AcST: Adaptive Human Motion Prediction Approach Based on Spatio-temporal Transformer

In response to increasing demand for accuracy and real-time applications in human pose prediction tasks, recent advancements in HMP (Human Motion Prediction) field have achieved precise short-term future pose sequence prediction and the generation of kinetically plausible sequences over longer time horizons. However, with increasing model complexity and expanding datasets sizes, generating future pose sequences in real-time remains a critical challenge for enhancing user experience. We propose a transformer-based model to generatively model pose sequences, incorporating a novel sparse Adaptive of components Spatio-temporal Transformer block (AcST). Aoc Block enables the model to adaptively adjust the complexity in the context of large-scale and complex parameters, maintaining relatively stable pose sequence generation speeds. It also addresses the issue of linear error accumulation inherent in autoregressive models, allowing for the real-time generation of accurate future predictions. Our method has been validated on benchmark datasets, demonstrating superior performance in comparison to existing approaches.