Selective directed graph convolutional network for skeleton-based action recognition

Abstract

Skeleton-based action recognition has gained significant attention due to the lightweight and robust nature of skeleton representations. However, the feature extraction process often misses subtle action cues, making it challenging to differentiate between similar actions and leading to misclassification. To address this issue, we propose a Selective Directed Graph Convolutional Network (SD-GCN) that decouples features at varying granularities to enhance sensitivity to subtle actions. Specifically, we introduce a Dynamic Topology Generation (DTG) module, which dynamically constructs a new topological structure by focusing on key local joints. This reduces the influence of dominant global features on subtle ones, thereby amplifying fine-grained motion features and improving the distinction between similar actions. Additionally, we present an Attention-guided Group Fusion (AGF) module that selectively evaluates and fuses local motion features of the skeleton while incorporating global skeletal features to capture contextual relationships among all joints. We validated the effectiveness of our method on three benchmark datasets, and experimental results demonstrate that our model not only outperforms existing methods in terms of accuracy but also excels at distinguishing similar actions.