Temporal-Spatial Time Series Self-Attention 2D & 3D Human Motion Forecasting
Andi Prademon Yunus, Kento Morita, Nobu C. Shirai, Tetsushi Wakabayashi · 2023
The ability to forecast human motion is crucial in increasing awareness of moving objects in the environment. To address this challenge, this study focuses on human motion forecasting based on annotated 2D and 3D data and the model’s usability on data obtained from pose estimation. This research presents the Temporal-Spatial Time Series Self-Attention method for human motion forecasting. The approach is evaluated using the Human 3.6M, 3DPW, and AMASS datasets based on standard evaluation protocols. Our method performed well in the 2D ground truth and pose estimation data compared to the other time series method. Our method did not yet outperform previous research in 3D input data. However, based on the quantitative and qualitative assessments, our approach demonstrated excellent performance in predicting human motion for short- and long-term objectives.