A Hypergraph Transformer With Adaptive EdgeConv for Skeleton-based Action Recognition

Skeleton-based action recognition is a crucial task in computer vision, aiming to perform pattern recognition on a skeleton sequence to identify the actions represented by the skeleton's semantics. However, current mainstream skeleton-based methods primarily rely on global edge information. Unimportant edges may influence this approach, limiting the model's performance in action recognition tasks. Additionally, because these methods treat all information equally, they may lead to resource wastage. This paper proposes a Hypergraph Transformer-based action recognition method using Adaptive EdgeConv Module(AEM) to address these issues effectively. In this method, we introduce Adaptive Edge Graph Convolution Networks(AEGCNs), which utilize the AEM to adaptive obtain edge sets and edge set weights, thereby adaptive identifying key edge sets. With this information, the model can dynamically adjust the importance of different edge sets based on various samples. This approach improves the accuracy of action recognition and reduces the loss of computational resources. Our experimental results on two challenging benchmark datasets (i.e., NTU RGB+D and NTU RGB+D 120) show that the proposed method is comparable to state-of-the-art methods.