrent Neural Network and the Generative Adversarial Networks,Recurrent Neural Network is able to learn the sequence relationship of the data,The Generative Adversarial Networks for the network can generate reasonable data and extend the human behavior data set.Therefore,relying on these two networks,this structure can well analyze the characteristics of the collected data,and based on these characteristics can generate a large number of reasonable data,and then through the finishing work,The data set can be formed,which greatly eases the deep learning The problem of scarcity of datasets at work.Secondly,this paper studies the algorithm of human action recognition.It is considered that it is necessary to fuse the temporal and spatial characteristics of human behavior to identify the behavior.Moreover,in order to reduce the influence of noise on the deep learning model,this paper chooses a combination of filtering algorithm and deep learning algorithm that can predict input,and finally proposes Long short-term memory algorithm that incorporates a noise filtering mechanism.The dataset is used to train the Long short-term memory network to ensure that the noise of the learning data set has the least influence on the calculation result,and it can also improve the accuracy of human behavior recognition.Finally,several data sets were used for experimentation,compared to several other excellent algorithms.Experimental results show that the proposed algorithm has higher accuracy.