Multi-filter dynamic graph convolutional networks for skeleton-based action recognition
Yating Yuan, Bo Yu, Bo Yu, Wei Wang, Bihui Yu, Bihui Yu · Procedia Computer Science · 2021
Action recognition plays an important role in video understanding, and the dynamic changes of human bones provide important information for it. In order to solve the problem that the source point can only obtain the characteristics of adjacent nodes and the static structure of the skeleton diagram can not meet the adaptive requirements, this paper proposes an Inception structure and dynamic skeleton diagram based on the convolutional neural network, namely, the multi-filter dynamic graph convolutional neural network. The method can not only enable the source point to obtain the characteristics of distant nodes, but also can use the dynamic skeleton diagram structure to generate different body associations for different actions, which improves the adaptability and accuracy of the model and achieves better results.