Spatial-Temporal Local Augmentation Graph Convolutional Networks

Abstract

Action recognition based on skeleton models has been widely focused in the field of computer vision in recent years. Most of the previous methods only focus on the change trajectory of the same joint point in the moving process, with less consideration of the correlation between joints in the moving process, and many of the current action recognition models lack sufficient consideration of local relationships, so this paper constructs a more generalized spatial-temporal skeleton graph considering the inter-frame dependence of neighboring skeletons, and introduces a local enhancement module, using the idea of local aggregation on each node for local aggregation, combining the node's own features with the aggregated features of neighboring nodes, so as to better capture the local relationships between nodes. The model can combine global and local information to provide a more comprehensive feature representation, thus improving the performance of the model. The introduction of local relationships can also increase the flexibility and sensitivity to the details of the model. Finally, we validate the proposed Spatial-Temporal Local Augmentation Graph Convolutional Networks (ST-LAGCN) model in two skeleton datasets, NTURGB+D and Kinetics, and compare it with several state-of-the-art graph neural network models for action recognition, both of which show improved performance.