Fine-Grained Unsupervised Temporal Action Segmentation and Distributed Representation for Skeleton-Based Human Motion...

Area of Science:

• Computer Science
• Biomedical Engineering
• Robotics

Background:

• Analyzing fine-grained temporal structures in human actions is crucial for applications like sports analysis and rehabilitation.
• Current action segmentation methods often lack interpretability and struggle with fine-grained analysis due to undefined subactions and datasets.
• Existing motion representations lack semantic and mathematical interpretability for quantitative action similarity evaluation.

Purpose of the Study:

• To develop a novel unsupervised framework for fine-grained, interpretable, scalable, and efficient action segmentation.
• To introduce a distributed representation based on intuitive motion primitives derived from pose data.
• To propose comprehensive evaluation metrics for unsupervised fine-grained action segmentation.

Main Methods:

• Proposed a novel unsupervised action segmentation and distributed representation framework.
• Utilized intuitive motion primitives defined on pose data for action analysis.
• Developed new metrics for evaluating unsupervised fine-grained action segmentation performance.

Main Results:

• The proposed method demonstrates good performance in fine-grained action segmentation.
• The framework shows generality across different subjects, datasets, and application scenarios.
• The approach provides interpretable motion representations for quantitative analysis.

Conclusions:

• The developed unsupervised framework effectively addresses the challenge of fine-grained action segmentation.
• The method offers interpretable and scalable solutions for analyzing human actions.
• The proposed approach shows promise for real-world applications in sports, rehabilitation, and daily activity analysis.