Towards cross-modal comparison of human motion data

Abstract

Analyzing human motion data has become an important strand of research in many fields such as computer animation, sport sciences, and medicine. In this paper, we discuss various motion representations that originate from different sensor modalities and investigate their discriminative power in the context of motion identification and retrieval scenarios. As one main contribution, we introduce various mid-level motion representations that allow for comparing motion data in a cross-modal fashion. In particular, we show that certain low-dimensional feature representations derived from inertial sensors are suited for specifying high-dimensional motion data. Our evaluation shows that features based on directional information outperform purely acceleration based features in the context of motion retrieval scenarios. © 2011 Springer-Verlag.