Spatio-temporal topological encoding via dual-pathway ST-GCN and attention-enhanced BiLSTM for robust human pose optimization

Accurately recognizing human poses across different environments remains a major challenge in computer vision due to issues such as occlusion and lighting. In this paper, we present multi-scenario adaptive pose recognition (MAPR), a framework that combines spatio-temporal graph convolutional networks (ST-GCN) with an attention-enhanced bidirectional LSTM (Attention-BiLSTM). The system uses two ST-GCN pathways to extract multi-scale spatial features from skeletal data, while the attention-BiLSTM captures long-term temporal patterns. This design helps the model understand both the relationships within a single frame and the changes between frames. On the MPII dataset, MAPR performs well, achieving a PCKh@0.5 of 0.892 and an area under the curve (AUC) of 0.919 under standard conditions. The model also remains robust in low-light settings, maintaining an AUC of 0.863 at 40% illumination, a 6.1% decrease. Ablation studies show that MAPR consistently outperforms HRNet baselines and versions with only one component. This framework supports reliable pose recognition for real-world applications, including sports analysis, rehabilitation, and smart environments.