Highly Condensed All-MLP Architecture for Long-Term Human Motion Prediction

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Constrained computational resources in AI, like autonomous driving, pose challenges for accurate, long-duration human motion prediction.
• Existing models often struggle to balance lightweight design with high predictive performance over extended timeframes.

Purpose of the Study:

• To introduce a highly condensed all-multilayer perceptron (HCMLP) architecture for efficient and accurate human motion prediction.
• To enable lightweight AI models capable of extended-range motion predictions without performance compromise.

Main Methods:

• Developed a spatiotemporal dynamic perception (STDP) block with parallel spatial and temporal multilayer perceptrons (SMLP and TMLP) for efficient feature extraction.
• Integrated dynamic aggregation (DA) and channel multilayer perceptron (CMLP) to refine spatial-temporal features.
• Introduced a multiterm union prediction (MTUP) block for direct, precise predictions up to 4000 ms, avoiding iterative short-term predictions.

Main Results:

• HCMLP demonstrated superior performance compared to state-of-the-art methods on Human3.6M, AMASS, 3DPW, and CMU-Mocap datasets.
• Achieved leading results in short-term (ST), long-term (LT), and extended long-term (ELT) human motion predictions.
• The HCMLP model utilizes significantly fewer parameters than existing approaches, highlighting its lightweight efficiency.

Conclusions:

• The HCMLP architecture offers a highly efficient solution for human motion prediction in resource-constrained AI applications.
• Its lightweight design and advanced feature processing enable accurate extended-range predictions, outperforming current state-of-the-art methods.
• HCMLP presents a significant advancement for real-time AI systems requiring precise human motion forecasting.