Human Motion Segmentation via Robust Kernel Sparse Subspace Clustering.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Human motion capture data offers precise movement recording, vital for applications like animation and robotics.
• Existing motion segmentation techniques struggle with the Riemannian manifold structure and non-Gaussian noise inherent in motion capture data.

Purpose of the Study:

• To develop a robust and efficient motion segmentation method for human motion capture data.
• To address limitations of current methods by incorporating the Riemannian manifold structure and handling non-Gaussian noise.

Main Methods:

• The proposed method frames motion segmentation as a temporal subspace clustering problem.
• It utilizes a geodesic exponential kernel for Riemannian manifold structure, correntropy for noise robustness, and triangle constraints for temporal continuity.
• A multi-view reconstruction approach captures inter-joint relationships, optimized via an efficient linear-complexity block coordinate descent algorithm.

Main Results:

• The new segmentation method demonstrates significant robustness against non-Gaussian noise.
• The optimized algorithm achieves linear complexity, an improvement over traditional quadratic sparse subspace clustering.
• Extensive experiments on simulated and real-world noisy datasets validate the method's effectiveness.

Conclusions:

• The proposed method offers a superior approach to segmenting human motion capture data.
• Its ability to handle complex data characteristics and noise makes it suitable for various motion analysis applications.