Multi-dimensional Feature Interaction Enabling Density-driven Contrastive Learning for Skeleton-based Action Recognition.

Skeleton-based action recognition has gained increasing attention due to its robustness to viewpoint variations and background clutter. However, existing approaches still face challenges in feature modelling and optimization, particularly in capturing inter-dimensional dependencies and adapting to complex sample distributions. To address these issues, we propose MFI-DCL, a novel graph convolutional network with multi-dimensional feature interaction-enabling density-driven contrastive learning strategy. Specifically, the model explicitly projects skeleton sequences into three independent feature spaces: spatial for modelling joint topology, temporal for capturing motion dynamics, and channel for extracting semantic features. Two dynamic interaction modules, dynamic spatial-channel interaction (DSCI) and dynamic spatial-temporal interaction (DSTI), are introduced to enhance the collaboration among these dimensions. Furthermore, a density-driven contrastive learning strategy (DCL) is designed to refine feature structures by dynamically optimizing misclassified samples based on local density estimation. We conduct experiments on NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets. MFI-DCL achieves 97.2% accuracy on the NTU RGB+D dataset X-View, 91.4% accuracy on the NTU RGB+D 120 dataset X-Set, and 97.7% accuracy on the NW-UCLA dataset. These results are highly competitive compared with existing state-of-the-art methods, demonstrating strong adaptability, especially in handling action categories with subtle differences or ambiguous boundaries. The source code is publicly available at https://github.com/silence-607/MFI-DCL.