DF-GCN: A Lightweight Dynamic Dual-Stream Graph Convolutional Network for Skeleton-Based Action Recognition

In view of the issues such as high computational cost and inadequate feature representation in GCN-based action recognition approaches, we propose a dynamic fusion graph neural network, DF-GCN. Initially, we develop a dual-stream fusion framework that bidirectionally fuses joint, bone, and motion information. By leveraging a dynamic adjacency matrix to model both large-range movement features and localized motion patterns, the parameter count is reduced by 50 % compared to conventional multi-stream methods. Then, we incorporate a lightweight channel attention module ECA, which employs adaptive 1D convolutions to capture inter-channel interactions. This approach lowers the parameter complexity to one-third of that of traditional attention mechanisms while avoiding dimensional reduction. Additionally, we present a multi-scale temporal convolution module, using parallel branches to combine features with varying temporal receptive fields. Under the same parameter budget, this leads to a 0.72 % increase in accuracy. Experimental results on the NTU60 and NTU120 datasets indicate that the proposed approach attains 91.3 % and 96.5 % accuracy under the X-Sub (cross-subject) and X-View (cross-view) settings respectively, while maintaining a parameter count of merely 2.42 million.