Learning Statistical and Physical Modeling for Consistency Human Motion Prediction

Abstract

Diffusion denoising models have great potential in generating diverse and realistic human motions. However, despite the impressive performance of existing methods, they still face some issues. The diffusion process often significantly overlooks physical laws, leading to physically implausible motions and a lack of consistency. Additionally, there are complexities involved in the training process. To address these limitations, inspired by Newton’s laws of motion, we propose a diffusion model—SPDiff that combines statistical modeling to fit the data with physical modeling to explore the true sources of motion. SPDiff introduces acceleration into the diffusion process using Taylor expansion, ensuring the plausibility of joint torques and motion velocities. By integrating statistical and physical modeling and employing cross-attention mechanisms to fuse their results, SPDiff creates an end-to-end human motion prediction diffusion model that achieves state-of-the-art performance.