Research On Skeleton-based Action Recognition Via Graph Neural Network

In recent years,skeleton-based human action recognition has received extensive attention from researchers.Skeleton data can robustly adapt to dynamic environmental changes such as camera viewpoint changes and background disturbances,and thus can help recognition methods focus on more robust action features.Current state-of-the-art methods model the human body as a topological graph and utilize a graph convolutional neural network(GCN)to extract action features.But there are still some problems:On the one hand,actions consist of time series of skeleton poses,and effective spatiotemporal features are easily hidden in redundant joint vertices or frames.How to make the model pay attention to more important spatiotemporal features is an unsolved problem.On the other hand,a robust recognition algorithm should have the ability to generate skeleton graph embeddings that can accurately and dynamically reflect the spatiotemporal correlation between joints.In addition,although GCN has a strong ability to learn spatial patterns,it ignores the varying degrees of higher-order dependencies,resulting in over-smoothing problems.Aiming at the above problems,this paper carries out the related research on skeleton action recognition.First,this work uses the kernel attention mechanism to design a spatiotemporal feature enhancer,the spatiotemporal kernel attention module,to help the model pay attention to more important spatiotemporal features.And innovatively apply the similarity function to aggregate the spatiotemporal features of the skeleton sequence,and propose a more robust adaptive graph strategy.Second,this work constructs a spatiotemporal kernel attention adaptive graph convolutional neural network(KA-AGCN)and exhibits excellent recognition performance.And use the multi-head(multi-head)self-attention graph Transformer operator to model higher-order spatial dependencies between joint points to alleviate the over-smoothing problem of graph convolution.Finally,based on KA-AGCN,a joint-bone two-stream strategy coupled with multiscale action features is applied,and a two-stream temporal kernel attention adaptive graph Transformer neural network(KA-AGTN)is further proposed.With the help of the dualstream strategy,KA-AGTN achieves a significant improvement in recognition accuracy on NTU-RGBD 60 and NTU-RGBD 120 datasets.On the Kinetics-Skeleton 400 dataset,the recognition accuracy has reached the state-of-the-art level.