Research On Human Skeleton Action Recognition Method Based On Graph Attention Mechanism

80%of the information comes from body language in human-to-human communication.It's of great significance for improving human-computer interaction to recognize human movements accurately.Skeleton data has information in two dimensions,time and space.And there are co-occurrence features between the two dimensions.How to better extract and use spatial-temporal features is a big challenge.multiple perspectives in the process of acquiring data is still a problem.This makes the same action may have very different performances.Once the model is constructed,it will not only increase the difficulty of model construction,but also have the problem of low model recognition rate.Aiming at solving the difficulty of extracting spatial-temporal features in skeleton action recognition,this paper apply graph attention network to realize skeleton action recognition.Firstly,construct a spatial-temporal graph of skeletal movements.According to the physical structure of the skeleton and the natural extension in time,the joint points are connected.Then each skeleton sequence is constructed into a spatial-temporal graph.Secondly,a single frame-based graph attention layer,and two functions are defined,namely neighbor function p and attention coefficient function ?.The spatial graph attention layer is extended to the spatial-temporal domain,and the multi-head attention mechanism is applied to implement the spatiotemporal graph attention model.Thirdly,the two different datasets are preprocessed and compared experiments are designed to determine the optimal number of attention heads.Finally,the experimental analysis and model comparison of the proposed method prove that the ST-GAT method proposed in this paper has a better framework and stronger generalization ability.Aiming at solving the problem of multiple perspectives in the data obtained by skeleton action recognition,this paper proposes view transform graph attention recurrent networks for skeleton-based action recognition.For a three-dimensional skeleton sequence in the camera coordinate system,it is helpful to make a sequence-based view conversion strategy by principal component analysis(PCA).The original data is converted into view-insensitive data.And the graph attention network(GAT)and the long-short-term memory network(LSTM)build a graph attention recurrent model.In addition,the paper also studies the experimental results about the input method based on human body parts.In order to further experimental research,five configuration models are proposed.The recognition effect is much better than that of the basic configuration model LSTM.The recognition accuracy has improved by 5.8%,9.1%,5.9%,9.3%and 11.8%respectively.Finally,full experiments and comparisons are carried out on three challenging datasets,and the possible reasons for the experimental results were discussed a lot.