Research On Human Action Recognition Based On Adaptive Graph Convolutional Networks

Human action recognition is one of the important research directions in the field of computer vision.There are broad application prospects in many fields such as human-computer interaction and automatic driving.In recent years,action recognition methods based on skeleton sequences have received extensive attention due to the lightweight and robustness to environmental changes of skeleton sequence data.There are inherent advantages of the graph convolutional network(GCN)in data modeling of non-Euclidean structures,which can fully utilize the spatial geometric structure information of graph data,and it has made great progress in action recognition based on skeleton sequences.This thesis mainly studies the action recognition method of the human skeleton sequence based on the adaptive graph convolutional network(AGCN).The main research work and innovations are as follows:Firstly,this thesis proposes an adaptive graph convolutional network(ME-AGCN)model with multi-partition and the external attention network.Aiming at the problem that the adaptive graph convolutional network cannot adequately establish joint connection relationships by performing graph convolution operations according to a fixed spatial configuration partition strategy,this thesis designs a multi-partition strategy,which explores the optimal number of partitions based on the spatial configuration partition strategy.It establishes richer joint connections and better adapts to the changing characteristics of different actions.Aiming at the problem that the temporal convolutional network(TCN)hierarchical aggregation of local temporal information cannot capture the long-term temporal dependence of the sequence,this thesis introduces the external attention network(EANet)into the model to obtain the correlation of any position of the sample through two memory modules to realize the long-term temporal dependence modeling of the sequence.Experiments on internationally recognized standard data sets show that the proposed ME-AGCN model has higher recognition accuracy.Secondly,this thesis proposes an adaptive graph convolutional network(TCPMSAGCN)model with temporal covariance pooling and multi-partition shared attention.Aiming at the problem that the use of global average pooling(GAP)at the end of the network to normalize the classification features ignores the spatiotemporal information of the target features,this thesis uses the temporal covariance matrix to retain the spatiotemporal information of the classification features and uses the fast matrix power normalization technology to fully use the spatial geometry information of the covariance matrix.At the same time,to avoid the repeated calculation of the attention mechanism of AGCN in each partition,which leads to data redundancy and increases the computational complexity,this thesis designs a shared attention mechanism,which calculates the joint attention graph outside the graph convolution operation and shares it in each partition.Experiments show that the proposed TCPMSAGCN model has a better recognition effect than the existing models.