Multimodal pose estimation and simulation modelling for real-time human motion analysis

To ensure safe and effective campus physical activities, this pioneering study proposes an innovative real-time sports pose recognition framework integrated with simulation-oriented process modelling, aligning with the core scope of dynamic motion analysis. The framework features a sophisticated multimodal architecture that fuses visual and inertial data across four interconnected layers, while embedding simulation-driven process modelling to capture the spatiotemporal dynamics of human motion. Enhanced spatiotemporal alignment mechanisms enable precise extraction of key biomechanical features, which are further refined through optimised Relief F algorithm for critical motion feature selection. A particle swarm-optimised graph convolutional network (PSO-AGCN) leverages simulated motion topology variations to process these features efficiently for pose classification. Evaluations on Human3.6M and a college sports dataset show 96.7% accuracy, 42.3% reduced occlusion errors, and 38 FPS operation, highlighting robustness and real-time performance, with simulation enhancing analysis interpretability.