Learning behavior aware features across spaces for improved 3D human motion prediction.

Journal: Scientific reports
Published Date:

Abstract

3D skeleton-based human motion prediction is an essential and challenging task for human-machine interactions, aiming to forecast future poses given a history of previous motions. However, most existing works model human motion dependencies exclusively in Euclidean space, neglecting the human motion representation in Euclidean space leads to distortions and loss of information when representation dimensions increase. In this paper, we propose Cross-space Behavior-aware Feature Learning Networks that can not only exploit the spatial-temporal kinematic correlations in Euclidean space, but also capture effective and compact dependencies and motion dynamics in Geometric algebra space. Specifically, we develop a Geometric Algebra Dependency-Aware Extractor, incorporating Geometric Algebra-based Fully Connected layers to adapt to geometric algebraic space, thus enabling the extraction of human action dependency representations. Additionally, we design an Euclidean Kinematic-Aware Extractor utilizing temporal-wise Kinematic-Aware Attention and spatial-wise Kinematic-Aware Feature Extraction. These two modules enhance and complement each other, leading to effective human motion prediction. Extensive experiments demonstrate that our proposed CBFL consistently improves performance, reducing MPJPE by an average of 4.3% on the Human3.6M dataset.

Authors

  • Ruiya Ji
    School of Electronic Engineering and Computer Science, Queen Mary University of London, London, UK.
  • Chengjie Lu
    College of Electronics and Information Engineering, Shenzhen University, 518060, Shenzhen, China.
  • Zhao Huang
    School of Optical and Electronic Information and Wuhan National Laboratory for Optoelectronics, Huazhong University of Science and Technology, Wuhan 430074, China.
  • Jianqi Zhong
    Guangdong Key Laboratory of Intelligent Information Processing and Shenzhen Key Laboratory of Media Security, Shenzhen University, Shenzhen 518060, China. Electronic address: zhongjianqi2017@email.szu.edu.cn.