Abstract
Trajectory segmentation refers to dividing a trajectory into meaningfulconsecutive sub-trajectories. This paper focuses on trajectory segmentation for3D rigid-body motions. Most segmentation approaches in the literature representthe body's trajectory as a point trajectory, considering only its translationand neglecting its rotation. We propose a novel trajectory representation forrigid-body motions that incorporates both translation and rotation, andadditionally exhibits several invariant properties. This representationconsists of a geometric progress rate and a third-order trajectory-shapedescriptor. Concepts from screw theory were used to make this representationtime-invariant and also invariant to the choice of body reference point. Thisnew representation is validated for a self-supervised segmentation approach,both in simulation and using real recordings of human-demonstrated pouringmotions. The results show a more robust detection of consecutive submotionswith distinct features and a more consistent segmentation compared toconventional representations. We believe that other existing segmentationmethods may benefit from using this trajectory representation to improve theirinvariance.