Spatiotemporal Fusion Algorithm Based on ST-GCN and Attention BiLSTM: A Method Study for Diverse Motion Posture Recognition

Posture recognition technology has broad application prospects in sports training, health monitoring, human-computer interaction, and other fields. However, it still faces challenges in accuracy and robustness in complex and changing real environments. This research proposes a multi-scenario adaptive posture recognition algorithm MAPR, which processes human posture data through spatial-temporal graph convolutional networks and attentionenhanced bidirectional long short-term memory networks, achieving effective expression of spatial structural features and temporal dynamic features. The algorithm was comprehensively evaluated on the MPII human pose dataset, comparing the performance of the complete MAPR algorithm, variants without ST-GCN, variants without Attention-BiLSTM, and the HRNet baseline model in normal and dim environments. Experimental results show that the MAPR algorithm performs excellently in normal environments, with [email protected] reaching 0.892 and AUC value reaching 0.919. More significantly, in dim environments with 40% illumination, the MAPR algorithm still maintains high performance (AUC=0.863), with a performance decrease of only 6.1%, clearly outperforming other comparison models. By analyzing the recognition difficulty of different types of actions, the research reveals the complementary value of spatial features and temporal features in posture recognition. The MAPR algorithm provides new ideas for posture recognition in diverse motion scenarios, showing significant robustness especially under environmental challenges such as changes in lighting conditions, laying a technical foundation for applications in motion analysis, sports training, and other fields.