Motion Segmentation via Generalized Curvatures

Area of Science:

• Computer Vision
• Robotics
• Biomechanical Analysis

Background:

• Depth sensors generate human pose data streams.
• These streams represent noisy samples of continuous trajectories in high-dimensional pose space.
• Characterizing and segmenting motion from this data is challenging.

Purpose of the Study:

• Introduce Generalized Curvature Analysis (GCA) for pose stream analysis.
• Develop tools to approximate generalized curvatures.
• Utilize GCA for motion characterization and segmentation.

Main Methods:

• Approximating generalized curvatures using singular values of local data.
• Applying GCA to synthetic and real-world datasets (Kinect II, MoCaP).
• Linear scaling of GCA with time series length for efficient analysis of large datasets.

Main Results:

• GCA features characterize trajectory changes and segment motion.
• Demonstrated GCA's effectiveness on synthetic and real pose data.
• Successfully segmented pose streams into motions and transitions.

Conclusions:

• GCA provides a robust method for analyzing human pose trajectories.
• The algorithm achieves high accuracy in identifying motion transitions.
• GCA enables efficient, large-scale motion segmentation without prior motion set knowledge.