Research On Human Action Recognition Based On Graph Neural Network

Human action recognition is an important research direction in the field of computer vision.Many action recognition models have been applied in smart home,human-computer interaction,security monitoring,automatic driving and other fields.Existing skeleton sequence action recognition models have problems such as single action feature modality and insufficient learning of temporal features.In view of the above problems,this paper takes the graph neural network as the theoretical basis and combines the joint stream and bone stream dual-stream network structure,and proposes a T-BiGRU-based dual-stream low-rank adaptive feature fusion graph convolution action recognition model to realize action recognition.A convolutional action recognition model based on two-stream adaptive feature fusion graph was constructed.The model uses two-streams adaptive garph convolutional network as the baseline network and combines multi-modal feature fusion to realize action recognition.Learn feature fusion parameters in adaptive feature fusion,complete adaptive dynamic adjustment of dual-stream features to fuse high-dimensional features of joint stream and bone stream;for splicing adaptive feature fusion,tensor adaptive feature fusion and low-rank adaptive feature fusion In-depth research on their respective integration methods.Compared with the 2SAGCN model,the proposed model increases the model receptive field and improves the action recognition ability.The performance of three adaptive feature fusion models with 2S-AGCN as the baseline network is analyzed on the X-View and XSub subsets of the NTU-RGB+D dataset.Constructed T-BiGRU unit,the structure of temporal convolution and bidirectional gated recurrent unit alternates is used to perform time-domain convolution feature extraction on the skeleton sequence to fully learn the timing characteristics of actions.The two-way gated loop unit simultaneously learns the sequence and reverse order features of the successive frame actions of the skeleton sequence,and deeply excavates the inner relationship between the successive frame actions in the long-frame video.At the same time,the temporal convolutional network is used to down-sample the skeleton sequence to realize the skeleton sequence The number of time frames is halved to achieve the purpose of reducing the amount of calculation.In the comparative experiment,using Accuracy as the evaluation index,the four methods of manual feature extraction,convolutional network,recurrent network and graph convolutional network are compared and analyzed in the two subsets X-View and X-Sub of NTU-RGB+D.Experimental results prove the effectiveness of the proposed method.Build an action recognition system based on human skeleton sequences.Pyqt is used to build an action recognition interface.The system uses the YOLOv5 algorithm for target detection and the sort algorithm for target tracking.The Alpha Pose human body pose estimation algorithm is used to extract the joint points of the characters in the video,and the collected key points of the human body are based on the proposed method.T-BiGRU’s dual-stream low-rank adaptive feature fusion graph convolutional action recognition model realizes action recognition,and the system is functionally tested.