Research On Human Action Recognition Based On Skeleton Data And Graph Convolution Network

In recent years,human action recognition research has been widely applied in fields such as human-computer interaction and medical health,and has achieved remarkable results.The research data of action recognition has undergone various morphological transformations,among which skeleton data has become a hot topic in human action recognition research due to its better compatibility with the natural physical connections of the human body and its strong robustness to complex environments.At the same time,the birth of graph convolutional network has greatly improved the performance of human action recognition.However,existing skeleton-based action recognition technologies have many shortcomings in exploring co-occurrence relationship in the spatiotemporal domain and identifying the long-range dependencies between joints,resulting in poor recognition performance.This study conducts in-depth exploration and research on extracting spatiotemporal features of joints and identifying remote logical dependencies between joints based on skeleton data and graph convolutional networks.Experimental verification is conducted on two large datasets,NTU RGB+D 60 and NTU RGB+D 120.Firstly,aiming at the inflexible of feature extraction caused by manually setting the graph topology structure according to the physical connection of human body and ignoring the semantic information that is beneficial to action recognition,a semantics-guided adaptive graph convolutional network is proposed.This model uses the end-to-end learning approach and the data-driven method to increase the flexibility of the modeling graph structure and the generalization ability of the model,enhance the expression ability of features,and improve the performance of action classification.Secondly,considering the differences of motion information among different channels in the sample,a channel-wise topology modeling adaptive graph convolutional network is proposed.This model uses the normalized embedded Gaussian function to model the correlation between all joints in each channel,and obtains the channel-wise graph topology.It sets the corresponding graph topology for different channels of different samples to achieve flexible and efficient feature extraction.Finally,based on the uneven distribution of joint features,a multi-attention mechanism is proposed and a graph convolutional network based on multi-attention modules is constructed.On the basis of channel-wise topology modeling,the network assigns different weight values to the spatiotemporal and the channel domains,so that the model can accurately allocate the importance according to the weight value,and focus on the key area to further improve the performance of the human action recognition algorithm.