Study On Human Action Recognition Method Based On Deep Learning

Human action recognition has gained a large number of applications in areas such as safety monitoring,sports training,road traffic,and healthcare.The skeleton sequence of human motion has the characteristics of smaller data amount and less background interference than the image,which makes the human action recognition method based on the skeleton sequence attract more attention.However,the features of space,time and spatiotemporal co-occurrence produced in the process of human movement are very complex,and extracting a variety of features generated by human movement is an urgent problem to be solved,so this thesis makes the following research on the construction and feature extraction of human action recognition model:(1)Aiming at the human action recognition model based on graph convolutional network(GCN),the global receptive field is limited and the number of parameters of the Transformer-based model is high.In this thesis,a human action recognition network(GCN-ATF)based on GCN and adaptive graph combined with Transformer is proposed.Firstly,the advantages of fewer GCN parameters and stronger ability to extract local features are used to extract local features of human movement in shallow networks.Secondly,taking advantage of Transformer’s stronger ability to extract global features,Transformer can extract more global features in the deep network,and after multiple down-sampling of the convolution part,the smaller feature scale makes the Transformer network have fewer parameters and reduces the complexity of the model.Finally,an adaptive graph is introduced at the Transformer layer,so that the human skeleton diagram breaks the limitation of the inherent connection of key points of the skeleton,and can be adapted to different network layers and different data samples in the network.Experiments show that the proposed method has high recognition accuracy in the field of human action recognition.(2)Aiming at the problem that the proposed GCN-ATF has insufficient ability to extract important features of human movement,and the model does not make full use of the features of each network layer,the method of dense residual and combined attention is proposed to improve GCN-ATF.Firstly,the combined attention of space and channel is adopted in the convolution part of the graph of GCN-ATF to improve the importance of important spatial features and channel features in the network,and the combined attention of time and channel is adopted in the time path to improve the importance of important time features and channel features in the network,and reduce the interference of redundant information on the network.Secondly,the densely connected network is used to enhance the transmission of information between network layers and the feature reuse of more shallow networks,so that the model converges faster and the influence of gradient disappearance or gradient explosion is reduced.Experiments show that the proposed method improves the recognition accuracy of the model GCN-ATF and has more resolving ability for action with similar characteristics.(3)In view of the lack of research on human action recognition in complex scenarios,this thesis focuses on coal mine scenarios to identify workers’ action.Due to the stronger background interference of coal mine environment,which will lead to the absence of human key points during attitude estimation,which makes the model training effect poor,a method of missing key point compensation is proposed,and a coal miner action dataset(MPA)is established.The dataset MPA is used to train the action recognition model proposed in this thesis,and the effectiveness and advanced nature of the model for coal miners’ action recognition are proved by experiments.The thesis has 31 figures,8 tables,and 85 references.