Projection graph convolution and variable-granularity feature shuffling with spatiotemporally separated feature enhancement: A lightweight 3D GCN for human action recognition

While the 3D Graph Convolution Networks (GCNs) have demonstrated outstanding results in human action recognition, the computational overhead of fully connected 3D graph convolution operations and feature enhancement with mixed spatio-temporal information greatly increases both the parameter size and computational complexity of the model, thereby hindering its deployment on resource-limited devices. The proposed lightweight model, VGST-PGCN, replaces the 3D graph convolution operation with projected graph convolution and variable-granularity feature shuffling, decouples spatiotemporal information into spatial and temporal dependencies for parallel enhancement and subsequent fusion, and employs multi-scale temporal convolution to extract motion features with different temporal spans. This design reduces the parameter count and computational load at the cost of discarding partial secondary spatial information during projection, while retaining competitive recognition performance. Comparisons with 14 representative methods on the NTU RGB+D 60 and NTU RGB+D 120 datasets show that VGST-PGCN achieves a competitive balance among recognition accuracy, a parameter count of 0.91M, and a computational cost of 1.85G FLOPs. Further evaluations on the NW-UCLA dataset using the Jetson Xavier NX show that VGST-PGCN achieves the highest accuracy of 95.32% and the lowest peak memory usage of 6.2G, while requiring 1.10G FLOPs and 24.2 ms per inference, highlighting its potential for application in resource-constrained scenarios. Code is available at https://github.com/LZ-KD/VGST-PGCN .