Web-Based Semantic Framework for Enhanced Human Motion Prediction With MSTIA-Net

Researchers in human motion prediction have focused on mathematical modeling of the human skeletal structure, often overlooking the spatio-temporal characteristics of human pose sequences. To address this, we propose the multi-scale spatio-temporal information aggregation net (MSTIA-Net), which includes two key modules: the graph convolutional spatio-temporal information aggregation (GCSTIA) module and the windowed discrete cosine transform (WDCT) temporal encoding module. GCSTIA extracts and integrates multi-scale temporal and spatial features of human motion sequences, while WDCT removes high-frequency noise and compresses data. model's efficacy is demonstrated on three datasets: Human3.6, CMU, and 3DPW, achieving performance improvements of 2.4%, 4.1%, and 1.7%, respectively.