Beyond Structural Symmetry: Elastic Spatio-Temporal Fluid Graph Convolutional Networks for Skeleton-Based Action Recognition

Skeleton-based human action recognition is a pivotal research area in computer vision. While conventional methods predominantly model human actions as rigid displacements strictly adhering to structural symmetry, the essence of authentic action is actually an elastic spatio-temporal fluid driven by highly asymmetric motion states. This over-reliance on symmetrical priors leads to severe representational bottlenecks: Spatially, traditional models solidify topological structures based on inherent anatomical symmetry, failing to capture the dynamic fluid coupling between cross-regional joints. Temporally, they employ fixed sliding windows for uniform, symmetrical sampling, which risks truncating core features during asymmetric action bursts while incorporating substantial noise during stationary periods. To address these issues, we propose a novel Elastic Spatio-Temporal Fluid Graph Convolutional Network (ESTF-GCN). The Spatial Fluid Topology (SFT) module introduces a motion-driven morphological principle to transcend rigid anatomical symmetry, adaptively reconstructing a task-driven dynamic fluid topology. Concurrently, the Temporal Flexible Evolution (TFE) module overcomes the limitations of symmetrical sampling by dynamically scaling the temporal receptive field, expanding during action bursts and retracting during redundant periods. Extensive experiments on the NTU RGB+D 60, NTU RGB+D 120, and NW-UCLA datasets demonstrate that ESTF-GCN significantly improves recognition accuracy while maintaining low computational complexity.