PoseRWGCN: an attention-free dual-stream RWKV–GCN architecture for real-time 3D human pose estimation

3D human pose estimation has a wide range of applications in fields such as healthcare and human–computer interaction. However, Transformer-based models often fall short of real-time requirements because self-attention incurs substantial computational and memory overhead. To overcome this limitation, we propose PoseRWGCN, an attention-free dual-stream architecture that integrates Receptance Weighted Key Value (RWKV) and Graph Convolutional Networks (GCNs). In this architecture, the RWKV stream captures global temporal features, and the proposed PoseRWKV module uses a recursive gating mechanism to efficiently model long-term temporal dependencies. Meanwhile, the GCN stream focuses on local spatial features. An adaptive fusion strategy unifies the two representations, thereby delivering comprehensive spatiotemporal modeling. For different application scenarios, we introduce causal inference for real-time inference and bidirectional inference for video. Extensive experiments show that PoseRWGCN achieves competitive results on two challenging datasets: Human3.6M and MPI-INF-3DHP. It also demonstrates robust performance in challenging conditions such as outdoor environments, occlusions, and motion blur. Compared to the baseline method, our approach achieves a 0.6 mm improvement in accuracy and reduces FLOPs by 62.1%, while supporting both causal and bidirectional inference without sacrificing accuracy. Notably, in causal mode, our model outperforms the baseline by 5.5 mm.