3D action recognition based on adaptive topological graph convolution with augmented skeletal features

This paper proposes the Adaptive Topological Modeling Graph Convolutional Network (ATM-GCN) for skeleton-based action recognition, addressing the limitations of fixed graph topologies and insufficient feature representation in existing methods. Traditional approaches rely on predefined skeleton graphs that only capture explicit physical connections between joints, neglecting implicit spatiotemporal correlations. To overcome this, ATM-GCN introduces two key innovations: (1) augmented skeletal features, which integrate multi-dimensional attributes such as joint coordinates, bone vectors, and their temporal motion dynamics, enriching human kinematic representation; and (2) adaptive topology modeling, dynamically inferring node correlations across motion features and network layers while balancing predefined and data-driven topologies. Extensive experiments on the NTU-RGB+D 120 and PKU-MMD datasets demonstrate the model’s superiority. ATM-GCN achieves state-of-the-art cross-subject accuracy of 88.9% and 93.1% on the two datasets, outperforming baseline methods like ST-GCN by significant margins. Ablation studies validate the incremental benefits of feature augmentation and adaptive topology learning, with full-feature configurations yielding optimal performance. Visualization further reveals the model’s evolving topology, transitioning from preserving predefined structures in early layers to learning task-specific connections in deeper layers. While the approach is effective for coarse-grained actions, challenges remain in fine-grained motion recognition and cross-scenario generalization. This work advances skeleton-based action recognition by harmonizing dynamic relational reasoning with robust feature representation, offering insights for applications requiring multi-perspective adaptability.