Multi-stage human motion prediction algorithm based on spatiotemporal graph convolution

Human motion prediction is a classic computer vision task, with the core objective of predicting future motion sequences based on given historical motion sequences. Multi-stage prediction strategies have achieved significant success in the field of human motion prediction. However, existing network models struggle to effectively integrate multi-stage frameworks, leading to limitations in dynamic adjustment of key features and long-term dependency modeling, ultimately constraining prediction accuracy and robustness. To address these issues, this paper proposes a multi-stage network model that fuses spatiotemporal features. Spatiotemporal convolution is utilized to enhance the long-term dependency of spatiotemporal information in motion sequences, while a dynamic gating mechanism is introduced to adaptively adjust feature weights. Finally, the proposed network model is integrated into each stage of multi-stage prediction to further optimize the efficiency of spatiotemporal feature extraction. Experimental results demonstrate that the proposed method outperforms existing approaches on multiple datasets, improving performance by 4%onHuman3.6M and 3% on CMU-MoCap.