Enhancing 3D Human Moiton Prediction with MSIGCN: A Novel Approach to Addressing Sensor Noise and State Accuracy

In the field of skeleton-based three-dimensional human motion prediction (HMP), recent advances in methods that model human motion through a series of intermediate states have demonstrated considerable potential. Nevertheless, a significant challenge with these approaches is the accurate acquisition of multiple intermediate states. In order to address this issue, we propose a method known as weight smoothing (WS). Moreover, current models for HMP frequently fail to address the issue of internal noise present in sensor-gathered human skeleton data. To address this issue, we propose an incremental information extraction module, which is designed to neutralise the effects of internal noise. Additionally, the incremental information extraction module is capable of extracting the underlying physical information of human motion. The combination of these two methodologies gives rise to a novel multi-intermediate state incremental graph convolutional network (MSIGCN). A comprehensive series of experiments has revealed that MSIGCN outperforms existing state-of-the-art methods by 1.95%, 1.70% and 2.44% in terms of 3D average joint position error on the Human3.6M, CMU Mocap and 3DPW datasets, respectively. The relevant code can be accessed at https://github.com/NirvanaYo/multi-intermediate-state-incremental-graph-convolution-networks.